0% found this document useful (0 votes)

1K views256 pages

Text 034

Uploaded by

leonhard

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

1K views256 pages

Text 034

Uploaded by

leonhard

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 256

Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)

Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
Teaching Statistics
Using Baseball
Second Edition

Jim Albert
Bowling Green State University

Published and Distributed by

The Mathematical Association of America

Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
MAA TEXTBOOKS
Bridge to Abstract Mathematics, Ralph W. Oberste-Vorth, Aristides Mouzakitis, and Bonita A.
Lawrence
Calculus Deconstructed: A Second Course in First-Year Calculus, Zbigniew H. Nitecki
Calculus for the Life Sciences: A Modeling Approach, James L. Cornette and Ralph A. Ackerman
Combinatorics: A Guided Tour, David R. Mazur
Combinatorics: A Problem Oriented Approach, Daniel A. Marcus
Common Sense Mathematics, Ethan D. Bolker and Maura B. Mast
Complex Numbers and Geometry, Liang-shin Hahn
A Course in Mathematical Modeling, Douglas Mooney and Randall Swift
Cryptological Mathematics, Robert Edward Lewand
Differential Geometry and its Applications, John Oprea
Distilling Ideas: An Introduction to Mathematical Thinking, Brian P. Katz and Michael Starbird
Elementary Cryptanalysis, Abraham Sinkov
Elementary Mathematical Models, Dan Kalman
An Episodic History of Mathematics: Mathematical Culture Through Problem Solving, Steven
G. Krantz
Essentials of Mathematics, Margie Hale
Field Theory and its Classical Problems, Charles Hadlock
Fourier Series, Rajendra Bhatia
Game Theory and Strategy, Philip D. Straffin
Geometry Illuminated: An Illustrated Introduction to Euclidean and Hyperbolic Plane Geome-
try, Matthew Harvey
Geometry Revisited, H. S. M. Coxeter and S. L. Greitzer
Graph Theory: A Problem Oriented Approach, Daniel Marcus
An Invitation to Real Analysis, Luis F. Moreno
Knot Theory, Charles Livingston
Learning Modern Algebra: From Early Attempts to Prove Fermat’s Last Theorem, Al Cuoco
and Joseph J. Rotman
The Lebesgue Integral for Undergraduates, William Johnston
Lie Groups: A Problem-Oriented Introduction via Matrix Groups, Harriet Pollatsek
Mathematical Connections: A Companion for Teachers and Others, Al Cuoco
Mathematical Interest Theory, Second Edition, Leslie Jane Federer Vaaler and James W. Daniel
Mathematical Modeling in the Environment, Charles Hadlock
Mathematics for Business Decisions Part 1: Probability and Simulation (electronic textbook),
Richard B. Thompson and Christopher G. Lamoureux
Mathematics for Business Decisions Part 2: Calculus and Optimization (electronic textbook),
Richard B. Thompson and Christopher G. Lamoureux
Mathematics for Secondary School Teachers, Elizabeth G. Bremigan, Ralph J. Bremigan, and
John D. Lorch
The Mathematics of Choice, Ivan Niven
The Mathematics of Games and Gambling, Edward Packel
Math Through the Ages, William Berlinghoff and Fernando Gouvea
Noncommutative Rings, I. N. Herstein

MAA Service Center

P.O. Box 91112
Washington, DC 20090-1112
1-800-331-1MAA FAX: 1-240-396-5647

Preface to the First Edition ix

Preface to the Second Edition xi

1 An Introduction to Baseball Statistics 1

2 Exploring a Single Batch of Baseball Data 13

2.1 Looking at Teams’ Offensive Statistics ..................................................... 13
2.2 A Tribute to Derek Jeter........................................................................ 17
2.3 A Tribute to Randy Johnson................................................................... 20
2.4 Analyzing Baseball Attendance............................................................... 23
2.5 Manager Statistics: the Use of Sacrifice Bunts ............................................ 26
2.6 Exercises ........................................................................................... 28

3 Comparing Batches and Standardization 45

3.1 Albert Pujols and Manny Ramirez ........................................................... 45
3.2 Robin Roberts and Whitey Ford .............................................................. 50
3.3 Home Runs: A Comparison of Four Seasons .............................................. 54
3.4 Slugging Percentages are Normal ............................................................ 57
3.5 Great Batting Averages ......................................................................... 59
3.6 Exercises ........................................................................................... 61

4 Relationships Between Measurement Variables 73

4.1 Relationships in Team Offensive Statistics................................................. 73
4.2 Runs and Offensive Statistics.................................................................. 78
4.3 Most Valuable Hitting Statistics .............................................................. 80
4.4 A New Measure of Offensive Performance ................................................ 86
4.5 How Important is a Run?....................................................................... 88
4.6 Baseball Players Regress to the Mean....................................................... 91
4.7 Exercises ........................................................................................... 94

vii
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
viii Contents

5 Introduction to Probability Using Tabletop Games 111

5.1 What is Chris Davis’ Home Run Probability?............................................. 111
5.2 Big League Baseball............................................................................. 114
5.3 All-Star Baseball ................................................................................. 116
5.4 Strat-O-Matic Baseball ......................................................................... 119
5.5 Exercises ........................................................................................... 124

6 Probability Distributions and Baseball 139

6.1 The Binomial Distribution and Hits per Game ............................................ 139
6.2 Modeling Runs Scored: Getting on Base ................................................... 142
6.3 Modeling Runs Scored: Advancing the Runners to Home.............................. 144
6.4 Exercises ........................................................................................... 148

7 Introduction to Statistical Inference 155

7.1 Ability and Performance........................................................................ 155
7.2 Simulating a Batter’s Performance if His Ability is Known ............................ 157
7.3 Learning About a Batter’s Ability ............................................................ 159
7.4 Interval Estimates for Ability.................................................................. 161
7.5 Comparing Wade Boggs and Tony Gwynn................................................. 165
7.6 Exercises ........................................................................................... 168

8 Topics in Statistical Inference 175

8.1 Situational Hitting Statistics for Mike Trout ............................................... 176
8.2 Observed Situational Effects for Many Players ........................................... 178
8.3 Modeling On-Base Percentages for Many Players........................................ 181
8.4 Models for Situational Effects................................................................. 185
8.5 Is Michael Brantley Streaky?.................................................................. 188
8.6 A Streaky Die ..................................................................................... 191
8.7 Exercises ........................................................................................... 193

9 Modeling Baseball Using a Markov Chain 211

9.1 Introduction to a Markov Chain............................................................... 211
9.2 A Half-inning of Baseball as a Markov Chain............................................. 215
9.3 Useful Markov Chain Calculations........................................................... 217
9.4 The Value of Different On-base Events ..................................................... 222
9.5 Answering Questions About Baseball Strategy ........................................... 224
9.6 Exercises ........................................................................................... 225

A An Introduction to Baseball 233

A.1 The Game of Baseball........................................................................... 233
A.2 One Half-Inning of Baseball................................................................... 234
A.3 The Boxscore: A Statistical Record of a Baseball Game................................ 235

Bibliography 239

Index 241

Over twenty years, this author has been in the enterprise of teaching introductory statistics to an
audience that is taking the class to satisfy their mathematics requirement. This is a challenging
endeavor because the students have little prior knowledge about the discipline of statistics
and many of them are anxious about mathematics and computation. Statistical concepts and
examples are usually presented in a particular context. However, one obstacle in teaching this
introductory class is that we often describe the statistical concepts in a context, such as medicine,
law, or agriculture that is completely foreign to the undergraduate student. The student has a
much better chance of understanding concepts in probability and statistics if they are described
in a familiar context.
Many students are familiar with sports either as a participant or a spectator. They know of the
popular athletes, such as Tiger Woods and Barry Bonds, and they are generally knowledgeable
with the rules of the major sports, such as baseball, football, and basketball. For many students
sports is a familiar context in which an instructor can describe statistical thinking.
The goal of this book is to provide a collection of examples and exercises applying proba-
bility and statistics to the sport of baseball. Why baseball instead of other sports?
Baseball is the great American game. Baseball is great in that it has a rich history of teams
and players, and many people are familiar with the basic rules of the game. The popularity of
baseball is reflected by the large number of movies that have been produced about baseball
teams and players.
Baseball is the most statistical of all sports. Hitters and pitchers are identified by their
corresponding hitting and pitching statistics. For example, Babe Ruth is forever identified by
the statistic 60, which was the number of home runs hit in his 1927 season. Bob Gibson is
famous for his record low earned run average of 1.12 during the 1968 season. A flood of
different statistical measures are used to rate players and salaries of players are determined in
part by these statistics. There is an active effort among baseball writers to learn more about
baseball issues by using statistics.
A wealth of baseball data is currently available over the Internet. Player and team hitting and
pitching statistics can be easily found. Comparisons between players of different eras can be
made using a downloadable dataset that gives hitting and pitching data for all players who
have ever played professional baseball.

ix
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
x Preface to the First Edition

This book is organized using the same basic organization structure presented in most
introductory statistics texts. After an introductory chapter, there is a chapter on the analysis
on a single batch of data, followed by chapters on the comparison of batches, and the analysis
of relationships. There are chapters on introductory and more advanced topics in probability,
followed by topics in statistical inference. Each chapter contains a number of essays or case
studies that describe the analysis of statistical or probabilistic methods to particular baseball
data sets. After the collection of case studies in each chapter, there is a set of activities and
exercises that suggest further exploration of baseball datasets similar to the analysis presented
in the case studies.
How can this book be used in teaching probability or statistics? We suggest several uses of
this material.
This book can be used as the framework for a one-semester introductory statistics class that
is focused on baseball. Such a class has been taught at the author’s home institution. This
course covers the basic topics of a beginning statistics course (data analysis, introductory
probability, and concepts of inference) using baseball as the primary source of applications.
This course is suitable for students who are interested or curious about the game of baseball.
It is also suitable for students with sports-related majors, such as sports management or sports
medicine.
This book can also be used as a resource for instructors who wish to infuse their present
course in probability or statistics with applications from baseball. The material in this book
has been presented at different levels to make it useable for introductory and more advanced
courses. The case studies can be used by the instructor to present the particular topic within
a baseball context and then the associated exercises and activities can be used for homework.
The case studies can serve as useful springboards for undergraduate students who wish to do
additional explorations on baseball data.

Acknowledgements
I am appreciative of the support given to this project by the Division of Undergraduate Education
of the National Science Foundation and by my colleagues in the Department of Mathematics and
Statistics at Bowling Green State University. The text was used for a number of experimental
sections of MATH 115 Introduction to Statistics at Bowling Green State University and I am
grateful for the valuable feedback from the students who enrolled in this course. In addition,
Chris Andrews, Jay Bennett, Eric Bradlow, Jim Cochran, Joe Gallian, Carl Morris, Jerome
Reiter, Ken Ross, Steve Samuels, Bob Wardrop, and Dex Whittinghill provided many helpful
suggestions in reviewing the book. I thank the editors of the MAA publications for their support,
particularly Zaven Karien and Dave Kullman. Last, but certainly not least, I thank my wife Anne,
and children Lynne, Bethany, and Steven for their understanding and great patience during the
completion of this text.
Jim Albert
December 2002

The author has been gratified with the reception to the first edition of this text. It has been
enjoyable teaching introductory statistics from a baseball viewpoint, as one can discuss statistical
concepts in the context of current and historical players and teams.
Sports provide a wonderful context for teaching introductory statistics and other texts have
been recently published using a sports theme. Rothman (2012) is a comprehensive statistics text
with a baseball theme, and Tabor and Franklin (2011) is a statistics text with examples from
a wide range of sports. There is a weath of publically available baseball data and Marchi and
Albert (2013) describe the use of these datasets together with the open-source statistics software
system R (R Core Team (2015)) to implement a variety of baseball studies.
Since the date of the first edition, there have been changes both in the players who play
baseball, and also in the development and use of analytics. So this motivates revisions to the
text reflecting these changes.
In this edition, many of the case studies and exercises have been revised to use data from
current teams and players. I encourage the instructor to always use current season data since the
modern players and teams are most familiar to students.
Also exercises have been added to the chapters reflecting some of the newer types of
baseball data. The pitchFX system has been tracking the trajectories of baseball pitches since
2006 and we know more about the breaks, locations, and speeds of pitches. Websites such
as fangraphs.com contain much of this pitchFX data for both batters and pitchers and
this website also gives information about the location, speed, and type (flyball, pop-up, and
grounders) of balls put in play. The tracking technology Statcast contains information about the
speed of runners and fielders and velocities of balls coming off a bat.
To facilitate the use of the data described in the examples and exercises, all of the datasets
are currently available using the StatCrunch statistical software system published by Pearson (in
StatCrunch search using the acronym TSUB and the example or exercise number). In addition,
the datasets are available as text files at https://bayesball.github.io/.
Jim Albert
December 2016

xi
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
1
An Introduction to Baseball Statistics

Leading Off
The baseball game has just started. The umpire has yelled “Play Ball!”, and the batter at the
top of the order is coming to bat. He’s the leadoff hitter and his job is to help produce runs by
getting on base. Who was the greatest leadoff hitter of all time? Most people believe that the
best leadoff man was Rickey Henderson. Bill James, a leading baseball statistician, says in The
New Bill James Historical Baseball Abstract that Henderson was

the greatest base stealer of all time,

the greatest power/speed combination of all time (except maybe Barry Bonds),
the greatest leadoff man of all time,
one of the top five players of all time in runs scored.

Moreover, James says: “You could find fifty Hall of Famers who, all taken together, don’t
own as many records, and as many important records, as Rickey Henderson.”
Here’s some biographical information about Rickey Henderson. He was born on Christmas
Day, 1958, in Chicago, one of seven children. His family moved to Oakland, California when
he was young and Henderson played baseball and football at Oakland Tech High School. When
he graduated, he received many football scholarships and also was selected by the Oakland
A’s in the fourth round of the 1976 baseball draft. Although Henderson preferred football,
his mother wanted him to play baseball, and Henderson agreed to go along with his mother’s
wishes. After a couple of years in the A’s minor league organization, he was promoted to
the Oakland major league team on June 23, 1979,1 and immediately was a starter on the
team. He has been a dominant player in the major leagues his entire career and was voted on
the All-Star Team for the years 1980, 1982, 1983, 1984, 1985, 1986, 1987, 1988, 1990, and
1991.
How can we demonstrate that Rickey Henderson was indeed the best leadoff man in
baseball? We look at his stats. Table 1.1 displays the season-to-season hitting statistics for
Henderson for his 25 seasons in Major League Baseball.

1
I have a kinship with Rickey Henderson since we both started professionally (me as a statistician and
Henderson as a ballplayer) in 1979. I lasted longer than Henderson in the professional ranks.

1
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
2 An Introduction to Baseball Statistics

Table 1.1. Batting statistics for Rickey Henderson’s career

Year Team G AB R H 2B 3B HR RBI SB CS BB SO
1979 OAK 89 351 49 96 13 3 1 26 33 11 34 39
1980 OAK 158 591 111 179 22 4 9 53 100 26 117 54
1981 OAK 108 423 89 135 18 7 6 35 56 22 64 68
1982 OAK 149 536 119 143 24 4 10 51 130 42 116 94
1983 OAK 145 513 105 150 25 7 9 48 108 19 103 80
1984 OAK 142 502 113 147 27 4 16 58 66 18 86 81
1985 NYY 143 547 146 172 28 5 24 72 80 10 99 65
1986 NYY 153 608 130 160 31 5 28 74 87 18 89 81
1987 NYY 95 358 78 104 17 3 17 37 41 8 80 52
1988 NYY 140 554 118 169 30 2 6 50 93 13 82 54
1989 TOT 150 541 113 148 26 3 12 57 77 14 126 68
1990 OAK 136 489 119 159 33 3 28 61 65 10 97 60
1991 OAK 134 470 105 126 17 1 18 57 58 18 98 73
1992 OAK 117 396 77 112 18 3 15 46 48 11 95 56
1993 TOT 134 481 114 139 22 2 21 59 53 8 120 65
1994 OAK 87 296 66 77 13 0 6 20 22 7 72 45
1995 OAK 112 407 67 122 31 1 9 54 32 10 72 66
1996 SDP 148 465 110 112 17 2 9 29 37 15 125 90
1997 TOT 120 403 84 100 14 0 8 34 45 8 97 85
1998 OAK 152 542 101 128 16 1 14 57 66 13 118 114
1999 NYM 121 438 89 138 30 0 12 42 37 14 82 82
2000 TOT 123 420 75 98 14 2 4 32 36 11 88 75
2001 SDP 123 379 70 86 17 3 8 42 25 7 81 84
2002 BOS 72 179 40 40 6 1 5 16 8 2 38 47
2003 LAD 30 72 7 15 1 0 2 5 3 0 11 16

The stats that most people talk about are Henderson’s career totals.
He stole the most bases (1395) of any baseball player in history.
He scored the most runs (2248) of any player in history.
He received the most walks (2141) of any player in history.
Those are great achievements and the statistics are a measure of these achievements. But
the goal of this book is to look deeper at baseball statistics.
There is a general confusion about the meaning of statistics. If you look at the American
Heritage Dictionary of the English Language, you will find two very different definitions of
statistics.
1. Statistics is a collection of numerical data.
2. Statistics is the mathematics of the collection, organization, and interpretation of numerical
data.
Let’s relate these two definitions to baseball. First, baseball statistics are the counts and
measures that we use to evaluate players and teams—this refers to the first definition of statistics.

When the announcer on a television broadcast of a baseball game thanks the statistician, he or
she is referring to the person who collects and gives baseball data to the announcers. Here the
focus is on the data.
But this book concentrates on the second definition of statistics: how can we interpret or
make sense of baseball stats? A professional statistician (to be distinguished from the person
who is collecting the data) is concerned with how we can use data to learn about some underlying
truth. In doing this, he or she has to think about several issues.

How should the data be collected to make it useful in drawing conclusions?

Once the data is collected, how do we organize and summarize it to learn about its general
features?
Last, how can we use the data to make our conclusions? (It turns out that probability or chance
plays an important role in decision-making.)

It might be helpful to distinguish the two meanings by capitalization—in this section I will
call numerical data “statistics,” and the science of learning from data “Statistics.”
The goal of this book is to introduce Statistical thinking and Statistical methods in the
context of baseball. We introduce the chapters of this book by looking at the statistics of Rickey
Henderson. Some questions will be raised in the following discussion and we’ll continue our
Statistical look at Rickey Henderson by “leadoff exercises” in each chapter.

Exploring a Single Batch of Baseball Data (Chapter 2)

The goal of a leadoff hitter is to get on base. The obvious measure of a player’s ability to get on
base is the on-base percentage (OBP), which is simply the fraction of plate appearances where
the player gets on base. (A precise definition of OBP will be given later.) Here are Henderson’s
season OBPs for all 25 seasons in the majors:

0.338 0.420 0.408 0.398 0.414 0.399 0.419 0.358 0.423

0.394 0.411 0.439 0.400 0.426 0.432 0.411 0.407 0.410
0.400 0.376 0.423 0.368 0.366 0.369 0.321

If we scan these numbers, we see variation—one season he had an OBP of :432 and another
season his OBP was :366. A Statistician will try to make sense of these data by constructing an
appropriate graph. Figure 1.1 shows a dotplot of the OBPs.

0.32 0.34 0.36 0.38 0.40 0.42 0.44

OBP

Figure 1.1. Dotplot of season on-base percentages for Rickey Henderson

We see from this graph that most of Henderson’s season OBPs are between :400 and :420.
There is a small cluster of values in the :340 to :380 range. Why? Was Henderson hurt these

particular seasons? Did these low OBPs correspond to the early or late periods of his career?
(We’ll answer these questions later.)
After we graph the data, we try to find suitable numbers to summarize the main features
of the distribution of OBPs. Looking at the graph, it seems that :410 might be a representative
season OBP for Henderson.

Comparing Batches and Standardization (Chapter 3)

We now have some handle on Henderson’s on-base ability: he generally got on base about 40%
of the time. But is an on-base percentage of :400 any good? How does his on-base performance
compare with other players?
Another good leadoff hitter, a contemporary of Henderson, was Tim Raines. How effective
was Tim in getting on base and how did he compare with Henderson?
Here are Tim’s on-base percentages for the 1981–1998 seasons where he had at least 200
at-bats.

0.391 0.353 0.393 0.393 0.405 0.413 0.429 0.350 0.395

0.379 0.359 0.380 0.401 0.365 0.374 0.383 0.403 0.395

How does this batch of OBPs compare to the batch of OBPs for Henderson? In Chapter 3, we’ll
talk about ways of comparing two batches of data. We will discuss methods for graphing the two
datasets and ways of stating Statistically (in this example) that Henderson is better at getting
on base than Tim. This chapter will also show how we can judge the greatness of Henderson’s
on-base performance in the context of all players that particular year.

Relationships between Measurement Variables (Chapter 4)

When we look at Rickey Henderson’s hitting statistics in Table 1.1, we notice that many of
the statistics are related. For example, if a batter gets many doubles and triples, he will have
a high slugging percentage, and a batter who rarely walks is likely to have a small on-base
percentage. In Chapter 4, we discuss ways of looking at relationships between variables. To
illustrate, consider the relationship between Henderson’s count of doubles and his count of
home runs for individual seasons. Doubles and home runs go hand-in-hand—one might think
that if Henderson is hitting a lot of deep fly balls one season then he would have many doubles
and home runs. We can graphically view the relationship between doubles and home runs by
the scatterplot shown in Figure 1.2. There is a pattern in the scatterplot as the points drift from
the lower left to the upper right sections—in seasons where Henderson hit many doubles, he
tended also to hit many home runs.
In Chapter 4, we discuss ways of measuring the pattern of association in the scatterplot, and
discuss how we can use a line to describe the relationship. The methods are helpful for finding
a good measure of the batting performance of a player.

Introduction to Probability Using Tabletop Games (Chapter 5)

Fans love to play baseball games; currently millions of people are playing fantasy and simulation
baseball. Before there were personal computers, Nintendo and fantasy baseball, there were a
number of tabletop baseball games that were very popular among fans. In Chapter 5, we
introduce the notion of a probability model by examining several tabletop baseball games. One

25 20
Home Runs
10 155
0

0 5 10 15 20 25 30
Doubles

Figure 1.2. Scatterplot of count of doubles and count of home runs for all seasons of Rickey Henderson’s
career.

of the first games played by the author as a child was All-Star Baseball where the performance
of a batter was represented by a random spinner with areas of the spinner corresponding to the
different outcomes of a plate appearance. A Rickey Henderson spinner is shown in Figure 1.3
where the areas of the regions are computed using his batting statistics from the 1990 season.
Note the large pie slice corresponding to a walk—this is a visual demonstration of Henderson’s
ability to draw walks this particular season.

TRIPLE
HOME RUN

E
UB L
DO

WALK

SINGLE

OUT

Figure 1.3. A Rickey Henderson spinner using his 1990 hitting statistics.

Probability Distributions and Baseball (Chapter 6)

Baseball has a very nice discrete structure that makes for convenient probability modeling.
The basic event in baseball is the outcome of a plate appearance. Rickey Henderson comes to
bat—either he will get on base or he won’t. If you assume that Henderson has a probability, say
:4, of getting on base for each plate appearance, and moreover, the chance of getting on base
does not depend on how he hit in earlier games, then one can make reasonable predictions about
the number of times he will get on base in a game, a week, or a month. In Chapter 6, we show
how some popular probability distributions (such as the binomial and negative binomial) can
be used to explain the random process of players getting on base, and then scoring runs.

Introduction to Statistical Inference (Chapter 7)

Rickey Henderson had a great ability to get on base. What does this mean? Well, to be an
effective leadoff hitter, Henderson had to be very able at making the bat hit the ball. (Such a
player is called a good contact hitter.) Also, he knew how to “work the count.” This means that
he would not swing at many bad pitches and would patiently wait for a good pitch to be thrown.
How do we know that Henderson had great leadoff hitter ability? We looked at his on-base
percentages over his career and saw a pattern of high numbers. In other words, Henderson exhib-
ited a high level of performance over many years, and there was no doubt that his performance
reflected great skill. It was very unlikely that he had average ability to get on base and, by luck
or chance variation, he happened to get high on-base percentages for all of those years.
In Chapter 7, we look at the connection between a player’s hitting ability and the performance
of the hitter over a season. Actually, if a player has an on-base percentage of OBP D :400 for
only a single season, we really don’t know if the player has a great ability to get on base. We
need to see a pattern of great hitting performance over many seasons to properly gauge the
player’s hitting ability.

Topics in Statistical Inference (Chapter 8)

Making judgments about a player’s hitting or pitching ability is relatively easy when you observe
the player’s performance over a 10–20 year career. But there are other aspects of ability that are
much harder to detect. To illustrate, consider the following situational statistics for Henderson
for the 1999 season. This season, Henderson’s on-base percentage was :423. However, we learn
that

his OBP was :376 for home games and :462 for away games,
his OBP was :490 for games played on artificial turf and :408 for games played on natural
grass,
his OBP was :472 for games played in a domed ballpark and :417 in open ballparks,
his OBP was :509 when the pitch count reached 3-2, and :338 when the pitch count reached
1-2.

How do we make sense of these situational hitting stats? Is Henderson really better at
getting on base when he is playing at home? Does Henderson really have better on-base ability
when the game is played on turf? Is it meaningful that his on-base percentage is over :500 when
the pitch count gets to 3 balls and 2 strikes? Does Henderson really have an ability to perform
extra well in particular situations?

One main topic of Chapter 8 is to try to interpret the significance of situational hitting data.
We will see that it is difficult to detect situational hitting ability and much of the variation in
situational hitting statistics is attributable to chance or random variation.

Modeling Baseball Using a Markov Chain (Chapter 9)

We say that Rickey Henderson is the greatest leadoff hitter of all time, but we haven’t talked
about how well Henderson performed when he was the leadoff hitter in an inning. From the
statistics in Table 1.1, we can compute Henderson had an OBP of :439 in the 1990 season; the
OBP tells us the fraction of plate appearances that Henderson got on base for all situations.
What was his on-base percentage in the innings when he was leading off?
Fortunately, this type of data is now readily available. Let’s focus on the home games in
Oakland in 1990. In these games, Henderson had 273 plate appearances, but only 108 of them
occurred at the beginning of an inning. So Henderson only led off 108 times. How did he do
in these lead-off opportunities? Table 1.2 shows the batting results. Note that he was out 63%
of the time, so his on-base percentage when he actually was leading off the inning was .370.
This on-base percentage appears to be a bit low, but one has be cautious about drawing a strong
conclusion since this table summarizes the results of only 108 plate appearances.

Table 1.2. Batting results for Rickey Henderson’s leadoff

plate appearances at home during the 1990 season
Play Count Percentage
Out 68 63%
Walk 18 16.7%
Single 13 12.0%
Double 5 4.6%
Triple 1 0.9%
Home run 3 2.8%

We can also see how Henderson performed when the bases were empty with one out, or
when the bases were loaded with two outs. We will use data such as this in Chapter 8 to construct
a sophisticated probability model for the sequence of batting events in baseball. This model will
be very useful for measuring the values of different types of hits such as a home run and for
evaluating the worth of different baseball strategies such as a sacrifice bunt.

Some Basic Measures of Baseball Performance

The effectiveness of batters and pitchers is typically assessed by particular numerical measures.
Here we define some basic measures for evaluating hitters and pitchers.

Measures for Batters

The classical measure of hitting effectiveness for a player is the batting average (AVG) that is
computed by dividing the number of hits (H) by the number of at-bats (AB):

H
AVG D :
AB

This statistic gives the proportion of time that a batter gets a hit among all at-bats. The
batter with the highest batting average during a baseball season is called the batting champion
that year. Batters are also evaluated on their ability to get singles (1B), doubles (2B), triples
(3B), and home runs (HR). The slugging percentage (SLG) is an average of a player’s “total
bases” (TB) where the hits are weighted by means of the number of bases reached:
1 1B C 2 2B C 3 3B C 4 HR
SLG D :
AB
This measure reflects the ability of a batter to hit the ball a long distance. A third measure
of hitting ability is the on-base percentage (OBP), which is defined as the proportion of plate
appearances where the player gets on base.
H C BB C HBP
OBP D
AB C BB C HBP C SF
In this formula, BB is the count of walks, HBP is the number of times the batter was hit by a
pitch, and SF is the number of sacrifice flies.

Measures for Pitchers

A number of statistics are also used in the evaluation of pitchers. For a particular pitcher, one
counts the number of games in which he was declared the winner (W) or loser (L) and the
number of runs allowed. Pitchers are usually rated by means of the earned run average (ERA)
the average number of earned runs (ER) allowed for a nine-inning game:
ER
ERA D 9 :
IP
(In this formula, IP is the number of innings pitched.) Other statistics are useful in understanding
pitching ability. A pitcher who can throw the ball very fast (such as Randy Johnson) can record
a high number of strikeouts (SO). A pitcher who is “wild” or relatively inaccurate in his pitching
will record a large number of walks (BB).

Better Measures of Hitting Ability

There is lively research to better interpret baseball statistics. Sabermetrics is the mathematical
and statistical analysis of baseball records. (The term sabermetrics was first used by Bill James in
honor of SABR, the Society of Baseball Research.) One interest of sabermetricians (the people
who analyze baseball statistics) is to find good measures of hitting and pitching performance.
Bill James compares in his 1982 Baseball Abstract the batting records of two players, Johnny
Pesky, who played from 1942 to 1954, and Dick Stuart, who played in the 1960s. Pesky was
a batter who hit for a high batting average but hit few home runs. Stuart, in contrast, had a
modest batting average, but hit a high number of home runs. Who was the more valuable hitter?
James argues that a hitter should be evaluated by his ability to create runs for his team. From
an empirical study of a large collection of team hitting data, he established the following runs
created (RC) formula for predicting the number of runs scored in a season based on the number
of hits, walks, at-bats, and total bases recorded in a season:
.H C BB/ TB
RC D :
AB C BB

This formula reflects two important aspects in scoring runs in baseball. The count of a team’s
hits and walks reflects the team’s ability to get runners on base. The count of a team’s total bases
reflects the team’s ability to move runners that are already on base. This runs created formula
can be used at an individual level to compute the number of runs that a player creates for his
team. In 1942, Johnny Pesky had 620 at-bats, 205 hits, 42 walks, and 258 total bases; according
to the formula, he created 96 runs for his team. Dick Stuart in 1961 had 532 at-bats with 160
hits, 34 walks, and 309 total bases for 106 runs created. The conclusion is that Stuart in 1961
was a slightly better hitter than Pesky in 1942 since he created a few more runs for his team.
There are a number of alternative measures that have been proposed to evaluate hitters.
Here we list some of the measures and they will be carefully compared and evaluated in later
chapters. Since OBP is a measure of a player’s ability to get on base and SLG is a measure of
a player’s ability to advance the runners, a simple measure OPS (for On-base percentage Plus
Slugging percentage), adds the two measures:
OPS D OBP C SLG:
In the chapters to follow, we will investigate the goodness of the measures RC and OPS in
predicting the number of runs scored by a team.

Baseball Data
Here we describe several common types of baseball data that we will analyze in this book.

Career statistics for a player

Probably the most familiar data set among baseball fans is the batting or pitching statistics of a
player over the seasons of his career. Many fans have collected baseball cards, either as children
or adults, and the back of these cards typically contain these career statistics. For each season,
the card gives statistics similar to what is displayed for Rickey Henderson in Table 1.1.

Player statistics for a given season

Another informative data set is the collection of all batting and pitcher statistics for all players in
a particular season. We will see that it is difficult to judge a single statistic, say Roger Maris’ 61
home run season in 1961, by itself. (Roger Maris is famous since he set the single season record
for home runs that particular year.) To understand the significance of this statistic, we need to
look at it in the context of all player statistics for that season. Although the basic rules of baseball
have not changed over the years, the athletic abilities of the players, the competitive balance, and
the equipment and environment have changed, and these changes have had a substantive impact
on the values of baseball statistics. To return to our Johnny Pesky and Dick Stuart comparison,
it really is not fair to compare the runs created by Pesky and Stuart at face value, since there
were many more runs scored in the 1961 season than the 1942 season.

Team statistics for a given season

Another interesting data set to work with is the batting and pitching statistics for all teams in
a particular season. One general problem of interest is to find a suitable measure of the hitting
ability of a player. Since the goal of batting is to score runs, one is interested in finding a hitting
statistic that is useful in predicting the number of runs scored. But teams, not individual players,
score runs, so one needs to look at team data in the development of good hitting measures.

Game logs
Baseball fans are fascinated with the day-to-day performance of their favorite players and teams.
Teams and individual players go through good and bad periods and these performances in short
time periods are often described in the media. So it is interesting to look at the performance of
players and teams for each game of the baseball season.

Situational statistics
Baseball fans are also fascinated with the performances of teams and players in given situations.
How does a batter perform at home and away games? How does he perform against different
pitchers? How does he do at night and day games? How well does he hit when he swings at the
first pitch, or when there are two strikes in the count? How well does a team perform when their
ace pitcher is starting? These situational or breakdown statistics are now reported on all of the
popular baseball news websites. One goal of this book is to try to make sense of the importance
of these statistics.

Statistics through the years

Baseball has a fascinating history. It is fun to read about the great players of the past, including
Ty Cobb, Walter Johnson, Shoeless Joe Jackson, Joe DiMaggio, and Babe Ruth. Also it is
interesting to look at the great teams of all time, including the 1927 New York Yankees, the
1929 Philadelphia Athletics, the 1975 Cincinnati Reds, and the 1998 New York Yankees. One
can explore this baseball history by means of baseball statistics. For example, we will look at
home runs hit by teams in the years 1927, 1961, 1998, and 2001, and the differences that we
see in this comparison tell us a lot about the relative difficulty of hitting a home run over time.

Miscellaneous statistics
Although we will focus much of our discussion on basic hitting and pitching statistics, there are
many other associated baseball statistics that are fun to explore. These include:
The salaries of the players.
The attendance counts at the games.
Statistics related to managerial strategy.
The duration of the game.
The number of pitches thrown.
The time needed to complete the game.

Collecting baseball data

The collection of baseball data is much easier today than in the past due to the internet. There
are many web sites that contain historical and current baseball data. Here we highlight several
internet sources that are particularly convenient for downloading data that can be easily entered
into a statistical computing package.
www.baseball-reference.com
This is a good site for historical data on teams and players. One can easily obtain career statistics
for any historical player and team data is available for any past season. The data sets are stored
as text files and it is relatively easy to import the data into a standard spreadsheet program such
as Microsoft Excel.

www.baseball1.com
The Baseball Archive is an especially good site for downloading large collections of baseball
data. In fact, one can download (either in text or Microsoft Access format) a single file that
contains player and team statistics for all years in baseball history. Many of the data sets used
in this book are taken from this database.
www.retrosheet.org
The Retrosheet organization is dedicated to collecting play-by-play baseball data. For each plate
appearance in a given game, a data file will record the name of the hitter, the name of the pitcher,
the game situation (score, runners on base, number of outs), the play, and other information.
One can currently download this type of data for entire baseball seasons. Using these data, one
can perform many interesting analyses. This type of data will be used in the Markov Chain
modeling described in Chapter 9.
www.fangraphs.com
Fangraphs is a good place to visit to collect some of the more modern types of baseball data. For
example, this site includes summaries of data collected from the PITCHf/x system, such as the
number of pitches thrown of different types such as fastballs, curveballs, sliders, and changeups,
and the velocities of these pitches. For batters, Fangraphs gives the percentages of batted balls
that are line drives, popups and flyballs, and also the percentages of batted balls hit to the left,
center, and right. This modern baseball data will be explored in some of the exercises.

What’s On-Deck?
In this chapter, we illustrate a number of graphs and summary statistics useful in exploring a
single batch of data. In Case Study 2.1, we begin with batting statistics for the 30 Major League
Teams for the 2014 season. We focus on the team home run numbers and use stemplots and five-
number summaries to compare the home run production of the National League and American
League teams. In Case Studies 2.2 and 2.3, we look at the career statistics for two current or
future Hall of Famers, Derek Jeter and Randy Johnson. When looking at an individual’s statistic,
it is helpful to construct a graph of the statistic against time. The patterns in this time series plot
are helpful for understanding how the player has matured as a baseball player. In Case Study
2.4, we study baseball attendance for the 30 teams in 2014. Although baseball teams would all
like to make a profit, we will see a wide disparity in the teams’ abilities to bring fans to the
ballpark. We conclude in Case Study 2.5 by looking at statistics for managers. One basic play in
baseball is the sacrifice bunt, and we will see that this is a popular strategy for some managers
and a very unpopular move for other managers.

2.1 Looking at Teams’ Offensive Statistics

Topics Covered: Stemplot, data distribution, five-number summary.
After the conclusion of a baseball season in November, teams begin to evaluate how well
they performed during the season. How effective was a particular team, say the Phillies, in
getting batters on base? Did the Phillies score a lot of runs this year? What teams were good
and bad at hitting home runs? How many home runs were hit by a representative major league
team this year?
Table 2.1 displays a number of offensive statistics for six major league teams for the
2014 season. Many of these statistics are counts, such as the number of hits, the number of
doubles, the number of home runs, and so on. Other offensive statistics are derived measures of
offensive performance, such as batting average (AVG), slugging percentage (SLG), and on-base
percentage (OBP), which are computed from the count statistics.

Home Run Totals

To explore the power hitting among the teams, let’s look at the number of home runs hit by the
30 teams. The first step in exploring a single batch of data, such as these 30 home run totals,

13
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
14 Exploring a Single Batch of Baseball Data

Table 2.1. Batting statistics for six Major League Baseball teams in 2014
Team AVG OBP SLG R.G H 2B 3B HR BB SO
Arizona .248 .302 .376 3.80 1379 259 47 118 398 1165
Atlanta .241 .305 .360 3.54 1316 240 22 123 472 1369
Baltimore .256 .311 .422 4.35 1434 264 16 211 401 1285
Boston .244 .316 .369 3.91 1355 282 20 123 535 1337
Chicago (NL) .239 .300 .385 3.79 1315 270 31 157 442 1477
Chicago (AL) .253 .310 .398 4.07 1400 279 32 155 417 1362

is to draw a suitable graph. An effective graph that is easy to draw by hand is the stemplot. To
construct a stemplot, we divide each home run total into two parts, called the stem and the leaf.
For example, Arizona’s home run total, 118, can be divided between the tens and units places
(see below)

118

stem leaf

to get a stem of 11 and a leaf of 8. We write down all of the possible stems, and record each
home run total by writing down its leaf on the line corresponding to the stem. If we do this for
all 30 home run totals, we get the stemplot shown in Figure 2.1.

Figure 2.1. Stemplot of team home run numbers from 2014 season.

The second line tells us that the totals 105, 109 were the number of home runs hit by two of
the 30 teams. To study this distribution of home run totals, it may be helpful to flip this stemplot
by a 90-degree turn, so that the small totals are on the left.

What observations can we make from this data distribution?

1. First, we look for the general shape of these home run totals. Although there is a general
mound shape, we see two clusters in these home run counts, one in the 120’s, and a second
in the 150’s.
2. After we think of the general shape, we look for an “average” home run total. Since exactly
half (15) of the totals fall below 135 and half fall above 135, we can regard 135 home runs
as a measure of the center of the distribution. (We call 135 the median of the observations.)
3. Next, we look at the spread or variation in these home run totals. Here the spread of the totals
is pretty large, the largest value 211 (number of home runs hit by Baltimore) is more than
twice as large as the smallest value 95 (hit by Kansas City). But most of the home run totals
fall between 110 and 157.
4. Last, we look for any unusual characteristics of the totals. There is clearly one large number
that is separated from the rest—Baltimore hit a lot of home runs in 2014.

If you are a baseball fan, you are probably interested in the relative standing of your team
in this distribution of home run totals. To better see the teams’ relative standing, we can redraw
this stemplot in Figure 2.2 using team labels instead of numerical leaves.

Figure 2.2. Stemplot of team home run numbers from 2014 season with teams identified.

My team, the Phillies, appears in the cluster of lower numbers. It is interesting to note that
home run production is not always associated with winning and losing. Colorado, the team with
the second highest home run total, finished next to last place in their division, whereas Kansas
City, a team who appeared in the 2014 World Series, had the lowest team home run total.
This stemplot display motivates a follow-up question: which league hit more home runs in
2014? To compare the team totals for the two leagues, in Figure 2.3 we will put the leaves of
the home run totals for the NL teams to the left of the stems and the leaves for the home run
totals for the AL on the right.
Comparing the left and right stemplots, I would conclude that the American League teams
tended to hit more home runs than the National League teams in 2014. Note that only one of

Figure 2.3. Back-to-back stemplots of team home run numbers from the National and American Leagues
for the 2014 season.

the 15 National League teams and three of the 15 American League teams hit more than 160
home runs. One possible explanation for this pattern is that the American League plays with
the designated hitter who is a substitute hitter for the pitcher (in contrast to the National League
where the pitcher comes to bat).

Figure 2.4. Stemplot of the OBPs for the 2014 Major League teams.

On-Base Percentages
Actually, the number of home runs hit is not really a good reflection of a team’s offensive
production. A better indicator of run production is the team’s on-base percentage (OBP) that
measures the fraction of plate appearances in which the team gets on-base. In constructing a
stemplot of the 30 team OBPs, we break the OBP between the 2nd and 3rd digits. Also we write
two lines for each possible stem, where the leaves 0–4 are written on the first line and the leaves
5–9 on the second line. The completed stemplot is displayed in Figure 2.4. The third line tells
us that two teams had an on-base percentage of :300 and two teams had an on-base percentage
of :302.

What do we see in these displays?

1. The shape of these team OBPs is pretty symmetric.
2. An average OBP is about :311 and the team OBPs range from :292 (San Diego) to :333 (LA
Dodgers).
3. One unusual feature that stands out is the three teams (Detroit, LA Dodgers, and Pittsburgh)
in the last line with the largest values of OBP.
One can summarize these team OBPs by the computation of a median and quartiles. The
median is the value that divides the data into a bottom half and a top half. Here we have 30
values—so the median is the average of the 15th and 16th largest OBPs:
Median D .:314 C :314/=2 D :314:
The quartiles divide the data into quarters. The lower quartile is the median of the lower half of
the data, and the upper quartile is the median of the upper half. In our example, we can divide
the team OBPs into an upper half of 15 values, and a lower half of 15 values. The median of
the lower half is the 8th lowest value (:307), and the median of the upper half is the 8th largest
value (:321). So
lower quartile D :307; upper quartile D :321:
A five-number summary of these data is (lowest value, lower quartile, median, upper quartile,
highest value) which is .:292; :307; :314; :321; :333/. These five numbers divide the data roughly
in quarters. So approximately 1=4 or 25% of the team OBPs fall between :292 and :307,
approximately 25% of the OBPs fall between :307 and :314, and so on.
Again, it’s interesting to note the relative standing of the OBP of your favorite team. My
team (the Phils) had a team OBP of :302 that puts it in the lower quarter of the dataset. It is likely
that the Phillies would like to improve their offensive production during the 2014–15 off-season
by signing free-agents or getting players by trade who appear to be effective in getting on-base.

2.2 A Tribute to Derek Jeter

Topics Covered: Dotplot, stemplot, time series plot, fitted line.
The 2014 baseball season was memorable for the retirement of one of baseball’s most
popular players, Derek Jeter. Jeter played shortstop for the New York Yankees for a 20-year
period. Table 2.2 displays Jeter’s batting statistics for his first six major league seasons. For all
of the seasons of Jeter’s career, we’ll focus on a couple of interesting statistics—the number

Table 2.2. Derek Jeter’s batting statistics for the first six seasons of his MLB career
Year AB R H 2B 3B HR BB SO AVG OBP SLG OPS
1995 48 5 12 4 1 0 3 11 .250 .294 .375 .669
1996 582 104 183 25 6 10 48 102 .314 .370 .430 .800
1997 654 116 190 31 7 10 74 125 .291 .370 .405 .775
1998 626 127 203 25 8 19 57 119 .324 .384 .481 .864
1999 627 134 219 37 9 24 91 116 .349 .438 .552 .989
2000 593 119 201 31 4 15 68 99 .339 .416 .481 .896

of home runs (HR) and the OPS statistic that is a good estimate of a player’s overall hitting
ability. In the following, we will not include the statistics for his 1995 rookie and his 2013 injury
seasons since he only had 48 and 63 at-bats, but we’ll analyze his batting data for the remaining
18 seasons.

Jeter’s Home Runs

Scanning over Jeter’s hitting statistics, we see that he displayed some power and hit a good
number of home runs in his career. We graph his season by season home runs using a dotplot in
Figure 2.5.

5 10 15 20
Home Runs

Figure 2.5. Dotplot of season home run numbers for Derek Jeter.

There appears to be much variation in these home run numbers—we see from the graph
that the numbers range from 4 to 24. There is one cluster of values in the 10–15 range. The
median number of home runs hit is 14:5 which seems to be a reasonable measure of the middle
of this home run distribution.
Maybe some of the variation of Jeter’s home run numbers can be explained by the age at
which he hit them. A ballplayer generally improves in hitting ability in the early part of his
career and declines in ability towards the end of his career. In Figure 2.6, we graph the home
run count against the year.
15 20
HR
10
5

2000 2005 2010

Year

Figure 2.6. Time series plot of home run numbers for Derek Jeter.

We see that Jeter’s home run numbers were in the 15–20 range during seasons around 2000,
but it appears that his season home run counts generally decreased over the later years of his

career. We can summarize this decrease by drawing a line through the points in Figure 2.7. (We
will discuss the use of one “best fitting” line in Chapter 4.)

1520
HR
10
5

2000 2005 2010

Year

Figure 2.7. Time series plot of home run numbers for Derek Jeter with a “good fitting” line drawn on top.

The equation of this line is

HR D 883:68 0:4337 Year

So Jeter’s home run count tended to decrease about .4 for each year of his career. There were
some exceptions to this general pattern such as his 18 home runs hit in 2009 and his 15 home
runs in 2012.

Jeter’s Season OPS Values

Although Jeter’s home runs decreased during his career, it’s not clear that his hitting ability
changed in a similar way. After all, good hitting is more than just hitting home runs. As explained
in Chapter 1, a good estimate of hitting effectiveness is the OPS statistic. In Figure 2.8 we display
Jeter’s season OPS values for his 18 seasons using a stemplot.

Figure 2.8. Stemplot of Jeter’s season OPS values.

From the stemplot, we see that Jeter’s OPS values are roughly bell-shaped centered about
:83. Most of his season values fall between :770 and :890 with a few large extreme values. To
see how Jeter’s OPS values changed over the seasons, we construct the scatterplot in Figure 2.9
and add a smoothing curve to help identify the main pattern in the plot.

1.0
0.9
OPS
0.8
0.7

2000 2005 2010

Year

Figure 2.9. Time series plot of Jeter’s OPS numbers

The general message from this graph is that Jeter’s batting performance as measured by OPS
was very consistent during the period 1996 through 2008, but then his performance dropped off
in the final seasons of his career. He had an unusually high value in 1999 when his OPS was
0:989. Since OPS values in the .0:8; 0:9/ range were high (relative to other hitters), this figure
reinforces the impression that Jeter was a consistently good hitter during his whole career.

2.3 A Tribute to Randy Johnson

Topics Covered: Stemplot, time series plot, summary statistics, comparison of distributions.
Randy Johnson was one of the greatest strikeout pitchers in modern baseball. He had a
lifetime win/loss record of 303-166 and he won the Cy Young award as the best pitcher in
the league for the 1995, 1999, 2000, 2001, and 2002 seasons. (In these seasons, a Cy Young
award was given in each league.) We collect Johnson’s pitching statistics for his 22 seasons in
Major League Baseball; the statistics for the first six seasons are presented in Table 2.3. (One
should note an unusual feature of innings pitched (IP) data typically presented. In 1989, Johnson
pitched for 160 and 2=3 of an inning—on Baseball-Reference this is represented as 160:2, but
here we represent it as the more traditional decimal representation of 160:67.)

Table 2.3. Pitching statistics for Randy Johnson for the first six seasons of his career
Season W L G IP H R SO BB ERA
1988 3 0 4 26 23 8 25 7 2.42
1989 7 13 29 160.67 147 100 130 96 4.82
1990 14 11 33 219.67 174 103 194 120 3.65
1991 13 10 33 201.33 151 96 228 152 3.98
1992 12 14 31 210.33 154 104 241 144 3.77
1993 19 8 35 255.33 185 97 308 99 3.24

Johnson’s Strikeouts
The strikeout numbers, 25, 130, and so on, are hard to interpret since the number of innings
pitched changes across seasons. A reasonable measure of strikeout ability that adjusts for the
number of innings is the strikeout rate defined by
SO
SO Rate D 9 :
IP
If we divide the count of strikeouts by the innings pitched, we get the number of strikeouts
per inning. By multiplying the ratio SO/IP by 9, we get the number of strikeouts for a standard
9-inning game. A strikeout rate of 9 is a useful reference value, since it means that the pitcher
struck out a batter per inning. (This particular rate value is quite rare.) Since there are 27 outs
for a team during a game, a strikeout rate of 9 means that one third of the outs were strikeouts. If
we compute the strikeout rate for all of Johnson’s seasons, we get the table shown in Table 2.4.
A stemplot of the rates is shown in Figure 2.10.

Table 2.4. Strikeout rates for Randy Johnson

Season SO.Rate Season SO.Rate
1988 8.65 1999 12.06
1989 7.28 2000 12.56
1990 7.95 2001 13.41
1991 10.19 2002 11.56
1992 10.31 2003 9.87
1993 10.86 2004 10.62
1994 10.67 2005 8.41
1995 12.35 2006 7.55
1996 12.47 2007 11.43
1997 12.30 2008 8.46
1998 12.12 2009 8.06

Figure 2.10. Stemplot of season strikeout rates for Randy Johnson.

Looking at the stemplot, we see that an average strikeout rate for Johnson is about 10.6.
That means that he strikes out about one batter per inning. We see two clusters of strikeout rates.
For a majority of his seasons, Johnson’s strikeout rate was in the 10–12 range, although there
were seven seasons where his rate was smaller than 9. To see when these high and low strikeout
seasons occurred, we plot the rates against the season year in Figure 2.11. We place a horizontal
line on our graph corresponding to the reference strikeout rate of 9.
13
12
SO.Rate
11
10
9
8

1990 1995 2000 2005

Year

Figure 2.11. Time series plot of Johnson’s season strikeout rates.

There is an interesting pattern in this plot of strikeout rates. We see that Johnson’s strikeout
rate exhibited a steady increase at the beginning of his career, hit a peak of about 12 around the
year 1998, and then decreased steadily until his retirement in 2009. This is a common pattern
of performance of baseball players.

Johnson’s Win/Loss Record in 1995

Randy Johnson had a remarkable win/loss record in 1995—he won 18 and lost only two games.
For this reason (and others), Johnson won the Cy Young award for the best pitcher in the
American League in 1995. But was Johnson really the best AL pitcher in 1995?
One problem with a pitcher’s win/loss record is that winning and losing is really a team
accomplishment. A team will win games if it scores more runs than its opponent. This requires
good pitching that allows few runs and good hitting that produces runs. It is possible that
Johnson won so many games in 1995 because his team, the Mariners, scored a lot of runs when
he was pitching. To see if this is true, we record the number of runs scored by the Mariners for
every one of the 30 games that Johnson started in the 1995 season. The run numbers are shown
below. We see that the Mariners scored three runs in the first game that Johnson started, 15 runs
in the second game he started, and so on.

3 15 3 6 4 5 8 11 2 3 1 9 3 5 4
5 3 4 8 2 2 6 6 7 4 7 8 7 6 9

We display the runs scored for Johnson in Figure 2.12 using a stemplot.

We see that the distribution of runs scored is right skewed. We see that most of the runs
scored are in the 3–7 range—it was unusual for the Mariners to score 0 or 1 runs, or to score
more than 9 runs. The median runs scored for Johnson was 5 and the mean was 5:53 runs.

Figure 2.12. Stemplot of runs scored by Figure 2.13. Back-to-back stemplots of runs scored
the Mariners during games in which by the 1995 Mariners for games where Randy John-
Johnson started in 1995. son was the starter (LEFT) and games where Johnson
was not the starter (RIGHT).

How does this distribution compare to the runs scored by the Mariners in the 1995 games
where Johnson was not a starter? Figure 2.13 displays back-to-back stemplots of the runs scored
for the two groups of games.
Comparing the two stemplots in Figure 2.13, we see that the Mariners tended to score about
the same number of runs in games in which Johnson did not start than they did in games in
which Johnson started. The median and mean runs scored for Mariners games when Johnson
was not starting are 5 and 5:48, respectively. If we compare medians or means of the two groups,
we see that the Mariners tended to average the same amount of runs when Johnson was starting
or when Johnson was not starting. So Johnson did not receive unusually good run support from
his team.

2.4 Analyzing Baseball Attendance

Topics Covered: Histogram, stemplot, data distribution.
Baseball is a business. Each of the thirty major league teams is privately owned and each
wants to make a profit. Much of the revenue produced by the teams comes from the money that
is made from ticket and concession sales. So it is desirable for a team to have high attendance at
the games played at its ballpark. We will see in this study that some teams are very successful
and other teams are not successful in getting fans to come to their games.

Table 2.5 shows the mean attendance at the home ballpark for all thirty major league teams
in 2014. We note that the teams who played in the 2014 World Series, San Francisco (SFN) and
Kansas City (KCA), had mean attendance of 41,589 and 24,154 per game, respectively. Are
these large or small values? We can answer this question by looking at the distribution of mean
attendance for all teams.

Table 2.5. Mean home attendance for all Major League Baseball teams in 2014
Team Attendance Team Attendance Team Attendance
ANA 38,221 DET 36,015 PHI 29,924
ARI 25,602 HOU 21,628 PIT 30,155
ATL 29,065 KCA 24,154 SDN 27,103
BAL 30,426 LAN 46,696 SEA 25,486
BOS 36,495 MIA 21,386 SFN 41,589
CHA 20,381 MIL 34,536 SLN 43,712
CHN 32,742 MIN 27,785 TBA 17,858
CIN 30,576 NYA 41,995 TEX 33,565
CLE 17,746 NYN 26,528 TOR 29,327
COL 33,090 OAK 24,736 WAS 31,844

We first construct a histogram of the attendance for all teams. We chose to group the data
using the bins .15;000; 20;000; .20;000; 25;000; : : : ; .45;000; 50;000 and count the number
of values in each bin. The graph of the grouped data, called a histogram, is shown in Figure 2.14.
8
6
Frequency
4
2
0

15000 20000 25000 30000 35000 40000 45000 50000

Mean Attendance

Figure 2.14. Histogram of average attendance for the Major League teams in 2014.

There are a number of interesting features about these data that we see from the histogram.

There is a large range in the average attendance, from about 15;000 to 50;000.
The shape of this distribution is symmetric about a typical average attendance of 30;000.
There are two teams that had small average attendance under 20;000.

To look further at these teams’ attendance, we construct a stemplot displayed in Figure 2.15.
In addition, we construct a Cleveland-style dotplot in Figure 2.16—this is a display of labeled
data where one lists the team names on the vertical axis and plots attendance values on the
horizontal scale.

Figure 2.15. Stemplot of average attendance of the Major League teams in 2014.

LAN
SLN
NYA
SFN
ANA
BOS
DET
MIL
TEX
COL
CHN
WAS
CIN
BAL
PIT
PHI
TOR
ATL
MIN
SDN
NYN
ARI
SEA
OAK
KCA
HOU
MIA
CHA
TBA
CLE

20000 25000 30000 35000 40000 45000

Attendance

Figure 2.16. Cleveland-style dotplot of average attendance of the Major League teams in 2014.

The graphs in Figure 2.15 and Figure 2.16 give us additional insight about the attendance
averages. The two weak attendance teams are Cleveland (CLE) and Tampa Bay (TBA). The
team with the highest average attendance in 2014 was the Los Angeles Dodgers who had a
successful 94-68 record in this season. But winning games and attendance don’t always go
hand in hand. Cleveland (CLE) had a 85-77 winning record but a small average attendance, but
Colorado (COL) and Texas (TEX), two teams with weak 66-96 and 67-95 records, had large
average attendance. It would be interesting for a future study to look into the impact of a number

of variables, such as the population of the surrounding region and the newness of the ballpark,
on the average home attendance of a baseball team.

2.5 Manager Statistics: the Use of Sacrifice Bunts

Topics Covered: Dotplot, comparison of distributions.
In this chapter, we’ve looked at statistics for teams and for individual players. Here we’ll look
at statistics for baseball managers. This might surprise you—how can we measure characteristics
of a manager? Managers do have different styles or strategies that they use in managing a game,
and we can contrast managers by measuring how often particular strategies are used.
Here we focus on the use of the sacrifice bunt, one of the more popular strategic plays in
baseball. Suppose there is a runner on first base with 0 or 1 out. Then the manager may elect to
have the batter bunt, which is a short hit in the infield. If the bunt is successful, the runner on
first advances to second. If the batter is out, this play is called a sacrifice bunt because the batter
is sacrificing his at-bat to advance the runner to second. Is the sacrifice bunt a good strategy? It
can be a good strategy when there is a relatively weak batter at the plate or the objective is to
score a single run.
Table 2.6 gives the number of sacrifice bunts for all 30 Major League teams in the 2013
season.

Table 2.6. Successive count of sacrifice hits for all Major League
teams in 2013
Team SH League Team SH League
ANA 37 AL MIL 77 NL
ARI 50 NL MIN 29 AL
ATL 58 NL NYA 36 AL
BAL 27 AL NYN 53 NL
BOS 24 AL OAK 21 AL
CHA 19 AL PHI 57 NL
CHN 43 NL PIT 62 NL
CIN 85 NL SDN 52 NL
CLE 31 AL SEA 26 AL
COL 65 NL SFN 66 NL
DET 32 AL SLN 56 NL
HOU 46 NL TBA 24 AL
KCA 37 AL TEX 45 AL
LAN 71 NL TOR 29 AL
MIA 57 NL WAS 68 NL

The dotplot in Figure 2.17 displays the number of sacrifice bunts for all teams. The median
number of “sac-bunts” was 45:5. But we note a wide spread—the Chicago White Sox only had
19 sacrifice bunts all season and other teams sacrificed over 70 times (in a 162 game season).
How can we explain the wide spread in the number of sacrifice bunts between teams?
Perhaps some managers don’t think that the sacrifice bunt is a good strategy. Or possibly a

20 30 40 50 60 70 80
Number of Sacrifice Bunts

Figure 2.17. Dotplot of number of sacrifice bunt attempts by the Major League managers in the 2013
season.

manager of a team with strong pitching uses a sacrifice more often than a manager of a team
with weak pitching.
Is it possible that the number of sacrifice bunts differs between National League and Amer-
ican League teams? We can check this by constructing two parallel dotplots in Figure 2.18—one
for the NL teams and a second for the AL teams.
NL
AL

20 30 40 50 60 70 80
Number of Sacrifice Bunts

Figure 2.18. Parallel dotplots of number of sacrifice bunt attempts by National and American League
managers.

We see in the Figure 2.18 display that there is quite a difference in the number of sacrifice
bunts attempted in the two leagues. The median number of attempts for NL teams is 57:5 (about
1 in every 3 games) contrasted with 29 (about 1 in every 5 games) for the AL teams. There
actually is one simple explanation for this discrepancy in sacrifice bunts—the designated hitter
rule. Because National League pitchers (typically weak hitters) come to bat, it is common for
them to attempt a sacrifice bunt when there is a runner on first base. They do this to avoid a
double play and to advance the runner to second. In contrast, in the AL, the pitcher doesn’t bat
(the designated hitter does instead) and there is much less reason to perform this strategy.
The league policy does explain many of the differences in attempts that we see. However,
even within one league, we still see large differences in sacrifice bunt attempts. It is interesting
to note that the Chicago Cubs 43 sac-bunt attempts is a small value even compared to other NL
teams, and the NL values ranged between 52 and 85. Why do these differences exist? Well, the
number of sacrifice bunts attempted depends partly on the batting abilities of the players and
the beliefs of the manager regarding the value of the sacrifice bunt. It would be interesting to
see if there is any relationship between the number of sacrifice bunts and other batting statistics.

Also, one could see if particular managers have tendencies over many years to attempt a large
or small number of sacrifice bunts.

2.6 Exercises
2.0. Table 2.7 displays the slugging percentages (SLGs) for Rickey Henderson for his first
23 years in the major leagues.

Table 2.7. Seasonal slugging percentages for Rickey Henderson

AGE SLG AGE SLG AGE SLG AGE SLG
20 .336 26 .516 32 .423 38 .342
21 .399 27 .469 33 .457 39 .347
22 .437 28 .497 34 .474 40 .466
23 .382 29 .399 35 .365 41 .305
24 .421 30 .399 36 .447 42 .351
25 .458 31 .577 37 .344

(a) Construct a stemplot of the slugging percentages.

(b) Compute a five-number summary of the SLGs. What was a representative slugging
percentage for Henderson?
(c) Construct a time series plot of the SLG against his age. Comment on any pattern
that you see in this plot.
(d) Based on your work in (c), do you have any explanations for the low slugging
percentages that Henderson had in his career?
2.1. Table 2.8 gives the career batting statistics for the great Yankee player Joe DiMaggio.

Table 2.8. Career Statistics for Joe DiMaggio

Year G AB H HR AVG OBP SLG
1936 138 637 206 29 0.323 0.352 0.576
1937 151 621 215 46 0.346 0.412 0.673
1938 145 599 194 32 0.324 0.386 0.581
1939 120 462 176 30 0.381 0.448 0.671
1940 132 508 179 31 0.352 0.425 0.626
1941 139 541 193 30 0.357 0.440 0.643
1942 154 610 186 21 0.305 0.376 0.498
1946 132 503 146 25 0.290 0.367 0.511
1947 141 534 168 20 0.315 0.391 0.522
1948 153 594 190 39 0.320 0.396 0.598
1949 76 272 94 14 0.346 0.459 0.596
1950 139 525 158 32 0.301 0.394 0.585
1951 116 415 109 12 0.263 0.365 0.422

(a) Construct a dotplot of DiMaggio’s yearly batting averages on the number line below.
+---------+---------+---------+---------+---------+-AVG
.200 .240 .280 .320 .360 .400
(b) Describe the basic features of this distribution (shape, average value, spread, and
unusual values).
(c) What proportion of the years was DiMaggio at least a .300 hitter?
(d) Construct a dotplot of DiMaggio’s yearly hits on the number line below.
+---------+---------+---------+---------+---------+-H
100 120 140 160 180 200
(e) What was a median number of yearly hits for DiMaggio? Were there any unusually
small or large hit values? What is a possible explanation for these unusual values?
2.2. (Continuation of Exercise 2.1)
(a) Construct a time series plot of DiMaggio’s AVG values.
(b) From the graph, DiMaggio appeared to peak in AVG two times during his career.
Which two years?
(c) From the graph, did you detect any gaps in DiMaggio’s career? What is a possible
explanation for these gaps?
(d) Generally, ballplayers mature slowly and reach their maximum performance during
the middle of their careers. Do you see any evidence for this maturation in DiMaggio’s
AVG plot? Explain.
(e) Also, ballplayers generally decline in ability during the last part of their career. Do
you see any decline in DiMaggio’s AVG plot? Explain.
2.3. (Joe DiMaggio data from Exercise 2.1)
(a) Construct a stemplot of DiMaggio’s on-base percentages (OBP) for his 13 seasons.
(b) Compute DiMaggio’s mean OBP and his median OBP.
(c) Compare the mean and median that you computed in (b). Which is a better measure
of a representative season OBP for DiMaggio?
2.4. (Joe DiMaggio data from Exercise 2.1)
A stemplot of DiMaggio’s season HR numbers is shown below:

1 | 24
1 |
2 | 01
2 | 59
3 | 00122
3 | 9
4 |
4 | 6

(a) Find a five-number summary of the HR numbers.

(b) Based on your calculations in part (a), half of all the HR numbers are smaller
than and approximately the middle 50 percent of the season HRs are between
and .

2.5. Table 2.9 shows the career hitting statistics for Babe Ruth, who was considered by ESPN
to be the second greatest athlete of the 20th century (after Michael Jordan).

Table 2.9. Career Statistics for Babe Ruth

Year G AB H HR AVG OBP SLG
1914 5 10 2 0 0.200 0.200 0.300
1915 42 92 29 4 0.315 0.376 0.576
1916 67 136 37 3 0.272 0.322 0.419
1917 52 123 40 2 0.325 0.385 0.472
1918 95 317 95 11 0.300 0.411 0.555
1919 130 432 139 29 0.322 0.456 0.657
1920 142 457 172 54 0.376 0.533 0.849
1921 152 540 204 59 0.378 0.512 0.846
1922 110 406 128 35 0.315 0.434 0.672
1923 152 522 205 41 0.393 0.545 0.764
1924 153 529 200 46 0.378 0.513 0.739
1925 98 359 104 25 0.290 0.393 0.543
1926 152 495 184 47 0.372 0.516 0.737
1927 151 540 192 60 0.356 0.486 0.772
1928 154 536 173 54 0.323 0.463 0.709
1929 135 499 172 46 0.345 0.430 0.697
1930 145 518 186 49 0.359 0.493 0.732
1931 145 534 199 46 0.373 0.495 0.700
1932 133 457 156 41 0.341 0.489 0.661
1933 137 459 138 34 0.301 0.442 0.582
1934 125 365 105 22 0.288 0.448 0.537
1935 28 72 13 6 0.181 0.359 0.431

(a) Construct a dotplot of the season home run numbers for Babe Ruth.
(b) Comment on the basic shape of the distribution of home run numbers.
(c) Can you think of some possible reasons why Ruth had so many seasons with a small
number of home runs?
(d) Since the number of at-bats is not constant across seasons, it may be better to consider
the home run rate, which is obtained by dividing the number of home runs by the
number of at-bats. So, for example, Ruth’s home run rate in 1915 is given by

4
HOME RUN RATE D D :0435:
92

The home run rates for Ruth in his 22 seasons are given below:
0 .043 .022 .016 .035 .067 .118 .109
.086 .079 .087 .070 .095 .111 .101 .092
.095 .086 .090 .074 .060 .083
(e) Construct a dotplot of the home run rates.

(f) Summarize the key features of the home run rates. Explain why it is better to look at
the home run rates instead of the home run numbers.
2.6. (Continuation of Exercise 2.5) Consider again the home rate rates for Babe Ruth displayed
in Exercise 2.5.
(a) Graph the home run rates with a time series plot.
(b) What pattern do you see in the graph you made in (a)? Draw a smooth curve over
the points.
(c) Using the smooth curve you constructed in (b), what was Babe Ruth’s peak year with
respect to home run rate? Is this the same year as the year when he hit the greatest
number of home runs?
(d) Repeat parts (a)–(c) using Ruth’s batting average statistic.
2.7. (Babe Ruth data from Exercise 2.5)
(a) Construct a dotplot of Ruth’s season SLG values.
(b) Find the mean and median SLG values.
(c) Which average computed in (b) seems to be most representative of Ruth’s slugging
ability? Explain.
(d) In the table of Ruth’s data, Ruth’s career slugging percentage is :69. Is this the same
as the mean SLG value that you computed in (b)? If the numbers are different, can
you explain why?
2.8. (Babe Ruth data from Exercise 2.5)
A stemplot of Ruth’s season batting averages is shown below. (Here the stem is the first
digit, so a batting average of :315 would have a stem of 3 and a leaf of 1.)
1 | 8
2 | 0
2 |
2 |
2 | 7
2 | 89
3 | 0011
3 | 222
3 | 4455
3 | 77777
3 | 9

(a) Compute a five-number summary.

(b) Looking back at the table of data in Exercise 2.3, find the years where Ruth’s batting
average was above the median.
(c) Find the years where Ruth’s average was below the median.
(d) Based on (b) and (c), do you think the median is a representative AVG for Ruth?
(e) In the table of Ruth’s data, Ruth’s career AVG is equal to :342. Compare this value
to the median and explain why they are different.
2.9. Table 2.10 shows the career pitching statistics for Hall of Famer Sandy Koufax. (Koufax
is the only pitcher to make ESPN’s list of the top 50 Greatest Athletes of the 20th century.)

Table 2.10. Pitching Statistics for Sandy Koufax

Year G W L IP H SO BB ERA
1955 12 2 2 41.70 33 30 28 3.02
1956 16 2 4 58.70 66 30 29 4.91
1957 34 5 4 104.30 83 122 51 3.88
1958 40 11 11 158.70 132 131 105 4.48
1959 35 8 6 153.30 136 173 92 4.05
1960 37 8 13 175.00 133 197 100 3.91
1961 42 18 13 255.70 212 269 96 3.52
1962 28 14 7 184.30 134 216 57 2.54
1963 40 25 5 311.00 214 306 58 1.88
1964 29 19 5 223.00 154 223 53 1.74
1965 43 26 8 335.70 216 382 71 2.04
1966 41 27 9 323.00 241 317 77 1.73

(a) Construct a stemplot of the season ERA values for Koufax (the breakpoint between
the stem and the leaf will be at the decimal point).
(b) There are two clusters in this dataset. Identify the two clusters and explain which
season years are identified with the two clusters.
(c) Construct a stemplot of Koufax winning percentages (PCT). Describe the features
of this dataset. Would it be accurate to say that Koufax was a successful pitcher?
Why? What proportion of seasons did Koufax have a winning record?
2.10. (Sandy Koufax data from Exercise 2.9.)
(a) Draw a stemplot with five leaves per stem for the season strikeout counts. Describe
the main features of this dataset. Are there any SO counts that appear unusually
small or large? Is there any explanation for these unusual values?
(b) Find the mean and median SO number.
(c) Which average in (b) is the best measure of the center of the distribution? Explain.
(It might help to look at the stemplot of the season strikeout counts from Exercise
2.9 (c).)
2.11. (Sandy Koufax data from Exercise 2.9)
(a) Compute the mean and median of Koufax season ERAs.
(b) Suppose Koufax played an additional season and his ERA was 6.5. Compute the
mean and median of Koufax’s ERAs with this additional data value.
(c) Which average (mean or median) changed the most when you added this new ERA
value? Explain why this average changed.
(d) Koufax’s career ERA was 2.76. Explain why this career ERA is different from the
averages you computed in part (a).
2.12. Table 2.11 gives career pitching statistics for the Hall of Famer Bob Feller. Note that there
are no statistics given for the three-year period 1942–1944—Feller served in the Navy
during World War II between 1942 and 1945.

Table 2.11. Pitching statistics for Bob Feller

Year G W L IP H BB SO ERA
1936 14 5 3 62.00 52 47 76 3.34
1937 26 9 7 148.67 116 106 150 3.39
1938 39 17 11 277.67 225 208 240 4.08
1939 39 24 9 296.67 227 142 246 2.85
1940 43 27 11 320.33 245 118 261 2.61
1941 44 25 13 343.00 284 194 260 3.15
1945 9 5 3 72.00 50 35 59 2.50
1946 48 26 15 371.33 277 153 348 2.18
1947 42 20 11 299.00 230 127 196 2.68
1948 44 19 15 280.33 255 116 164 3.56
1949 36 15 14 211.00 198 84 108 3.75
1950 35 16 11 247.00 230 103 119 3.43
1951 33 22 8 249.67 239 95 111 3.50
1952 30 9 13 191.67 219 83 81 4.74
1953 25 10 7 175.67 163 60 60 3.59
1954 19 13 3 140.00 127 39 59 3.09
1955 25 4 4 83.00 71 31 25 3.47
1956 19 0 4 58.00 63 23 18 4.97

(a) Construct a stemplot of Feller’s strikeout counts for his 18 seasons. What is the basic
shape of this distribution? Are there any unusually small or large SO counts?
(b) It is difficult to tell which years Feller was especially good at striking out batters
since the innings pitched (IP) is not constant across years. To adjust for the innings
pitched, one can compute the strikeout ratio, the number of strikeouts divided by the
innings pitched. Here are the strikeout ratios for Feller’s 18 seasons:
1.23 1.01 .87 .83 .82 .76 .82 .94 .66
.59 .51 .48 .45 .42 .34 .42 .30 .31
Graph these SO ratios against year number. Describe any pattern that you see in
the graph. What does that tell you about Feller’s ability to strike out batters over
time?
(c) Use Feller’s lifetime average number of wins, losses, and strikeouts to fill in the
missing years 1942–1944 and the partial year 1945. What are his new career totals
in these categories?
2.13. (Bob Feller data from Exercise 2.12)
(a) Construct a dotplot of Feller’s strikeout ratios for the 18 seasons. (These ratios are
listed in part (b) of Exercise 2.12.)
(b) Compute the mean and median strikeout ratio. Explain why the mean is larger than
the median for this dataset.
2.14. Figure 2.19 displays a histogram of the on-base percentages (OBP) for the 944 MLB
players who batted during the 2014 baseball season.

300
250
Frequency
200
150
100
50
0

0.0 0.2 0.4 0.6 0.8 1.0

OBP

Figure 2.19. Histogram of OBPs of 2014 MLB players.

(a) Describe the basic shape of the histogram and any unusual features.
(b) What percentage of players had an OBP value between :3 and :4?
(c) What percentage of players had an OBP smaller than :2?
(d) There were a few players that had an OBP value of 1 during the 2014 season. (This
means that they reached base 100% of the time.) This may seem surprising—can
you offer any possible explanation for these large values?
2.15. (Exercise 2.14 continued.) If one graphs the OBPs for only those National League hitters
who had at least 300 at-bats (AB), one obtains the histogram in Figure 2.20. (In the
following, we will refer to the players with at least 300 AB as the “regulars.”)
60
50
Frequency
40
30
20
10
0

0.25 0.30 0.35 0.40

OBP

Figure 2.20. Histogram of OBPs of 2014 MLB players with at least 300 at-bats.

(a) Describe the shape of this histogram. Why does this histogram look so different from
the histogram of the OBP of all NL players in Figure 2.18?
(b) What percent of NL regulars had an OBP value exceeding :3?
(c) What percent of NL regulars had an OBP value between :3 and :4?
(d) From the histogram, what OBP value would you consider outstanding? Why?
2.16. Figure 2.21 shows a histogram of the number of home runs for all regular players in the
2014 season with at least 300 at-bats.
(a) Describe the basic shape of the histogram and any unusual characteristics.
(b) What percentage of NL regular players hit fewer than ten home runs?
(c) What percentage of NL regular players hit more than 20 home runs?
(d) How many home runs did an NL regular player hit, on average?
70
60
50
Frequency
40
30
20
10
0

0 10 20 30 40
HR

Figure 2.21. Histogram of count of home runs of 2014 MLB players with at least 300 at-bats.

2.17. (Exercise 2.16 continued.) One explanation for the wide variation of home run totals
shown in Figure 2.21 is that players had different numbers of at-bats. If we divide a
player’s home run count by his number of at-bats, one obtains a player’s home run rate.
(A rate of :08 means a batter hit a home run in 8% of his at-bats.) Figure 2.22 shows a
histogram of the home run rates for the 2014 MLB regulars.
(a) Describe the basic shape of this distribution of home run rates. Are there any unusual
values?
(b) What is an average home run rate among these NL regulars?
(c) If a batter comes to bat 100 times, how many home runs do you expect him to hit?
(d) What percent of NL regulars had a home run rate under :02?
(e) A hitter is considered to be an unusually good home run hitter if his rate of hitting
home runs exceeds 10% or :1. What percent of regular NL hitters fall into this
classification? Can you guess which players are in this class?

60
50
40
Frequency
30
20
10
0

0.00 0.02 0.04 0.06 0.08

HR Rate

Figure 2.22. Histogram of count of home run rates of 2014 MLB players with at least 300 at-bats.

2.18. Table 2.12 displays the length (in minutes) of a sample of 60 baseball games in 2014.

Table 2.12. Length in minutes of a selection of 2014 baseball games

169 181 178 212 159 213 156 197 201 217
241 195 165 194 163 175 180 242 159 190
169 184 227 166 259 182 355 201 199 206
197 216 196 144 202 169 182 191 181 196
173 205 185 154 198 165 175 221 187 193
155 176 171 218 185 203 222 194 215 205

(a) Construct a stemplot of these game times.

(b) What is the general shape of the distribution of times? Are there any unusually short
or long games?
(c) What is a median length of a baseball game?
(d) What fraction of games are over three hours?
(e) Why is there such a large spread of times? Can you think of some variables that
might influence the length of a baseball game?
2.19. (Length of baseball games from Exercise 2.18.)
A stemplot of the times of games (in minutes) is shown below.

1 | 2: represents 12
leaf unit: 1
n: 60
14 | 4
15 | 45699
16 | 3556999
17 | 135568

18 | 011224557
19 | 013445667789
20 | 1123556
21 | 235678
22 | 127
23 |
24 | 12
25 | 9
HI: 355

(a) Find the five-number summary of the lengths of the 2014 baseball games.
(b) Approximately the middle half of the game lengths is between and .
(c) The mean and standard deviation of the game lengths are 193:5 minutes and 31:6
minutes, respectively. (The standard deviation is a measure of spread defined in
Chapter 3.) Find an interval defined by (mean standard deviation, mean C standard
deviation) .
(d) Find the proportion of games that fall in the interval you found in part (c).
(e) If the data is bell-shaped, then one would expect 68% of the data to fall in the interval
in (c). Here you should find that the proportion in (d) is significantly larger than :68.
Can you explain why?
2.20. For each of 60 baseball games in 2014, the total number of home runs was recorded—the
numbers are shown in Table 2.13.

Table 2.13. Number of home runs hit in a sample

of 2014 baseball games
1 1 2 1 3 1 1 0 1 0
1 0 0 0 3 2 2 1 1 2
1 1 5 1 1 1 1 1 2 1
2 4 1 0 1 0 0 5 3 3
2 3 3 4 2 0 0 4 3 3
2 1 2 2 1 1 4 1 3 2

(a) Construct a frequency table for the number of home runs.

(b) Graph the frequencies using a bar chart.
(c) Find the proportion of games where exactly two home runs are hit.
(d) Find the proportion of games where four or more home runs are hit.
(e) What is an average number of home runs hit during a game? Explain how you
computed this average number.
2.21. Table 2.14 displays basic batting statistics for the 2014 American League teams.
(a) Draw a stemplot of the batting averages (AVG) for the 15 AL teams.
(b) Find the median batting average and find a team that has a batting average close to
the median value.
(c) Find the team that has the highest batting average and the team with the lowest
batting average.

Table 2.14. Batting statistics for the 2014 American League teams
Team AVG OBP SLG AB R H 2B 3B HR
Baltimore 0.256 0.311 0.422 5596 705 1434 264 16 211
Boston 0.244 0.316 0.369 5551 634 1355 282 20 123
Chicago 0.253 0.310 0.398 5543 660 1400 279 32 155
Cleveland 0.253 0.317 0.389 5575 669 1411 284 23 142
Detroit 0.277 0.331 0.426 5630 757 1557 325 26 155
Houston 0.242 0.309 0.383 5447 629 1317 240 19 163
Kansas City 0.263 0.314 0.376 5545 651 1456 286 29 95
Los Angeles 0.259 0.322 0.406 5652 773 1464 304 31 155
Minnesota 0.254 0.324 0.389 5567 715 1412 316 27 128
New York 0.245 0.307 0.380 5497 633 1349 247 26 147
Oakland 0.244 0.320 0.381 5545 729 1354 253 33 146
Seattle 0.244 0.300 0.376 5450 634 1328 247 32 136
Tampa Bay 0.247 0.317 0.367 5516 612 1361 263 24 117
Texas 0.256 0.314 0.375 5460 637 1400 260 28 111
Toronto 0.259 0.323 0.414 5549 723 1435 282 24 177

(d) For a second statistic of your choice, construct a stemplot. Find a team that has an
average value of the statistic. Also find a team that has the smallest value, and a team
that has the largest value of the statistic that you chose.
2.22. (Batting statistics for 2014 AL teams from Exercise 2.21)
(a) Draw a stemplot of the HR numbers for the 15 AL teams.
(b) Find the five-number summary.
(c) Approximately percent of the HR numbers are smaller than 155.
(d) Approximately percent of the HR numbers are larger than 146.
(e) Find a team that has an HR number that is close to the average.
2.23. Table 2.15 gives the year of birth and earned run average (ERA) for all 57 pitchers who
have been inducted into the Baseball Hall of Fame.
(a) Draw a stemplot of the ERAs of these Hall of Fame pitchers.
(b) Discuss the general shape of the distribution of ERAs and any unusual characteristics
of this data.
(c) Find an average ERA and a Hall of Fame pitcher who has (approximately) this
average ERA.
(d) Find the pitchers who have the lowest and highest ERAs.
(e) What might explain this wide range of ERAs if all of these pitchers are in the Hall
of Fame?
2.24. (Exercise 2.23 continued.) Consider the years of birth of the pitchers in the Hall of Fame.
(a) Construct a frequency table of the years of birth, using the intervals in the table
below. Place the counts and the proportions in the table.
(b) Construct a histogram from the above frequency table.
(c) Describe the basic shape of the distribution.

Table 2.15. Year of birth and ERA for pitchers who have been selected in the Baseball
Hall of Fame
Pitcher Birthyear ERA Pitcher Birthyear ERA
Grover Alexander 1887 2.56 Sandy Koufax 1935 2.76
Charles Bender 1884 2.46 Bob Lemon 1920 3.23
Mordecai Brown 1876 2.06 Juan Marichal 1937 2.89
Jim Bunning 1931 3.27 Richard Marquard 1889 3.08
Steve Carlton 1944 3.22 Christy Mathewson 1880 2.13
Jack Chesbro 1874 2.68 Joe McGinnity 1871 2.66
John Clarkson 1861 2.81 Hal Newhouser 1921 3.06
Stan Coveleski 1889 2.89 Charles Nichols 1869 2.95
W.A. Cummings 1848 2.78 Phil Niekro 1939 3.35
Jay Hanna Dean 1911 3.02 Satchel Paige 1906 3.29
Don Drysdale 1936 2.95 Jim Palmer 1945 2.86
Urban Faber 1888 3.15 Herb Pennock 1894 3.60
Bob Feller 1918 3.25 Gaylord Perry 1938 3.11
Rollie Fingers 1946 2.90 Eddie Plank 1875 2.35
Edward Ford 1928 2.75 Eppa Rixey 1891 3.15
Rube Foster 1888 2.36 Robin Roberts 1926 3.41
James Galvin 1856 2.87 Amos Rusie 1871 3.07
Bob Gibson 1935 2.91 Nolan Ryan 1947 3.17
Vernon Gomez 1908 3.34 Tom Seaver 1944 2.86
Burleigh Grimes 1893 3.53 Warren Spahn 1921 3.09
Robert Grove 1900 3.06 Don Sutton 1945 3.26
Jesse Haines 1878 3.64 Dazzy Vance 1891 3.24
Waite Hoyt 1899 3.59 George Waddell 1876 2.16
Carl Hubbell 1903 2.98 Ed Walsh 1881 1.82
Jim Hunter 1946 3.26 Mickey Welch 1859 2.71
Ferguson Jenkins 1943 3.34 Hoyt Wilhelm 1923 2.52
Walter Johnson 1887 2.16 Vic Willis 1876 2.63
Addie Joss 1880 1.89 Early Wynn 1920 3.54
Tim Keefe 1857 2.62 Cy Young 1867 2.63

(d) You should note that the Hall of Fame pitchers are not uniformly distributed over
the six eras 1840–1859, 1860–1879, . . . , 1940–1959. Can you explain this pattern?
Were pitchers better during particular years of baseball?

Interval for year of birth Count Proportion

1840 to 1859
1860 to 1879
1880 to 1899
1900 to 1919
1920 to 1939
1940 to 1959

2.25. The salaries of the players on the 2014 Atlanta Braves are displayed in Table 2.16.

Table 2.16. Salaries of the 2014 Atlanta Braves

Name Salary Name Salary
Freddie Freeman $8,500,000 Eric Young $1,000,000
Andrelton Simmons $3,000,000 Josh Outman $925,000
Nick Markakis $11,000,000 Shelby Miller $535,000
Julio Teheran $1,000,000 Luis Avilan $530,000
Trevor Cahill $12,000,000 Chris Withrow $522,500
Cameron Maybin $7,000,000 Alex Wood $520,000
Chris Johnson $6,000,000 Shae Simmons $508,750
Juan Uribe $6,500,000 Mike Foltynewicz $508,750
Jason Grilli $4,250,000 Joey Terdoslavich $507,500
Mike Minor $5,600,000 Philip Gosselin $507,500
Jonny Gomes $4,000,000 Christian Bethancourt $507,500
A.J. Pierzynski $2,000,000 Jace Peterson $507,500
Jim Johnson $1,600,000 Todd Cunningham $507,500
Kelly Johnson $1,500,000

(a) Construct a stemplot of the salaries.

(b) Describe the basic shape of the data. Are there any unusually small or large salaries?
(c) Find the median salary and a player who has this median salary.
(d) Can you explain why there are no salaries below $200,000?
2.26. (Exercise 2.25 continued). Collect from the internet the salaries of your favorite MLB in
the current season.
(a) Answer questions (a), (b), (c) of Exercise 2.25 for this salary dataset.
(b) Compare the distributions of salaries of the 2014 Braves with your team.

2 4 6 8 10 12

Salary (millions of dollars)

2.27. (Exercise 2.25 continued) A dotplot of the salaries of the 2014 Atlanta Braves is shown
above.
(a) Based on the shape of the distribution of these salaries, do you expect the mean to
be larger than, equal to, or smaller than the median?
(b) Compute the mean and median of the salaries of the 2014 Atlanta Braves.
(c) Do the values you computed in (b) agree with your expectation in (a)?
2.28. In baseball, there are four famous (or notable) years with respect to home runs: 1927 when
Babe Ruth hit 60 home runs, 1961 when Roger Maris hit 61, 1998 when Mark McGwire

hit 70 home runs, and 2001 when Barry Bonds hit 73 home runs. For the three years 1927,
1961, 1998, the number of home runs was collected for each player who had at least 400
at-bats. A player’s home run count was classified into one of the six groupings, 0, 1–5,
6–10, 11–20, 21–40, 41 and more. The frequency table for the home runs for each of the
four years is presented in Tables 2.17–2.20.

Table 2.17. Classification of home run numbers for 1927 players

Number of home runs 0 1–5 6–10 11–20 21–40 41 and more
Count 7 50 19 12 3 2
Proportion

Table 2.18. Classification of home run numbers for 1961 players

Number of home runs 0 1–5 6–10 11–20 21–40 41 and more
Count 0 18 12 31 28 7
Proportion

Table 2.19. Classification of home run numbers for 1998 players

Number of home runs 0 1–5 6–10 11–20 21–40 41 and more
Count 0 21 34 60 63 12
Proportion

Table 2.20. Classification of home run numbers for 2001 players

Number of home runs 0 1–5 6–10 11–20 21–40 41 and more
Count 1 15 34 63 66 12
Proportion

(a) For each table, find the proportion of players who fall in each of the six categories.
Place the proportions in the rows labeled “Proportion.”
(b) In 1927, what proportion of players hit more than 20 home runs? Repeat this com-
putation for the 1961 and 1998 years. Place your answers in the table below.

Proportion hitting more than

YEAR 20 home runs
1927
1961
1998
2001

(c) Based on your computations in (a) and (b), would you say that in some years it was
more difficult to hit a home run than in other years? Explain.

2.29. Table 2.21 gives average home ballpark attendance for each of the major league teams
in 2014.
Table 2.21. Average attendance figures for all ML teams in 2014
American League National League
Team Attendance Team Attendance
Los Angeles 38,221 Arizona 25,602
Baltimore 30,426 Atlanta 29,065
Boston 36,495 Chicago 32,742
Chicago 20,381 Cincinnati 30,576
Cleveland 17,746 Colorado 33,090
Detroit 36,015 Los Angeles 46,696
Houston 21,628 Miami 21,386
Kansas City 24,154 Milwaukee 34,536
Minnesota 27,785 New York 26,528
New York 41,995 Philadelphia 29,924
Oakland 24,736 Pittsburgh 30,155
Seattle 25,486 San Diego 27,103
Tampa Bay 17,858 San Francisco 41,589
Texas 33,565 St. Louis 43,712
Toronto 29,327 Washington 31,844

(a) Construct a dotplot of the home attendance averages for all AL teams. Comment on
the basic shape of the data and note any unusual values.
(b) Repeat (a) for the home averages for the NL teams.
(c) Find the AL team that has the smallest average home attendance, the team with the
largest average, and a team that has a median average.
(d) Repeat (c) for the NL teams.
(e) Can you explain the great variation in home attendance averages? Do you think the
variation can be explained solely by the differences in the sizes of the cities?
2.30. (Ballpark attendance data from Exercise 2.29.)
(a) Find the five-number summary of the home average attendance for the AL teams.
(b) Find the five-number summary of the home average attendance for the NL teams.
(c) By comparing the medians you found in (a) and (b), which league tends to draw
more fans at their home games?
(d) Which league seems to have more variation in home attendance among its teams?
Which number should you be computing to measure the spread of the AL and NL
datasets?
2.31. Table 2.22 gives the batting side (L D left, R D right, B D both sides) and the throwing
hand (L D left, R D right) for 40 randomly selected ballplayers who were born in 1950
or later.
(a) Construct a frequency table for the batting side of the 40 players.
(b) Graph the proportions of left, right, and both sides batters using a bar chart.
(c) Repeat (a) and (b) for the throwing hand of the 40 players.

Table 2.22. Batting side and throwing hand for forty randomly
selected players
Batting Side
L R L B R R L R R B
R R R L R L R L R L
R L R R R R L R R R
L R R L R R B L R B
Throwing Hand
R R R R R R R R L L
R R R L R R R R R R
R L R R R R R R R R
R L R R L L R R R R

(d) Do you think that the proportion of left-handed batters is less than, equal to, or
greater than the proportion of lefties in the American population (which is about
10%)? Explain.
(e) Do you think that the proportion of left-handed throwers is less than, equal to, or
greater than the proportion of lefties in the American population? Explain.
2.32. Table 2.23 displays the speed (in miles per hour) of 40 four-seam fastballs thrown by Zack
Greinke during a game on July 19, 2015.

Table 2.23. Speeds of fastballs (in mph) thrown by Zack Greinke during a specific
game during the 2015 season
92.3 94.3 93.0 93.4 93.4 94.1 94.4 93.3 93.5 93.5
94.3 93.7 92.7 94.1 93.0 92.4 92.0 92.5 92.9 91.3
92.9 92.6 92.4 92.9 91.8 91.6 93.4 93.0 93.5 93.0
91.9 93.5 92.4 93.6 94.0 92.2 92.9 93.3 93.2 93.3

(a) Construct a suitable graph of these pitch speeds.

(b) Find a five-number summary of these speeds.
(c) These speeds are listed in order of when they were thrown in the game—the first row
contains the speeds of the first 10 pitches, the second row contains the speeds of the
second 10, and so on. Construct a graph of the speeds against the order they were
thrown—do you see any pattern?
2.33. Table 2.24 displays the pitch type of the first 40 pitches thrown by Zack Greinke during
the game played on July 19, 2015. (FF is a four-seam fastball, SL is a slider, CH is a
changeup, FT is a two-seam fastball, and CU is a curveball.)
(a) Construct a frequency table of the pitch types.
(b) Construct a suitable graph of these data.
(c) For each row of 10 pitches, compute the number of different pitch types that were
thrown. What does this say about Greinke’s pitching pattern during this game?

Table 2.24. Pitch types of the first 40 pitches thrown by Zack Greinke during
a specific game during the 2015 season
FF SL FT FF FF FF CH SL FF FF
FT FF CU FT CU FF CH SL FT SL
FT FF CH FF FF SL FF FT FF SL
FF SL FF FF SL SL FF FT CH FT

2.34. Table 2.25 shows batted ball data for Albert Pujols for the seasons 2002 through 2015.
The “Direction” statistics give the proportion of batted balls that were pulled, and hit in
the center and opposite sides of the field. The “Hardness” statistics give the proportion of
batted balls that were hit softly, medium hardness, or hard.
Table 2.25. Batted ball statistics for Albert Pujols for seasons 2002 through 2015
Direction Hardness
Season Pull Center Opposite Soft Medium Hard
2002 0.474 0.265 0.261 0.126 0.601 0.273
2003 0.508 0.258 0.234 0.142 0.526 0.332
2004 0.490 0.277 0.233 0.097 0.552 0.352
2005 0.469 0.285 0.246 0.127 0.484 0.389
2006 0.453 0.340 0.207 0.078 0.580 0.342
2007 0.404 0.373 0.223 0.136 0.464 0.400
2008 0.475 0.318 0.207 0.155 0.416 0.429
2009 0.496 0.342 0.162 0.137 0.457 0.406
2010 0.507 0.348 0.145 0.161 0.416 0.424
2011 0.422 0.400 0.178 0.206 0.489 0.305
2012 0.520 0.317 0.164 0.142 0.523 0.335
2013 0.431 0.420 0.149 0.120 0.519 0.362
2014 0.510 0.298 0.193 0.145 0.494 0.361
2015 0.460 0.347 0.193 0.151 0.510 0.338

(a) Construct a suitable graph of the proportions of batted balls that were pulled for all
seasons. Write a short paragraph describing the main features of this distribution.
(b) Construct a graph of the proportions of “hard” batted balls and summarize this
distribution.
(c) Construct a graph of the proportion of “hard” batted balls against season. How has
the proportion of hard batted balls changed over time?

Further Reading
Devore and Peck (2011) and Moore, McCabe and Craig (2012) provide good descriptions of the
exploratory methods used in this chapter to describe a single batch of data. Chapter 2 of Albert
and Bennett (2003) illustrates the use of data analysis methods on baseball data.

What’s On-Deck?
In this chapter, we illustrate some basic data analysis tools for comparing two or more datasets.
It is popular among baseball fans to make comparisons between individual players. In Case
Study 3.1, we compare two popular sluggers in baseball, Albert Pujols and Manny Ramirez.
By using parallel boxplots and time series plots, we compare the batting performance of the
two players. In Case Study 3.2, we compare Robin Roberts and Whitey Ford, two Hall of Fame
pitchers who pitched in the opening game of the 1950 World Series. When you think of great
individual home run accomplishments, one naturally thinks of Babe Ruth, who hit 60 home
runs in 1927, Roger Maris, who hit 61 in 1961, Mark McGwire, who hit 70 in 1998, and Barry
Bonds, who hit 73 home runs in 2001. To better understand these hitting accomplishments, one
should look at the general pattern of home run hitting during these four seasons, and Case Study
3.3 compares the team home run rates for these seasons. Case Study 3.4 looks at the slugging
percentages of all players in the 2014 season that had at least 400 at-bats. We will see that the
distribution of slugging percentages has a distinctive bell-shape and this pattern makes it easy to
find intervals that contain a given percentage of the data. We conclude the chapter in Case Study
3.5 by looking at four of the highest season batting averages in recent baseball history. One can
assess the greatness of each hitting accomplishment by looking at each batting average in the
context of all batting averages for that particular season. By computing standardized scores, we
can compare these hitting accomplishments and say which player had the best average relative
to his peers.

3.1 Albert Pujols and Manny Ramirez

Topics Covered: Stemplot, five-number summary, time series plot, boxplot.
In this first case study, we statistically compare Albert Pujols and Manny Ramirez, two of
the best sluggers in modern baseball history.
Let’s make some initial comments about these two players:
Both players were born in the Dominican Republic and are both known for their great batting
skill and power.
Ramirez debuted in Major League Baseball in 1993 and played parts of 19 seasons. Pujols
debuted with the St. Louis Cardinals in 2001 and is currently (in 2015) in his 15th season
playing for the Los Angeles Angels of Anaheim.

45
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
46 Comparing Batches and Standardization

Both players are especially proficient in hitting home runs. Ramirez hit 555 in his career and
Pujols is approaching that career home run mark in the 2015 season.

Who is the Better Hitter?

Before we compare Pujols and Ramirez, some general truths about evaluating hitting should be
stated.

1. The objective for a baseball team is to score RUNS.

2. How does a team score runs?
batters get ON-BASE
other batters ADVANCE the runners to home by hits, walks, HBP, or errors
3. A measure of a hitter’s ability to get on-base is the On-Base Percentage (OBP):
H C BB C HBP
OBP D :
AB C BB C HBP C SF
The total number of plate appearances is PA D AB C BB C HBP C SF and the number of
times on-base is H C BB C HBP. So OBP is the fraction of PAs that are on-base.
4. A measure of a hitter’s ability to advance runners is the Slugging Percentage (SLG):
TB
SLG D :
AB
Here TB is total bases D # of Singles C 2 .# of Doubles/ C 3 .# of Triples/ C 4
.# of Home Runs/.
5. As discussed in Chapter 1, a simple measure that combines ON-BASE ability with AD-
VANCEMENT ability is the OPS statistic:

OPS D OBP C SLG:

(For example, Ramirez in 2000 had an OBP of :457 and an SLG of :697, so his 2000 OPS
was OPS D :457 C :697 D 1:154.)

Comparing OPS
Table 3.1 gives the season OPS values for Pujols and Ramirez for their careers (through the
2014 season).
We use back-to-back stemplots in Figure 3.1 to make a comparison—we break an OPS
value like 1:106 into a stem of 11 and a leaf of 06. We use one-digit leaves, so the OPS value of
1:106 is represented by a 0 leaf on the 11 stem line.
We summarize each dataset by a five-number summary—this is

.LO; QL ; M; QU ; HI /

where LO, HI are the lowest and highest values in the dataset, QL , QU are the lower and
upper quartiles, and M is the median.
We illustrate these calculations for Pujols’ data. Here there are n D 14 values. The position
of the median is pos.M / D .14 C 1/=2 D 7:5. The median is the average of the 7th and 8th
values which is .1:011 C 1:013/=2 D 1:012. To find the quartiles, we divide the 14 remaining
observations into two groups of 7, and find the median of the lower 7 and the median of the

Table 3.1. Career hitting statistics for Albert Pujols and Manny Ramirez
Pujols Ramirez
Age SLG OBP OPS SLG OBP OPS
21 0.610 0.403 1.013
22 0.561 0.394 0.955 0.521 0.357 0.878
23 0.667 0.439 1.106 0.558 0.402 0.960
24 0.657 0.415 1.072 0.582 0.399 0.981
25 0.609 0.430 1.039 0.538 0.415 0.953
26 0.671 0.431 1.102 0.599 0.377 0.976
27 0.568 0.429 0.997 0.663 0.442 1.105
28 0.653 0.462 1.114 0.697 0.457 1.154
29 0.658 0.443 1.101 0.609 0.405 1.014
30 0.596 0.414 1.011 0.647 0.450 1.097
31 0.541 0.366 0.906 0.587 0.427 1.014
32 0.516 0.343 0.859 0.613 0.397 1.010
33 0.437 0.330 0.767 0.594 0.388 0.982
34 0.466 0.324 0.790 0.619 0.439 1.058
35 0.493 0.388 0.881
36 0.601 0.430 1.031
37 0.531 0.418 0.949
38 0.460 0.409 0.870

1 | 2: represents 0.12, leaf unit: 0.01

Pujols OPS Ramirez OPS
___________________________
| 7* |
96| 7. |
| 8* |
5| 8. |778
0| 9* |4
95| 9. |56788
311| 10* |0113
7| 10. |59
1000| 11* |0
| 11. |5
| 12* |
Figure 3.1. Back-to-back stemplots for Albert Pujols and Manny Ramirez’s season OPS values.

upper 7. We get

QL D :906; QU D 1:101:

The smallest value is 0:767 and the largest is 1:114.

So Pujols’ OPS values 5-number summary is .0:767; 0:906; 1:012; 1:101; 1:114/. In
a similar fashion, we compute Ramirez’s OPS values 5-number summary: .0:870; 0:953;
0:982; 1:031; 1:154/.
Essentially, a 5-number summary divides the data into quarters. For Pujols’ data, we can
say that (approximately) 25% of the data falls between 0:767 and 0:906, 25% falls between
0:906 and 1:012, and so on.
Before we graph these two datasets, we look for possible outliers. Using a standard rule,
we say that an extreme observation is worthy of special attention if it falls outside one step from
the lower and upper quartiles, where a step is defined to be

STEP D 1:5 .QU QL /:

To illustrate for Ramirez’s data,

STEP D 1:5 .1:031 0:953/ D 0:117

Outliers are observations that have values smaller than

QL STEP D 0:953 0:117 D 0:836:

and larger than

QU C STEP D 1:031 C 0:117 D 1:148:

Looking back at the stemplot of Ramirez’s OPS values, we see that there is one outlier in this
dataset at the high end.
If one does a similar analysis for Pujols’ OPS values, one finds that there are no outliers.

A boxplot is a graph of a five-number summary. One draws this by

drawing a number line covering all of the values,

drawing a box, where the ends of the box correspond to the quartiles QL and QU and a
vertical line is drawn through the box at the median M ,
drawing lines (whiskers) out to the most extreme values that are not considered outliers,
indicating outliers by plotting a special symbol.

Side-by-side boxplots are most useful in comparing datasets. In Figure 3.2 we show boxplots
for the Pujols data and the Ramirez data on the same scale.

Ramirez

Pujols

0.8 0.9 1.0 1.1

OPS

Figure 3.2. Parallel boxplots of the OPS season values for Albert Pujols and Manny Ramirez.

What have we learned in this comparision?

On the average, Pujols is a slightly better hitter than Ramirez on the OPS scale. The median
OPS for Pujols is 1:012 and the median OPS for Ramirez is 0:982, so Pujols is 1:012 0:982 D
0:030 better using this measure.
On the other hand, the spread or variability of Ramirez’s OPS values is smaller than the spread
of Pujols’ OPS values. This is pretty obvious from the boxplot as the width of the box in the
boxplot of Ramirez’s OPS values is clearly smaller than the width of the box of the boxplot
of Pujols’ OPS values. One could conclude that Ramirez is a more consistent in the pattern
of his hitting than Pujols.

Adjusting Comparison for Ages

But this comparison may be viewed as unfair, since Pujols is six years younger than Ramirez
and has more baseball seasons left. Maybe the two players would be viewed differently after
both have finished their careers.
Let’s consider a different type of comparison that accounts for the ages of the two players.
What does it mean for Player A to have more ability than Player B? A popular way to
model the relationship between ability and age is with the curve shown in Figure 3.3. In this
curve, ability is low at the start of one’s career, grows until mid-career, and then declines with
advancing years. The peak ability, age at peak, the steepness of incline and decline, would be
expected to vary from player-to-player.
ABILITY

AGE

Figure 3.3. Expected ability curve for a Major League player.

Looking at the graph, we can distinguish between two measures of ability.

peak ability—this is the player’s ability at the top of the curve,

career ability—this is the total accumulation of a player’s ability over time—you can think
of the area under the above curve as representing career performance.

So when you say Pujols is “better than” Ramirez, you should be clear what you’re talking
about. One player might play better at his peak, and the second player might be better in his
total performance over his Major League career. We can look at Pujols and Ramirez’s hitting
performances by plotting the season OPS against age—Figure 3.4 graphs the OPS against age
for both players.

1.10
Pujols
Ramirez

1.00
OPS
0.90
0.80

20 25 30 35 40
Age

Figure 3.4. Time series plots of Albert Pujols’ and Manny Ramirez’s OPS values; the o’s correspond to
Pujols and the x’s correspond to Ramirez. Smooth curves are drawn through the points to show the basic
patterns.

Drawing separate smooth curves over the points to show the basic patterns, what do we
see?

It appears that Pujols peaked earlier than Ramirez—Pujols’ peak in OPS is about at age 26
and Ramirez peaked at closer to 30 years.
Ramirez’s deterioration in batting performance was more gradual than Pujols. It is interesting
that Pujols’ drop-off in OPS occurred as he was changing teams from the Cardinals to the
Angels.
As I am writing, Pujols is having a much better 2015 hitting season, so perhaps it is premature
to talk about Pujols’ slump towards the end of his career.

3.2 Robin Roberts and Whitey Ford

Topics Covered: Stemplot, time series plot.
In this case study, we discuss two great pitchers, Robin Roberts and Whitey Ford. In Phillies
history, there were four great seasons:

1. 1950 when the Whiz Kids won the NL pennant—they were a young team that won the
pennant on the last game of the season. (They lost to the Yankees in the World Series in four
games.)
2. 1980 when they won the World Series against the Royals.
3. 1993 when they won the NL pennant and lost in six games to Toronto. (Joe Carter who hit
the series-ending home run comes to mind.)
4. 2008 when they won the World Series against the Rays.

We compare two pitchers, both in the Hall of Fame, who played in the 1950 World Series.
Robin Roberts, born September 30, 1926, was a great pitcher for the 1950 Whiz Kids. He
pitched for Philadelphia between 1948 and 1961. He won 286 games—many seasons he won

Table 3.2. Career pitching statistics for Robin Roberts

Year Age W L G IP H HR BB SO ERA lgERA
1948 21 7 9 20 146.67 148 10 61 84 3.19 3.96
1949 22 15 15 43 226.67 229 15 75 95 3.69 3.96
1950 23 20 11 40 304.33 282 29 77 146 3.02 4.06
1951 24 21 15 44 315 284 20 64 127 3.03 3.84
1952 25 28 7 39 330 292 22 45 148 2.59 3.66
1953 26 23 16 44 346.67 324 30 61 198 2.75 4.20
1954 27 23 15 45 336.67 289 35 56 185 2.97 4.03
1955 28 23 14 41 305 292 41 53 160 3.28 3.96
1956 29 19 18 43 297.33 328 46 40 157 4.45 3.73
1957 30 10 22 39 249.67 246 40 43 128 4.07 3.80
1958 31 17 14 35 269.67 270 30 51 130 3.24 3.95
1959 32 15 17 35 257.33 267 34 35 137 4.27 4.11
1960 33 12 16 35 237.33 256 31 34 122 4.02 3.88
1961 34 1 10 26 117 154 19 23 54 5.85 4.07
1962 35 10 9 27 191.33 176 17 41 102 2.78 3.77
1963 36 14 13 35 251.33 230 35 40 124 3.33 3.52
1964 37 13 7 31 204 203 18 52 109 2.91 3.59
1965 38 10 9 30 190.67 171 18 30 97 2.78 3.42
1966 39 5 8 24 112 141 15 21 54 4.82 3.54

over 20 games. His pitching was notable for his great control. We will compare Roberts with a
great Yankees pitcher, Whitey Ford. Ford, born October 21, 1928, is notable for having 10 World
Series wins. He won the Cy Young award in 1961 with a record of 25-4. (He was overshadowed
that year by Roger Maris who hit 61 home runs.) His career win/loss record was 236-106. Like
Roberts, Ford was a young pitcher in 1950.
Comparing two pitchers is a little tougher than comparing two hitters since some of the
standard pitching statistics are team-dependent. For example, if a pitcher has a great win/loss
record such as 20-5, this could mean that he was an outstanding pitcher or it could mean that he
was just a good pitcher whose team scored a lot of runs when he was pitching. Here we consider
the ERA pitching statistic that is the mean number of runs earned by the opposing team in a
9-inning game. An “earned” run is one that is not produced by means of an error by the team’s
fielders. This is one of the most common measures that is used to rate pitchers. Tables 3.2 and
3.3 displays the season by season pitching statistics for both Roberts and Ford.
Figure 3.5 displays back-to-back stemplots of the ERAs for Ford and Roberts. In construct-
ing the stemplot, we break an ERA such as 2:81 at the decimal point, so the stem is 2 and the
(one-digit) leaf is 8.
What do we see in comparing Ford’s and Roberts’ ERAs?
Both pitchers had several seasons with ERAs in the 2:5–3:3 range. But Ford’s worst season
with respect to ERA was only 3:2 and he had five seasons where his ERA fell under 2:5. In
contrast, Roberts’ best season was 2:5 and he had six seasons where his ERA was over 4:0.
Ford generally appears to be the better pitcher with respect to this pitching statistic.

Table 3.3. Career pitching statistics for Whitey Ford

Year Age W L G IP H HR BB SO ERA lgERA
1950 21 9 1 20 112 87 7 52 59 2.81 4.30
1953 24 18 6 32 207 187 13 110 110 3 3.68
1954 25 16 8 34 210.67 170 10 101 125 2.82 3.42
1955 26 18 7 39 253.67 188 20 113 137 2.63 3.76
1956 27 19 6 31 225.67 187 13 84 141 2.47 3.87
1957 28 11 5 24 129.33 114 10 53 84 2.57 3.60
1958 29 14 7 30 219.33 174 14 62 145 2.01 3.54
1959 30 16 10 35 204 194 13 89 114 3.04 3.63
1960 31 12 9 33 192.67 168 15 65 85 3.08 3.60
1961 32 25 4 39 283 242 23 92 209 3.21 3.70
1962 33 17 8 38 257.67 243 22 69 160 2.90 3.73
1963 34 24 7 38 269.33 240 26 56 189 2.74 3.52
1964 35 17 6 39 244.67 212 10 57 172 2.13 3.62
1965 36 16 13 37 244.33 241 22 50 162 3.24 3.39
1966 37 2 5 22 73 79 8 24 43 2.47 3.33
1967 38 2 4 7 44 40 2 9 21 1.64 3.13

But we have to be careful about making this conclusion. Why? Well, the measurement of a
pitcher’s ability, such as an ERA, should be made in the context of the season and league and
ballpark in which the player played. Ford and Roberts played during roughly the same seasons.
But Ford pitched primarily in the American League, Roberts in the National League, and they
pitched in different ballparks. Maybe Ford’s ERAs looked better than Roberts’ ERAs because
of the differences in league and ballpark.

1 | 2: represents 1.2, leaf unit: 0.1

Ford$ERA Roberts$ERA
__________________________
| 1* |
6| 1. |
4410| 2* |
988765| 2. |577799
22000| 3* |001223
| 3. |6
| 4* |0024
| 4. |8
| 5* |
| 5. |8
| 6* |
Figure 3.5. Back-to-back stemplots of the season ERAs for Ford and Roberts.

To illustrate how much ERAs can be different across seasons and leagues, Figure 3.6
displays a time series plot of the ERA for all pitchers. The season ERAs of the National League
are plotted using a light line and the AL ERAs by a darker line. Note the great rises and falls in
the mean ERA across seasons. In the “dead-ball” era in the early 1900s, pitching was dominant
and the mean ERA was low. Then pitching got worse over time and the AL ERA peaked at a
value close to 5:0 around the year 1940. In the late 1960s, pitching again was dominant—the
league ERAs dipped to about 3:0—and recently the ERAs have been rising. Also, note that the
NL and AL ERAs were quite different for some seasons. Since the season ERAs are so variable,
it is reasonable to view a pitcher’s season ERA in the context of the average ERA that particular
year.
5.0

AL
NL
4.5
4.0
ERA
3.5
3.0
2.5

1900 1920 1940 1960 1980 2000

Year

Figure 3.6. Time series plot of the season ERAs for all pitchers in the National and American Leagues.

Actually, the baseball-reference.com site goes one step further than a season ad-
justment. In the data tables shown above, the “lgERA” statistic is the average ERA of all pitchers
in the same league and the same ballparks. We adjust a pitcher’s ERA by computing

lgERA
ERAC D 100;
ERA
which is called the “adjusted ERA”.
To illustrate the computation and interpretation of ERAC, Robin Roberts’ ERA in 1950
was 3:02. The average ERA in the National League in the ballparks that Roberts pitched was
4:06. So Roberts’ adjusted ERA was

4:06
ERAC D 100 D 135:
3:02
An adjusted ERA over 100 means that the pitcher performed above-average. Roberts had a good
year in 1950—the league/ballpark average ERA was 35% higher than his ERA.
Suppose that we compare the adjusted ERAs of Roberts and Ford by stemplots in Figure 3.7.

1 | 2: represents 12, leaf unit: 1

Ford$aERA Roberts$aERA
__________________________
| 6 |9
| 7 |3
| 8 |3
| 9 |366
4| 10 |57
965| 11 |
8821| 12 |013346
4| 13 |455
20| 14 |1
63| 15 |2
9| 16 |
6| 17 |
| 18 |
0| 19 |
| 20 |
Figure 3.7. Back-to-back stemplots of the adjusted ERAs of Roberts and Ford.

We get some additional insight about the quality of these pitchers. Ford had an adjusted
ERA of over 100 for every year that he pitched—that means that he was above average every
year. In contrast, Roberts had 13 years where he was above-average and six years where he was
below average. (It is interesting to note that five of his bad seasons occurred between 1956 and
1961.) Ford had three seasons where his adjusted ERA was 170 or higher. Ford does appear to
be the superior pitcher in this study, but Ford pitched for a much stronger team (the Yankees)
than Roberts (the Phillies), and it is possible that the difference in team strengths might explain
part of the difference in ERAs.

3.3 Home Runs: A Comparison of Four Seasons

Topics Covered: Stemplot, five-number summary, time series plot, boxplot.
When a baseball fan thinks about home runs from a historical perspective, four seasons
should come to mind: 1927, 1961, 1998, and 2001. The year 1927 was the year in which Babe
Ruth hit 60 home runs and broke his own season record for home runs. This record stood for 34
years until 1961, when Roger Maris hit his 61 home runs. Another 37 years went by until Mark
McGwire broke the season record with 70 home runs. It was generally thought that McGwire’s
record would last for a while, but Barry Bonds broke the record in 2001 (only three years later)
with his great feat of 73 home runs.
Is Bonds really the greatest home run slugger of all time? Well, it is not clear. It is difficult
to compare Bonds’ accomplishment in 2001 with Ruth’s accomplishment in 1927 since the
difficulty of hitting a home run may have been different in the two years. Maybe we can better
understand the magnitude of these players’ accomplishments if we compare the home run hitting
of all players for these four seasons.

Table 3.4. Number of home runs (HR) hit by every team in the 1927, 1961, 1998, and 2001
seasons
Year
1927 1961 1998 2001
HR G Rate HR G Rate HR G Rate HR G Rate
158 155 1.02 240 163 1.47 198 162 1.22 212 162 1.31
56 155 0.36 180 163 1.10 198 163 1.21 164 162 1.01
29 157 0.18 149 163 0.91 134 161 0.83 214 162 1.32
51 156 0.33 138 163 0.85 115 162 0.71 139 162 0.86
36 153 0.24 150 161 0.93 165 162 1.02 152 162 0.94
26 153 0.17 112 163 0.69 207 162 1.28 203 161 1.26
55 155 0.35 167 161 1.04 205 162 1.27 198 161 1.23
28 154 0.18 189 162 1.17 221 163 1.36 195 162 1.20
54 156 0.35 90 162 0.56 214 162 1.32 136 162 0.84
84 153 0.55 119 161 0.74 111 162 0.69 121 162 0.75
109 155 0.70 158 154 1.03 201 162 1.24 169 162 1.04
74 153 0.48 157 154 1.02 147 162 0.91 199 162 1.23
29 153 0.19 183 155 1.18 234 161 1.45 158 162 0.98
39 154 0.25 188 155 1.21 149 162 0.92 246 162 1.52
37 155 0.24 103 155 0.66 166 162 1.02 208 162 1.28
57 155 0.37 128 154 0.83 212 163 1.30 199 162 1.23
176 156 1.13 223 163 1.37 194 162 1.20
103 155 0.66 138 162 0.85 209 162 1.29
152 162 0.94 176 162 1.09
107 163 0.66 161 162 0.99
215 162 1.33 174 162 1.07
136 162 0.84 164 162 1.01
126 162 0.78 147 162 0.91
147 162 0.91 166 162 1.02
114 162 0.70 131 162 0.81
167 162 1.03 208 162 1.28
161 163 0.99 235 162 1.45
159 162 0.98 206 162 1.27
183 162 1.13 161 162 0.99
159 162 0.98 213 162 1.31

Table 3.4 gives the number of home runs (HR) hit by every team in the 1927, 1961, 1998,
and 2001 baseball seasons. The table also gives the number of games (G) for each team each
year. Note that the number of games in a season has changed over the years—the 1927 teams and
the 1961 teams had season lengths that were about 7–9 games shorter than the current seasons.
To make a reasonable comparison, it is helpful to compute the home run rate (RATE)—the
number of home runs hit divided by the number of games (G).
This home run rate is also given in the table. If a team’s RATE =1, then it hits, on average,
one home run each game.

One effective way of graphically comparing the team home run rates across seasons is by
means of parallel dotplots shown in Figure 3.8. What do we see when we compare team home
run rates across these four seasons?

2001
Season

1998

1961

1927

0.2 0.4 0.6 0.8 1.0 1.2 1.4

Rate

Figure 3.8. Parallel dotplots of the team home run rates for the four seasons.

The pattern of home run hitting in 1927 is very interesting. Babe Ruth’s team, the Yankees,
averaged about a home run every game. But the Yankees were an outlier. A typical team in
1927 had a home run rate about :3—this average team hit a home run every three games. It
is pretty clear that Ruth’s 60 home run total was quite an accomplishment in the year 1927
where the rate of hitting home runs was small.
The year 1961 was a sharp contrast to 1927 with respect to home run hitting. Most of the
team home run rates are in the :8–1:1 range which was comparable to the home run rate of
the great 1927 Yankees. It was much more common to hit a home run in 1961 when Roger
Maris hit his 61 roundtrippers.
We see that the years 1998 and 2001 saw even greater home run hitting than 1961. Looking
at the dotplots in Figure 3.8, it appears that a typical home run rate has increased from 1961
to 1998 and from 1998 to 2001. Half of the 2001 baseball teams had home run rates of 1:19
or higher.

To get a better handle on the differences in home run hitting between seasons, we can
compute summary statistics. Table 3.5 gives five-number summaries for each batch of team
home run rates. Figure 3.9 draws parallel boxplots using these summaries.

Table 3.5. Five-number summaries of the team home rates for the four seasons
Year N LO QL Median QU HI
1927 16 0.16 0.20 0.34 0.45 1.01
1961 18 0.55 0.73 0.98 1.14 1.47
1998 30 0.65 0.85 1 1.27 1.45
2001 30 0.74 0.99 1.14 1.28 1.51

2001

1998
Season

1961

1927

0.2 0.4 0.6 0.8 1.0 1.2 1.4

Rate

Figure 3.9. Parallel boxplots of the team home run rates for the four seasons.

This display reinforces the conclusions that were drawn earlier. Home runs were generally
not hit in 1927 and there was a small spread in the team home run rates. The dot in the 1927
boxplot indicates the outlying team, the Yankees. From 1961 to 2001 there has been a gradual
increase in the rate of hitting home runs. Comparing the median home run rate (:98) in 1961
with the median rate (1:14) in 2001, we see that a typical team in 2001 hit approximately
1:14 :98 D :16 more home runs than a typical team in 1961—that translates to about an
additional home run each six games.

3.4 Slugging Percentages are Normal

Topics Covered: Stemplot, mean and standard deviation, normal distribution probabilities.
The standard deviation, denoted s, is a measure of dispersion about the center (mean) and
has a useful interpretation when our data is bell-shaped. We illustrate the use of the empirical
rule or “68-95-99.7 rule” that tells us about the fraction of observations that fall within three
intervals. Let’s look at all hitters in 2014 that had at least 400 at-bats—we can regard these hitters
as the “regular” players, since they played most of the games during the season. Figure 3.10
displays a stemplot of the slugging percentages (SLG) for the 177 regular players. In this
stemplot, a slugging percentage of :402 is divided between the first and second digits, so 4 is
the stem and 0 is the leaf. Also the 10 possible leaf values is divided into five parts—leaves of 0
and 1 appear next to the “*” symbol, leaves of 2 and 3 appear next to the “t” symbol, the leaves
of 4 and 5 appear next to “f ”, leaves of 6 and 7 appear next to “s”, and leaves of 8 and 9 appear
next to “.”.
This set of slugging percentages looks pretty symmetric; most of the values are in the :350
to :450 range and there is a large spread—the best hitter in 2014 had a slugging percentage
of .581 and the worst was only about :280. (Who are the players with the best and worst SLG
values in 2014? It might be easier to guess the largest value rather than the smallest.) We use R
to compute the mean and standard deviation for these SLGs. Rounding, we see that

xN D 0:416 and s D 0:060:

1 | 2: represents 0.12
leaf unit: 0.01
n: 177

2. | 8
3* | 0111
3t | 222333333
3f | 44444555555
3s | 66666666666677777777777777777
3. | 8888888888888888999999999999
4* | 0000000000000011111111111
4t | 22222333333
4f | 44444444555555555555555
4s | 6666677777
4. | 888899999
5* | 0001
5t | 2222
5f | 44455
5s | 666
5. | 8
Figure 3.10. Stemplot of slugging percentages for the 177 regular players in the 2014 season.

When the distribution of the data is bell-shaped (as it is here), we expect

68% of the data to fall within one standard deviation of the mean (that is, between xN s and
xN C s),
95% of the data to fall within 2 standard deviations of the mean (between xN 2s, xN C 2s),
99.7% of the data to fall within 3 standard deviations of the mean (between xN 3s, xN C 3s).

Applying this 68-95-99.7 rule, we compute

xN s and xN C s
:416 :060 and :416 C :060
:356 and :476:

We expect 68% of the SLGs to fall between :356 and :476.

Let’s check how many actually fall between :356 and :476: there are 128 SLGs between :356
and :476 (using the original dataset) and the proportion of SLGs in this interval is 128=177 D
:723, which is pretty close to :68.
Next, we compute

xN 2 s and xN C 2 s
:416 2.:060/ and :416 C 2.:060/
:296 and :536:

We expect 95% of the SLGs to fall between :296 and :536. Checking the data again, we obtain
167 (out of 177) in this interval which corresponds to 167=177 D :943, which is very close
to :95.
Last, we compute

xN 3 s and xN C 3 s
:416 3.:060/ and :416 C :3.:060/
:236 and :596:

We expect practically all (99.7%) of the SLGs to fall between :236 and :596. In fact all fall in
this range which is 177=177 D 1 which is very close to :997, so again this rule seems to work.
Generally, most “derived” baseball statistics that are computed by a division or a multipli-
cation will be bell-shaped and well-suited for applying the 68-95-99.7 rule. Examples of such
derived statistics are AVG, SLG, OBP, and ERA.

3.5 Great Batting Averages

Topics Covered: Stemplot, mean and standard deviation, standardized score. In this case study,
we examine Great Batting Averages. We identify four players who had unusually high batting
averages (AVG) in recent history:

1994—Tony Gwynn (SD) hit :394

1980—George Brett (KC) hit :390
1977—Rod Carew (MIN) hit :388
1941—Ted Williams (BOS) hit :406

Now, on face value, any baseball statistics like these high averages are meaningless because we
don’t know the context in which these statistics were achieved. If I tell you that Joe Schmoe (a
hypothetical player) batted :420, you should ask:

When? You have to know when Joe Schmoe got this AVG. For example, batting averages
in 1900 were generally much higher than batting averages in 1968.
Where? It matters where Joe played. Some parks (like Coors Field) are easier to hit in, and
others (like Dodger Stadium) are harder to hit in.
How many at-bats? We’ll talk about this in a later chapter, but it is easier to hit :420 based
on 100 AB than 500 AB.

We can make better sense of a batting average, like Carew’s :388 AVG in 1977, when we see it
compared to the AVGs of all players in the year 1977. Let’s focus on the 1977 hitters who had
at least 400 at-bats. Figure 3.11 displays a stemplot of the 168 AVGs.
Here are some features of this dataset. (1) The distribution looks symmetric and bell-shaped.
(2) A typical AVG is about :275. (3) Most of the averages fall between :230 and :300. (4) We
can’t miss the one large AVG at :388 which corresponds to Carew. We would like a measure of
relative standing that tells us how good Carew’s batting average is. We compute a standardized
score of an average by subtracting the mean and then dividing by the standard deviation s.
Calculations reveal that xN D :27742 and s D :02717 for these 1977 AVGs. So the standardized

1 | 2: represents 0.012
leaf unit: 0.001
n: 173

20 | 06
21 | 5
22 | 018
23 | 012335899
24 | 01344666779
25 | 0001111223345678899
26 | 0111122333333456667788
27 | 00011123334445555777789999
28 | 00011112222334444566677788999
29 | 0001122235677888
30 | 000022344777888
31 | 00112556779
32 | 02558
33 | 557
34 |
35 |
36 |
37 |
38 | 7
Figure 3.11. Stemplot of the batting averages of the 1977 Major League players with at least 400 at-bats.

score (or z score) for Carew’s :388 is

:388 :27742
zD D 4:07 (WOW!):
:02717
Carew’s z-score is 4:07 which means that his AVG was about four standard deviations above
the mean. Let’s compute the standardized score for the weakest hitter in 1977 who only batted
.200. His z-score would be
:200 :27742
zD D 2:85:
:02717
In other words, his AVG was 2:85 standard deviations below the mean. If a player has a z-score
of 0, this means that his AVG is equal to the mean 277. So

A positive z-score corresponds to an AVG above the mean.

A negative z-score corresponds to an AVG below the mean.
A z-score of 0 corresponds to an AVG that is equal to the mean.

Recall, for bell-shaped data, practically all of the z-scores for the data will fall between 3 and
3. (This is a restatement of the 99:7 rule.) Using the concept of standardized scores, we can
compare these great batting averages:

We already saw that Carew’s .388 corresponded to a z-score of 4:07.

Tony Gwynn’s :394—that year (1994), the mean xN D :29311 and s D :03201, so his z-score
would be
:394 :29311
zD D 3:15:
:03201
Ted Williams’ :406—that year (1941), xN D :28063 and s D :03281, and
:406 :28063
zD D 3:82:
:03281
George Brett’s :390—that year (1980), xN D :27882 and s D :02757, and
:390 :27882
zD D 4:03:
:02757
Comparing the four, Carew’s AVG was best (relative to his peers) since he had the largest
standardized score. In this comparison, we see that batting averages have changed over the years.
Table 3.6 shows the mean and standard deviation for five years.

Table 3.6. Mean and standard deviation of

players’ batting averages with at least 400
at-bats for five years
Year Mean Standard Deviation
1900 :295 :037
1941 :280 :033
1977 :277 :027
1980 :279 :028
1994 :293 :032

A couple of general comments from this table:

Looking at the means, there doesn’t seem to be any general trend in batting averages—they
have gone up and down. Higher batting averages could mean better hitting or weaker pitching,
or both.
If you look at the standard deviations over years, you will see a general decreasing pattern.
The standard deviation of AVGs today is smaller than it used to be. This means that the hitting
abilities of current players are more similar than they used to be.

3.6 Exercises
3.0a. Rickey Henderson and Tim Raines were both great leadoff hitters who played in the
1980’s and 1990’s. Table 3.7 shows the career on-base percentage (OBP) and the slugging
percentage (SLG) for each player—the seasons where Raines had fewer than 200 at-bats
have been removed.
(a) Construct back-to-back stemplots of the OBPs for Henderson and Raines.
(b) Find five-number summaries of the two batches.
(c) Based on your work in (a) and (b), which hitter was more effective in getting on-base?
On average, how much better was the superior player?

Table 3.7. On-base and slugging percentages for Rickey Henderson

and Tim Raines for the seasons of their careers
Henderson Raines
Age OBP SLG Age OBP SLG
20 .338 .336
21 .420 .399 21 .391 .438
22 .408 .437 22 .353 .369
23 .398 .382 23 .393 .429
24 .414 .421 24 .393 .437
25 .399 .458 25 .405 .475
26 .419 .516 26 .413 .476
27 .358 .469 27 .429 .526
28 .423 .497 28 .350 .431
29 .394 .399 29 .395 .418
30 .411 .399 30 .379 .392
31 .439 .577 31 .359 .345
32 .400 .423 32 .380 .405
33 .426 .457 33 .401 .480
34 .432 .474 34 .365 .409
35 .411 .365 35 .374 .422
36 .407 .447 36 .383 .468
37 .410 .344 37 .403 .454
38 .400 .342 38 .395 .383
39 .376 .347
40 .423 .466
41 .368 .305
42 .366 .351

(d) Construct one time series plot of the slugging percentages of Henderson and Raines
graphing against age. Compare the patterns of change (across time) for each player.
Was one player consistently better than the other with respect to SLG? Did both
players display the typical pattern in which one gets better in performance, hits a
peak, and then declines towards the end of his career?
3.0b. If you look at Rickey Henderson’s batting statistics in Table 1.1, we see that his highest
season OBP was :439 in the 1990 season. How impressive was that :439 OBP? To check,
Figure 3.12 displays a stemplot of the OBPs of all players that had at least 400 plate
appearances in the 1990 season.
(a) Describe the general shape of the distribution of OBPs and discuss any unusual
features.
(b) The mean and standard deviation of the OBPs are given by xN D :337 and s D :036,
respectively. Find the standardized score of Henderson’s OBP of :439.
(c) Find an interval that contains approximately 95% of the OBPs.

25 | 88
26 | 7
27 | 466999
28 | 2222589
29 | 0012236799
30 | 00112334444445557778
31 | 223466889
32 | 000122234466677888999
33 | 000002233345666778899
34 | 0001112222234677888899
35 | 001223466777899
36 | 00444466679
37 | 012345555566789
38 | 13455679
39 | 77
40 | 0457
41 | 377
42 |
43 | 9
44 | 1
Figure 3.12. Stemplot of the OBPs of all players with at least 400 plate appearances in 1990.

(d) How does Henderson’s :439 OBP compare with the “greatness” of the batting av-
erages discussed in Case Study 3.5. Is this study sufficient to convince you that
Henderson was the greatest leadoff hitter of all time? Explain.
3.1. (Comparing batting averages across eras). The stemplots in Figure 3.13 display the batting
averages of all players with at least 400 at-bats in the years 1897 and 1997:
(a) Find the five-number summaries of each dataset
(b) Which group of players tends to have the higher batting averages? Explain.
(c) Which group of players had the greater spread in batting averages? Explain.
3.2. (Continuation of Exercise 3.1.)
(a) For the 1897 batting averages, one can compute the mean and standard deviation to
be :318 and :037, respectively. Find an interval of values where you expect 68% of
the batting averages to fall.
(b) Count the number of players who had a batting average in the interval in (a).
(c) Find the proportion of players who had an average in the interval in (a). Is it close to
what you expect?
(d) For the 1997 batting averages, the mean D :281 and the standard deviation D :028.
Find an interval where you think 95% of the averages will fall.
(e) Find the proportion of 1997 players who had an average in the interval you found in
(d). Check if it is close to what you expect.

1897 players 1997 players

0 | 20 | 2
| 21 |
| 22 | 389
| 23 | 35799
| 24 | 012334567889
| 25 | 000000222234678
76 | 26 | 01112333445555557789999
73 | 27 | 0112333345555677789
9997554322 | 28 | 000112234456666667788
210 | 29 | 0000011112344444556677888999
876520 | 30 | 0223445678888
8865320 | 31 | 01344889
9998521 | 32 | 1337789
75 | 33 | 022
53000 | 34 | 7
54322 | 35 |
21 | 36 | 16
7 | 37 | 1
92 | 38 |
| 39 |
| 40 |
| 41 |
3 | 42 |
Figure 3.13. Stemplots of batting averages of all players with at least 400 at-bats from 1897 and 1997
seasons.

3.3. (Continuation of Exercise 3.1)

(a) In 1897, Willie Keeler won the batting crown with an average of :424. Using the
mean and standard deviation of batting averages given in the previous exercise, find
the standardized score for Keeler’s average.
(b) In 1997, Tony Gwynn won the batting crown with an average of :372. Find the
standardized score for Gwynn’s average.
(c) Based on your computations in (a) and (b), which batter was better relative to his
contemporaries? Why?
(d) In the book Full House, Stephen Jay Gould argues that ballplayers are generally
getting less variable over time and it will be difficult for a player to hit for a :400
season batting average in the future. On the basis of your computations, do you agree
with Gould’s assertion? Why?
3.4. (Comparison of hit profiles) Table 3.8 gives the number of at-bats, hits, doubles, triples
and home runs for the players during the 1900, 1945 and 1990 baseball seasons.
(a) For each year, compute the number of singles and put the numbers in the 1B column.

Table 3.8. Batting statistics for all players in the 1900, 1945 and
1990 baseball seasons
YEAR AB H 1B 2B 3B HR
1900 39132 10925 1432 607 254
1945 84447 21977 3497 728 1007
1990 142768 36817 6526 865 3317

(b) For each year, compute the proportion of hits that are doubles. (For 1900, for example,
you would divide the number of doubles 1432 by the number of hits 10925.) Put your
answers in the table below.
Proportion of hits that are
YEAR 1B 2B 3B HR
1900
1945
1990

(c) For each year, compute the proportion of hits that are singles, proportion of hits that
are triples, and the proportion of hits that are home runs.
(d) Based on the proportions you computed in (b) and (c), comment on how the game
of baseball has changed from 1900 to 1990. Would you say that the game is more or
less exciting than the games in the past? Why?
3.5. (Comparisons of rates of strikeouts and walks.) Table 3.9 gives the number of at-bats,
strikeouts, and walks for the three baseball seasons 1900, 1945, and 1990.
Table 3.9. Strikeout and walk statistics for all players
in the 1900, 1945 and 1990 baseball seasons
YEAR AB BB SO
1900 39132 3034 2697
1945 84447 8295 8050
1990 142768 13852 24390

(a) For each season, compute the number of plate appearances by adding the at-bats to
the walks. Put your answers in the PA column.
Proportion of
YEAR PA BB SO
1900
1945
1990
(b) For each year, compute the proportion of PAs that are strikeouts.
(c) For each year, compute the proportion of PAs that are walks.
(d) Explain how pitching has changed from 1900 to 1990 by looking at the proportions
you computed in (b) and (c). Have pitchers become more or less dominating over the
years? Do pitchers today have more or less control?

3.6. (Pujols vs Ramirez, continued.) Refer back to the Case Study 3.1 comparing the hitting
of Albert Pujols and Manny Ramirez.
(a) Construct back-to-back stemplots of the season on-base percentages (OBP) for Pujols
and Ramirez.
(b) Compare the two datasets. Which hitter generally has a higher OBP? Which hitter
has the larger variation in OBP from season to season?
(c) Repeat (a) and (b) using the season slugging percentages (SLG).
(d) Based on your work, who is the more valuable hitter, Pujols or Ramirez? Explain
why.
(e) Although Pujols may have better hitting statistics than Ramirez, some people still
claim that Ramirez is the more valuable player? Why might they say that?
3.7. Table 3.10 gives the number of wins (W), losses (L), winning proportion (PCT), and
Earned Run Average (ERA) for two great modern pitchers, Greg Maddux and Tom
Glavine.

Table 3.10. Career pitching statistics for Greg Maddux and Tom Glavine
Greg Maddux Tom Glavine
YEAR W L PCT ERA W L PCT ERA
1986 2 4 0.333 5.52
1987 6 14 0.300 5.61 2 4 0.333 5.54
1988 18 8 0.692 3.18 7 17 0.291 4.56
1989 19 12 0.612 2.95 14 8 0.636 3.68
1990 15 15 0.500 3.46 10 12 0.454 4.28
1991 15 11 0.576 3.35 20 11 0.645 2.55
1992 20 11 0.645 2.18 20 8 0.714 2.76
1993 20 10 0.666 2.36 22 6 0.785 3.2
1994 16 6 0.727 1.56 13 9 0.590 3.97
1995 19 2 0.904 1.63 16 7 0.695 3.08
1996 15 11 0.576 2.72 15 10 0.600 2.98
1997 19 4 0.826 2.2 14 7 0.666 2.96
1998 18 9 0.666 2.22 20 6 0.769 2.47
1999 19 9 0.678 3.57 14 11 0.560 4.12
2000 19 9 0.678 3 21 9 0.700 3.40
2001 17 11 0.607 3.05 16 7 0.696 3.57

(a) Using back-to-back stemplots, compare the season winning percentages of Maddux
and Glavine.
(b) Based on your comparison, which pitcher tended to win a greater percentage of
games?
(c) Compare the season ERAs of Maddux and Glavine using stemplots.
(d) Which pitcher tended to have a lower season ERA?
3.8. Table 3.11 shows the slugging percentage (SLG) for the “regular” shortstops and “regular”
3rd basemen in 2014.

Table 3.11. Slugging percentages for regular shortstops and

3rd basemen in 2014
Shortstops 3rd Basemen
Name SLG Name SLG
Andrus 0.333 Beltre 0.492
Aybar 0.379 Carpenter 0.375
Castro 0.438 Castellanos 0.394
Cozart 0.300 Dominguez 0.330
Crawford 0.389 Donaldson 0.456
Desmond 0.430 Johnson 0.361
Escobar 0.377 Longoria 0.404
Escobar 0.340 McGehee 0.357
Hardy 0.372 Moustakas 0.361
Hechavarria 0.356 Rendon 0.473
Mercer 0.387 Sandoval 0.415
Peralta 0.443 Seager 0.454
Ramirez 0.408 Wright 0.374
Reyes 0.398
Rollins 0.394
Segura 0.326
Simmons 0.331

(a) Construct parallel boxplots of the slugging percentages of the shortstops and the 3rd
basemen.
(b) Based on your work in (a), which type of player is more likely to be a slugger?
(c) On the average, how much superior is one group over the other group with respect
to slugging percentage?
(d) Is the general pattern you found in (b) and (c) true for all players? Find some players
that don’t agree with the general pattern.
3.9. (Triple rates for 1899 and 1999 players.) Old-timers would argue that there is less use
of speed in baseball today than in the old days. One indication of the lack of speed in
baseball today is the relatively small number of triples hit. (There are other explanations
for the small number of triples, including the liveliness of the ball and the change in ball
park design.) For each of the “regular” players (with at least 400 at-bats) in 1899 and
1999, we compute the TRIPLE RATE D 3B=AB. Stemplots of the triple rates for each
group of players are shown in Figure 3.14. For the stemplots, the break point is between
the hundredths and thousandths places, so for a triple rate of :048, the stem is 4 and the
leaf is 8.
(a) Compute five-number summaries of each dataset and graph parallel boxplots.
(b) Which group of players tended to get a higher rate of triples? What is the difference
between the two average triple rates?
(c) Does your comparison support the claim that there is less speed in baseball nowadays?
What other variables could you use to measure speed of ballplayers?

1899 players
0 | 114
0 | 6667999
1 | 00011222233334444
1 | 55555666788889999
2 | 01122233334
2 | 68
3 | 0012
3 | 7
4 | 23
1999 players
0 | 0000000000000000000000000000011111111111111111111
0 | 222222222222333333333333333
0 | 444444444444445555555
0 | 666666666666667777777
0 | 88888888888888889999999
1 | 0000000000111
1 | 222222333
1 | 44444455
1 | 666
1 | 889
2 | 0
Figure 3.14. Stemplots of triple rates of players in 1899 and 1999 seasons with at least 400 at-bats.

3.10. Table 3.12 displays the total number of home runs (HR) hit by each of the 30 major
league teams in 2014.

Table 3.12. Home run numbers of all Major

League teams in 2014
Team HR Team HR
BAL 211 ARI 118
BOS 123 ATL 123
CHA 155 CHN 157
CLE 142 CIN 131
DET 155 COL 186
HOU 163 LAN 134
KCA 95 MIA 122
LAA 155 MIL 150
MIN 128 NYN 125
NYA 147 PHI 125
OAK 146 PIT 156
SEA 136 SDN 109
TBA 117 SFN 132
TEX 111 SLN 105
TOR 177 WAS 152

(a) Using the stems shown below, construct back-to-back stemplots of the HR totals of
the American League and the National League.
American League National League

10
11
12
13
14
15
16
17
18
19
20
21
22

(b) Find five-number summaries of the two datasets.

(c) Using your work in (a) and (b), compare the two datasets. Did one league tend to hit
more home runs than the other league? Which teams hit an unusually small or large
number of home runs in 2014?
3.11. (Comparing two hitters.) Find two contemporary hitters who have each played ten or
more major league seasons. Compare the two hitters on the basis of season data using one
batting statistic such as AVG, SLG, HR, or OBP. You should (1) list the season data for
both players, (2) compare the datasets using appropriate graphs, (3) compute summaries
of both datasets, and (4) describe what you’ve learned from this comparison.
3.12. (Comparing two pitchers.) Find two contemporary pitchers who have each played ten
or more major league seasons. Compare the two pitchers on the basis of season data
using one pitching statistic such as ERA, PCT, SO, or BB. You should (1) list the
season data for both players, (2) compare the datasets using an appropriate graph, (3)
compute summaries of both datasets, and (4) describe what you’ve learned from this
comparison.
3.13. (Comparing two teams.) Find two teams that you are interested in comparing with respect
to some batting statistic, such as AVG, SLG, OBP, or HR. For each of the two teams, list
all of the regular players on the team for a particular season (the ones with at least 400
at-bats) and the value of the batting statistic for each player. Compare the two batches of
data using an appropriate graph, compute summaries of both batches, and describe which
team appears to be superior from the viewpoint of this batting statistic.
3.14. (Comparing NL and AL teams.) Compare the NL and AL teams for any season with
respect to a particular hitting, pitching, or fielding statistic. List the 30 teams and the
value of the statistic for each team. Compare the NL and AL batches using a graph,
compute statistics, and describe which league appears to be superior with respect to this
statistic.

3.15. (Comparing players from different eras.) Compare baseball hitters or pitchers from two
different years. Choose two years of interest (not too close together) and one baseball
statistic that you are interested in. Randomly pick at least 30 “regular” players from each
year, and list the name and the baseball statistic for each player. Compare the two years
of data using an appropriate graph, compute summary statistics, and describe what you
have learned in this comparison.
3.16. There has been much talk about the recent surge in home run hitting. How has the rate
of home run hitting changed since 1990? From a baseball web site, find the total number
of home runs and at-bats for each of the 30 teams for the current year and compute the
home run rate for the teams. Find the home run rates for all the 1990 teams. (This data is
available on the book website described in Appendix 2.) Comparing the 1990 and current
datasets, has there been a significant change in the rate of home run hitting?
3.17. Figure 3.15 displays parallel boxplots of the speeds of four-seam fastballs thrown by
Clayton Kershaw and Zack Greinke during the 2015 season.

Kershaw

Greinke

92 93 94 95
Fastball Speed

Figure 3.15. Speeds of four-seam fastballs (in miles per hour) by Clayton Kershaw and Zack Greinke
during 2015 season.

(a) By reading from the graph, find the five-number summaries of the pitch speeds for
the two pitchers.
(b) Which pitcher shows more variation in the fastball speeds? Explain what measure
was helpful in answering this question.
(c) Which pitcher, on average, throws a faster fastball? By how many mph?
3.18. Table 3.13 displays the number of changeups (CH), curve balls (CU), four-seam fastballs
(FF), two-seam fastballs (FT), and sliders (SL) thrown by Clayton Kershaw and Zack
Greinke during a single game for each pitcher in the 2015 season.

Table 3.13. Frequency of pitches of different types of Clayton Kershaw and

Zack Greinke during a game pitched by each pitcher in the 2015 season
CH CU FF FT SL
Clayton Kershaw 2 20 48 0 26
Zack Greinke 18 12 43 23 23

(a) For each pitcher, compute the percentages of pitches of each type.
(b) Describe, using several sentences, how the pitch selection differs between the two
pitchers.
3.19. Table 3.14 shows batted ball data for Ichiro Suzuki for 13 seasons of his career.
Table 3.14. Batted ball statistics for Ichiro Suzuki for seasons 2002 through
2015
Direction Hardness

Season Pull Center Opposite Soft Medium Hard

2002 0.307 0.298 0.395 0.196 0.611 0.193
2003 0.383 0.270 0.347 0.176 0.601 0.223
2004 0.337 0.313 0.350 0.161 0.639 0.200
2005 0.375 0.280 0.345 0.171 0.593 0.237
2006 0.331 0.351 0.318 0.161 0.613 0.225
2007 0.293 0.379 0.328 0.201 0.506 0.293
2008 0.341 0.344 0.314 0.207 0.587 0.206
2009 0.319 0.333 0.349 0.198 0.615 0.187
2010 0.229 0.405 0.366 0.232 0.545 0.222
2011 0.299 0.392 0.310 0.323 0.558 0.119
2012 0.296 0.396 0.308 0.182 0.640 0.178
2013 0.299 0.426 0.275 0.245 0.578 0.176
2014 0.244 0.437 0.319 0.298 0.512 0.190
2015 0.227 0.382 0.391 0.290 0.628 0.082

(a) Suzuki has a very different pattern of batted balls than Albert Pujols whose batted
ball statistics are displayed in Table 2.25 (Exercise 2.34). Use a suitable graph to
compare the proportions of “Pull” batted balls for Suzuki and Pujols. Write a short
paragraph explaining how the two batters differ.
(b) Do a similar comparison of Suzuki and Pujols using the proportion of “Hard” batted
balls.

Further Reading
Devore and Peck (2011) and Moore, McCabe and Craig (2012) provide good descriptions of
the exploratory methods used in this chapter to compare several batches of data. Chapter 2 of
Albert and Bennett (2003) illustrates the use of these methods to compare batches of baseball
data.

What’s On-Deck?
This chapter illustrates statistical methods for studying the association between different baseball
statistics. A large number of statistics, such as hits, doubles, triples, home runs, batting average,
slugging percentage, and on-base percentage are used to evaluate hitters. Case Study 4.1 explores
the relationships between these different batting statistics for the teams in the 2014 baseball
season. The basic tool used in this case study is the scatterplot, and the patterns in the scatterplots
are informative in assessing the strength and direction of association between pairs of variables.
The most valuable team batting statistic is the one that is most highly associated with runs
scored per game. In Case Study 4.2, we rank the different batting statistics with respect to their
correlation with runs scored. In Case Study 4.3, we go one step further and evaluate a number
of batting statistics, such as batting average, slugging percentage, on-base percentage, and OPS
(on-base percentage plus slugging percentage) by how close each statistic can be used to predict
teams’ runs scored per game. Multiple regression is a useful tool for finding the “best” batting
statistic for predicting team runs based on the number of singles, doubles, triples, home runs,
walks, and hit-by-pitch by the team. In Case Study 4.4, we find the best linear combination of
these batting events using the least-squares criterion, and the weights of this linear combination
are useful for comparing the worth of a single, double, triple, and home run from the standpoint
of producing runs.
The remaining case studies illustrate some interesting relationships in baseball data. In
Case Study 4.5 we illustrate Bill James’ Pythagorean Formula which is an empirical relationship
between the ratio of a team’s wins and losses and the square of the ratio of a team’s runs scored
and runs allowed. In Case Study 4.6, we see that there is a natural tendency for a player’s statistic
one year to regress towards the average value of the statistic for the next year. Despite popular
opinion, it is natural for a hot-hitting rookie to have a less-than-hot hitting performance the
following year.

4.1 Relationships in Team Offensive Statistics

Topics Covered: Relationships between two measurement variables, scatterplot, looking for
association.

73
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
74 Relationships Between Measurement Variables

Table 4.1. Team names, league, and batting statistics for the teams in the 2014 baseball season
Team G AB R H 2B 3B HR AVG SLG OBP
ARI 162 5552 615 1379 259 47 118 0.248 0.376 0.302
ATL 162 5468 573 1316 240 22 123 0.241 0.360 0.305
BAL 162 5596 705 1434 264 16 211 0.256 0.422 0.311
BOS 162 5551 634 1355 282 20 123 0.244 0.369 0.316
CHA 162 5543 660 1400 279 32 155 0.253 0.398 0.310
CHN 162 5508 614 1315 270 31 157 0.239 0.385 0.300
CIN 162 5395 595 1282 254 20 131 0.238 0.365 0.296
CLE 162 5575 669 1411 284 23 142 0.253 0.389 0.317
COL 162 5612 755 1551 307 41 186 0.276 0.445 0.327
DET 162 5630 757 1557 325 26 155 0.277 0.426 0.331
HOU 162 5447 629 1317 240 19 163 0.242 0.383 0.309
KCA 162 5545 651 1456 286 29 95 0.263 0.376 0.314
LAA 162 5652 773 1464 304 31 155 0.259 0.406 0.322
LAN 162 5560 718 1476 302 38 134 0.265 0.406 0.333
MIA 162 5538 645 1399 254 36 122 0.253 0.378 0.317
MIL 162 5462 650 1366 297 28 150 0.250 0.397 0.311
MIN 162 5567 715 1412 316 27 128 0.254 0.389 0.324
NYA 162 5497 633 1349 247 26 147 0.245 0.380 0.307
NYN 162 5472 629 1306 275 19 125 0.239 0.364 0.308
OAK 162 5545 729 1354 253 33 146 0.244 0.381 0.320
PHI 162 5603 619 1356 251 27 125 0.242 0.363 0.302
PIT 162 5536 682 1436 275 30 156 0.259 0.404 0.330
SDN 162 5294 535 1199 224 30 109 0.226 0.342 0.292
SEA 162 5450 634 1328 247 32 136 0.244 0.376 0.300
SFN 162 5523 665 1407 257 42 132 0.255 0.388 0.311
SLN 162 5426 619 1371 275 21 105 0.253 0.369 0.320
TBA 162 5516 612 1361 263 24 117 0.247 0.367 0.317
TEX 162 5460 637 1400 260 28 111 0.256 0.375 0.314
TOR 162 5549 723 1435 282 24 177 0.259 0.414 0.323
WAS 162 5542 686 1403 265 27 152 0.253 0.393 0.321

In this case study we discuss relationships in baseball hitting data. Table 4.1 lists batting
statistics for all 30 teams for the 2014 baseball season. These data give G, AB, R, H, 2B, 3B,
HR, RBI, AVG, TB, SLG, and OBP for all teams. When there are many variables measured on
each team, we are interested in exploring relationships between them.

Relating Slugging Percentage and On-Base Percentage

A first step in examining the relationship between two numerical variables, say SLG and OBP, is
to draw a scatterplot. This graph plots each pair of values on a grid, and we look for association
by finding particular patterns in the display.
Note from Table 4.1 that Arizona (ARI) has an SLG value of .376 and an OBP value of
.302. Here we’ve decided to use OBP on the vertical scale and SLG on the horizontal. (It would
not have mattered if we had switched what variables were on which axis.) We plot (SLGD.376,

0.33
0.32
OBP
0.31
0.30

0.34 0.36 0.38 0.40 0.42 0.44

SLG

Figure 4.1. Scatterplot of slugging percentage and on-base percentage for the 2014 Major League teams.

OBPD.302) by plotting the point (.376, .302) on a Cartesian grid. We continue for the remaining
29 teams, getting the display shown in Figure 4.1.
We see a general tendency for the points to drift from (left, low) to (right, high), indicating
that teams with a low OBP tend also to have a low SLG, and teams with high OBP tend to have
a high SLG. This is a positive association in the scatterplot.

Relating Triples and Doubles

As a second example, we draw a labeled scatterplot, shown in Figure 4.2, of the number of triples
and number of doubles hit by the 30 teams. Each plotting point has a two-letter abbreviation
label for the team.

ARI
45

SFN
COL
40

LAN
MIA
35
Triples

OAK
SEA CHA
CHN LAA
30

SDN PIT
KCA
TEX MIL
PHI WAS MIN
NYA DET
25

TBA TOR
CLE
ATL
SLN
20

CIN BOS
HOU NYN

BAL
15

220 240 260 280 300 320

Doubles

Figure 4.2. Scatterplot of numbers of doubles and triples for the 2014 Major League teams.

The team labels give some interesting information: Colorado (COL) was high in both triples
and doubles and Houston (HOU) was low on both variables. Looking at the pattern in the graph,
we don’t see a very strong association between the numbers of doubles and triples hit. The points
may be drifting slightly upward as one goes from left to right, but the positive relationship is
clearly weaker than the relationship we saw above between OBP and SLG.

Relating Home Runs and Triples for Historical Teams

To illustrate a different type of relationship, 30 teams were randomly selected from the group
of all major league baseball teams in the years from 1901 to 2013. For each team, we collected
two hitting statistics:

Home run rate (HR/G): the number of home runs hit per game. This statistic is computed by
dividing the number of team home runs by the number of games played.
Triple rate (3B/G): the number of triples hit per game. This is computed by dividing the
number of triples by the number of games.

Table 4.2 gives the team names, the years, and the two hitting statistics.

Table 4.2. Team names, years, home runs per game, and triples per game for 30 historical
baseball teams
Team Year HR Rate T Rate Team Year HR Rate T Rate
WS1 1916 0.0023 0.0117 NYA 2001 0.0364 0.0036
CHN 1983 0.0254 0.0076 NY1 1942 0.0209 0.0067
ATL 1982 0.0265 0.0040 NYA 1939 0.0313 0.0104
CLE 1983 0.0157 0.0057 PHI 1927 0.0107 0.0087
SDN 2003 0.0231 0.0058 DET 1934 0.0135 0.0097
SDN 1984 0.0198 0.0076 WS1 1946 0.0112 0.0118
CIN 1902 0.0037 0.0157 SLA 1944 0.0137 0.0085
PHA 1934 0.0271 0.0094 NY1 1923 0.0156 0.0139
BAL 1987 0.0378 0.0036 SLN 1906 0.0020 0.0136
PIT 1972 0.0200 0.0086 CLE 1932 0.0144 0.0137
MON 1989 0.0182 0.0055 CHN 1999 0.0345 0.0064
PIT 1975 0.0251 0.0086 ML4 1973 0.0262 0.0072
BOS 1942 0.0196 0.0105 HOU 2007 0.0298 0.0054
LAN 2008 0.0249 0.0053 SDN 2000 0.0282 0.0067
DET 1943 0.0144 0.0088 PIT 1907 0.0038 0.0157

In the scatterplot in Figure 4.3, we plot the home run rate against the triple rate for the
30 teams. Here the points drift from top left to bottom right, which corresponds to a negative
relationship in the graph. Teams that hit a high rate of home runs tended to hit a low rate of
triples, and teams that hit a low rate of home runs tended to have a high triple rate. We can explain
this relationship if we look at the years of the teams in the table. In Figure 4.4, we have redrawn
the scatterplot, where the plotting symbol corresponds to the time when the team played.
Note that the points in the upper left portion of the scatterplot generally correspond to
teams that played in the first quarter of the 20th century, and the points in the lower right section

0.016
0.014
0.012
Triples.Rate
0.010
0.008
0.006
0.004

0.005 0.010 0.015 0.020 0.025 0.030 0.035

HomeRun.Rate

Figure 4.3. Scatterplot of home run rates against triple rates for 30 historical baseball teams.

correspond to teams that played in the most recent period from 1991–2014. In the early days
of baseball, relatively few home runs were hit and it was important to use players’ speed to hit
triples to score runs. In recent years, speed has played much less of a factor in scoring runs, and
the home run is an important offensive weapon. So the association structure in this scatterplot
tells us how the game of baseball has changed in the last 100 years.
0.016

1901-1930
1931-1960
1961-1990
0.014

1991-2014
0.012
Triples.Rate
0.010
0.008
0.006
0.004

0.005 0.010 0.015 0.020 0.025 0.030 0.035

HomeRun.Rate

Figure 4.4. Scatterplot of home run rates against triple rates for 30 historical baseball teams where the
plotting symbol gives the era of the team.

4.2 Runs and Offensive Statistics

Topics Covered: Scatterplots and looking for association, correlation. Let’s talk more about
baseball offensive statistics:

Since the objective of a baseball team is to score more runs than its opponent, the most
important offensive statistic is runs.
The value of any offensive statistic depends on its relationship with runs.
We are interested in finding the offensive statistic that has the strongest association with runs.

Here is a list of the basic offensive statistics.

H 2B 3B HR RBI AVG TB SLG OBP

How strong is the relationship of these statistics with runs scored (R) by a team? We make some
preliminary observations based on our knowledge of baseball.

Clearly, RBI will have a strong positive relationship with runs, since RBI counts the number
of runs that are batted in by a team.
Triples (3B), in contrast, seem to have a relatively weak relationship with runs. Teams that
get a lot of triples don’t necessarily score more runs. This is true since triples are a relatively
rare event and a large number of triples might just be a reflection of the team’s speed or the
configuration of the ballpark.
Home runs (HR) and doubles (2B) have stronger relationships with runs since they occur
more frequently than triples and do typically result in runs. (Home runs actually cause at least
one run to be scored.)
TB (total bases) and SLG (slugging percentage) are essentially the same stat (you divide TB
by AB to get SLG), so they both have about the same relationship with runs.
AVG is probably less associated than OBP or SLG with runs. Unlike SLG, AVG doesn’t give
added weight to extra base hits (doubles, triples, and home runs) over singles; and unlike OBP,
AVG doesn’t reflect walks or HBP, which both contribute to runs scored by a team.

From this discussion, we get the following list of the statistics ranked with respect to our
expected strength of the association with runs scored.

Statistic Rank
RBI High association with runs
OBP
SLG, TB
AVG
HR
2B
3B Low association with runs

At this point, we are thinking about association by means of the pattern that we see in the
scatterplot. We can summarize these visual associations by means of correlations. Briefly, a
correlation is a number between 1 and 1 that summarizes the straight-line assocation between

0.30 0.32 220 260 300 100 160

550 650 750

R
0.32

OBP
0.30

0.34 0.38 0.42

SLG
220 260 300

X2B

15 25 35 45
X3B
160

HR
100

550 650 750 0.34 0.38 0.42 15 25 35 45

Figure 4.5. Scatterplot matrix of different offensive statistics for 30 Major League teams in the 2014
season.

two numerical variables. A stronger association corresponds to a correlation value that is closer
to the extreme values 1 and 1.
We return to the team offensive statistics from the 2014 season discussed previously. We’re
interested in the offensive statistics that are most associated with runs scored. Look at the
scatterplot matrix in Figure 4.5 and focus on the plots in the first column; these are the ones that
plot runs (R) against OBP, SLG, 2B, 3B, HR.
Some comments:

Runs are strongly positively associated with OBP and SLG, although the relationship with
SLG looks a bit stronger.
Runs are somewhat positively associated with HR; the association is less strong than the
association with OBP and SLG.
The relationship between Runs and 2B, and Runs and 3B looks weak.

We can support these comments by computing correlations. The matrix in Table 4.3 gives
correlations between the different offensive team statistics. Using these correlation values, we
rank the most valuable offensive statistics in Table 4.4.

Table 4.3. Correlation matrix between a number of different offensive team statistics
R OBP SLG 2B 3B HR
R 1 0.797 0.857 0.737 0.202 0.578
OBP 0.797 1 0.668 0.724 0.104 0.260
SLG 0.857 0.668 1 0.676 0.209 0.788
2B 0.737 0.724 0.676 1 0.090 0.238
3B 0.202 0.104 0.209 0.090 1 0.094
HR 0.578 0.260 0.788 0.238 0.094 1

Table 4.4. Correlations of different

offensive statistics with runs scored,
ordered by the size of the correlation
Statistic Correlation with Runs
SLG 0.857
OBP 0.797
2B 0.737
HR 0.578
3B 0.202

Of the two “averages”, while SLG is the best OBP is nearly as good. Of the three count
variables, HR is by far the most strongly correlated with runs scored, but it lags well behind the
two averages. Doubles (2B) and triples (3B) are inferior to HR, with 3B being virtually useless.

4.3 Most Valuable Hitting Statistics

Topics Covered: Linear regression, prediction.
There is a group of baseball fans who are members of SABR, the Society of American
Baseball Research. These people are interested in the history of the game and write many books
and articles about the game. Sabermetrics is the mathematical and statistical study of baseball
records.
We focus on one of the most important topics to sabermetricians: What is the most valuable
hitting statistic?
Who is the better hitter Mike Trout or Miguel Cabrera? This can be a difficult question to
answer since batting has two different characteristics, the ability to get on-base and the ability
to advance runners on base to score. Your decision about who is the better hitter may depend
partly on how you value power hitting relative to on-base ability. We want to find statistics that
can distinguish the “total” hitting accomplishments of Trout and Cabrera.
Let’s review the basic hitting statistics recognized by Major League Baseball.

AVG, the batting average, is the most commonly quoted hitting statistic of Major League
Baseball. The player who wins the batting crown is the one with the highest batting average.
OBP, the on-base percentage.
SLG, the slugging percentage.

Starting in the 1960’s, a large number of new statistics have been introduced as “improve-
ments” to the basic three statistics:
OPS, the sum of OBP and SLG.
Runs created (RC) introduced by Bill James. It is defined as
.H C BB/TB
RC D :
AB C BB
Total average (TA). It is like SLG, but it is the ratio of bases to the number of outs. (An
approximate formula is given below which ignores rare events such as stealing, hit-by-pitches,
and ground balls resulting in double plays.)
TB C BB
TA D :
AB H
Batter’s runs average (BRA), the product of OBP and SLG.
How do we evaluate all of these hitting statistics?
First, we have to understand that the goal of hitting is to create runs, and a single player
can’t create a run (unless he hits a home run). Teams create runs and so we have to look at team
data to understand the usefulness of these various statistics.
Let’s consider team statistics for the 2014 American League teams displayed in Table 4.5.
(We present data for only one league in this case study to make it easier to present the method-
ology for evaluating a hitting statistic.)

Table 4.5. Offensive statistics for the 2014 American League teams
Team R G R/G AVG OBP SLG OPS RC TA BRA
BAL 705 162 4.35 0.256 0.311 0.422 0.733 723 0.664 0.131
BOS 634 162 3.91 0.244 0.316 0.369 0.685 635 0.615 0.117
CHA 660 162 4.07 0.253 0.310 0.398 0.708 673 0.634 0.123
CLE 669 162 4.13 0.253 0.317 0.389 0.706 683 0.641 0.123
DET 757 162 4.67 0.277 0.331 0.426 0.757 790 0.698 0.141
HOU 629 162 3.88 0.242 0.309 0.383 0.692 636 0.624 0.118
KCA 651 162 4.02 0.263 0.314 0.376 0.690 646 0.603 0.118
LAA 773 162 4.77 0.259 0.322 0.406 0.728 731 0.665 0.131
MIN 715 162 4.41 0.254 0.324 0.389 0.713 693 0.652 0.126
NYA 633 162 3.91 0.245 0.307 0.380 0.687 632 0.613 0.117
OAK 729 162 4.50 0.244 0.320 0.381 0.701 668 0.644 0.122
SEA 634 162 3.91 0.244 0.300 0.376 0.676 604 0.593 0.113
TBA 612 162 3.78 0.247 0.317 0.367 0.684 632 0.614 0.116
TEX 637 162 3.93 0.256 0.314 0.375 0.689 633 0.607 0.118
TOR 723 162 4.46 0.259 0.323 0.414 0.737 735 0.680 0.134

The important statistic for a team is the number of runs scored. If we divide this number
by the number of games, we get the RUNS/GAME statistic (or R/G). For example, Baltimore
scored 705 runs in 162 games, so
705
R/G D D 4:35:
162

4.8
4.6
4.4
R.G
4.2
4.0
3.8

0.245 0.250 0.255 0.260 0.265 0.270 0.275

AVG

Figure 4.6. Scatterplot of batting average and runs scored per game for 2014 American League teams.

We are interested in seeing the effectiveness of different hitting statistics in predicting R/G.
Let’s start with the traditional AVG measure. How well can we predict a team’s runs per game
(R/G) using AVG? If we do a scatterplot of R/G against AVG (shown in Figure 4.6) we see a
positive association—teams that hit for high averages tend to score more runs.
The standard “least-squares” fit to these data gives the relationship

R/G D 20:48 AVG 1:00:

Is this a good fit? In other words, how close are the points in the scatterplot to the fitted line?
We evaluate the goodness of fit of this prediction equation in several steps. The calculations are
summarized in the following table.

1. First, we find the predicted values of R/G for each team. For example, we see that Baltimore
had a batting average of :256. We would predict its R/G to be R/G D 20:48.:256/ 1:00 D
4:24. This value is placed in the “Predicted” column.
2. The residual is the difference between the actual R/G for a team and its predicted R/G. In the
case of Anaheim, it actually scored an average of 4.35 runs per game and we predicted 4:24,
so

RESIDUAL D ACTUAL R/G PRED R/G D 4:35 4:24 D 0:11:

In Figure 4.7, we have drawn the least-squares line on the scatterplot. The vertical lines
drawn from each point represent the residuals. (This least-squares line is actually the line that
minimizes the sum of the squared residuals.) We have identified two teams with unusually
large residuals. LAA and KC both had a season batting average close to 260. However,
LAA’s R/G in 2014 was much greater than the R/G predicted by using batting average and
the residual is a large positive value. In contrast, KC scored a very small number of runs
given its batting average and has a large negative residual.

4.8
LAA

4.6
4.4
R.G
4.2
4.0

KC
3.8

0.245 0.250 0.255 0.260 0.265 0.270 0.275

AVG

Figure 4.7. Scatterplot of batting average and runs scored per game for 2014 American League teams,
with least-squares fit and residuals displayed.

3. We summarize the sizes of these residuals by use of a Root Mean Square Error (RMSE)
criterion. Table 4.6 illustrates the calculation of the RMSE. We first square all the residuals
and put the answers in the “Residual2 ” column.
The Sum of Squared Errors is the sum of values in the “Residual2 ” column. Here it is
0.9092. The mean of the Squared Residuals, called Mean Square Error or MSE, is
MSE D 0:9092=15 D 0:0606:
The Root Mean Square Error, RMSE, is the square root of MSE:
p
RMSE D 0:0606 D 0:246:
The RMSE is a measure of the size of the residuals from the model that uses AVG to predict
R/G. Here is a nice interpretation of RMSE: Generally, if you graph all of the residuals,
you will find them approximately normally distributed with mean 0 and standard deviation
RMSE. Since we have a normal distribution, we can apply the 68-95-99.7 rule and say that
68% (roughly 2=3) of the residuals fall between RMSE and CRMSE To illustrate this
interpretation, suppose we use AVG to predict R/G for the American League teams. We
found that RMSE = 0.246. This means if you use AVG to predict R/G, then roughly 2=3
(about 9) of the residuals will fall between 0:246 and C0:246.
How good are other batting measures? Now let’s try using the batting statistic OBP to
predict R/G. The least-squares line is
R/G D 29:6.OBP/ 5:16:
As we did above, we find the predicted values of R/G for each team and compute the residuals
and squared residuals and put the results in Table 4.7. The sum of squared residuals is equal
to .6798 and the RMSE is equal to
r
0:6798
RMSE D D :213:
15
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
84 Relationships Between Measurement Variables

Table 4.6. Calculation of the Root Mean Square Error criterion for the least-squares fit with
AVG as the predicting variable
Team AVG R.G Predicted Residual Residual2
BAL 0.256 4.35 4.24 0.11 0.0121
BOS 0.244 3.91 3.99 0.08 0.0064
CHA 0.253 4.07 4.18 0.11 0.0121
CLE 0.253 4.13 4.18 0.05 0.0025
DET 0.277 4.67 4.67 0 0.0000
HOU 0.242 3.88 3.95 0.07 0.0049
KCA 0.263 4.02 4.38 0.36 0.1296
LAA 0.259 4.77 4.30 0.47 0.2209
MIN 0.254 4.41 4.20 0.21 0.0441
NYA 0.245 3.91 4.01 0.10 0.0100
OAK 0.244 4.50 3.99 0.51 0.2601
SEA 0.244 3.91 3.99 0.08 0.0064
TBA 0.247 3.78 4.06 0.28 0.0784
TEX 0.256 3.93 4.24 0.31 0.0961
TOR 0.259 4.46 4.30 0.16 0.0256
Sum 0.9092

Let’s summarize what we learned. The RMSE is a measure of the size of the residuals in our
prediction. When we used AVG to predict Runs/Game, the RMSE was equal to 0:246; when we
used OBP, the RMSE is equal to 0:213. This tells us that the residuals are generally much larger
using AVG instead of using OBP as a predictor. This means that OBP is a much better predictor

Table 4.7. Calculation of the Root Mean Square Error criterion for the least-squares fit with
OBP as the predicting variable
Team OBP R.G Predicted Residual Residual2
BAL 0.311 4.35 4.04 0.31 0.0961
BOS 0.316 3.91 4.19 0.28 0.0784
CHA 0.310 4.07 4.01 0.06 0.0036
CLE 0.317 4.13 4.22 0.09 0.0081
DET 0.331 4.67 4.63 0.04 0.0016
HOU 0.309 3.88 3.98 0.10 0.0100
KCA 0.314 4.02 4.13 0.11 0.0121
LAA 0.322 4.77 4.37 0.40 0.1600
MIN 0.324 4.41 4.43 0.02 0.0004
NYA 0.307 3.91 3.92 0.01 0.0001
OAK 0.320 4.50 4.31 0.19 0.0361
SEA 0.300 3.91 3.72 0.19 0.0361
TBA 0.317 3.78 4.22 0.44 0.1936
TEX 0.314 3.93 4.13 0.20 0.0400
TOR 0.323 4.46 4.40 0.06 0.0036
Sum 0.6798

of Runs Scored than AVG using the RMSE criterion. (In general, the smaller the RMSE, the
closer the predictions.)

A Prediction Contest
We used each of our seven batting statistics to predict runs per game for teams for our American
League data. Table 4.8 displays the corresponding values of RMSE:

Table 4.8. Values of the root mean

squared criterion using each of
seven batting statistics to predict
the runs scored per game for the
2014 American League team data
Statistic RMSE
AVG 0.246
OBP 0.213
SLG 0.199
OPS 0.159
TA 0.151
RC 0.147
BRA 0.153

We see that OPS (OBP C SLG), TA, RC, and BRA (OBP SLG) are all pretty close at the
top, OBP and SLG are a bit behind, and AVG is pulling up the rear. Of course, we just looked
at 2014 data. Is it possible that another batting statistic would do better a different year?’ Albert
and Bennett (2001) looked at all years from 1952 to 1999. Each year, they used each statistic
to predict runs per game for the teams data, and found RMSE for each statistic. Albert and
Bennett’s general conclusions were:

AVG performs terribly in predicting runs scored. This measure should be thrown out. (But it
probably won’t be.)
Both SLG and OBP are relatively mediocre measures.
The “brainstorming” statistics TA, BRA, RC, OPS are all good measures and they are pretty
close. You can’t say for sure that one statistic in this group is better than another in the
group.

Remember our motivation for finding the best hitting statistic? We want to use a good
statistic to compare the relative worth of two hitters.
To illustrate the use of good hitting statistics to compare players, let’s compare Mike Trout
and Miguel Cabrera who were both candidates for Most Valuable Player (MVP) in the 2013
season. Table 4.9 gives the batting average and the six alternative batting measures for both
Trout and Cabrera for this season.

Table 4.9. Batting measures for Mike Trout and Miguel Cabrera for the 2013 season
AVG OBP SLG OPS TA RC BRA
Mike Trout .323 .432 .557 .988 1.098 141 .241
Miguel Cabrera .348 .442 .636 1.078 1.224 155 .281

Note from the table that, although Cabrera and Trout had similar on-base performances, Cabrera
was substantially better than Trout using each of the four “new” batting statistics. Cabrera clearly
had a much better hitting year in 2013.

4.4 A New Measure of Offensive Performance

Topics Covered: Multiple linear regression, root mean square error, least-squares criterion.
In the earlier case studies, we introduced a number of batting measures, some old and
some new, for evaluating the value of a hitter. We used the 2014 American League team data
to evaluate these different measures. Specifically, we used each measure to predict the average
runs per game for the 15 AL teams, and compared the measures by computing the average size
of a residual (RMSE). We found that many of the modern batting measures, such as OPS, and
RC did substantially better compared to the traditional measures AVG and SLG for predicting
run production.
There is actually a straightforward way of finding a “best” measure of hitting performance
using the tool of multiple linear regression.
The basic batting counts are the numbers of singles, doubles, triples, and home runs (1B,
2B, 3B, and HR, respectively), and the number of walks (BB) and hit-by-pitch (HBP). The goal
is to combine these different counts in some way to obtain an accurate prediction of the runs
scored per game.
Many of the standard batting measures combine these batting counts in a linear way. For
example, the batting average AVG can be expressed as
1 1 1 1
AVG D 1B C 2B C 3B C HR;
AB AB AB AB
where singles, doubles, triples, and home runs are given equal weights. The slugging percentage
SLG is also a linear measure of the form
1 2 3 4
SLG D 1B C 2B C 3B C HR;
AB AB AB AB
where a hit is weighted by the number of bases. The on-base percentage, OBP, weights all
on-base events (hits, walks, and hit-by-pitch) equally:
1 1 1 1 1 1
OBP D BB C HPB C 1B C 2B C 3B C HR;
PA PA PA PA PA PA
where PA D AB C BB C HBP is the number of plate appearances.
Suppose that we consider an alternative batting measure, called OPTAVG (for optimal
average), that combines all of the batting events in a linear way with arbitrary weights:
OPTAVG D w0 C w1 1B C w2 2B C w3 3B C w4 HR C w5 BB C w6 HBP:
Can we find values of the weights w0 ; : : : ; w6 that give a “best” batting measure?

Fortunately, there is an easy way to find optimal weights for this batting measure using
a least-squares criterion. As in the earlier case study, suppose that we have team hitting data.
For each team, we observe the runs scored per game (R/G) and the counts of the six batting
events (1B, 2B, 3B, HR, BB, HBP). The goal is to predict R/G based on the linear measure
OPTAVG. Using the least-squares criterion, we wish to find values of the weights w0 ; : : : ; w6
that minimize the sum of squared residuals. Values of these weights, called the least-squares
estimates, are easily available using any standard statistical computing package. We apply this
method to estimate R/G for the 30 Major League Teams in 2014. Using the R package, we find
the regression equation to be

R/G D 1:83144 C 0:00251 1B C 0:00563 2B C 0:00761 3B

C 0:00725 HR C 0:00175 BB 0:00092 HBP:

Before we interpret the weights of this measure, let’s explain why this is the best linear batting
measure. Among all batting statistics that combine the different batting events (single, double,
etc.) in a linear way, this measure will give the smallest value of RMSE, which is an average
size of a residual when this measure is used as a predictor. Since this “best linear” measure has
the smallest RMSE, it will have a smaller RMSE than other linear measures such as AVG, SLG,
OBP, and OPS.
To demonstrate this fact, we also tried using AVG, SLG, OBP, OPS, and RC, each alone,
as predictors for this 2014 Major League team dataset. Table 4.10 gives values of RMSE for all
30 Major League teams in 2014 using the best linear and the more familiar measures.

Table 4.10. Values of the root mean

squared criterion using the best linear
and other measures for the 2014 Major
League team data
Statistic RMSE
OPTAVG 0.116
AVG 0.205
OBP 0.206
SLG 0.176
OPS 0.144
TA 0.138
RC 0.129
BRA 0.141

We see that the best linear measure OPTAVG has an RMSE value, 0:116, that appears to be
much smaller than its nearest competitor RC (0:129) and the popular OPS statistic (0:144). But
this difference is a bit deceptive since we found the best weights using the same 2014 dataset—it
is not clear that this best measure will also be good in predicting the runs scored per game for
team data for 2015 year.
The weights of the best linear measure OPTAVG tell about the worth of each type of batting
event. When we compute a batting average, we give each type of hit the same weight and we
ignore walks and hit-by-pitches. A slugging percentage weights each hit by the number of bases

Table 4.11. Weights of batting events for the best linear

measure OPTAVG
Event Weight Wt/Wt (Single)
Single 0.00251 1
Double 0.0056 2.25
Triple 0.00761 3.04
Home run 0.00725 2.90
Walk 0.00175 0.70
Hit by Pitch 0.00092 0.37

produced. What does our best linear measure do for weights? Table 4.11 shows the values of
the weights for each base event, and, to make the comparison easy, it standardizes the weights
by dividing each by the weight of a single.
Some of the values of these weights are a bit surprising. The weight for a home run is
actually smaller than the weight given to a triple! A hit by pitch is given a negative weight and
a triple and home run are given roughly the same weights.

How can we explain these weights?

The first comment is that we are using a relatively small amount of data and these unusual
weight values are a consequence of the fact that we are fitting a relatively complicated model to
a small dataset. We could get more reasonable estimates at these weights by using, say, 10 years
of team data instead of just one year. Albert and Bennett (2001) (Chapter 5) do exactly that and
get more reasonable weights for this linear batting measure.
However, some of these weights do make sense. It is well known by quantitative baseball
people that a home run is not four times the worth of a single—its actual worth is closer to
three times a single which is the value in this table. A walk and single should have weights of
a similar size since all these events get the batter to first base. But the single is a more valuable
event than a walk since the runners are typically advanced more than one base. In other words,
the larger weight for a single reflects the fact that the single is more effective than a walk in
producing runs.

4.5 How Important is a Run?

Topics Covered: Nonlinear regression, transformation, residuals.
In this chapter, we have discussed how it is important for a baseball team to score runs
and we evaluated the goodness of different batting measures by their relationship with runs
scored. But of course the objective of a baseball team is not solely to score runs—it wins
games by scoring more runs than their opponent. That raises the interesting question. What
is the importance of a single run towards the goal of winning a baseball game? If a player is
responsible for scoring, say 20 runs, then how many wins for his team has he contributed? Bill
James discovered a special relationship between the number of wins (W) and losses (L) for a
baseball team and the number of runs scored (R) and number of runs allowed (RA). He called
this relationship “The Pythagorean Method.” This result says that the ratio between a team’s
wins and losses is approximately equal to the square of the ratio of runs scored to runs allowed.

That is, approximately,

2
W R
D :
L RA

If we take logs of both sides, we get the equivalent relationship

W R
log D 2 log :
L RA

(We take logs to convert a nonlinear equation in the runs ratio to a linear equation.) Can we
demonstrate that the Pythagorean Method gives a good description of the relationship between
the win/loss pattern and the runs scored/allowed for current Major League Baseball teams? To
answer this question, we look at the relevant team statistics (wins, losses, runs scored, and runs
allowed) for the 30 teams for the 2014 season. Looking at Table 4.12, we see that if a team has a
winning record, then it generally scores more runs than its opponent. But there is one interesting
exception—the New York Yankees (NYA) had a win/loss record of 84-78 but actually allowed
664 633 D 31 more runs this season than they scored. (We suppose that the Yankees won a
lot of close games in 2014.)

Table 4.12. Team statistics for the Major League teams in the 2014 season
Team W L R RA Team W L R RA
ARI 64 98 615 742 MIL 82 80 650 657
ATL 79 83 573 597 MIN 70 92 715 777
BAL 96 66 705 593 NYA 84 78 633 664
BOS 71 91 634 715 NYN 79 83 629 618
CHA 73 89 660 758 OAK 88 74 729 572
CHN 73 89 614 707 PHI 73 89 619 687
CIN 76 86 595 612 PIT 88 74 682 631
CLE 85 77 669 653 SDN 77 85 535 577
COL 66 96 755 818 SEA 87 75 634 554
DET 90 72 757 705 SFN 88 74 665 614
HOU 70 92 629 723 SLN 90 72 619 603
KCA 89 73 651 624 TBA 77 85 612 625
LAA 98 64 773 630 TEX 67 95 637 773
LAN 94 68 718 617 TOR 83 79 723 686
MIA 77 85 645 674 WAS 96 66 686 555

To look for the Pythagorean relationship, we compute log(W/L) and log(R/RA) for all
teams and construct a scatterplot of the two quantities in Figure 4.8.
We see a linear positive association in this graph, indicating that there is indeed a linear
association between log.W/L/ and log.R/RA/. Next we want to fit a “best line” to this graph. It
seems natural to restrict this line to pass through one point. If a team scores the same number
of runs against its opponents (R D RA), then we expect the team to win half of its games
(W D L). In other words, the point .log.R/RA/; log.W/L// D .0; 0/ should fall on the line.

0.4
0.2
log(W/L)
0.0
-0.2
-0.4

-0.2 -0.1 0.0 0.1 0.2

log(R/RA)

Figure 4.8. Scatterplot of log runs ratio against log of ratio of wins to losses for Major League team data
from the 2014 season.

With this restriction, we look at line fits of the form

W R
log D k log :
L RA
We choose k by using a least-squares criterion. It turns out that the sum of squared residuals is
minimized when k D 1:81. Figure 4.9 shows this best line on the scatterplot and a display of
-0.4 -0.2 0.0 0.2 0.4
log(W/L)

-0.2 -0.1 0.0 0.1 0.2

log(R/RA)
-0.2 -0.1 0.0 0.1
Residual

-0.2 -0.1 0.0 0.1 0.2

log(R/RA)

Figure 4.9. Least-squares fit (top) and residual plot (bottom) for (R/RA, W/L) data.

the corresponding residuals. We do not see any linear trend or any other pattern in the residual
plot, so it appears that our fit is satisfactory.
So based on our analysis, we arrive at the relationship

W R 1:81
D :
L RA
which is pretty close to James’ Pythagorean relationship which uses the power of 2. How useful
is this rule in predicting a team’s win numbers? To check the accuracy of this relationship in
prediction, Table 4.13 gives the actual number of wins, the predicted number of wins (using the
above model) and the residual (actual minus predicted). One can see from the table that 24 of
the sizes of the 30 residuals are smaller than 4. This indicates that for 80% of the teams, we can
predict the number of wins to within four games using this formula.

Table 4.13. Number of wins, predicted number of wins, and residuals using James’
Pythagorean relationship
Team W expected residual Team W expected residual
ARI 64 67.4 3.4 MIL 82 80.2 1.8
ATL 79 78.0 1.0 MIN 70 74.9 4.9
BAL 96 93.5 2.5 NYA 84 77.5 6.5
BOS 71 72.2 1.2 NYN 79 82.3 3.3
CHA 73 70.9 2.1 OAK 88 98.5 10.5
CHN 73 70.7 2.3 PHI 73 73.4 0.4
CIN 76 78.9 2.9 PIT 88 86.7 1.3
CLE 85 82.8 2.2 SDN 77 75.5 1.5
COL 66 75.1 9.1 SEA 87 90.8 3.8
DET 90 86.2 3.8 SFN 88 86.8 1.2
HOU 70 70.9 0.9 SLN 90 82.9 7.1
KCA 89 84.1 4.9 TBA 77 79.5 2.5
LAA 98 95.8 2.2 TEX 67 67.0 0.0
LAN 94 92.0 2.0 TOR 83 84.8 1.8
MIA 77 77.8 0.8 WAS 96 96.3 0.3

4.6 Baseball Players Regress to the Mean

Topics Covered: Least-squares regression, prediction.
Here we examine least-squares regression. Specifically, we describe the “regression effect”
that is generally unknown to most people in baseball.
If you look up the word “regress” in the dictionary, it will tell you the word means to “go
back”. We will see that there is a general tendency for a player’s baseball statistics from one
year to the next to go back, or regress to the mean.
Many of you have heard about the so-called “sophomore slumps” in sports. This happens
when someone does well in his/her rookie year and then slumps in the sophomore year. This
tendency is commonly discussed among baseball people. Here we see that there is a natural
tendency for best performing rookies to slump their sophomore years.

Table 4.14. 2013 and 2014 slugging percentages for 20 randomly selected players
Name SLG.2013 SLG.2014 Improvement
Domonic Brown 0.494 0.349 0.145
Melky Cabrera 0.360 0.458 0.098
Miguel Cabrera 0.636 0.524 0.112
Alberto Callaspo 0.369 0.290 0.079
Zack Cozart 0.381 0.300 0.081
Chris Davis 0.634 0.404 0.230
Matt Dominguez 0.403 0.330 0.073
Brian Dozier 0.414 0.416 02
Conor Gillaspie 0.390 0.416 0.026
Alex Gordon 0.422 0.432 0.010
Chase Headley 0.400 0.372 0.028
Aaron Hill 0.462 0.367 0.095
Matt Holliday 0.490 0.441 0.049
Brandon Moss 0.522 0.438 0.084
Albert Pujols 0.437 0.466 0.029
Pablo Sandoval 0.417 0.415 02
Denard Span 0.380 0.416 0.036
Ichiro Suzuki 0.342 0.340 02
Mark Trumbo 0.453 0.415 0.038
Jayson Werth 0.532 0.455 0.077

We look at the SLG data for twenty randomly selected players who had at least 300 at-bats
in the 2013 and 2014 seasons. Table 4.14 gives the 2013 and 2014 SLG values for these players.
In addition, in the column labeled “Improvement”, we show how many SLG points a player
improved from 2013 to 2014; if this improvement is negative, this means that the player had a
lower SLG in 2014.
Suppose that some player has a 500 slugging percentage in 2013. What do you predict his
SLG value to be in 2014? (Assuming that you don’t know his 2014 data yet.) If you don’t know
anything about the hitter, it would seem reasonable to predict his 2014 SLG also to be 500.
Is this the best prediction? To investigate this, we first plot the players’ 2014 SLG against the
players’ 2013 values in Figure 4.10.
We see a relatively strong positive association in this graph. Players that were good sluggers
in 2013 (like Miguel Cabrera) tended to be good in 2014. Likewise, players with small SLG
values in 2013 (Ichiro Suzuki comes to mind) tended to be poor also in 2004.
The least-squares line (plotted on the scatterplot) is

SLG:2014 D 0:358.SLG:2013/ C 0:242:

Let’s illustrate using this least-squares line to predict a player’s 2014 SLG. Let us select Chris
Davis of the Orioles.
His 2013 SLG value was 0.634. Using the line, we would predict his 2014 SLG value to be

SLG:2014 D 0:358 0:634 C 0:242 D 0:469:

0.50
0.45
SLG.2014
0.40
0.35
0.30

0.35 0.40 0.45 0.50 0.55 0.60

SLG.2013

Figure 4.10. Scatterplot of 2013 and 2014 slugging percentages for 20 players. A least-squares line is
shown on the graph.

Let’s interpret this prediction:

1. In 2013, Davis’ SLG value was 0:634. Since the mean SLG (among these 20 players) was
0:469, Chris was 165 points above average in 2013.
2. We would predict Davis’ 2014 SLG value to be 0:469. Since the mean SLG in 2014 was
0:402, our prediction is 67 points above average.

So his performance in 2013 was 165 points above average, and we predict his 2014 performance
to be 67 points above average. In other words,

We predict that his 2014 slugging percentage will be closer to the mean than his 2013
slugging percentage.

This demonstrates a general tendency for a player’s SLG to regress towards the mean.
Let’s explain this regression effect a different way. Recall that we defined a player’s im-
provement as

IMPROVEMENT D .SLG:2014/ .SLG:2013/:

For example, Melky Cabrera’s improvement would be 0:458 0:360 D 0:098—he was a better
slugger in 2014. Looking at the data, we see that

the best improvement was Melky Cabrera at C0:098,

the most negative improvement was Chris Davis at 0:230.

Suppose that we compute the improvement for all players and graph the Improvement
against the 2013 SLG. The scatterplot is shown in Figure 4.11. We see that good players in 2013

0.10
0.05
0.00
Improvement
-0.20 -0.15 -0.10 -0.05

0.35 0.40 0.45 0.50 0.55 0.60

SLG.2013

Figure 4.11. Scatterplot of 2013 slugging percentages and improvement in SLG from 2013 to 2014 for
20 players. A least-squares fit is drawn on the graph.

(with large SLGs) tend to have negative improvements, and poor players in 2013 (with small
SLGs) tend to improve in 2014. This demonstrates a negative association in this graph.
The line in the graph is a least-squares fit. The correlation r D 0:72 and the least-squares
line is

IMPROVEMENT D 0:242 0:39 .SLG:2013/:

This “regression effect” actually is very common. If you take statistics for a group of players for
two consecutive years, you will find that a player’s improvement is negatively associated with
his first year performance. Players’ statistics (from one year to the next) tend to move towards
the average value.

4.7 Exercises
4.0. Table 4.15 shows career on-base percentage (OBP) and slugging percentage (SLG) for
Rickey Henderson for the first 23 seasons of his career.
(a) Construct a scatterplot of Henderson’s OBP and SLG values where OBP is the
horizontal variable and SLG is the vertical variable.
(b) There are five unusual points on the left side of the plot that don’t follow the general
pattern. This is where Henderson had a small slugging percentage. Find the ages that
correspond to these unusual points.
(c) Circle the point that corresponds to Henderson’s best season with respect to both
OBP and SLG. What was Henderson’s age this particular year?
(d) If you ignore the five unusual points, what is the general pattern in the scatterplot?

Table 4.15. On-base and slugging percentages for Rickey Henderson for
his first 23 seasons of his career
Age OBP SLG Age OBP SLG
20 .338 .336 32 .400 .423
21 .420 .399 33 .426 .457
22 .408 .437 34 .432 .474
23 .398 .382 35 .411 .365
24 .414 .421 36 .407 .447
25 .399 .458 37 .410 .344
26 .419 .516 38 .400 .342
27 .358 .469 39 .376 .347
28 .423 .497 40 .423 .466
29 .394 .399 41 .368 .305
30 .411 .399 42 .366 .351
31 .439 .577 .345

4.1. Table 4.16 shows the batting average and number of runs per game for all of the National
League teams in the 2014 season.

Table 4.16. Batting average and runs scored per game for 2014
National League teams
Team AVG R.G Team AVG R.G
ARI 0.248 3.80 NYN 0.239 3.88
ATL 0.241 3.54 PHI 0.242 3.82
CHN 0.239 3.79 PIT 0.259 4.21
CIN 0.238 3.67 SDN 0.226 3.30
COL 0.276 4.66 SFN 0.255 4.10
LAN 0.265 4.43 SLN 0.253 3.82
MIA 0.253 3.98 WAS 0.253 4.23
MIL 0.250 4.01

(a) On the grid in Figure 4.12, construct a scatterplot of Runs per Game (vertical axis)
against Batting Average (horizontal).
(b) What does the scatterplot say about the relationship between Runs per Game and
Batting Average? If you know that a team has a high batting average, what does that
say about the number of runs it scores?
4.2. (Exercise 4.1 continued.) A least-squares fit to the (Runs per Game, Batting Average) data
for the 2014 NL teams gives the relationship.
RUNS D 2:57 C 26:2AVG:
(a) Suppose that a team has a :250 batting average. How many runs do you predict the
team will score in a game?
(b) Suppose that a team has a :260 batting average. Predict the number of runs the team
will score.

5.0
4.5
RUNS PER GAME
4.0
3.5
3.0

0.22 0.23 0.24 0.25 0.26 0.27 0.28

AVG

Figure 4.12

(c) Suppose a team hits for a 10 point higher batting average (say from :250 to :260).
How many additional runs per game could we have predicted the team to score?
(d) Arizona in 2014 had a :248 batting average and scored an average of 3:80 runs per
game. For Arizona, find the predicted runs scored and the residual. Can you explain
why the residual is negative in this case?
4.3. Table 4.17 gives a number of pitching statistics for the 2014 National League teams. The
abbreviations used in the table are
SOR D number of pitching strikeouts per game
BOR D number of pitching walks per game
ERA D earned run average

Table 4.17. Team pitching statistics for 2014 National League teams
Team SOR BOR ERA PCT HG HRG RG
ARI 7.89 2.90 4.26 0.40 9.06 0.73 4.58
ATL 8.03 2.91 3.38 0.49 8.45 0.76 3.69
CHN 8.09 3.11 3.91 0.45 8.63 0.97 4.36
CIN 7.96 3.13 3.59 0.47 7.91 0.81 3.78
COL 6.63 3.28 4.84 0.41 9.43 1.15 5.05
LAN 8.48 2.65 3.40 0.58 8.26 0.83 3.81
MIA 7.35 2.83 3.78 0.47 9.14 0.75 4.16
MIL 7.69 2.66 3.67 0.51 8.56 0.93 4.06
NYN 8.04 3.14 3.49 0.49 8.46 0.77 3.81
PHI 7.75 3.22 3.79 0.45 8.62 0.77 4.24
PIT 7.58 3.08 3.47 0.54 8.28 0.96 3.90
SDN 7.93 2.85 3.27 0.47 8.02 0.67 3.56
SFN 7.48 2.40 3.50 0.54 8.06 0.81 3.79
SLN 7.54 2.90 3.50 0.56 8.15 0.65 3.72
WAS 7.95 2.17 3.03 0.59 8.34 0.94 3.43

PCT D winning percentage

HG D number of hits allowed per game
HRG D number of home runs allowed per game
RG D number of runs allowed per game
(a) Circle the variables below that you believe have a positive association with runs
allowed per game (RG).
SOR BOR ERA PCT HG HRG
(b) Circle the variables below that you believe have a negative association with runs
allowed per game (RG).
SOR BOR ERA PCT HG HRG
(c) Do you think a team’s pitched strikeouts per game (SOR) is related to a team’s pitched
walks per game (BOR)? Explain what type of relationship you would expect to find
and why.
(d) What variable among the above do you think has the strongest relationship with a
team’s winning percentage (PCT)?
4.4. (2014 NL team pitching data) Figure 4.13 displays scatterplots of runs allowed per game
(RG) and each of the variables SOR, BOR, ERA, PCT, HG, HRG.
5.0

5.0
4.5

4.5
RG

RG
4.0
4.0
3.5

3.5

7.0 7.5 8.0 8.5 2.2 2.4 2.6 2.8 3.0 3.2
SOR BOR
5.0
5.0

4.5
4.5
RG

RG
4.0

4.0
3.5
3.5

3.0 3.5 4.0 4.5 0.40 0.45 0.50 0.55 0.60

ERA PCT
5.0

5.0
4.5

4.5
RG

RG
4.0

4.0
3.5

3.5

8.0 8.5 9.0 0.7 0.8 0.9 1.0 1.1

HG HRG

Figure 4.13. Scatterplots of runs allowed and different pitching statistics for 2014 National League teams.

(a) From looking at the scatterplots, list two variables that have a negative association
with runs allowed per game.
(b) List two variables that have a positive association with runs allowed per game.
(c) List the variables in order in terms of their association with runs allowed, from most
negatively associated to most positively associated.
4.5. (2014 NL team pitching data) Table 4.18 gives the correlations between all of the variables
in Exercise 4.6 for the 2014 NL team pitching data. To help interpret the table, note that
the correlation between the walk rate (BOR) and the strikeout rate (SOR) is 0:230. This
means that there is a positive association between the strikeout and walk rates—teams
that struck out a lot of batters tended also to walk a lot of batters, and likewise teams with
small strikeout rates tended also to have small walk rates.

Table 4.18. Correlation matrix for a number of pitching variables for 2014
National League teams
SOR BOR ERA PCT HG HRG
BOR 0:230
ERA 0:627 0:562
PCT 0:320 0:679 0:781
HG 0:539 0:315 0:805 0:626
HRG 0:413 0:097 0:453 0:084 0:404
RG 0:567 0:543 0:980 0:759 0:849 0:498

(a) From looking at the correlation matrix, which variable has the strongest relationship
with the average runs per game allowed (RG)? Why do you think the correlation
value is so close to one?
(b) Which variable has the weakest association with average runs per game? Does this
variable have a positive or negative association with RG?
(c) What is the correlation between home runs allowed per game (HRG) and strikeouts
pitched per game (SOR)? Explain in words what this correlation says about the
relationship between HRG and SOR.
(d) By the correlation matrix, the correlation between home runs allowed per game
and walks allowed per game is 0:097 Explain in words what this means about the
relationship between home runs and walks allowed.
4.6. (2014 NL team pitching data). We saw earlier that there was a negative relationship
between a team’s winning percentage (PCT) and the team’s earned run average (ERA). A
least-squares line through the data has the equation

PCT D 0:880 0:105ERA:

(a) Suppose a team allows, on average, four earned runs per game. Use the line to predict
its winning percentage.
(b) Suppose a team’s ERA jumps from 4 to 5. Using the least-squares line, do you predict
its winning percentage would increase or decrease? By how much?
(c) Arizona in 2014 had a 4:26 ERA and a :395 winning percentage. Find Arizona’s
predicted PCT and the residual. Can you explain why the residual is large in this
case?

4.7. Two hundred teams were randomly selected from baseball history. For each team, the
season and the number of triples hit per game were recorded. Figure 4.14 plots the triples
per game (vertical) against the season (horizontal).
(a) Describe the basic pattern in the graph? Has the number of triples per game increased
or decreased over time?
(b) Would it be appropriate to find a best line to these data? Why or why not?
(c) Can you offer any explanation for the pattern in this graph? In other words, why has
the frequency of triples changed over the years?
1.2
1.0
Triples Per Game
0.8
0.6
0.4
0.2

1880 1900 1920 1940 1960 1980 2000

Year

Figure 4.14. Scatterplot of triples per game against season for 200 randomly selected baseball teams.

4.8. Two hundred teams were randomly selected from baseball history. For each team, the
season and the number of stolen bases (SB) per game were recorded. In Figure 4.15, we
plot the SB per game (vertical) against the season (horizontal).
2.5
2.0
Stolen Bases Per Game
1.5
1.0
0.5

1880 1900 1920 1940 1960 1980 2000

Year

Figure 4.15. Scatterplot of stolen bases per game against season for 200 randomly selected baseball
teams.

(a) Describe the basic pattern in the graph. Is the number of SBs per game increasing or
decreasing over time?
(b) Is there a “straight-line” relationship between SB per game and year?
(c) What does this graph say about the importance of stolen bases in baseball today
compared to the past?
4.9. For all pitchers in MLB history, we collected the year of birth and the throwing hand
(left or right). For each birth year, the fraction of left-handed throwers is computed
and Figure 4.16 displays a time series plot of the fraction as a function of the birth
year.
0.34
0.32
Fraction of Left-Handers
0.30
0.28
0.26
0.24
0.22

1900 1920 1940 1960 1980 2000

Year of Birth

Figure 4.16. Time series plot of the fraction of left-handed throwers.

(a) Draw vertical lines on the scatterplot at the years 1900, 1920, 1940, 1960, 1980, and
2000. For each of the periods defined by the lines (1900–1920, 1920–1940, etc.),
find the approximate average fraction of left-handers. Put your averages in the table
below.

Average Fraction of
Era Left-Handed Pitchers
1900–1920
1920–1940
1940–1960
1960–1980
1980–2000

(b) Based on your work, do you see any general pattern in the fraction of left-handers
over time?

4.10. For each of a sample of 100 games played during the 2014 regular season, we recorded:
TIME: the time of game (in minutes)
RUNS: the number of runs scored in the game
HITS: the number of hits in the game
PITCHES: the number of pitches thrown in the game
We are interested in the relationship between the time of game and the three other
variables. Scatterplots of TIME with RUNS, HITS, and PITCHES are displayed in Fig-
ure 4.17.
300

300

300
250

250
250
Time

Time

Time
200

200

200
150

150

150
10 15 20 25 30 5 10 15 20 200 250 300 350 400 450 500
Hits Runs Pitches

Figure 4.17. Scatterplots of time of game with runs scored, hits, and pitches for 100 games played during
the 2014 season.

(a) From the scatterplots, describe each of the relationships (TIME and HITS, TIME and
RUNS, TIME and PITCHES) as either positive, negative, or small. Which variable
has the strongest relationship with time of game?
(b) The least-squares fit to the (TIME, PITCHES) data is

TIME D 14:2523 C 0:6061 PITCHES

Suppose that a game has 250 pitches. Use this least-square line to predict the length
of the game (in minutes).
(c) Predict the length of the game if there are 350 pitches.
4.11. Two individual measures of pitching performance are the ERA (earned run average) and
PCT (the percentage of pitching decisions won). A scatterplot of the ERA and PCT for
Hall of Famer Walter Johnson for each of his 21 seasons in the major leagues is displayed
in Figure 4.18.
(a) Describe the direction and strength of the association between PCT and ERA.
(b) Circle two points that correspond to two years where Johnson had the higher ERA
and also the higher PCT in both years.
(c) Based on this graph, do you think that PCT is a good measure of a pitcher’s perfor-
mance? Why or why not?
(d) A famous pitcher who played primarily for Texas teams was inducted into the Hall
of Fame but had a winning percentage that was close to 50%. Name this pitcher.
Explain why he was elected in spite of his low winning percentage.

0.8
0.7
0.6
Pct
0.5
0.4

1 2 3 4 5
ERA

Figure 4.18. Scatterplot of season ERA and winning percentage for Walter Johnson.

4.12. Over the last hundred years, there has been a substantial growth in the United States
population, and a corresponding large growth in the number of players and fans of Major
League baseball. Table 4.19 gives the U.S. population and total MLB baseball attendance
for the years 1900, 1910, . . . , 2010. The third column divides the baseball attendance by

Table 4.19. United States population, Major League baseball

attendance, and ratio for the years 1900 through 2010
Baseball Ratio =
Year Population Attendance Attend. / Pop.
1900 76:2 1:8 0:024
1910 92:2 6:2 0:067
1920 106 9:1 0:086
1930 123:2 10:1 0:082
1940 132:2 9:8 0:074
1950 151:3 17:5 0:116
1960 179:3 19:9 0:111
1970 203:3 28:7 0:141
1980 226:5 43 0:190
1990 248:7 54:8 0:220
2000 281:4 71:4 0:254
2010 309:3 73:1 0:236

the population size. We see that in 2010, baseball attendance was equal to 23:6 per cent
of the American population. Figure 4.19 displays a scatterplot of year and ratio.

0.25
0.20
0.15
Ratio
0.10
0.05

1900 1920 1940 1960 1980 2000

Year

Figure 4.19. Time series plot of the ratio of Major League baseball attendance to population.

(a) Describe, using a few sentences, how the attendance/population ratio has changed
over the years. Would it be accurate to say that the ratio has consistently increased
from 1900 to 2010?
(b) Would it be reasonable to fit a least-squares line to these data? Why or why not?
(c) What would explain the dip in the ratio at the year 1940? (Hint: What was happening
in the world scene at this time?)
(d) If one fits a least-squares line only for the 1940–2010 data, one obtains the equation

RATIO D 54869 C 02619 YEAR.

Use this equation to predict the attendance/population ratio for the year 2020.
(e) Would you expect the same straight-line relationship between RATIO and YEAR for
the next 50 years? Why or why not?
4.13. Suppose you are interested in comparing the batting abilities of Jose Altuve and Nelson
Cruz for the 2014 baseball season; the batting statistics for both players are given in the
following table.

Player AB H 2B 3B HR BB HBP SF 1B TB
Jose Altuve 660 225 47 3 7 36 5 5
Nelson Cruz 613 166 32 2 40 55 5 5

(a) Compute the number of singles (1B) and total bases (TB) for each player and place
the values in the table.
(b) Compute the batting average (AVG), slugging percentage (SLG), on-base percentage
(OBP), on-base plus slugging (OPS), and runs created (RC) measures for each player.
(c) Compare the two players with respect to ability to get on base, slugging ability, and
total offensive contribution for this season.

(d) Altuve was the singles leader and Cruz was a home run leader in 2014. Which player
would you rather have on your team? Why?
4.14. Find the batting statistics (AB, R, H, 2B, 3B, HR, BB) for the most recent year for two
good hitters, one from the National League and one from the American League. Find the
batting average (AVG), slugging percentage (SLG), on-base percentage (OBP), on-base
plus slugging (OPS), and runs created (RC) measures for each player. Which player had
a better offensive year? Why?
4.15. (Least-squares criterion) Suppose you are interested in a typical number of home runs
hit by a National League team in 2014, and that you guess that the typical number
is 150 home runs. We can evaluate the goodness of the guess 150 by means of the
Root Mean Squared Error (RMSE) criterion. Table 4.20 summarizes the calculations of
RMSE using the 2014 National League team home run data. (The “Residual” column
contains the difference between the HR value and the guess, and the “Residual2 ” column
contains the square of the residual.) Fill in the missing cells in the table and compute the
RMSE.

Table 4.20. Summary of RMSE calculations for the 2014

National League team home run data using the guess 150
Team HR Guess Residual Residual2
ARI 118 150
ATL 123 150 27 729
CHN 157 150 7 49
CIN 131 150
COL 186 150 36 1296
LAN 134 150 16 256
MIA 122 150 28 784
MIL 150 150 0 0
NYN 125 150
PHI 125 150 25 625
PIT 156 150
SDN 109 150 41 1681
SFN 132 150 18 324
SLN 105 150 45 2025
WAS 152 150 2 4
Sum 9819

4.16. (Exercise 4.15 continued.) In the previous exercise, it turns out that the best guess at
this typical number of home runs is the mean, which is found by adding up all of
the home run numbers and dividing by the number of teams (15). Here the mean is
xN D .118 C 123 C 157 C C 105 D 152/=15 D 135:
(a) Compute the RMSE of the mean by completing Table 4.21.
(b) Compare the RMSE of the mean with the RMSE of the guess 150 that you found in
Exercise 4.16. Which is a better estimate at a typical home run total?

Table 4.21. Summary of RMSE calculations for the 2014 National

League team home run data using the mean estimate 135
Team HR Guess Residual Residual2
ARI 118 135 17 289
ATL 123 135 12 144
CHN 157 135 22 484
CIN 131 135
COL 186 135 51 2601
LAN 134 135 1 1
MIA 122 135
MIL 150 135
NYN 125 135 10 100
PHI 125 135 10 100
PIT 156 135 21 441
SDN 109 135
SFN 132 135 3 9
SLN 105 135 30 900
WAS 152 135 17 289
Sum 6444

4.17. (Exercise 4.1 continued.) The “AVG” and “R.G” columns of Table 4.22 give the batting
averages and runs per game for the 2014 NL teams. The “Predicted” column gives the

Table 4.22. RMSE calculations using AVG to predict runs per game
Team R.G AVG Predicted Residual Residual2
ARI 3.80 0.248 4.24
ATL 3.54 0.241 3.99
CHN 3.79 0.239 4.18 0:39 0.1521
CIN 3.67 0.238 4.18 0:51 0.2601
COL 4.66 0.276 4.67 0:01 0.0001
LAN 4.43 0.265 3.95
MIA 3.98 0.253 4.38
MIL 4.01 0.250 4.30 0:29 0.0841
NYN 3.88 0.239 4.20 0:32 0.1024
PHI 3.82 0.242 4.01 0:19 0.0361
PIT 4.21 0.259 3.99 0.22 0.0484
SDN 3.30 0.226 3.99 0:69 0.4761
SFN 4.10 0.255 4.06 0.04 0.0016
SLN 3.82 0.253 4.24 0:42 0.1764
WAS 4.23 0.253 4.30 0:07 0.0049

predicted runs per game using the least-squares fit RUNS D 1:00 C 20:48AVG. The
“Residual” column gives the residual for each team, and the “Residual2 ” column contains
the squared value of the residual.

Table 4.23. RMSE calculations using SLG to predict runs per game
Team R.G SLG Predicted Residual Residual2
ARI 3.80 0.376 3.87
ATL 3.54 0.360 3.66
CHN 3.79 0.385 3.98
CIN 3.67 0.365 3.73
COL 4.66 0.445 4.76 0:10 0.0100
LAN 4.43 0.406 4.26 0.17 0.0289
MIA 3.98 0.378 3.89 0.09 0.0081
MIL 4.01 0.397 4.14 0:13 0.0169
NYN 3.88 0.364 3.71 0.17 0.0289
PHI 3.82 0.363 3.70 0.12 0.0144
PIT 4.21 0.404 4.23 0:02 0.0004
SDN 3.30 0.342 3.43 0:13 0.0169
SFN 4.10 0.388 4.02 0.08 0.0064
SLN 3.82 0.369 3.78 0.04 0.0016
WAS 4.23 0.393 4.09 0.14 0.0196

(a) Fill in the missing cells in the “Residual” and “Residual2 ” columns.
(b) Compute the Root Mean Square (RMSE) value. This is a measure of the goodness
of using AVG as a predictor of runs scored per game.
4.18. (Exercise 4.1 continued.) Suppose that the slugging percentage (SLG) is used, instead of
AVG, to predict runs per game. The least-squares fit is given by
RUNS D 0:99 C 12:92SLG:
Table 4.23 contains the predicted runs per game, the residuals, and the residuals squared
using SLG as a predictor.
(a) Fill in the missing cells in the “Residual” and “Residual2 ” columns.
(b) Compute the RMSE value for the SLG fit. Compare this value with the RMSE value
from the AVG fit that you computed in Exercise 4.17.
(c) Construct parallel boxplots of the residuals from the SLG fit and the residual from the
AVG fit (Exercise 4.17). Looking at the boxplot display, which is the better predictor
AVG or SLG? Why?
4.19. (Exercise 4.1 continued.) Suppose that we estimate the runs per game using the derived
statistic OPS D OBP C SLG as a predictor. The least-squares fit is given by RUNS D
2:56 C 9:38OPS:
(a) Suppose a team has an OBP D :360 and SLG D :700. What is the value of the team’s
OPS?
(b) For each team in the 2014 NL, Table 4.24 computes the OPS statistic, the predicted
runs per game (using the least-squares formula), the residual and the residual squared.
(c) Fill in the blank cells in the table.
(d) Compute the RMSE. Compare the RMSE value with the RMSE values using the
predictors AVG and SLG. (Exercises 4.3 and 4.4.) Which is the best predictor among
AVG, SLG, and OPS? Why?

Table 4.24. RMSE calculations using OPS to predict runs per game
Team R.G OPS Predicted Residual Residual2
ARI 3.80 0.678 3.80 0.00 0.0000
ATL 3.54 0.665 3.68 0:14 0.0196
CHN 3.79 0.685 3.86 0:07 0.0049
CIN 3.67 0.661 3.64
COL 4.66 0.772 4.68
LAN 4.43 0.739 4.37
MIA 3.98 0.695 3.96
MIL 4.01 0.708 4.08 0:07 0.0049
NYN 3.88 0.672 3.74 0.14 0.0196
PHI 3.82 0.665 3.68 0.14 0.0196
PIT 4.21 0.734 4.32 0:11 0.0121
SDN 3.30 0.634 3.39 0:09 0.0081
SFN 4.10 0.699 4.00 0.10 0.0100
SLN 3.82 0.689 3.90 0:08 0.0064
WAS 4.23 0.714 4.14 0.09 0.0081

4.20. Table 4.25 contains the proportion of batted balls that were “hard” and the slugging
percentage (SLG) for 14 seasons of Albert Pujols career.

Table 4.25. Proportion of hard batted balls and the slugging

percentages for 14 seasons of the career of Albert Pujols
Season Hard SLG Season Hard SLG
2002 0.273 0.561 2009 0.406 0.658
2003 0.332 0.667 2010 0.424 0.596
2004 0.352 0.657 2011 0.305 0.541
2005 0.389 0.609 2012 0.335 0.516
2006 0.342 0.671 2013 0.362 0.437
2007 0.400 0.568 2014 0.361 0.466
2008 0.429 0.653 2015 0.338 0.527

(a) Before any exploration, do you believe there is a relationship between the two vari-
ables? Explain.
(b) Construct a scatterplot of “Hard” (horizontal) against SLG. Does the pattern in this
plot confirm with your beliefs about the relationship in part (a)?
(c) Find a best (least-squares) fit to these data. If Pujols hits 40% hard batted balls next
season, predict the value of his slugging percentage.
4.21. Table 4.26 contains the number of pitches and the game duration (in minutes) for 20
randomly selected games from the 2014 season.
(a) Construct a scatterplot of Pitches (horizontal) against Duration (vertical) to confirm
that there is a strong relationship.

Table 4.26. Number of pitches and game duration in minutes

for twenty games in the 2015 season
Game Pitches Duration Game Pitches Duration
1 278 182 11 318 203
2 316 197 12 273 216
3 287 168 13 267 163
4 225 142 14 254 168
5 270 170 15 297 186
6 266 173 16 284 173
7 339 221 17 292 177
8 348 235 18 292 186
9 279 181 19 250 167
10 343 226 20 276 185

(b) A best-fitting line has the equation

Duration D 0:65 C 0:65 Pitches.

If a game has 50 more pitches, how much longer do you predict the game will be?
(c) Predict the length of a game with 250 pitches.
4.22. Table 4.27 presents a table which categorizes all balls put in play in the 2014 season by
type of hit (popup, groundball, fly ball, and liner) and the outcome.

Table 4.27. All balls put in play in the 2014 season categorized by the
type of batted ball and the outcome
out single double triple home run
Popup 8823 146 42 1 0
Groundball 46003 13877 1073 70 0
Fly Ball 23111 957 1356 240 2873
Liner 11261 13440 5641 537 1313

(a) Find the proportion of batted balls that are popups, groundballs, fly balls, and liners.
(b) For each type of hit, find the proportion of outs. Which type of batted ball is most
likely to result in an out?
(c) Which type of batted ball is most likely to be a home run? Support your answer by a
suitable calculation.

Further Reading
Devore and Peck (2011) and Moore, McCabe and Craig (2012) describe basic descriptive tools
for understanding their relationship between two measurement variables. Albert (1998), Bennett
(1998), Thorn and Palmer (1985), and Albert and Bennett (2003), Chapters 6, 7, 8, describe
a number of ways of measuring offensive performance. The mean squared error criterion for

evaluating a particular batting measure is used in Chapter 6 of Albert and Bennett (2003). The
Pythagorean Formula for relating a team’s win/losses with the number of runs scored/allowed
is described in James (1982, 2001). A nice discussion of the regression-to-the-mean effect is
given in Berry (1996).

What’s On-Deck?
In this chapter we introduce basic concepts of probability by using tabletop games. To begin, we
introduce the relative frequency notion of probability by using Chris Davis’ hitting log for the
2013 season. During a plate appearance, Davis can either hit a home run or not, and we assume
that the chance that he hits a home run is p. We learn about the value of p by observing his
hitting performance over a number of games and we can estimate his home run probability by
calculating the relative frequency of home runs. In Case Study 5.2, we explore a tabletop dice
game Big League Baseball, and find probabilities of different events, say home runs, singles,
and outs, by finding probabilities of the sum of two rolls of the dice. The game All-Star Baseball
is a more elaborate game in that the batting performance of each hitter is modeled using a
separate spinner. In Case Study 5.3, we show how career statistics for a player can be used
to compute probabilities that the player gets a single, double, etc. and these probabilities are
used to compute areas of the random spinner. We conclude our discussion by describing a more
sophisticated game, Strat-O-Matic Baseball, that uses four dice and models the abilities of each
pitcher and each hitter by a separate card. We consider a classic matchup in this game Mark
McGwire against Greg Maddux and show how one can compute the probability of McGwire
hitting a home run using the theorem of total probabilities.

5.1 What is Chris Davis’ Home Run Probability?

Topics Covered: Relative frequency interpretation of probability, law of large numbers.
In this case study, we begin our discussion on probability.

What Is a Probability?
First, we recognize that life is full of uncertain events. For example

Who will win the next World Series?

Will you retire before the age of 60?
Will a Major League player ever break Joe DiMaggio’s record of hitting in 56 consecutive
games?

111
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
112 Introduction to Probability Using Tabletop Games

A probability is a way of measuring the uncertainty that we see. We can define a probability
scale from 0 to 1 and any event can be assigned a number in this range.

We assign a probability of 0 to an event that we are certain will not occur. For example, the
probability that the Phillies will meet the Mets in the World Series is zero since the World
Series matches the winners of the National League and American League playoffs and the
Phillies and Mets both play in the National League.
On the other hand, we assign a probability of 1 to an event that we are sure will occur, such
as “a Major League Baseball game will finish in ten hours.”
What if we assign a probability of .5? Suppose we toss a coin. We give the event “heads” a
probability of .5 if we think that the events “heads” and tails, or “not heads”, have the same
chance of occurring.

One way of thinking about probabilities is the following relative frequency interpretation.
An experiment is a process where the outcome (or result) is unknown. We let an event be a
collection of outcomes, and we are interested in computing the probability of the event.
Say that we can repeat the experiment many times under similar conditions. (For example,
if the experiment is tossing a coin, then suppose that we can toss the coin repeatedly under
similar conditions.)
Then the Prob.event/ is approximately the relative frequency of the event. If N is the
number of replications, then
# of times event occurs
Prob.event/ :
N
To illustrate this interpretation, suppose we are interested in estimating the probability that
Chris Davis in 2013 would hit a home run in a single plate appearance. We assume that Davis
comes to bat multiple times during the season under the same conditions. (Note that this is
a questionable assumption, but it simplifies our discussion.) For our purposes, there are two
results of this plate appearance he either hits a home run or he doesn’t, and the chance that he
hits a home run in this particular season is measured by a probability.
We don’t know the value of Davis’ home run probability, but we can learn about it by
watching Davis perform during the 2013 baseball season. Table 5.1 gives the number of plate
appearances (PA) and the number of home runs (HR) Davis hit for each of the 160 games he
played in 2013.
In the first game, Davis came to bat five times and had a single home run. At that point, the
current estimate of Davis’ home run probability is
1
pOHR D D :200:
5
Certainly, this is not a great estimate of Davis’ home run probability since it is based on only
five plate appearances. Next, suppose it is the middle of April and we ve now watched Davis
play ten games. In games 1–10, he has 42 PA and 6 HR our new estimate of Davis’ home run
probability is
6
pOHR D D :143:
42

Table 5.1. Plate appearances and home runs for Chris Davis for each game played during
the 2013 baseball season
Game PA HR Game PA HR Game PA HR Game PA HR
1 5 1 41 5 1 81 4 2 121 5 0
2 4 1 42 4 0 82 4 1 122 5 1
3 4 1 43 5 1 83 4 0 123 5 0
4 5 1 44 4 0 84 4 1 124 4 0
5 4 0 45 4 1 85 4 0 125 4 1
6 4 0 46 4 1 86 3 0 126 5 0
7 4 0 47 5 1 87 4 1 127 4 0
8 4 1 48 4 0 88 4 0 128 4 0
9 4 1 49 5 0 89 5 0 129 3 0
10 4 0 50 5 0 90 4 0 130 4 1
11 4 0 51 4 1 91 4 0 131 5 0
12 4 0 52 4 2 92 3 1 132 4 0
13 4 0 53 3 0 93 4 1 133 4 0
14 4 0 54 4 0 94 4 1 134 4 0
15 4 0 55 4 0 95 4 1 135 5 0
16 4 0 56 4 1 96 4 0 136 5 0
17 4 1 57 4 0 97 5 0 137 4 0
18 4 0 58 5 0 98 5 0 138 3 0
19 4 0 59 5 0 99 5 0 139 4 1
20 4 0 60 4 0 100 4 0 140 5 0
21 5 0 61 3 0 101 4 0 141 4 0
22 5 1 62 5 0 102 4 0 142 4 0
23 4 0 63 3 0 103 4 0 143 4 1
24 4 0 64 4 0 104 4 0 144 4 0
25 5 1 65 4 1 105 4 0 145 4 0
26 3 0 66 7 0 106 4 1 146 5 1
27 5 0 67 4 1 107 4 0 147 4 0
28 4 0 68 4 0 108 4 1 148 4 0
29 4 0 69 4 1 109 4 1 149 4 1
30 2 0 70 4 1 110 4 0 150 6 0
31 4 0 71 4 0 111 4 0 151 4 0
32 4 0 72 5 2 112 4 0 152 8 0
33 4 0 73 4 1 113 5 1 153 4 0
34 4 1 74 4 0 114 5 0 154 4 0
35 6 0 75 4 0 115 5 1 155 5 1
36 4 0 76 4 0 116 5 0 156 5 0
37 4 1 77 4 1 117 4 1 157 4 0
38 4 0 78 4 0 118 4 1 158 3 1
39 5 0 79 4 0 119 6 0 159 4 0
40 4 0 80 3 0 120 4 0 160 1 0

This is likely a better estimate of Davis’ ability since it is based on a greater number of plate
appearances. Suppose that we compute Davis’ current 2013 home run rate after each of his 153
games. Figure 5.1 plots the home run rate (our probability estimate) against the game number.

0.20
Estimate
0.15
0.10

0 50 100 150
Game

Figure 5.1. Plot of home run rate against game number for Chris Davis during the 2013 baseball season.

Note that for early game numbers, the probability estimate shows a lot of fluctuation. But
after about game 80, the home run rate appears to settle down to about the value 0:08. This
illustrates the relative frequency of probability. As Davis gets more and more plate appearances,
the relative frequency of his home runs settles down and approaches his probability of a home
run for 2013. After 160 games, he hit 53 home runs in 673 plate appearances for an estimate
of 53=673 D 0:079. This value is a good estimate of Davis’ home run ability for the 2013
season.

5.2 Big League Baseball

Topics Covered: Sample space, experiments with equally likely outcomes, experiment of rolling
two dice, finding probabilities of events.
Assigning probabilities is generally hard to do if we are not able to repeat the experiment
many times. But probabilities can be assigned for simple experiments, such as those involving
dice and cards, and we talk about several of these experiments here. We describe a simple
tabletop game, Big League Baseball, made in the 1960’s based on dice rolls.

Suppose We Roll a Fair Die

We first think of the possible outcomes of rolling a single die. There are six possibilities:

ROLL 1 2 3 4 5 6

We call the collection of all possible outcomes the sample space. To assign probabilities,
we assume that each roll outcome has the same probability. (In other words, we assume that the

outcomes are equally likely. We want to assign a positive number to each outcome such that the
sum of all the probabilities assigned will be equal to 1, since we are certain there will be one
of the six outcomes in each roll. It should be clear that we should assign a probability of 1/6 to
each outcome:
ROLL 1 2 3 4 5 6
Probability 1=6 1=6 1=6 1=6 1=6 1=6
Next suppose we roll two dice. It will be convenient to distinguish the dice, so we will take
one die to be brown and one to be orange. (Brown and orange are the colors of the author’s
university.)

1. How many outcomes are there in this experiment? We know from above that there are
six possibilities for the result on the brown die. For each result on the brown die, say
BROWN = 2, there are six possible outcomes for the orange die. So the number of possible
outcomes for two dice is

6 6 D 36:

2. If we assume that each of the 36 possible rolls of two dice have the same probability, then
(by the same logic as before) we assign a probability of 1=36 to each outcome. So, for
example,

Prob.5 on brown and 2 on orange/ D 1=36:

Big League Baseball was a dice tabletop baseball game made by Sycamore Games of Lima,
Ohio in the 1960’s. This game is based on rolling 1 red die and 2 white dice. One first rolls the
red die to get the pitch:

If 1 or 6 is rolled, a fair ball is hit.

If 2 or 3 is rolled, a ball is pitched
If 4 or 5 is rolled, a strike is pitched.

Let’s find the probabilities of some outcomes:

1. The probability of pitching a ball, Prob(ball), is the same as the probability of rolling a 2
or 3. We find the probability of a set of outcomes by adding the probabilities of the out-
comes. So

Prob.ball/ D Prob.2 or 3/ D Prob.2/ C Prob.3/ D 1=6 C 1=6 D 2=6:

2. The probability of pitching a ball or a strike is found the same way:

Prob.ball or strike/ D Prob.2; 3; 4; 5/ D Prob.2/ C Prob.3/ C Prob.4/ C Prob.5/ D 4=6:

3. What is the probability that a strike is not thrown? We note that a strike is not thrown if a
ball is pitched or a fair ball is hit, so

Prob.no strike/ D Prob.ball or fair ball/ D Prob.1; 6; 2; 3/ D 4=6:

In Big League Baseball, if a ball is in play, then two white dice are rolled. Table 5.2 shows
the outcomes for each possibility of rolling two dice.

Table 5.2. Outcomes of the rolls of two dice

2nd Die
1 2 3 4 5 6
1 Single Out Out Out Out Error
2 Out Double Single Out Single Out
1st Die 3 Out Single Triple Out Out Out
4 Out Out Out Out Out Out
5 Out Single Out Out Out Single
6 Error Out Out Out Single Home Run

Since each cell in the table is assigned a probability of 1=36, we can compute several
probabilities of interest:
1. Prob.Home run/ D 1=36. (There is just one way to roll a home run.)
2. Prob.Single/. We see from the table that there are 7 ways of getting a single each outcome
has probability 1=36, so Prob.Single/ D 7=36.
3. Prob.Hit/ D 10=36. (From the table, we see 10 ways of getting a hit.)
4. Prob.Out/ D 24=36. (It is easiest to note that there are 12 ways of getting on base by a hit or
error, and therefore 36 12 D 24 ways of getting an out.)
Is Big League Baseball a realistic baseball game? In other words, does this game provide
a good representation of real baseball? Of course not—it would be insulting to the game to
think that we could simulate real baseball by only using three dice. This game assumes many
unrealistic things, including
that all batters have the same ability,
that all pitchers have the same ability,
that it is equally likely to add a strike or a ball to the pitch count. (This is not realistic in real
baseball, it is more common to pitch a strike than a ball.)
Although this game is a bit unrealistic, it is a nice first attempt to simulate a baseball game.
The results of a game are partly due to chance variation, and this game introduces chance
variation by the use of dice. This game provides a useful comparison to the more sophisticated
baseball games that are described in the next two case studies.

5.3 All-Star Baseball

Topics Covered: Spinner as a randomization device, probabilities represented by areas of the
spinner, multinomial experiment.
We next consider a more sophisticated baseball game, All-Star Baseball. This game was
popular in the 1960’s and 1970’s and was played by the author when he was young. This game is
based on using a spinner. Each hitter is represented by a circular spinner where different areas
of the spinner correspond to different play outcomes. An All-Star Baseball spinner for the great
baseball player Babe Ruth is displayed in Figure 5.2. The areas of the spinner correspond to
the probabilities of the different outcomes. This randomization device has more flexibility than
dice, since we can represent a greater range of probabilities in the outcomes.

3 1
9 10
2
10 5
8
4 Babe Ruth
13
7
12
6
11 9
14

1: HR 7, 13: 1B 11: 2B 5: 3B 9: BB
10: K 2, 6, 12: Ground Ball 3, 4, 8, 14: Fly Ball

Figure 5.2. All-Star Baseball spinner for Babe Ruth.

We illustrate constructing a spinner for one of my favorite players, Mike Schmidt. By the
way, it is doubtful that one spinner for an entire career is realistic, but we need to start somewhere.
We start with a very simple model one spinner for an entire career, regardless of age, ballpark,
opposing pitcher and defense. In later chapters, we will examine the benefits and drawbacks
of using more complicated models. We make this spinner in two steps. First, we calculate
approximate probabilities for the different outcomes (1B, 2B, 3B, HR, BB, Out) that can happen
when Schmidt comes to bat. Then we shade regions on the spinner where the areas of the regions
correspond to the probabilities. We start with Schmidt’s career statistics shown below.
AB R H 2B 3B HR SO BB AVG OBP SLG
8352 1506 2234 408 59 548 1883 1507 .267 .380 .527
We want to classify all plate appearances (PAs) into the outcomes 1B, 2B, 3B, HR, BB,
and Outs. (We ignore events like HBP and Sacrifices, since they are relatively insignificant
compared to the other outcomes.) We put the counts we know from the career data in Table 5.3.

Table 5.3. Counts of career offensive

statistics for Mike Schmidt
PLAY COUNT
PA
AB 8352
H 2234
1B
2B 408
3B 59
HR 548
BB 1507
OUTS

We complete Table 5.3 in Table 5.4, using the computations below.

We get PAs by adding at-bats (AB) and walks (BB).
PA D AB C BB D 8352 C 1507 D 9859:
We get singles (1B) by adding the doubles (2B), triples (3B), and home runs (HR), and
subtracting the total from hits (H).
1B D H .2B C 3B C HR/ D 2234 .408 C 59 C 548/ D 1219:
We get OUTS by subtracting hits (H) from at-bats (AB).
OUTS D AB H D 8352 2234 D 6118:
We next change these counts to approximate probabilities in Table 5.5 by dividing each
count (1B, 2B, 3B, HR, BB, OUTS) by the number of PAs. We check if we did this right by
seeing if the sum of probabilities is equal to one.
Table 5.4. Counts of career offensive statistics Table 5.5. Computation of event probabilities
for Mike Schmidt with PA, singles, and outs from Mike Schmidt’s career offensive
included statistics
PLAY COUNT PLAY COUNT PROPORTION
PA 9859 PA 9859 xxxx
AB 8352 AB 8352 xxxx
H 2234 H 2234 xxxx
1B 1219 1B 1219 0.124
2B 408 2B 408 0.041
3B 59 3B 59 0.006
HR 548 HR 548 0.056
BB 1507 BB 1507 0.153
OUTS 6118 OUTS 6118 0.621
1

To make our spinner, we start with a blank circle and shade areas of the regions of single
(1B), double (2B), etc., that correspond to these probabilities. To make the construction process
easier, one can divide the circular region into 36 equal regions. We convert the probabilities into
number of regions by multiplying the probability by 36 and rounding the result to the nearest
whole number:
Number of Regions D round .36 probability/:
To illustrate, we multiply the probability of a single by 36 to get
Number of regions D 36 0:124 D 4:46:
We round this to the nearest integer, getting four regions. If we do this calculation for all events,
we get the region numbers in Table 5.6.
This didn’t quite work, since the total number of regions is 35, not 36. So we make a small
adjustment I changed the number of regions of 1B from 4 to 5 to make the sum of regions add

Table 5.6. Computation of spinner region numbers from Mike

Schmidt’s career offensive statistics
PLAY COUNT PROPORTION REGIONS
PA 9859 xxxx xxxx
AB 8352 xxxx xxxx
H 2234 xxxx xxxx
1B 1219 0.124 4
2B 408 0.041 1
3B 59 0.006 0
HR 548 0.056 2
BB 1507 0.153 6
OUTS 6118 0.621 22
1 35

up to 36. (I adjusted the number of regions of 1B instead of HR, say, since this adjustment has
a modest change in the single probability.)

Do Single
ub
le
Hom
e R un

Out
Walk

Figure 5.3. Blank spinner divided into 36 Figure 5.4. Spinner using Mike Schmidt’s career
equal-size regions. hitting statistics.

We completed the calculations for our spinner and now can begin the construction process.
We take a blank spinner, shown in Figure 5.3, and color-code the events according to the
work in Table 5.6. So we color five regions black (corresponding to single), one region purple
(corresponding to double), two regions light purple, six regions green (walk) and 22 regions
white (out). When we are done we get the spinner in Figure 5.4. We will be using a spinner as a
basic probability model in our study of inference in Chapter 7.

5.4 Strat-O-Matic Baseball

Topics Covered: Probabilities of the sum of two dice, theorem of total probabilities, conditional
probability.
We conclude our discussion of tabletop games by briefly describing my favorite game, Strat-
O-Matic Baseball. It’s a game played with three dice (like Big League Baseball). The game is

much more realistic than the two previous games discussed, since it models the different abilities
of pitchers as well as hitters. The All-Star Baseball game described in Case Study 5.3 doesn’t
take into account the abilities of pitchers.
We introduce this game by considering a classic matchup between Mark McGwire and
Greg Maddux. (McGwire was a great power hitter during the so-called “Steroid Era” of baseball
and Maddux was a great control pitcher in recent baseball history.) Each player is represented
by a card—the Mark McGwire and Greg Maddux cards are shown in Tables 5.7 and 5.8. We
first roll a single white die if the result is 1, 2, 3, we look at McGwire’s card; otherwise we
look at Maddux’s card. Then we roll two red dice and observe the sum. The play is determined
by reading the line (corresponding to the sum) on the pitcher’s or hitter’s card. Sometimes the
outcome is not determined by the roll of the white and red dice and a twenty-sided die must be
rolled to determine the play result.

Table 5.7. Mark McGwire’s Strat-O-Matic Table 5.8. Greg Maddux’s Strat-O-Matic
card from the 1998 baseball season card from the 1998 baseball season
1 2 3 4 5 6
2-lineout 2-flyball 2-WALK 2-lineout 2-flyball 2-strikeout
3-strikeout 3-WALK 3-WALK 3-groundball 3-flyball 3-groundball
4-flyball 4-HOMERUN 4-strikeout 4-flyball 4-groundball 4-flyball
5-WALK 5-HOMERUN 5-strikeout 5-groundball 5-strikeout 5-HOMERUN
6-WALK 6-HOMERUN 6-strikeout 6-popout 6-strikeout 1
7-WALK 7-HOMERUN 7-strikeout 7-groundball 7-DOUBLE flyball
8-WALK 8-HOMERUN 8-strikeout 8-flyball 1-7 2-20
9-WALK 1-10 9-strikeout 9-strikeout SINGLE 6-SINGLE
10-flyball flyball 10-strikeout 10-groundball 8-20 7-SINGLE
11-groundball 12-20 11-WALK 11-flyball 8-flyball 1-12
12-flyball 9-WALK 12-flyball 12-flyball 9-strikeout lineout
10-DOUBLE 10-catcher 13-20
1-11 11-groundball 8-strikeout
SINGLE 12-groundball 9-strikeout
12-20 10-groundball
11-SINGLE 11-groundball
1-6 12-lineout
lineout
7-20
12-WALK

Let’s illustrate playing this game for two plate appearances.

1. For the first plate appearance, we roll a 2 on the white die and so we refer to the 2 column of
McGwire’s card. We roll the two red dice and get a 2 and 3 for a sum of 5. We look at the
number 5 in the 2 column and read the result HOMERUN! Mac has hit a home run against
Greg Maddux.
2. For the second plate appearance, we roll a 5 on the white die and we refer to the 5 column of
Maddux’s card. The roll of the red dice is 5 and 2 for a sum of 7. Looking at the 7 line, we see

that the result is DOUBLE if the roll of the twenty-sided die is between 1 and 7 and SINGLE
if the die roll is between 8 and 20. We roll the twenty-sided die and get a 10—McGwire has
hit a single against Maddux.

To really get an understanding how the game works, we need to calculate probabilities for
the sum of two dice. As in Case Study 5.2, we distinguish between the two red dice. The 36
possible outcomes are displayed in Table 5.9.

Table 5.9. Outcomes of the rolls of two dice

2nd Die
1 2 3 4 5 6
1 x x x x x x
2 x x x x x x
1st Die 3 x x x x x x
4 x x x x x x
5 x x x x x x
6 x x x x x x

Since each possible outcome has the same chance, we assign a probability of 1=36 to each
outcome. So

Prob.1st die is 4, 2nd die is 3/ D 1=36; Prob.1st die is 5, 2nd die is 6/ D 1=36:

We are interested in probabilities about the sum of the two dice.

First, we think of possible values of the sum. Table 5.10 shows the value for the sum for
each possible rolls of the two dice. Looking at the table, we see that the possible sums are

2; 3; 4; : : : ; 12:
Table 5.10. Table of the sum of the rolls of two dice for all possible
outcomes
2nd Die
1 2 3 4 5 6
1 2 3 4 5 6 7
2 3 4 5 6 7 8
1st Die 3 4 5 6 7 8 9
4 5 6 7 8 9 10
5 6 7 8 9 10 11
6 7 8 9 10 11 12

We find the probabilities of the different sums by adding the probabilities of the individual
outcomes for the two dice. For example, suppose we wish to compute the probability that the
sum is 4. We first note that we can get a sum of 4 three ways:

(1st die is 1, 2nd die is 3)

(1st die is 2, 2nd die is 2)
(1st die is 3, 2nd die is 1)

and since the probability of each outcome is 1=36,

Prob.sum is 4/ D 1=36 C 1=36 C 1=36 D 3=36:

If we continue this way, we obtain the probability table for the sum shown in Table 5.11.

Table 5.11. Probability distribution for

the sum of two fair dice
Sum Probability
2 1/36
3 2/36
4 3/36
5 4/36
6 5/36
7 6/36
8 5/36
9 4/36
10 3/36
11 2/36
12 1/36

Now we can compute some probabilities of the Strat-O-Matic game:

Question: Looking at McGwire’s card, if we roll a 1 on the white die, what is the probability
that he walks?
Answer: Looking at McGwire’s card, we see that he walks if the sum of the dice is 5, 6, 7, 8,
9 so

Prob.walk/ D Prob.sum is 5/ C Prob.sum is 6/ C C Prob.sum is 9/ D 24=36:

Question: If we roll a 3 on the white die, what is the probability that he strikes out?
Answer: Prob.strikeout/ D Prob.sum is 4 through 10/ D 30=36.

We use similar logic for finding probabilities off of Maddux’s card. For example, if we roll
a 4 on the white die (look at Maddux’s card), we see that a fly ball results if we roll a sum equal
to 4, 8, 11. So

Prob.flyball/ D Prob.sum is 4/ C Prob.sum is 8/ C Prob.sum is 11/ D 10=36:

Now, actually we are really interested in the probability that McGwire hits a home run if
Mac is facing Greg Maddux. Figure 5.5 shows, using what is commonly called a tree diagram,
all of the ways McGwire can hit a home run off of Maddux.

He can hit a home run if one rolls a 2 on the white die (look at column 2 of McGwire’s card),
and rolls a sum of 4, 5, 6, 7, or 8 on the red dice. If one rolls an 8, then one uses the 20-sided
die and a roll between 1 and 10 results in a home run.
Mac also hits a home run if one rolls 6 on the white die (look at column 6 of Maddux’s card),
rolls a sum of 5 on the red dice, and rolls a 1 on the 20-sided die. In the diagram, the branches
of the tree are labeled with the corresponding probabilities.

WHITE SUM OF RED 20 SIDED

3/36
5
4/36

5/36
McGwire 6
6/36

7
(2)
5/36

6 1 through 10
10/20
(6)

1/6

Maddux 5 1
4/36 1/20

Figure 5.5. Tree diagram to illustrate the computation that Mark McGwire hits a home run against Greg
Maddux.
By use of the multiplication rule, we multiply the probabilities across a particular branch
to find the probability of a particular outcome of the white, red, and twenty-sided dice. For
example, the probability

Prob.White die is 6 AND the sum of the red dice is 5 AND the twenty-sided die is 1/

is found by multiplying the conditional probabilities

Prob.White die is 6/ Prob.Sum of red dice is 5 IF white die is 6/

Prob.Twenty-sided die is 1 IF white die is 6 and sum of red dice is 5/
D .1=6/ .4=36/ .1=20/:

We find the probability of Mac hitting a home run off of Greg Maddux by
multiplying the probabilities along each branch for each possible way of Mac hitting a home
run,
adding the products.
We then find the probability to be

1 3 1 4 1 5 1 6
Prob.Home Run/ D C C C
6 36 6 36 6 36 6 36

1 5 10 1 4 1
C C
6 36 20 6 36 20
D 0:0958:

5.5 Exercises
5.0. Table 5.12 shows the basic batting statistics for Rickey Henderson for the 1990 season.
Table 5.12. Batting statistics for Rickey Henderson for the 1990 season
AB R H 2B 3B HR RBI BB SO
489 119 159 33 3 28 61 97 60

Construct a random spinner for Henderson using his 1990 season. Compute the number
of plate appearances in the 1990 season. Find the approximate probability that Henderson
gets (1) a single, (2) a double, (3) a triple, (4) a home run, (5) a walk, and (6) an out.
Make a spinner like the one described in Case Study 5.3 where the areas of the regions
correspond to the probabilities of the batting events that you computed.
5.1. Table 5.13 shows the number of at-bats (AB) and hits (H) for Ichiro Suzuki during each
month of the 2004 season.
Table 5.13. Number of hits and at-bats for Ichiro
Suzuki for each month of the 2004 season
Month H AB
April 26 102
May 50 125
June 29 106
July 51 118
August 56 121
September 44 118
October 6 14

(a) Compute Suzuki’s batting average at the end of April. Do you think this is a good
estimate of Suzuki’s ability to get a base hit?
(b) Compute Suzuki’s batting average at the end of May. Do you think this average is
a better estimate of Suzuki’s “true” batting average than the value you computed in
part (a)? Why?
(c) Compute Suzuki’s batting average at the end of each month and at the end of the
baseball season. Put your batting averages in the table below.

Month Total H Total AB AVG

End of April
End of May
End of June
End of July
End of August
End of October

(d) Plot the batting averages against the month.

(e) What pattern do you see in your graph? What do you think is Suzuki’s true batting
average? Why?

5.2. Angel Pagan, the centerfielder for the 2012 San Francisco Giants, is coming to bat.
(a) Without using any data, guess at the probability that Pagan will get a triple on this
at-bat.
(b) Table 5.14 shows the number of at-bats (AB) and the number of triples (3B) for
Pagan for each month of the 2012 season. For each month, compute the proportion
of triples for all games through that month and place your answers in the last column
of the table.
Table 5.14. Number of at-bats and triples for Angel Pagan for each month of the
2012 season
Month Triples AB Proportion of Triples
April 3 92
May 1 104
June 0 98
July 1 81
August 4 114
September 6 107
October 0 9

(c) Based on your calculations, what is your best estimate of the probability that Pagan
will get a triple? Explain why this is your best guess.
(d) Suppose a hitter gets no triples in 400 at-bats. Does this mean that the probability
that he hits a triple is exactly zero? Why or why not?
5.3. If you had watched Mark Reynolds bat, you would have noticed that he tends to strike out
a lot. After each game of the 2009 season, I computed the strikeout rate
SO
SO RATE D :
AB
Figure 5.6 graphs Reynolds’ strikeout rate (SO RATE) after each game of the 2009 season.
0.40
0.35
Strikeout Rate
0.30
0.25
0.20

0 50 100 150
Game

Figure 5.6. Plot of strikeout rate against game number of Mark Reynolds for the 2009 baseball season.

(a) Using the graph, estimate Reynolds’ probability of striking out in an at-bat after
30 games.
(b) Estimate the chance of striking out in an at-bat after 50 games.
(c) What is your best estimate at Reynolds’ strikeout probability based on this data? (Use
the graph.)
5.4. Table 5.15 shows the number of innings pitched (IP) and the number of strikeouts for
each game that Madison Bumgarner started during the 2014 season.

Table 5.15. Number of strikeouts and innings pitched for all games Madison
Bumgarner started in the 2014 season
Game SO IP Game SO IP Game SO IP
1 3 4.00 12 10 7.00 23 2 4.00
2 10 6.33 13 5 8.00 24 10 9.00
3 7 6.00 14 5 7.00 25 5 8.00
4 6 4.33 15 9 7.00 26 9 7.00
5 6 8.00 16 7 8.00 27 12 7.00
6 5 5.00 17 3 6.00 28 13 9.00
7 9 6.00 18 6 5.00 29 7 6.00
8 8 8.00 19 3 7.00 30 0 6.00
9 5 5.00 20 5 6.33 31 9 7.00
10 6 6.00 21 7 6.00 32 6 6.00
11 10 7.00 22 6 8.00 33 5 7.33

(a) Define the strikeout rate per inning as SO.Rate D SO=IP. Find Bumgarner’s strikeout
rate after the first game in the 2014 season.
(b) Find Bumgarner’s strikeout rate after 11 games, 22 games, and at the end of the
season. Put your answers in the table below.

SO IP Strikeout Rate
After 11 Games
After 22 Games
After 33 Games

Figure 5.7 plots Bamgarner’s strikeout rate after each game pitched in the 2012
season.
(c) Describe the pattern of the graph from left to right. What does this pattern mean in
terms of Bumgarner’s pitching performance during the 2014 season?
(d) Suppose that Bumgarner pitched one more game in 2014 and strike out 16 batters
in 8 innings. Without calculating anything, do you think Bumgarner’s strikeout rate
would go up? Now calculate what the strikeout rate would be. Is this more or less
than you expected?
5.5. In Case Study 5.2, a basic tabletop baseball game, Big League Baseball, is described.
This game was played on a computer for 1000 games. The number of home runs hit in
each game was recorded and these data are summarized in Table 5.16.

1.2
1.1
Strikeout Rate
1.0
0.9
0.8

0 5 10 15 20 25 30
Game

Figure 5.7. Plot of season strikeout rate against game number for Madison Bumgarner after each game
he pitched in the 2014 season.

Table 5.16. Frequency table of the number of home runs hit in 1000
simulated games of Big League Baseball
HR 0 1 2 3 4 5 6 7
Count 165 289 263 170 75 29 5 4

(a) Find the probability that no home runs are hit in a game.
(b) Find the probability that between two and four home runs are hit.
(c) Find the probability that at least one home run is hit.
(d) If you play a game of Big League Baseball, what is the most likely number of home
runs that will be hit?
5.6. (Exercise 5.5 continued.) For each of the 1000 games played of Big League Baseball,
the number of pitches was recorded. Table 5.17 gives a grouped frequency table of these
data.
Table 5.17. Grouped frequency table of the number of pitches thrown in 1000 simulated
games of Big League Baseball
# Pitches 1–150 151–200 201–250 251–300 301–350 351–400 401–450
Count 20 661 278 29 8 3 1

(a) What is the probability that between 151 and 200 pitches are thrown in a game?
(b) What is the probability that at most 250 pitches are thrown?
(c) What would be a typical number of pitches thrown in a game?
(d) Would you be surprised to see a game with over 300 pitches thrown? Why?
(e) Can you think of circumstances where over 300 pitches are likely to to occur?
5.7. (Exercise 5.5 continued.) For each of the 1000 games played of Big League Baseball, the
margin of victory (winning team score—losing team score) was recorded. A frequency
table of these data is given in Table 5.18.

Table 5.18. Grouped frequency table of the number of pitches thrown in 1000 simulated
games of Big League Baseball
Margin of victory 1 2 3 4 5 6 7 8 9 10 or more
Count 333 209 142 108 66 48 33 21 21 19

(a) What is the probability that a game of Big League Baseball is decided by one run?
(b) Suppose you define a blowout as a game where a team wins by six runs or more.
What is the probability a game is a blowout?
(c) What is the probability a game is not a blowout?
(d) Is it unusual that a team would win by ten or more runs? Why?
5.8. (Exercise 5.5 continued.) Suppose you are interested in exploring the relationship between
the number of runners to reach base and the runs scored in a half-inning of baseball. For
1000 games of Big League Baseball, we record for each half-inning
RUNNERS the total number of runners to reach base in the half-inning,
RUNS the number of runs scored.
Table 5.19 classifies 18,200 half-innings with respect to RUNNERS and RUNS.

Table 5.19. Two-way count table of the number of runners on base and the runs
scored in half-innings from 1000 simulated games of Big League Baseball
Runs
Runners 0 1 2 3 4 5 6 >D 7 SUM
0 5481 0 0 0 0 0 0 0 5481
1 5441 539 0 0 0 0 0 0 5980
2 2408 896 244 0 0 0 0 0 3548
3 463 778 434 108 0 0 0 0 1783
4 13 215 369 176 52 0 0 0 825
5 0 3 81 159 71 19 0 0 333
6 0 0 1 35 66 35 11 0 148
7 0 0 0 0 15 32 14 2 63
8 0 0 0 0 0 6 11 6 23
9 0 0 0 0 0 0 3 4 7
10 0 0 0 0 0 0 0 6 6
11 0 0 0 0 0 0 0 3 3
SUM 13806 2431 1129 478 204 92 39 21 18200

(a) Find the probability that there are no runners on base in a particular half-inning.
(b) Find the probability that a team doesn’t score any runs during their at-bat.
(c) Suppose that a team has one runner on base during their half-inning what is the
probability that the runner scores?
(d) Suppose that a team has three runners on base find the probability that at least one
run is scored.

5.9. (Exercise 5.8 continued.) For ten consecutive games played by the Phillies in 1999, the
number of runners and the number of runs scored are recorded for each half-inning.
Table 5.20 classifies all half-innings by RUNS and RUNNERS.

Table 5.20. Two-way count table of the number of runners on base and the
runs scored in half-innings from ten consecutive Phillies games in 1999
Runs
RUNNERS 0 1 2 3 4 5 8 SUM
0 57 0 0 0 0 0 0 57
1 46 5 0 0 0 0 0 51
2 17 12 5 0 0 0 0 34
3 3 5 9 0 0 0 0 17
4 0 1 6 1 0 0 0 8
5 0 0 1 1 1 2 0 5
6 0 0 0 1 0 1 0 2
7 0 0 0 0 1 0 0 1
10 0 0 0 0 0 0 1 1
SUM 123 23 21 3 2 3 1 176

(a) Compute the same probabilities asked in parts (a)–(d) of Exercise 5.8.
(b) Compare your answers with those of Exercise 5.8. Are the results from the game Big
League Baseball similar to those from real baseball?
5.10. In the game Big League Baseball described in Case Study 5.2, a red die is thrown to
represent the pitch. The possible rolls of the die and the pitch results are shown in the
table below.
Red die roll Pitch result
1, 6 batter hits fair ball
2, 3 ball
4, 5 strike

(a) What is the probability that a pitch results in a strike?

(b) What is the probability that the pitch results in the batter hitting a fair ball?
(c) What is the probability that the pitch results in a ball or a strike?
5.11. (Big League Baseball, continued.) Suppose two pitches are thrown to a batter. Figure 5.8
displays a tree diagram that shows the possible outcomes, where a ball is denoted by B
and a strike by S.
(a) Assign probabilities to each branch of the tree using the table of dice rolls in
Exercise 5.10.
(b) Find the probability that two balls are thrown in a row. (To find a probability of Ball
on first pitch and Ball on second pitch or Ball 1, Ball 2, you multiply the probabilities
along the branches of the tree.)
(c) Find the probability that exactly one of the two pitches is a strike. (Find the probability
of Ball 1, Strike 2 and Strike 1, Ball 2 and then add the probabilities of the two ways
of getting one strike.)
(d) Find the probability that at least one of the pitches is a strike.

1st Pitch 2nd Pitch

B
S

Figure 5.8. Outcomes of two pitches thrown to a batter.

5.12. (Big League Baseball, continued.) Use a tree diagram like the one in Figure 5.8 to answer
the questions about what happens on three or more pitches.
(a) Find the probability that a batter strikes out on three consecutive pitches.
(b) Suppose the pitch count is 2-1 (that is 2 balls and 1 strike) after three pitches. List
below all the possible pitch sequences (like BSB) which would result in a 2-1 count.
(c) Find the probability that the pitch count is 2-1 after three pitches. (Hint: Find the
probability of each pitch sequence in (b) and then add the probabilities of all the
sequences to obtain the probability you want.)
(d) Find the probability that a batter strikes out on four pitches.
(e) Find the probability that a batter strikes out on at most four pitches.
5.13. (Big League Baseball, continued.) If the roll of the red die is 1 or 6, then the batter hits
a fair ball. Two white dice are rolled and the outcome of the play depends on the roll of
the dice as shown in Table 5.21. Using the table,

Table 5.21. Play outcomes from the rolls of two dice from Big League Baseball
White Die 2
1 2 3 4 5 6
1 Single Out Out Out Out Error
2 Out Double Single Out Single Out
White Die 1 3 Out Single Triple Out Out Out
4 Out Out Out Out Out Out
5 Out Single Out Out Out Single
6 Error Out Out Out Single Home Run

(a) Find the probability that the batter gets a double.

(b) Find the probability that the batter gets a triple.
(c) Find the probability the batter gets a hit (single, double, triple, or home run).
(d) Find the probability that the batter gets on base. (Note that getting on base is different
from getting a hit.)
(e) Find the probability that the batter gets on base as a result of an error.

5.14. (Big League Baseball, continued.) In the roll of the white dice, suppose that we consider
only the outcomes that result in hits. The results of the rolls of the white dice that result in
hits are shown in Table 5.22, all other results are left blank. Suppose that each hit shown
in the table has the same probability.
(a) How many outcomes of the two dice result in a hit?
(b) Find the probability that a hit is a single.
(c) Find the probability that a hit is a double.
(d) Find the probability that a hit is not a single.

Table 5.22. Play outcomes from the rolls of two dice from Big League Baseball
that result in hits
White Die 2
1 2 3 4 5 6
1 Single
2 Double Single Single
White Die 1 3 Single Triple
4
5 Single Single
6 Single Home Run

5.15. Table 5.23 gives total at-bats, singles, doubles, etc. for all Major League games played in
the 2014 season.
(a) Compute the number of plate appearances (PA) for the 2014 season.
(b) Find the probability that a hitter gets a single in a PA.
(c) Find the probability the hitter gets an extra base hit in a PA.
(d) In a PA, what is the most likely outcome: a hit, a strikeout, a walk, or an out? Explain.

Table 5.23. Total offensive statistics for all games played in the 2014 season
AB 1B 2B 3B HR BB SO HBP
165614 28423 8137 849 4186 14020 37441 1652

5.16. (2014 hitting data, continued.)

(a) Consider only the PAs that get on base (including home runs) in the 2014 offensive
data. In the table below, put the number of hits of different types and the number of
walks and HBPs, and find the proportion of each type.
On-base Profile
Type Count Proportion
Single
Double
Triple
Home run
Walk
HBP
TOTAL

(b) Find the probability that an on-base event is for extra-bases (an extra-base is a hit for
more than one base; it includes doubles, triples, and home runs).
(c) Find the probability that a person on-base has reached there by a walk or an HBP.
(d) Compare your proportions in the table above with the on-base
profile from the tabletop game Big League Baseball. Are there any similarities? Any
differences?
5.17. (All-Star Baseball) The spinner shown in Figure 5.9 was created using Ty Cobb’s batting
statistics for the year 1911. As in Case Study 5.3, the spinner has 36 areas of the same
size. By spinning the spinner, one is simulating the result of a single plate appearance by
Ty Cobb during the 1911 season.

1B 1B OUT OUT
1B OUT
1B OUT
1B
OUT
1B
OUT
1B OUT
1B OUT
1B OUT
1B OUT
2B OUT
2B OUT
2B OUT
2B OUT
HR OUT
BB OUT
BB OUT OUT OUT

Figure 5.9. Spinner for Ty Cobb based on 1911 batting statistics.

(a) Find the probability that Cobb gets a home run on a single plate appearance.
(b) Find the probability that Cobb gets a single.
(c) Find the probability that Cobb gets on-base.
(d) Find the most likely outcome for Cobb on a single plate appearance.
(e) Find the probability that Cobb gets an extra base hit.
5.18. (Creating a random spinner for All-Star Baseball for a player from a season of batting
data.) Table 5.24 gives basic hitting statistics for Mickey Mantle for the 1956 baseball
season. For simplicity, we assume that plate appearances (PA) are recorded as either at-
bats (AB) or walks (BB). (Hit by pitches and sacrifice hits are ignored) Also we assume
at-bats that are not hits are either strikeouts (SO), or groundouts or flyouts.
(a) Compute the number of Plate Appearances (PA) by adding the AB and BB.
(b) Compute the number of singles (1B) by adding up doubles (2B), triples (3B), and
home runs (HR), and subtracting this sum from the number of hits (H).
(c) Find the number of groundouts or flyouts by adding H and SO, and subtracting this
sum from the number of at-bats (AB).

Table 5.24. Mickey Mantle’s 1956 batting statistics

Groundballs
PA AB H 1B 2B 3B HR BB SO and Flyouts
Count 533 188 22 5 52 99 112
Probability xxx xxx xxx
Regions = xxx xxx xxx
36 Prob

(d) Find the proportion of PAs that are 1B, 2B, 3B, HR, BB, SO, and Groundouts or
Flyouts put the results in the Probability row.
(e) Convert the probabilities to spinner regions by multiplying by 36 and rounding to the
nearest whole number as was done in Case Study 5.3.

(f) Construct the random spinner (from the blank one shown below) using the Region
values computed in part (e).
5.19. (Group activity) Suppose you are interested in playing a game of All-Star Baseball between
the greatest players of the National League (NL) and the greatest players in the American
League (AL). Using the career statistics for the 18 great players given in Table 5.25,
construct a set of random spinners using the same method as Exercise 5.17.
5.20. (Probabilities of sum of two dice.) Suppose you toss two dice. Assume that the dice are
distinguishable (think of one die as white and the other die as red) and there are 36 possible
outcomes shown in Table 5.26. The outcomes are written in the table as roll of white, roll
of red. So, for example, a 4,3 indicates that the roll of the white die was 4 and the roll
of the red die was 3, a 3, 5 indicates that the white die was 3 and the red die was 5, and
so on.
(a) List the outcomes in the table where the roll of the white die is equal to 2. (One of
these outcomes is 2, 3.)
(b) List the outcomes where the roll of the white die is equal to the roll on the red die.
(c) List the outcomes where the roll of the white die is greater than the roll on the
red die.
(d) Find the probabilities of the events

Table 5.25. Career batting statistics for 18 great players

Catcher G AB H 2B 3B HR SO BB
(NL) Johnny Bench 2158 7658 2048 381 24 389 1278 891
(AL) Yogi Berra 2120 7555 2150 321 49 358 414 704
FIrst Base G AB H 2B 3B HR SO BB
(AL) Lou Gehrig 2164 8001 2721 534 163 493 790 1508
(NL) Mark McGwire 1745 5830 1553 247 6 547 1461 1247
Second Base G AB H 2B 3B HR SO BB
(NL) Jackie Robinson 1382 4877 1518 273 54 137 291 740
(AL) Eddie Collins 2826 9949 3315 438 187 47 286 1499
Shortstop G AB H 2B 3B HR SO BB
(AL) Cal Ripken, Jr. 2848 10982 3046 580 44 415 1228 1095
(NL) Honus Wagner 2792 10430 3415 640 252 101 327 963
Third Base G AB H 2B 3B HR SO BB
(NL) Mike Schmidt 2404 8352 2234 408 59 548 1883 1507
(AL) Brooks Robinson 2896 10654 2848 482 68 268 990 886
Outfielders G AB H 2B 3B HR SO BB
(AL) Babe Ruth 2503 8399 2873 506 136 714 1330 2062
(AL) Ted Williams 2292 7706 2654 525 71 521 709 2019
(NL) Willie Mays 2992 10881 3283 523 140 660 1526 1464
(NL) Hank Aaron 3298 12364 3771 624 98 755 1383 1402
(AL) Joe DiMaggio 1736 6821 2214 389 131 361 369 790
(AL) Mickey Mantle 2401 8102 2415 344 72 536 1710 1733
(NL) Pete Rose 3562 14053 4256 746 135 160 1143 1566
(NL) Stan Musial 3026 10972 3630 725 177 475 696 1599

Table 5.26. Outcomes of rolling a white die and a red die

Red Die
1 2 3 4 5 6
1 1, 1 1, 2 1, 3 1, 4 1, 5 1, 6
2 2, 1 2, 2 2, 3 2, 4 2, 5 2, 6
White Die 3 3, 1 3, 2 3, 3 3, 4 3, 5 3, 6
4 4, 1 4, 2 4, 3 4, 4 4, 5 4, 6
5 5, 1 5, 2 5, 3 5, 4 5, 5 5, 6
6 6, 1 6, 2 6, 3 6, 4 6, 5 6, 6

i. roll of white die is equal to 2,

ii. roll of white die is equal to roll of red die,
iii. roll of white die is greater than roll of red die.
5.21. (Strat-O-Matic baseball) Suppose the 1998 Tony Gwynn is facing the 1998 Randy John-
son. (A great hitter against a great pitcher.) The Gwynn and Johnson Strat-O-Matic cards

are displayed in Table 5.27 and Table 5.28. Three dice, one red and two white, are tossed
to determine the result. The red die determines which card to look at: if the red die is 1, 2,
3, Gwynn’s card is used and if the die is 4, 5, 6, Johnson’s card is used (these numbers are
shown at the top of the cards). The sum of the two white dice determines which number
is checked under the red dice number. Suppose that the roll of the red die is 3 and the
sum of the white dice is 9. We check the number 9 under Gwynn’s 3 column and read that
Gwynn has walked on this particular at-bat.

Table 5.27. Tony Gwynn’s Strat-O-Matic Table 5.28. Randy Johnson’s Strat-O-Matic
card from the 1998 baseball season card from the 1998 baseball season
1 2 3 4 5 6
2-lineout 2-lineout 2-lineout 2-flyball 2-strikeout 2-groundball
3-groundball 3-groundball 3-groundball 3-groundball 3-WALK 3-flyball
4-HOMERUN 4-groundball 4-flyball 4-catcher 4-strikeout 4-groundball
5.HOMERUN 5.flyball 5.SINGLE 5.strikeout 5.strikeout 5.strikeout
1-6 6-popout 6-SINGLE 6-strikeout 6-DOUBLE 6-strikeout
DOUBLE 7-popout 7-groundball 7-WALK 1-13 7-SINGLE
7-20 8-lineout 8-groundball 8-strikeout SINGLE 8-strikeout
6-DOUBLE 9-flyball 9-WALK 9-strikeout 14-20 9-strikeout
7-DOUBLE 10-groundball 10-SINGLE 10-groundball 7-groundball 10-groundball
1-7 11-groundball 1-4 11-strikeout 8-strikeout 11-groundball
SINGLE 12-lineout lineout 12-flyball 9-HOMERUN 12-groundball
8-20 5.20 1-15
8-SINGLE 11-groundball DOUBLE
9-groundball 12-foulout 16-20
10-SINGLE 10-flyball
11-SINGLE 11-WALK
12-popout 12-SINGLE
1-12
lineout
13-20

(a) What is the chance that the Gwynn card is used (when the red die is rolled 1, 2, or
3)?
(b) Suppose that the roll of the red die is 2 (so the middle column of Gwynn’s card is
used). What rolls of the sum of the white dice will result in a home run?
(c) If the roll of the red die is 6, which rolls of the sum of the white dice will result in a
strikeout for Gwynn?
(d) If the roll of the red die is 1, which rolls of the sum of the white die will result in a
hit?
5.22. (Strat-O-Matic continued). The 1998 Tony Gwynn is facing the 1998 Randy Johnson. If
the roll of the red die is 6, Table 5.29 gives the possible rolls of the sum of white dice, the
probabilities, and the outcomes (copied from the Johnson card above).

(a) In this case (the roll of the red die is 6)

i. What is the probability that Gwynn will single?
ii. What is the probability that Gwynn will strikeout?
iii. What is the probability that Gwynn will hit a groundball?
(b) If the red die is rolled 2, what is the probability that Gwynn will popout? (Look at
the Gwynn card above.)
Table 5.29. Rolls of the sum of the white dice, probabilities, and
outcomes from the “6” column of Johnson’s card
Sum of
white dice Probability Outcome
2 1=36 Groundball
3 2=36 Flyball
4 3=36 Groundball
5 4=36 Strikeout
6 5=36 Strikeout
7 6=36 Single
8 5=36 Strikeout
9 4=36 Strikeout
10 3=36 Groundball
11 2=36 Groundball
12 1=36 Groundball

5.23. (Strat-O-Matic continued). The 1998 Tony Gwynn is facing the 1998 Randy Johnson.
Suppose that we are interested in the probability that Gwynn strikes out against Johnson.
The event that Gwynn strikes out can be divided into six different events.
The roll of the red die is 1 and Gwynn strikes out.
The roll of the red die is 2 and Gwynn strikes out.
The roll of the red die is 3 and Gwynn strikes out.
The roll of the red die is 4 and Gwynn strikes out.
The roll of the red die is 5 and Gwynn strikes out.
The roll of the red die is 6 and Gwynn strikes out.
These events are represented as the six branches of the tree diagram below. To find the
probability that Gwynn strikes out, we first find the probabilities of all of the subbranches

Red die
1
Strikes out

2
Strikes out
3
Strikes out

4 Strikes out

5 Strikes out

6 Strikes out

and then add the products of the probabilities along the subbranches to get the desired
result. We outline these calculations below.
(a) First, find the probability that the roll of the red die is 1, 2, . . . , 6, and place these
probabilities at the six branches under the label Red die in the diagram.
(b) If the roll of the red die is 1 (so that you look at Gwynn’s card in the 1 column), find
the probability that Gwynn strikes out. Place this probability above the first horizontal
line in the second set of branches.
(c) If the roll of the red die is 2, find the probability of a strikeout. Similarly, find the
probability of a strikeout if the red die falls 3, 4, 5, and 6. Place these five probabilities
of a strikeout above the corresponding horizontal lines to the right of 2, 3, 4, 5, 6 in
the diagram.
(d) The probability of a strikeout is found in two steps:
Multiply (for each possible red die roll) the probability of the red die roll and the
probability of a strikeout for that red die roll, placing them in the blank lines to the
right of the word strikes out.
Add the six products you found above.
You have now found the probability of Gwynn striking out.
5.24. (Strat-O-Matic continued.) Use the tree diagram method described in Exercise 5.23 to
find the probability that Tony Gwynn gets a walk against Randy Johnson.
5.25. (Win Probabilities) Probabilities can be used to describe the certainty that a team will
win at different times during a game. One assumes that the probability of the home team
winning is 0.5 at the beginning of a game, and this probability will increase or decrease
during the game based on the runs scored by either team. The table below gives the inning,
score, and probability the home team wins for particular instances during the Royals at
Astros playoff game on October 12, 2015. (The information was collected from the box
score of this game posted at baseball-reference.com.)

Score
Inning Royals Astros P(Astros win)
Start of game 0 0 0.50
End of bottom of 5th 2 3 0.67
End of bottom of 7th 2 6 0.97
End of bottom of 8th 7 6 0.30
End of top of 9th 9 6 0.04
End of game 9 6 0

(a) What team was likely to win at the end of the 7th inning?
(b) What team was likely to win at the end of the 8th inning? What happened during the
8th inning (runs scored) which caused a dramatic change in win probabilities?
(c) For a different baseball game of interest, create a similar table (using information
from baseball-reference.com) displaying the probability the home team wins
during different times during the game.

Further Reading
Basic concepts of probability are presented in Devore and Peck (2011), Moore, McCabe and
Craig (2012), and Scheaffer and Young (2009). Albert and Bennett (2003), Chapter 1, discuss the
probability models behind some popular tabletop baseball games, including All-Star Baseball,
Strat-O-Matic Baseball, and APBA Baseball.

What’s On-Deck?
In this chapter, we show how some basic discrete probability distributions can be used to model
events in baseball. One nice feature of baseball is its discrete structure. A player comes to bat
during an inning, and there is a result of this plate appearance which might be a hit, a walk,
an out, an error, or a hit-by-pitch. The number of hits by a player in a given number of at-bats
can be represented by a binomial distribution where the probability of a hit is the player’s true
batting average. We show in Case Study 6.1 that binomial probabilities can be very useful in
predicting the number of games with 0 hits, 1 hit, and so on for a particular hitter. The remainder
of the chapter focuses on modeling of run production for a team. To score runs, the batters need
to get on base, and Case Study 6.2 considers a probability distribution for the number of batters
that come to bat in a half-inning. If the result of each batting appearance is “out” or “get on
base” and we are interested in the number of hitters B until three outs, then we can model B
with a negative binomial distribution. If we can estimate a team’s on-base probability, then we
can find a probability distribution for the number of batters that come to bat during an inning.
The second step of the run scoring process is advancing runners that get on-base. Suppose that
we know the proportion of on-base events of a team that are walks, singles, doubles, triples,
and home runs. (We call this set of proportions the team’s on-base profile.) If we know that a
team has a particular number of on-base events, then Case Study 6.3 shows how we can use a
team’s on-base profile to compute the probability that the team scores a given number of runs.
By looking at the actual runs scored by the 2014 Boston Red Sox, we will see that this model
does a reasonable job in explaining a team’s run production.

6.1 The Binomial Distribution and Hits per Game

TOPICS COVERED: Binomial probabilities, independence, expected counts, simulation.
The binomial distribution is one of the most useful probability distributions in statistics.
It’s a most useful distribution since it is applicable in a wide variety of situations, including
baseball. Suppose we have an experiment consisting of a sequence of n identical trials, where
each trial can result in one of two possibilities, called a success and a failure. The chance of
a success, p, is assumed constant for each trial, and the results of different trials are assumed
independent. Then X , the number of successes in n trials, has a binomial probability function

139
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
140 Probability Distributions and Baseball

given by
!
n x
Pr.X D x/ D p .1 p/nx ;
x
n
where x
is the binomial coefficient
!
n nŠ
D ;
x xŠ.n x/Š

x is an integer and the values of x can range from 0 to n. (The notation nŠ, called n factorial, is
the product of integers n .n 1/ .n 2/ 1.)
Suppose a baseball hitter comes to bat n times during a game. (For this example, we will
only consider official at-bats, where the hitter gets a hit or produces an out. A walk or hit-by-
pitch or a sacrifice fly are not considered official at-bats.) For each at-bat, we will call a base
hit a success, and an out a failure. Suppose that p, the probability of a hit remains constant for
all at-bats, and the results of different at-bats are independent. Then X , the number of hits in n
at-bats during a game, will have a binomial distribution with parameters n and p.
Before we try to fit a binomial distribution to hitting data, we should ask ourselves if the
above assumptions make sense. One important assumption is that the probability of a hit for a
batter doesn’t change across a game. In our modeling, it is convenient to go one step further and
assume that the probability of a hit for a batter doesn’t change across a season. The probability
p represents the hitting ability of the particular player, and so we are saying that a batter’s ability
stays relatively constant across games during the season. The second important assumption is
that the chance of a player getting a hit doesn’t depend on his performance in previous at-bats.
This means that the player can’t be streaky in his batting ability the result of a particular at-bat
(hit or out) doesn’t depend on how he did in his recent at-bats.
One might argue that the binomial assumptions are too simplistic you might think that a
batter’s ability does change over the season, and you may have heard that particular players are
streaky who tend to go through stretches of good hitting and bad hitting. Also, a batter may
have different abilities to hit against different pitchers and in different ballparks. But, although
the assumptions seem a bit restrictive, the binomial model will be shown to give reasonable
predictions of batting performance.
Let’s illustrate the use of the binomial distribution to model the game-to-game hitting
performance of Melky Cabrera. In the 2014 season, Cabrera got 171 hits in 568 official at-bats
for a batting average of 171=568 D :301. Table 6.1 categorizes the games where he had at least
one official at-bat by the number of at-bats (AB) and the number of hits (H).
Let’s focus on the games where Cabrera had exactly four at-bats. We notice a lot of variation
in the number of game hits. Sixteen games Cabrera was hitless, 32 games Cabrera was 1 for 4,
22 games he was 2 for 4, three games he was 3 for 4, and he never went 4 for 4.
Can this variation in the number of hits be explained by the binomial distribution?
To find probabilities using the binomial formula, we need to specify n and p. Since
we’re only considering games where Cabrera has four (official) opportunities to bat, n D 4. A
reasonable guess at p is :301, the batting average of Cabrera for the entire 2014 season. Here
X denotes the number of hits for Cabrera for this four-AB game.

Table 6.1. Categorization of the 2014 game batting results of Melky Cabrera
by the number of at-bats and the number of hits
Hits
0 1 2 3 4 Total
1 1 1 0 0 0 2
2 3 1 0 0 0 4
Game 3 5 8 2 1 0 16
AB 4 16 32 22 3 0 73
5 7 13 12 6 2 40
6 0 0 3 0 0 3

Using these values of n and p, we compute the probabilities for the five possible values of
X in Table 6.2. Also, since Cabrera has 73 of these games, we can also compute the expected
number of games where he has different numbers of hits by multiplying these probabilities
by 73.

Table 6.2. Probability and expected number of games (in 73 games) for Melky Cabrera
to have different hit numbers using the binomial formula with n D 4 and p D :3
X Binomial probability Expected number Observed
4
0 :3010 .1 :301/40 D :2387 73 :2387 D 17:4 16
04
1 :3011 .1 :301/41 D :4112 73 :4112 D 30:0 32
14
2 :3012 .1 :301/42 D :2656 73 :3656 D 19:4 22
24
3 :3013 .1 :301/43 D :0762 73 :0762 D 5:6 3
34
4 4
:3014 .1 :301/44 D :0082 73 :0082 D :6 0

To see if the binomial distribution is a good fit to Cabrera’s hitting data, we compare
the expected counts using the binomial formula with the actual observed counts in the table.
Comparing the two columns, we see that:

Cabrera had more 1-hit and 2-hit games than we would expect.
Cabrera had fewer no-hit and 3-hit games than we expect.

These observations don’t necessarily mean that the binomial formula is a poor fit to Cabrera
data. For example, we expect 50 heads in 100 tosses of a fair coin, but the probability of getting
exactly 50 heads is very small. The question is if the differences between Cabrera’s observed
numbers and the expected numbers can be explained by chance, or if the differences really
reflect a misfit of the binomial model.
To answer this question, we perform a simple simulation. We assume that Cabrera’s hitting
probability is p D :301 and we simulate many sequences of 73 four-AB games using the
binomial model with n D 4 and p D :301. Each time, we simulate the season data, we keep
track of the number of no-hit games, one-hit games, and so on. In Figure 6.1, we graph the
numbers of game hits for the 100 simulated seasons using dots, and graph Cabrera’s numbers
by a solid line. Comparing Cabrera with the dotplots, we see that his numbers of 0-hit, 1-hit,

40
30
Count
20
10
0

0 1 2 3 4 5
Number of Game Hits

Figure 6.1. Graph of count of games (out of 73) with different number of game hits from the simulated
binomial distribution with n D 4 and p D :301. The observed number of 0-hit, 1-hit, 2-hit, 3-hit, and 4-hit
games for Melky Cabrera is displayed by a solid line.

2-hit, 3-hit, and 4-hit games are in the middle of the simulated distributions. For this particular
player, we conclude that the binomial formula is a good fit for that game-to-game variation in
hitting data.
What if we looked at a large number of batters and tried to fit a binomial distribution to
the daily hit numbers for each player. What would we find? In my experience, the binomial
probability distribution tends to be a pretty good fit to hitting data for most players. Although
it might be insulting to say this to a baseball hitter, his variation of hitting performance across
games resembles the same type of chance variation you get from tossing a coin many times
where the probability of heads matches the batter’s true average.

6.2 Modeling Runs Scored: Getting on Base

Topics Covered: Negative binomial distribution, expected counts, Pearson residuals.
In the next two case studies, we discuss the use of probability distributions to model the
number of runs that a team scores.
How does a team score runs? It is a two-step process. First batters need to get on-base
by means of hits, walks, errors, or hit-by-pitches. Second, once players get on-base, runs are
generally scored by hits by other batters that move the base runners to home plate. Hitters are
valuable if they are successful in getting on-base, or if they are effective in driving home runners.
One measure of the ability of a batter to get on base is the on-base percentage (OBP ). The
ability to drive runners home is typically measured by the runs batted in or RBI .
Let R denote the number of runs scored by a particular team during a half-inning of a
baseball game. We are interested in modeling the variation in the variable R by means of a
probability distribution. We will do this in two steps:

1. We first construct a model for B, the number of players that come to bat during a half-inning.

2. We then find a suitable model for the runs scored (R) given that we know how many batters
came to bat in the inning.

Here we focus on the number of hitters (B) that come to bat. The random variation in B can be
modeled by a well-known probability distribution for experiments consisting of a sequence of
yes or no outcomes (so-called Bernoulli trials).
During a half-inning, a number of players come to bat. Each player will either

OB: get on-base without creating an out

or
OUT: create an out.

Assume that the probability that any player creates an out is equal to p and the outcomes
of different players in the inning are independent. Then the results of the batters in the inning
can be regarded as independent Bernoulli trials with probability of creating an out p. Players
will continue to come to bat until the number of OUTS, or the number of successes, is equal to
3. (Here we are referring to a “success” as getting an OUT.) If the batter results are independent
Bernoulli trials, then the number of trials until the 3rd success, B, is distributed according to a
negative binomial distribution where
!
b1 3
Pr.B D b/ D p .1 p/b3 ; b D 3; 4; : : :
2

Let’s apply this formula to model the number of batters per inning for the 2014 Houston
Astros. To compute this negative binomial formula, we need only estimate p, the chance that a
batter creates an out. For the 2014 Astros, the team on-base percentage is

OBP D :309

and so the probability that a player creates an out can be estimated by

p D 1 :309 D :691:

In Table 6.3, we show a table of these negative binomial probabilities. In the 2014 baseball
season, the Astros batted in 1450 innings. We can obtain the expected number of innings with
three batters, with four batters, and so on by multiplying the negative binomial probabilities by
1450.
Table 6.3. Probability of having different numbers of batters in a half-inning from a negative
binomial distribution where b is the number of batters until the third out with p D :691
b 3 4 5 6 7 8 9
Probability .3299 .3059 .1890 .0973 .0451 .0195 .0080
Expected 478.4 443.3 274.1 141.1 65.4 28.3 11.6

Are these probabilities and expected counts a reasonable match to the actual on-base
production of the Astros during the 2014 season? To check, Table 6.4 also includes the observed
number of innings where the Astros had different numbers of batters. A standard way of

gauging the difference between the observed and expected counts is by means of the Pearson
residual

.observed count expected count/2

rD :
expected count

When the observed count is far from its corresponding expected count, the residual will be
large. A value of the residual that is 4 or larger indicates a significant discrepancy between the
observed count and the fitted count assuming the negative binomial model. Looking at Table 6.4,
we see two large residuals corresponding to B D 3. and B D 6. We would expect (from the
model) that the Astros would have three batters (that is, a one-two-three inning) for 478 innings
when in actuality they had three batters for 554 innings, 76 more. Also, the model predicts that
the Astros would have 141 six-batters innings when they actually had only 107 of these type of
innings. Generally the negative binomial distribution does not appear to be a good match to the
Astros on-base data.

Table 6.4. Probability of having different numbers of batters, the observed numbers from the
2014 Astros season, and the Pearson residuals comparing the observed and expected counts
b 3 4 5 6 7 8 9
Probability .3299 .3059 .1890 .0973 .0451 .0195 .0080
Expected 478.4 443.3 274.1 141.1 65.4 28.3 11.6
Observed 554 430 257 107 60 27 13
Residual 11.96 0.41 1.06 8.23 0.45 0.06 0.17

Can we offer any explanation for the lack-of-fit of our model? We are assuming that the
probability that a player gets on-base is constant for all players. We know that players are not
equally proficient in getting on-base and the batters at the bottom of the order are weak hitters
with relatively small values of p. So the Astros large number of three-batter innings may be a
reflection of the innings where the bottom of the batting order is hitting.

6.3 Modeling Runs Scored: Advancing the Runners to Home

TOPICS COVERED: Multinomial probability distribution, independence, expected counts,
Pearson residuals.
In the previous study, we focused on the number of hitters that come to bat during a
particular half-inning. In this case study, we consider the problem of modeling the number of
runs scored in a half-inning. The runs scored depends significantly on the number of players B
that bat during an inning. So we focus on modeling, with a simple probability distribution, the
number of runs scored conditional on the fact that exactly B batters come to bat.
If six hitters come to bat in an inning, we know that three hitters created outs (there are
three outs in an inning), and so the remaining 6 3 D 3 hitters get on-base. How many runs
can the team score when 3 runners get on-base? Well, the team could leave the bases loaded
and score no runs. Or maybe all of the runners will score, resulting in three runs. Actually, the
number of runs scored can be 0, 1, 2, or 3.

The exact number of runs scored depends heavily on the type of hit or non-hit of the batters
that reach base. There are five possibilities for this on-base event:

On-base event Abbreviation

walk or hit-by-pitch. W
single 1
double 2
triple 3
home run H

(We place a walk and a hit-by-pitch in the same classification since both events have the same
effect on runners on base.) Each team will tend to hit different proportions for these five
events. Some teams will rely on power and hit a high fraction of doubles and home runs; other
teams may rely on their ability to draw a walk and have a high proportion of walk/hit-by-pitch
events. We let f0 ; f1 ; f2 ; f3 ; f4 denote the probabilities that the on-base event of the team is
a walk/hit-by-pitch, single, double, triple, and home run, respectively. We call the proportions
.f0 ; f1 ; f2 ; f3 ; f4 / the on-base profile of the team.
Suppose that you are given a particular sequence of on-base events. For example, suppose
you are told that the first batter on-base gets a double, the next one singles, the next one draws
a walk, and the last on-base person singles. (We abbreviate this sequence as “21W1”.) Then by
making some assumptions about runner advancement, we can figure out how many runs will
score.
We will use the following runner advancement assumptions based on what is typical in a
baseball game.

1. A single will move a runner from first base to third base, and score a runner from second or
third base.
2. A double or a triple will score all runners from first, second, and third bases.
3. A home run will score all runners.
4. An out does not advance a runner and does not eliminate a runner. (That is, all outs are
treated as strikeouts.)

Using these assumptions, we can compute the runs scored for any sequence of on-base
events. To illustrate, Table 6.5 shows the base situation after every event in the sequence 21W1.
For this particular sequence, a total of two runs were scored in the inning.

Table 6.5. Illustration of runs scored in the play sequence

“21W1” using our advancement assumptions
Event Description Bases Runs Scored
“2” Double 0
“1” Single 1
“W” Walk 0
“1” Single 1

Suppose that each on-base event can be one of the five possibilities with probabilities f0 ,
f1 , f2 , f3 , f4 , and the on-base events for different hitters are independent. Then we can compute
the probability of any sequence of events by simply multiplying the corresponding probabilities.
For example, the probability of the sequence “21W1” will be
Prob.“21W1”/ D Prob.2/ Prob.1/ Prob.W / Prob.1/
D f2 f1 f0 f1 :
Let’s return to our original question suppose a team has three on-base events. Note that at
most three runs can score from three on-base sequences, since a runner can only score if he gets
on base. What is the probability that the team scores 0, 1, 2 or 3 runs? We compute this by
finding all three on-base event sequences that result in the particular number of runs scored,
finding the probability of each on-base sequence by multiplying the probabilities as we did
above,
adding the probabilities of all of the on-base sequences.
We illustrate this computation for one case. What is the probability of scoring exactly one
run if you have three on-base events?
It turns out there are 24 sequences of three on-base events that result in one run scored.
ww1 12w 23w hww
w11 13w 3w1 hw1
w2w 2w1 31w h1w
w3w 21w 311 h11
1w1 211 32w h2w
111 22w 33w h3w
So, the probability of scoring one run (given three people on-base) is
Prob D f0 f0 f1 C f1 f2 f0 C f2 f3 f0 C f4 f0 f0
C f0 f1 f1 C f1 f3 f0 C f3 f0 f1 C f4 f0 f1
C f0 f2 f0 C f2 f0 f1 C f3 f1 f0 C f4 f1 f0
C f0 f3 f0 C f2 f1 f0 C f3 f1 f1 C f4 f1 f1
C f1 f0 f1 C f2 f1 f1 C f3 f2 f0 C f4 f2 f0
C f1 f1 f1 C f2 f2 f0 C f3 f3 f0 C f4 f3 f0 :
Using this method, we can compute the probability of scoring any number of runs given that we
know how many batters get on base.
Does this probability model explain the variation in the runs scored in an inning in baseball?
To check, we look at the 2014 Boston Red Sox. To compute the probabilities of runs scored
(given a particular number of runners on base), we need only to estimate the on-base profile for
the Red Sox shown in Table 6.6.
Table 6.6. On-base profile of the 2014 Boston Red Sox
Event Singles Doubles Triples Home runs Walks
Count 930 282 20 123 603
Proportion .4750 .1440 .0102 .0628 .3080

We see that, of the on-base events of the Red Sox, about half (47.50%) were singles, 14%
were doubles, 1% were triples, 6% were home runs, and 31% were walks. We can use this
on-base profile to compute the chance that the Red Sox will score any number of runs given
a particular number of base runners. Actually, it is a bit tedious to compute these probabilities
using formulas—it is much simpler to simulate this process and use the simulated output to
compute the probabilities.
We consider four scenarios one, two, three, and four runners on-base. Table 6.7 shows, in
each case, the computation of

Table 6.7. Probability of scoring different numbers of runs from the model, the
observed count of run numbers from the 2014 Red Sox season, and the Pearson
residuals. These quantities are given for each of four runner on-base scenarios
One runner on-base
Runs Scored
0 1
Probability 0.938 0.062
Expected count 389.3 25.7
Observed 374 41
Residual 0.60 9.06
Two runners on-base
Runs Scored
0 1 2
Probability 0.660 0.277 0.063
Expected count 169.0 70.9 16.1
Observed 149 90 17
Residual 2.36 5.14 0.05
Three runners on-base
Runs Scored
0 1 2 3
Probability 0.202 0.459 0.276 0.063
Expected count 22.5 50.6 30.7 7.0
Observed 21 61 22 7
Residual 0.01 2.13 2.49 0.00
Four runners on-base
Runs Scored
1 2 3 4
Probability 0.203 0.456 0.277 0.063
Expected count 12.6 28.3 17.2 3.9
Observed 7 29 21 5
Residual 2.48 0.02 0.85 0.31

the probability of scoring different numbers of runs from the model,

the expected counts found by multiplying the probabilities by the number of times the Red
Sox had that number of batters on base,

the observed counts of run numbers from the 2014 Red Sox season,
the Pearson residual

.o e/2
rD
e
that compares the expected (e) and observed (o) numbers.

How do the expected counts from our model compare with the observed counts? Generally
the model works very well, especially when there are three or four runners on-base in the inning.
Looking at the largest Pearson residuals, it seems that the model is not working well in two
situations: scoring one run when there is one runner (r D 9:06), and scoring one run when there
are two runners in the inning (r D 5:14). The Red Sox actually scored one run 41 times where
they had only one runner in the inning and we would predict 25:7 from the model. Also, when
the Red Sox had two runners, they scored a run 90 times compared to 70:9 predicted from the
model.
Why doesn’t the model fit well in these two situations? Remember that the run probabilities
are based on our model, which assumed that runner advancement was a function only of a team’s
on-base profile of walks and hits. There is no allowance for base stealing, sacrifice hits, or errors
that are also helpful in advancing runners. Strategies like base stealing and sacrificing are often
used in baseball when the game is close and the team is trying to score a single run. These other
types of run advancement may explain why a team is more likely to score a run than what we
predict.
But the model produces estimates that are generally close to the actual numbers of runs
scored. This is a bit surprising since we are assuming that

each player on the team has the same on-base profile (this clearly is not true),
on-base outcomes by different players in an inning are independent,
as we said above, the only way to advance a runner is by means of a hit or a walk.

We could change our model to make it more realistic. (Certainly people who construct
baseball simulation games such as the ones discussed in Chapter 5 would want to design a
model to make it as realistic as possible.) But the above model is attractive in that it is fairly
simple and helps us understand the importance of on-base events towards the goal of scoring
runs.

6.4 Exercises
6.0. Here we use the Rickey Henderson spinner that we constructed for the leadoff exercise of
Chapter 5.
(a) Suppose that Henderson plays 50 games and each game he has five plate appearances.
Spin the Henderson spinner a total of 250 times, keeping track of the number of times
he gets on-base for each game. Record your counts in the following table.

Number of times on-base 0 1 2 3 4 5

Count

(b) Find the probabilities that Henderson gets on-base 0, 1, 2, 3, 4, and 5 times, if the
number of times on-base has a binomial distribution with 5 trials and probability of
success p (Estimate p by the fraction of times he gets on-base for the 1990 season.)
(c) Find the expected number of games where Henderson gets on base 0; 1; : : : ; 5 times.
Compare your expected counts with the simulated counts using the spinner.
6.1. Barry Bonds had a remarkable 2001 season when he hit 73 home runs. Suppose we focus
on the games in which Bonds had at least three plate appearances. Table 6.8 gives a
frequency distribution of the number of home runs that Bonds hit in these games.

Table 6.8. Frequency distribution of the number of home runs

hit by Barry Bonds in 2001 in all games with at least three plate
appearances
Number of home runs hit 0 1 2 3
Count 87 50 8 2

Overall in these games he hit 72 home runs an average of :4898 home runs per game.
(a) If x denotes the number of home runs hit per game, estimate the probability that x is
equal to 0, 1, 2, 3 using a Poisson density with mean equal to :4898.
A Poisson density with mean L has probabilities given by

e L Lx
Pr.x/ D ; x D 0; 1; 2; : : :
xŠ

Put your probabilities in the table below.

Number of home runs hit x 0 1 2 3

Count 87 50 8 2
Probability
Expected count

(b) Find the expected number of games that Bonds would hit 0, 1, 2, 3 home runs put
these expected counts in the same table.
(c) Use Pearson residuals to compare the observed and expected counts. Is the Poisson
distribution a good fit to these data?
6.2. (Exercise 6.1 continued.) As an alternative to the Poisson distribution, perhaps the dis-
tribution of x can be modeled using a binomial distribution with n D 4 and p D :1099.
(Here n D 4 represents a typical number of opportunities for Bonds to bat each game
and p is the probability that he will hit a home run in a single plate appearance in that
particular season. Find the binomial probabilities, expected counts, and Pearson residuals
using this binomial fit. Comment on the suitability of the fit and contrast the fit with the
Poisson fit in Exercise 6.1.
6.3. A baseball team typically experiences a number of winning and losing streaks during
a season. Should we be surprised by these streaks? The 162 games played by the 2001
World Champion Arizona Diamondbacks were divided into 54 groups of three consecutive

games. In each group of three games, the number of wins w was recorded. A frequency
distribution for w is shown in Table 6.9.

Table 6.9. Frequency distribution of games won in groups of

three consecutive games by the Arizona Diamondbacks in 2001
w 0 1 2 3
Count 5 16 23 10

(a) Find the probabilities that Arizona wins 0, 1, 2, and 3 games in a three game period
if w is given a binomial distribution with n D 3 and p D :5679. (In the 2001 season,
the Diamondbacks had a 92-70 record for a winning fraction of 92=162 D :5679.)
(b) Find the expected number of groups that Arizona wins 0, 1, 2, and 3 in a 162-game
season assuming the binomial model.
(c) Compare the observed and expected counts by the computation of Pearson residuals.
Does a binomial fit seem reasonable for these data?
6.4. Table 6.10 gives the distribution of game hits for all 4-AB games in the 2001 baseball
season for Shawn Green and Brian Jordan

Table 6.10. Distribution of game hits in all 4-AB games in the 2001
season by Shawn Green (2001 AVG D :297) and Brian Jordan (2001
AVG D :295)
Number of hits 0 1 2 3 4
Count for Shawn Green 21 21 30 2 1
Count for Brian Jordan 16 35 12 6 1

As in Case Study 6.1, investigate if a binomial distribution is a reasonable model of the

number of game hits for each player. In each case, use the season batting average (AVG)
for the binomial probability of success.
6.5. Suppose that we are interested in the random variable X that is the number of the half-
inning in which the first run is scored in a baseball game. In the game

R H E LOB
Texas 040 210 000 7 8 0 7
Anaheim 012 001 001 5 15 0 10

the first run scored was in the top of the 2nd inning, so X = 3. In the game

R H E LOB
Milwaukee 001100010 3 9 1 9
Houston 20600030x 11 11 0 5

the first run scored was in the bottom of the 1st, so X D 2. We recorded X for a sample
of 150 games in the 2001 season and the observed frequency distribution of X is shown
in Table 6.11.

Table 6.11. Frequency distribution of the number of half-innings until the

first run scored for 150 games in the 2001 season
X 1 2 3 4 5 6
Count 39 39 22 14 9 4
X 7 8 9 10 11 19
Count 9 5 4 2 2 1

One can model X by a geometric distribution where p is the probability that a team scores
in a single inning. Here X represents the number of trials (half-innings) until the first
success (half-inning with a run scored). In a sample of 2400 half-innings from the 2001
baseball season, the offensive team scored in 703 of the half-innings, so the probability
p can be estimated by 703=2400 D :297.
(a) Find the probabilities that X takes on the values 1, 2, 3, using the geometric probability
distribution with formula

Pr.X D x/ D p.1 p/x1 ; x D 1; 2; : : :

where p D :297.
(b) Find the expected number of games (out of 150) in which X D 1; 2; 3; : : : ; 19.
(c) Compare the observed and expected counts. Is the geometric distribution a good fit
to these data?
6.6. Suppose a team has the on-base profile .f0 ; f1 ; f2 ; f3 ; f4 /, where the fractions f0 , f1 ,
f2 , f3 , f4 represent the probabilities of the events walk (w), single (1), double (2), triple
(3), and home run (h), respectively.
(a) Suppose that the team has three on-base events during a particular half-inning. Find
the probability of not scoring any runs that half-inning. (Hint: Using the runner
advancement rules described in this chapter, the sequences of events that will not
produce a run are fwww; w1w; 1ww; 11w; 2ww; 3wwg.)
(b) Suppose that you have 2 on-base events during a half-inning. Find the probability
of scoring 0, 1, and 2 runs in that inning. (Hint: There are 25 possible sequences of
two on-base events—six of these sequences will not score a run and five of these
sequences will score exactly two runs.)
6.7. Table 6.12 gives the observed distribution of inning runs scored for 150 games in the
2001 season (the first through eighth innings).

Table 6.12. Frequency distribution of the number of runs scored in an inning for
150 games in the 2001 season
Runs scored 0 1 2 3 4 5 6 or more Total
Count 1687 372 179 90 40 21 11 2400
Expected count

Suppose that we apply the run scoring model described in the case studies to these data.
For the 2001 teams, a typical on-base fraction is :3315, so we can estimate the probability
of an out to be p D 1 :3315 D :6895. Also for the 2001 teams, the on-base fractions

(of a walk, single, double, triple, and home run) are given by

f0 D :2649; f1 D :4805; f2 D :1477; f3 D :0155; f4 D :0914:

By simulation, we used this model to estimate the probability that a team scores 0; 1; 2; : : :
runs in the half-inning. The probabilities from the model are displayed in Table 6.13.

Table 6.13. Probability distribution of the number of inning runs scored for
using the run-scoring model
Runs scored 0 1 2 3 4 5 6 or more
Probability .697 .149 .080 .043 .018 .009 .005

Use these probabilities to find the expected number of half-innings (out of 2400) where
0; 1; 2; : : : runs would score. Put the expected counts in the first table. Comparing the
observed and expected counts by the use of Pearson residuals, is the model effective in
predicting the number of inning runs scored in major league games?
6.8. (Exercise 6.7 continued.) One assumption in our run scoring model is that the run pro-
duction ability of the team doesn’t change across innings. Are there particular innings
during a game where a team is more or less likely to score runs? For our sample of 150
games during the 2001 season, the mean runs scored for each inning (first through eighth)
was computed—a graph of the mean runs scored is shown in Figure 6.2. This figure
illustrates that teams are indeed more or less likely to score in particular innings. Describe
the variation that you see in the plot and explain why some innings are particularly good
(or bad) for run production.
0.60
Mean Runs Scored
0.55
0.50

1 2 3 4 5 6 7 8
Inning

Figure 6.2. Plot of the mean number of runs scored in different innings for a sample of 150 games during
the 2001 season.

6.9. In the 2001 baseball season, the Colorado Rockies were very effective in scoring runs and
the Tampa Bay Devil Rays were relatively ineffective in scoring. Here are some hitting
statistics for the two teams.

Team H 2B 3B HR BB OBP
COL 1663 324 61 213 511 :354
TBD 1426 311 21 121 456 :320

(a) Compare the on-base profile of the two teams.

(b) Compute the probability that the team scores at least one run with two base runners
for the two teams.
(c) Compute the average number of runners per inning for the two teams.
(d) If you were the General Manager of Tampa Bay, what type of hitters would you try
to sign for the following season? Explain.
6.10. Of all of the pitches thrown in the 2014 season, 26.4% ended the plate appearance with a
ball hit in-play, a strikeout, a walk, or a hit-by-pitch. If X denotes the number of pitches
thrown in a plate appearance, we can represent X with a geometric distribution where
Pr.X D x/ D p.1 p/x1 ; x D 1; 2; 3; : : :
where p D :264:
(a) Use this formula to find the probability the plate appearance ends in 1, 2, 3, 4, or
more than 4 pitches. Put your probabilities in the following table.
(b) Of the 189,792 plate appearances in the 2014 season, find the expected number that
end in 1, 2, 3, 4, and more than pitches.
(c) The actual number of plate appearances that end in 1, 2, 3, 4, or more than 4 pitches
is displayed in the table. Compute Pearson residuals and conclude if the geometric
distribution is a good fit to these data.

Number of Pitches Probability Expected Observed

1 22248
2 31124
3 35032
4 36030
4 or more 65358

Further Reading
Basic discrete probability distributions, such as the binomial, negative binomial, and Poisson,
are presented in Scheaffer and Young (2009). Mosteller (1952) uses probability modeling to
understand the World Series baseball competition. D’Esopo and Lefkowitz (1977), Cover and
Keilers (1977) and Albert and Bennett (2003), Chapter 8, discuss probabilistic models for run
production.

What’s On-Deck?
This chapter describes some fundamental notions about statistical inference in the context of
baseball. One important idea in inference discussed in the introduction is the distinction between
a player’s ability and his performance. We are interested in a player’s hitting ability, which can
be measured by a probability p that represents the player’s chance of getting on base in a single
plate appearance. We don’t know a player’s ability p, but we learn about this value when we
see the player perform in a series of games. In Case Study 7.2, we first consider the situation
where you know a player’s on-base probability p, and we look for basic patterns in his hitting
performance in ten plate appearances. In Case Study 7.3, we use a simple simulation to describe
how one can learn about a player’s batting ability when one observes his performance in ten
at-bats. We initially suppose that the player’s on-base ability p is equally likely to be :2, :3,
or :4. In the simulation, we choose a player’s ability at random, and then simulate the process
of having this player have ten plate appearances. We categorize the simulated players abilities
and performances in a two-way table, and we perform inference by looking at the abilities of
the players corresponding to a given number of on-base events in ten plate appearances. Case
Study 7.4 illustrates the use of a basic formula for an interval estimate for a batter’s ability p
of a particular probability content. We use this interval formula in Case Study 7.5 to compare
the hitting abilities of two great hitters, Wade Boggs and Tony Gwynn. We will see that it is
difficult to distinguish the hitting abilities of two players based on only one season of data, but
one can make a clearer distinction by looking at the pattern of hitting of the two players over all
of the seasons of their careers.

7.1 Ability and Performance

Topics Covered: Distinction between ability and performance.
Recently a statistics class of 28 students played a spinner baseball game between Hall of
Fame players of the American League and the National League. In this game, the AL defeated
the NL 10-0. Why? I gave the students four possible explanations in a test question:
1. The AL was lucky the spins went the right way.
2. The AL players were better than the NL players.
3. The win was a result of luck AND the fact that the AL players were better.
4. There was some cheating going on.

155
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
156 Introduction to Statistical Inference

How did the students answer this question? Here’s the tally:

Answer Tally
Luck 12
Ability 1
Luck & Ability 14
Cheating 1

Practically all of the students thought that luck played a role in the AL win.
Now what if I asked the students the same question regarding the Giants 2014 World Series
win against the Royals. Would they also say that the Giants’ win was partly due to luck? I’m
guessing that most of the students would say that the Giants won because they were more skilled
than the Royals. But actually luck or chance variation does play a big role in baseball games.
Why did the students think differently about the spinner game and the real World Series?
Well, it is obvious that chance plays a big role in the spinner game since all of the outcomes
depend on the spins of the spinners. We don’t see any obvious dice or spinners in baseball
games. But, if you think about it, there are a lot of chance elements in a baseball game (how the
ball moves through the infield grass, how the bat hits the ball, how a player fields a groundball,
etc.) that have random or unpredictable elements and this randomness can affect the outcome
of the game.

One primary role of the statistician is to understand how much of the

variation in baseball data (and other data as well) can be explained by
chance and how much is due to some real cause, like the skill of a player.

Baseball Hitting—Ability and Performance

It’s helpful to look at the dictionary’s definition of these two words.

Ability is (1) the power to do or act (2) skill (3) power to do some special thing, natural gift,
talent.
Performance is (1) the act of carrying out, doing; performing (2) a thing performed; act;
deed.

When we say a player has great ability to hit, we are talking about his skill or his gift to
hit the ball. This player might have a great eye for the ball, have a nice swing, and make good
contact with the ball. He may be a muscular individual with the capability to hit the ball a
long way. In contrast, the performance of a player is actually how he hits during games. Here a
batter’s performance is simply the record of hitting that you observe in the box scores. If Albert
Pujols hit two home runs today, he had a great batting performance.
These two words are connected. If a player has great batting ability, he will generally exhibit
great batting performances. But it is important to distinguish ability and performance. Barry
Bonds may be 1 for 10 in two playoff games. Although he had a weak batting performance, it
doesn’t mean that he’s turned into a bad hitter. Likewise, a hitter who is 4-4 for one game isn’t
necessarily a much better hitter. He may actually be a mediocre hitter and happened to be lucky
or fortunate that particular game.

We say that Mickey Mantle was a great hitter. Most baseball fans would agree that Mantle
had great batting ability. Why? Because he had one great season? No. He had a number of
great seasons. In other words, he exhibited a pattern of great batting behavior. Roger Maris,
in contrast, had only a few great hitting seasons (especially 1961 when he hit 61 home runs).
Since Maris didn’t show a consistent behavior of great hitting for many seasons, there is some
doubt that he really had great batting ability. In fact, some baseball people think that he was a
bit lucky in 1961; he just happened to be the right hitter at the right time.
In statistical inference, the objective is to learn about a player’s ability based on his per-
formance in the field. We will see that it is difficult to learn a lot about a player’s ability based
on the performance of the player in a single season. In the next case study, we’ll look at the
relationship between performance and ability more carefully.

7.2 Simulating a Batter’s Performance if His Ability is Known

Topics Covered: Random spinner as a probability model, simulation of a binomial experiment
using dice.
We represent ability by means of a probability model. This is a simple randomization
device with known properties. This model is said to be realistic if it generates baseball data
similar to what is actually observed. A basic probability model is one that we have already seen,
a spinner. Figure 7.1 displays a spinner that represents the outcome of a plate appearance:

ON.BASE

DONT.GET.ON.BASE

Figure 7.1. Random spinner to represent the outcome of a plate appearance.

In this spinner there are two possible outcomes: “on-base” and “not on-base”. The ratio
of the area of the on-base region to the area of the entire circle in the spinner represents the
probability that the player will get on-base.
Another convenient probability model is based on a die. Imagine that you have a die with
ten sides, labeled 0; 1; : : : ; 9. Each side has the same chance of being rolled and each side has
probability 1=10. Then we can represent the result of a plate appearance by a roll of this die.
Suppose that we decide to let

rolls 1, 2, 3, 4 correspond to “on-base”,

rolls 0, 5, 6, 7, 8, 9 correspond to “not on-base”.

The 10-sided die represents the player’s ability to get on-base. A roll of this 10-sided die
represents the performance of the player on a single at bat.
Let’s say we specify a player’s true ability. For example, for Mike Trout, we specify a
probability of .400 of getting on base. What kind of performance can we expect, say in ten plate
appearances?
In a statistics class, 10-sided dice were handed out to the students and a number of simu-
lations were performed. Each simulation, we rolled the die ten times, representing Trout’s ten
PAs. Remember an on-base event corresponds to a 1, 2, 3, 4 on the die, and we keep track of the
number of rolls that were either 1, 2, 3, 4. Each of the 30 students performed five simulations;
the results of the 150 simulations are shown in Table 7.1. This table represents the performance
of the batter over 150 10-plate appearance periods.

Table 7.1. Frequency distribution of number of times on-base

using dice with 150 simulations
Number of Times on Base Count Probability
0 1 0.007
1 7 0.047
2 18 0.120
3 40 0.267
4 31 0.207
5 32 0.213
6 11 0.073
7 8 0.053
8 2 0.013
9 0 0
10 0 0

After we do this a sufficient number of times, we answer the following questions:

What was the most likely number of times on-base for Trout?
Here the most likely outcome in our 150 simulations was 3 times on-base; this occurred
40 times.
What is the chance of this most likely number?
The estimated probability of three times on-base is 40=150 D :267.
What is the probability that Trout will get on-base two or more times?
In our 150 simulations, we see that Trout got on-base once or never 8 times. So, by subtraction,
the frequency of 2 or more times is 150 8 D 142. The estimated probability of 2 or more
times is 142=150 D :947.
What is the probability that Trout will not get on-base during these ten PAs?
In our 150 simulations, Trout didn’t get on-base one time. So the estimated probability of not
getting on-base is 1=150 D :007.

In statistical inference, we’re actually interested in looking at ability and performance the
opposite way. We observe a hitter’s performance. What does that tell us about a player’s ability?
This is what we’ll talk about in the next section.

7.3 Learning About a Batter’s Ability

Topics Covered: Modeling ability by using a spinner with probability p, simulating hitting data
for a given ability, simulating abilities and performance, Bayes thinking, finding the most likely
ability for a given performance.
We represent a player’s ability by means of a probability model. Imagine that a player’s
talent to get on-base is represented by the spinner displayed in Figure 7.1. For a circle of area
1, the area of the on-base region is equal to the player’s ability to get on-base. We call this area
p—this is our measure of a player’s hitting ability.
Suppose we know a player’s hitting ability. Specifically, suppose we know his on-base
probability p is equal to :4. This player comes to bat ten times during a doubleheader. How
many times will he get on-base?
We did this simulation in a statistics class in the previous case study using dice. These
probabilities aren’t too precise since the simulation was performed only 150 times. On the
computer I performed this experiment 1000 times. We assume that the player has an on-base
probability of p D :4, and the results of 1000 doubleheaders were simulated where the player
had ten opportunities to hit in each doubleheader. A frequency distribution of the times on-base
is shown in Table 7.2.

Table 7.2. Frequency distribution of number of times on-base

using a computer with 1000 simulations
# of Times on Base Count Probability
0 4 0.004
1 39 0.039
2 130 0.130
3 208 0.208
4 246 0.246
5 209 0.209
6 98 0.098
7 49 0.049
8 14 0.014
9 3 0.003
10 0 0
Sum 1000 1

Note that the most likely outcome for this hitter is four times on-base. This makes sense: if
a player with probability of :4 of getting on-base has ten chances, one would expect him to get
on-base 10.:4/ D 4 times.
Above we assumed that our player had an on-base probability of :4. What if his on-base
probability was p D :3? On the computer I simulated the result of 1000 doubleheaders assuming
p D :3. The results are shown in Table 7.3 and compared with the results when p D :4.
If we compare the results when p D :4 with the results when p D :3, we see some differ-
ences. If p D :4, the most likely outcome is four times on-base; in the case when p D :3 the
most likely value is three times on-base.

Table 7.3. Frequency distribution of number of times on-base using computer

with 1000 simulations for the cases where p D :4 and where p D :3
p D 0:4 p D 0:3
# of Times on Base Count Prob Count Prob
0 4 0.004 35 0.035
1 39 0.039 131 0.131
2 130 0.130 208 0.208
3 208 0.208 261 0.261
4 246 0.246 210 0.210
5 209 0.209 101 0.101
6 98 0.098 39 0.039
7 49 0.049 13 0.013
8 14 0.014 2 0.002
9 3 0.003 0 0
10 0 0 0 0
Sum 1000 1 1000 1

If a batter has a true :3 on-base probability, is it accurate to say that he’s sure to get on-base
three times (out of ten)? No. We see that the probability of three times on-base is only :261—it’s
actually more likely (:739) that he won’t get on-base exactly three times.
In our simulations above, we assumed that we knew the player’s ability (the value of p)
and we looked at possible outcomes in a doubleheader (the batter’s performance in ten plate
appearances).
We actually want to solve the inverse problem. If a guy gets on-base four times (out of ten),
what does that say about the guy’s ability (value of p)? We solve this problem by an application
of Bayesian thinking.
We use a simple simulation to see how we can learn about a batter’s hitting ability. Suppose
there is a manager named Casey who has a dugout of players who are equally divided between
hitters of three abilities: the crummy hitters who have a true on-base probability of p of :2, the
mediocre hitters who have p D :3, and the good hitters who have p D :4. Suppose Casey picks
a player from the dugout at random and the player gets 10 chances to hit. Casey observes

the player’s ability (value of p),

the player’s performance the value of x = # of times on-base.

Three spinners were used in this simulation: one with an on-base probability of :2, another
with an on-base probability of :3, and the third with an on-base probability of :4. We first choose
a spinner at random. We rolled a single die; if the die roll was 1 or 2, we used the p D :2 spinner;
if we rolled 3 or 4, we used the p D :3 spinner; if the roll was 5 or 6, we used the p D :4 spinner.
We then spin the chosen spinner ten times.
In one particular simulation, we chose the bad spinner (p D :2). We spun it ten times with
the results

NNNNNBNNNB

(N means not on-base, B means on-base.) So we observed p D :2 and x D 2. We repeat this

simulation 1000 times we classify all of the results in Table 7.4 by ability (value of p) and
performance (value of x):

Table 7.4. Two-way table of simulation results classified by the player’s

ability (value of p) and his performance (value of x)
Performance (value of x)
0 1 2 3 4 5 6 7 8 9 10
Ability 0.2 38 105 113 61 22 5 0 0 0 0 0
(value 0.3 14 35 74 75 83 36 12 3 1 0 0
of p) 0.4 3 9 34 76 90 54 46 4 6 1 0

Before we observed any hitting data, what is the chance that the spinner chosen is a p D :2
spinner? Each spinner has the same chance of being chosen, so the probability that a p D :2
spinner is chosen is 1=3. That is,

Prob.p D :2/ D Prob.p D :3/ D Prob.p D :4/ D 1=3:

Now suppose that we observe four times on-base for our player? (That is, x D 4.) In
Table 7.5, we focus on the column where x D 4. We see that we observed four on-base a total
of 22 C 83 C 90 D 195 times. Of these 195 times, 22 corresponded to a hitting probability of
p D :2, 83 corresponded to a probability of p D :3, and 90 corresponded to a hitting probability
of p D :4. Converting these counts to probabilities

Count Probability
Ability 0.2 22 22=195 D 0:11
(value 0.3 83 83=195 D 0:43
of p) 0.4 90 90=195 D 0:46

We see that this player is likely to be either a p D :3 hitter or a p D :4 hitter. The probability
that he is a p D :2 hitter is only about 11%.
What if the hitter got on-base six times? (That is, you observed x D 6.) Now you would
focus on the column of the table corresponding to x D 6.

Count Probability
Ability 0.2 0 0=58 D 0
(value 0.3 12 12=58 D 0:21
of p) 0.4 46 46=58 D 0:79

Here it is highly likely (probability :79) that p D :4. So if you observe six times on-base, you
can conclude that the hitter is likely a true 40% hitter. Also you are pretty sure that the hitter is
not a :2 hitter, since Prob.p D :2/ D 0.

7.4 Interval Estimates for Ability

Topics Covered: Learning about a true ability (value of p) by means of an interval estimate,
subjective interpretation of probability.

A Different Way of Thinking About a Probability

Al Gore said during a particular moment during the 2001 Presidential campaign that he had a
50% chance of winning the presidential election. What does 50% mean? It’s a probability, but
not in the relative frequency sense—the election is a one-time event and it doesn’t make sense
to imagine having many elections between Bush and Gore. Al Gore’s 50% is a probability, but
it represents Gore’s degree of belief in winning the election. This interpretation of probability is
subjective; different people can assign different probabilities to the event “Gore will win”, since
different people will have different opinions about the likelihood of the event. This interpretation
of probability is relevant here. We will use probability to represent our beliefs about different
batting abilities of a player.
Suppose we are interested in learning about the batting ability of Alex Rodriguez, who is
called A-Rod. Recall that we represent the hitting ability of a player by a spinner, where there
are two events, HIT and OUT, and the area of the HIT area represents the probability of a HIT.
We denote this hitting probability by p this is a player’s true batting average.
The first step in learning about A-Rod’s hitting ability is to list some possible values of p.
Let’s assume that A-Rod’s hitting probability could be

p D :1; :2; :3; :4; :5; :6; :7; :8; :9

and each of these nine values are equally likely. So Prob.A-Rod is a .100 hitter/ D
Prob.A-Rod is a .200 hitter/ D D Prob.A-Rod is a .900 hitter/ D 1=9.
Now you are probably thinking these assumptions are silly. A-Rod can’t be a .100 or .900
hitter, and even if there were nine possible batting abilities, it doesn’t make sense to assign
each value of p the same probability. (A-Rod is more likely to be a .300 or .400 hitter.) You’re
right that these are unrealistic assumptions, but it makes the calculations to follow easy to
explain.
Next, we let A-Rod bat 20 times and we observe x D # of hits. Suppose we observe x D 6.
(A-Rod gets six hits out of 20 at-bats.)
What can we say about A-Rod’s true batting ability p? We use the simulation scheme that
we introduced in Case Study 7.3 to learn about A-Rod’s hitting ability.

1. We first choose an ability at random from the values p D :1; :2; : : : ; :9. Imagine nine spinners
corresponding to the nine possible hitting probabilities and we choose one spinner at random.
2. We spin the chosen spinner (from part 1) 20 times and we count x D # of hits.

We repeat this process (choose a spinner and spin 20 times) a total of 10,000 times (on a
computer).
The results are displayed in Table 7.5.
Remember that each time we do the simulation, we select an ability p (a spinner) and
observe a hit number x. Look at the first count in the table, 136, in the upper left corner. This
means that 136 times we chose the p D :1 spinner and observed no hits (x D 0).
Remember A-Rod got six hits (out of 20 AB). To see what we have learned about A-Rod’s
batting ability, we focus on the x D 6 column of the table, shown in Table 7.6.
We convert the counts to probabilities by dividing each count by the Total. These probabil-
ities represent the likelihoods of A-Rod having different batting abilities.

Table 7.5. Two-way table of counts from simulation where there are nine possible abilities
p D :1; : : : ; :9, and the batter comes to bat 20 times and gets x hits
Ability (p)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0 136 17 0 0 0 0 0 0 0
1 276 66 11 0 0 0 0 0 0
2 311 140 32 4 0 0 0 0 0
3 211 211 87 16 1 0 0 0 0
4 104 211 156 42 5 0 0 0 0
5 42 189 194 92 11 1 0 0 0
6 7 128 212 137 37 4 1 0 0
7 2 62 196 209 76 20 0 0 0
8 0 24 134 190 121 44 3 0 0
9 0 11 63 177 182 66 20 1 0
x 10 0 2 34 126 216 122 46 4 0
11 0 2 9 82 176 175 69 11 0
12 0 0 5 39 153 189 130 29 0
13 0 0 2 19 87 169 168 56 0
14 0 0 1 6 42 125 216 116 9
15 0 0 0 0 16 79 208 223 28
16 0 0 0 0 3 35 140 247 94
17 0 0 0 0 0 11 61 250 224
18 0 0 0 0 1 3 27 165 328
19 0 0 0 0 0 2 13 68 293
20 0 0 0 0 0 0 0 7 146

Table 7.6. Simulated values of ability p when the hitter gets x D 6 hits
Ability (p)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Total
Count 7 128 212 137 37 4 1 0 0 526
Probability .013 .243 .403 .260 .070 .008 .002 .000 .000 1

From the table we see

Prob(A-Rod has a true .200 AVG) = 128 / 526 = .243.
Prob(A-Rod’s ability is between p = .2 and .4) = Prob(p = .2, .3, .4) = .243 + .403 + .260 =
.906.

We’d like to find an interval of ability values that are very likely. We call this interval a
probability interval for the true batting average p.
To find a probability interval, we use the following table. In the left column, we put values of
p, from most likely to least likely, and the second column contains the total probability content
of these ability values.
1. We note from the table below that p D :3 is the most likely ability for A-Rod we put the
value (:3) in the first column and the associated probability (:380) in the second column.

2. The next most likely value of p is p D :4 we put this value in the first column and the
probability (:260) in the second column.
3. We continue doing this until the total probability (the sum of the probabilities) in the second
column is a large number, say between 80 and 95 percent.
Values of p Probability
.3 .403
.4 .260
.2 .243
TOTAL .906
We see that the values :3; :4; :2 or the interval Œ:2; :4 is an 90.6% probability interval for
A-Rod’s ability p. The chance that A-Rod actually has one of these abilities is 90.6%. In other
words, we are 90.6% confident that A-Rod’s true batting average is between :2 and :4.
Actually this is very little information since we know nearly every player’s batting average
falls between :2 and :4. We really haven’t learned much about ability based on only 20 AB. We
need much more data to get a better handle on a player’s ability.
There is a simpler recipe for constructing this interval estimate for a player’s true batting
average p.

1. We first compute a player’s observed AVG pO D x=AB (this is his reported batting average).
2. We compute the standard error (SE) which is a measure of the accuracy of pO to estimate a
true batting average p:

SE D square root of .p.1

O p/=AB/:
O

3. A probability interval of content PROB will have the general form

.pO z SE; pO C z SE/:

where z is a value of a standard normal density such that the upper tail area is equal to
.1 PROB/=2.
Table 7.7 gives values of z for several choices of PROB.

Table 7.7. Value z of a standard normal

density corresponding to various
values of the probability content PROB
PROB Upper tail area z
.8 .1 1.28
.9 .05 1.645
.95 .025 1.96

We illustrate using this recipe. A-Rod got six hits in 20 at-bats, so

6
pO D D :3:
20
The standard error is

SE D square root of .:3.1 :3/=20/ D :102:

From Table 7.7, we see that if we are interested in a probability content of PROB D :900, we
should use a value of z D 1:645. So a 90% interval estimate for A-Rod’s batting ability p is

Œ:3 1:645 :102 to :3 C 1:645 :102 D Œ:132; :468:

(This gives a similar answer to what we got using our simulation method.)
Generally, you learn more about a player’s true batting average by observing more AB.
Suppose A-Rod plays for LA next year and has a good season: 210 hits in 600 AB.
His observed batting average is D 210=600 D :350. We compute SE D square root of
.:35.1 :35/=600/ D :019. A 90% probability interval for A-Rod’s p would be

:350 1:645 :019 to :350 C 1:645 :019 D .:319; :381/:

So actually, A-Rod’s batting ability in 2001 could be as low as :319 or as high as :381. So we
don’t learn as much about a player’s true batting average as we might expect.

7.5 Comparing Wade Boggs and Tony Gwynn

Topics Covered: Interval estimates for a proportion, comparing proportions by use of interval
estimates, time-series plots.
In this case study we compare two of the greatest “hitters in average”, Wade Boggs and
Tony Gwynn. Boggs played many years for the Red Sox. He was very effective in getting his bat
on the ball (a so-called contact hitter) and won many batting crowns for the best batting average.
(He’s also known for his preference in diet: he had chicken before every game.) Gwynn played
for the Padres his entire career and is also considered a great contact hitter. Michael Schell, in
Baseball’s All-Time Greatest Hitters, rates Gwynn as the best hitter of all time with respect to
batting average.
In Table 7.8, I present the batting averages for Boggs and Gwynn for all years that they
played in the major leagues. I give the ages for each player each season; we’ll make comparisons
of the players for different ages.
Let’s focus on the batting averages when both players were 31. At this age, Gwynn had 168
hits in 530 at-bats for a batting average of :317; Boggs had 205 hits in 621 at-bats for an average
of :330.

Who Was a Better Hitter at Age 31?

I didn’t ask who performed better at age 31. That question is simple to answer: Boggs had the
higher batting average (.330 compared to :317). I’m actually asking about the hitting abilities
of Gwynn and Boggs. Can we say that Boggs had a better hitting ability this year?
Remember we represent a hitter’s ability by means of a spinner where the HIT area is equal
to p. So Gwynn and Boggs this year have abilities (or spinners) with hitting probabilities of pG
and pB . (We call these the true batting averages.)

Can We Say That pG Is Larger Than pB ?

We answer this question by computing probability intervals for the two true batting averages.
We use 95% intervals (recall the formula from Case Study 7.4).

Table 7.8. Career batting statistics for Tony Gwynn and Wade Boggs
Tony Gwynn Wade Boggs
Age AB H AVG AB H AVG
22 190 55 0.289
23 304 94 0.309
24 606 213 0.351 338 118 0.349
25 622 197 0.317 582 210 0.361
26 642 211 0.329 625 203 0.325
27 589 218 0.370 653 240 0.368
28 521 163 0.313 580 207 0.357
29 604 203 0.336 551 200 0.363
30 573 177 0.309 584 214 0.366
31 530 168 0.317 621 205 0.330
32 520 165 0.317 619 187 0.302
33 489 175 0.358 546 181 0.332
34 419 165 0.394 514 133 0.259
35 535 197 0.368 560 169 0.302
36 451 159 0.353 366 125 0.342
37 592 220 0.372 460 149 0.324
38 461 148 0.321 501 156 0.311
39 411 139 0.338 353 103 0.292
40 127 41 0.323 435 122 0.280
41 102 33 0.324 292 88 0.301

Boggs: pOB D :330 with a sample size nB D 621; we compute a 95% probability interval for
to be .:293; :369/.
Gwynn: pOG D :317 with a sample size nG D 530; we compute a 95% probability interval for
to be .:278; :359/.

Let’s interpret what these mean. We are 95% confident that Boggs’ true batting average (at
age 31) is between :293 and :369; that is, the probability that falls between :293 and :369 is
95%. Likewise, we are pretty sure (with confidence :95) that Gwynn’s true batting average that
year is between :278 and :359. We have plotted the two intervals in Figure 7.2.

Gwynn

Boggs

0.25 0.30 0.35 0.40

True Batting Average

Figure 7.2. Display of probability intervals for Gwynn’s and Boggs’ true batting abilities.

These two intervals overlap so it is possible that Gwynn really had a higher batting ability
that year and by chance or luck variation, Boggs just happened to perform better. So really a
13 point difference in season batting averages is not enough to say that one hitter is better than
another hitter.
Of course, we know much more about these two hitters than their performance in a single
year. In Figure 7.3, we plot the two players’ batting averages across all years. (We are plotting
AVG against AGE.)

0.40 Player
Gwynn
Boggs

0.35
AVG

0.30

25 30 35 40
Age

Figure 7.3. Graph of Tony Gwynn and Wade Boggs season batting averages against age.

This graph is hard to interpret. Why? First, you should notice a lot of up and down variation
in both players’ batting averages. This is very common. Let me explain why. Suppose that we
assume that Gwynn is a true p D :338 hitter for all of the years of his career (note that :338 is
Gwynn’s career batting average). I simulated hitting data for Gwynn using his at-bat numbers.
Figure 7.4 shows the plot of one simulation.
0.38
0.36
AVG
0.34
0.32
0.30

25 30 35 40
Age

Figure 7.4. Simulated batting average for a true p D :338 hitter using the at-bat numbers of Tony Gwynn.

This is an interesting plot. We’re assuming that Gwynn has the same batting ability across
his whole career. His hitting probability is .338 his first year, :338 his second year, etc. But his
actual batting performance (his season batting average) shows great variability. The first year
he bats :374, the second year he bats :322 this is the natural variation even when Gwynn is
assumed to have the same true batting average for all years. So you will typically see a lot of
variation in season batting averages for any player that you look at.
Does this mean that we can’t tell if Gwynn or Boggs is the better hitter? No—we can draw
a conclusion by looking at the performance of both hitters for all years.
Let’s look at Figure 7.3 again. Despite the great up and down fluctuation in the batting
averages, we see some general patterns.

Boggs’s season batting averages were high relatively early in his career (pre 30). After 30 he
appears to be a weaker hitter despite his good year at age 36.
Gwynn, in contrast, has maintained a high (around :350) batting average for most of his career.
Boggs and Gwynn are similar hitters before 32; after 32, Gwynn was consistently higher than
Boggs in batting average

Based on the above observations, I think Gwynn has been the better hitter on the basis of
batting average. You can only make this conclusion based on 20 years of data, not just a single
season.

7.6 Exercises
7.0. In the following table, some batting statistics for Rickey Henderson for the 1990 and 1991
seasons are displayed.

Season AB BB PA OBP
1990 489 97 .439
1991 470 98 .400

(a) For each season, compute the number of (approximate) plate appearances by adding
at-bats (AB) to walks (BB). Put these values in the table.
(b) Let denote Henderson’s true OBP in 1990. Construct a 95% probability interval for
using the number of plate appearances (PA) and his on-base fraction OBP.
(c) Construct a 95% probability interval for Henderson’s true OBP in 1991.
(d) Based on your work in (b) and (c), can you say that Henderson’s on-base ability was
better in 1990 than 1991? Explain how you reached this conclusion.
7.1. Table 7.9 shows the batting average of twelve ballplayers for the years 2013 and 2014.
(a) Draw a scatterplot of these data, plotting the 2013 AVG on the horizontal axis and
the 2014 AVG on the vertical axis.
(b) Comment on any relationship between the 2013 and 2014 AVGs that you see from
the scatterplot.
(c) Guess at the value of the correlation coefficient.
(d) The least-squares line to these data is

AVG.2014 D 0:190 C 0:269 AVG.2013:

Table 7.9. Batting averages for twelve players for the

2013 and 2014 seasons
Name AVG.2013 AVG.2014
Adrian Beltre 0.315 0.324
Michael Bourn 0.263 0.257
Starlin Castro 0.245 0.292
Shin-Soo Choo 0.285 0.242
Zack Cozart 0.254 0.221
Josh Donaldson 0.301 0.255
Todd Frazier 0.234 0.273
Evan Longoria 0.269 0.253
Jed Lowrie 0.290 0.249
David Ortiz 0.309 0.263
Buster Posey 0.294 0.311
Jean Segura 0.294 0.246

Suppose that a player bats :320 in 2013. Use this line to predict his AVG in 2014.
(e) Explain why there is a positive correlation between a player’s 2013 AVG and his
2014 AVG.
7.2. Consider the simple experiment of tossing a fair coin 20 times and recording the number
of heads.
(a) What is the probability of a head on a single toss? Why?
(b) Toss a fair coin 20 times and record the number of heads. Put your sequence of
Heads and Tails and the number of heads in the following table.

Toss 1 2 3 4 5 6 7 8 9 10
Result (H or T)
Toss 11 12 13 14 15 16 17 18 19 20
Result (H or T)

(c) Combine your result (Number of Heads) with the results of other students in your
class. Plot the results using a dotplot on the number line below.

+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Number of Heads

(d) Describe the basic features of the dotplot of Number of Heads that you constructed
in (c).
(e) Suppose that you get seven heads in 20 tosses. Since the fraction of heads is only
7=20 D :35, does that mean that the coin is not fair? Why or why not?
7.3. Suppose that Dustin Pedroia is really a :300 hitter. That is, his true probability of getting
a hit on a single at-bat is :3. We represent his batting ability by means of a spinner shown
in Figure 7.5 with a HIT area of :3.

HIT

.300

.700

OUT

Figure 7.5. Spinner to simulate an at-bat for a hitter with a probability of a hit equal to p D :3.

Suppose Pedroia comes to bat 12 times over the weekend. We can simulate 12 at-bats for
Pedroia by spinning the spinner 12 times and counting the number of times the spinner
lands in the HIT region. We repeated this process 20 times and below are the number of
hits that we observed for these 20 weekends.

4, 2, 5, 8, 1, 5, 5, 0, 2, 4, 5, 1, 2, 7, 3, 2, 3, 3, 2, 3

(a) Construct a frequency table of these hit numbers and put your counts in the following
table.

Number of hits 0 1 2 3 4 5 6 7 8 9 10 11 12
Count

(b) What is the most frequent number of hits that Pedroia gets?
(c) Find the probability that Pedroia gets five or more hits over the weekend.
7.4. (Exercise 7.3 continued.) Suppose that Pedroia really is a :400 hitter, which means that
the chance that he gets a hit is :4. We think of a spinner with a HIT area of :4 (instead of
:3 as in the previous exercise) and we simulate the results of 12 at-bats by spinning this
:4 spinner 12 times. We repeated this (spinning the spinner 12 times) 20 times, obtaining
the following number of hits.

5, 3, 7, 7, 1, 6, 4, 3, 2, 5, 2, 5, 6, 3, 7, 4, 6, 4, 6, 4

Construct a frequency table of these hit numbers using the table below and answer
questions (b) and (c) from Exercise 7.3.

Number of hits 0 1 2 3 4 5 6 7 8 9 10 11 12
Count

7.5. (Exercise 7.3 continued.) Is it possible to distinguish a true :300 hitter from a true :400
hitter on the basis of 12 at-bats? First, using a spinner, we let a true :300 hitter bat 12
times (a weekend of hitting), and we repeat this simulation for 1000 weekends to get 1000
hit numbers. Likewise, we let a true :400 hitter bat for 1000 weekends, obtaining 1000

simulated hit numbers. We construct count tables in Table 7.10 for the number of hits for
both types of hitter.

Table 7.10. Simulated number of hits in 12 at-bats for true .300 and .400 hitters
Number of hits
0 1 2 3 4 5 6 7 8 9 10
.300 hitter 11 65 166 229 241 161 85 34 8 0 0
.400 hitter 3 24 70 125 208 211 173 117 47 19 3

(a) Find the probability that a true :300 hitter will get exactly five hits during a weekend.
(b) Find the probability that a true :400 hitter will get exactly five hits during a weekend.
(c) Suppose that you don’t know the batter’s ability: he either could be a :300 or :400
hitter. Given that you observe this batter get exactly five hits over the weekend, use
Table 7.10 to find the probability that the hitter has a :300 true batting average and a
:400 true batting average.
(d) From your computation in (c), are you pretty sure that the hitter has a :400 batting
average? Can you really learn much about a batter’s ability on the basis of a weekend
of hitting (12 at-bats)?
7.6. Suppose that a manager has 11 different types of hitters on his team. One player hits with
a true probability of :200, another hits with a true probability of :210, the third player hits
with a true probability of :220; : : : ; and the last player hits with a true probability of :300.
Consider this hypothetical experiment the manager chooses one player at random from
the 11, and then this player comes to bat 20 times. Each time, the manager records

p D the true batting average of the player selected,

x D the number of hits of this player in 20 at-bats.

Suppose that this hypothetical experiment is repeated 10,000 times, obtaining 10,000
values of p (the true batting average) and x (the number of hits). These values are
organized by means of the two-way table shown in Table 7.11.
(a) How many times was a player with a :22 ability chosen and four hits were observed?
Find the probability that a player with a :22 ability is chosen and four hits are
observed.
(b) What is the most likely number of hits observed? What is the probability of this
number of hits?
(c) Suppose that a hitter only gets two hits (out of 20). What is the chance that he is a
:200 ability hitter?
(d) If a hitter only gets two hits, find the probability that the hitters ability (p) is at most
:25.
(e) If a hitter gets two hits, find the smallest interval of values that contains the ability
p with a probability at least :9.
7.7. (Exercise 7.6 continued.)
(a) Suppose that the player gets five hits. Use the two-way table to find the probability
that the player has a true batting average (p) greater than :25, and the probability

Table 7.11. Two-way table of simulated values of true batting average and number of
hits in simulation
No. of True Batting Average
hits 0.2 0.21 0.22 0.23 0.24 0.25 0.26 0.27 0.28 0.29 0.3 Total
0 13 8 8 6 2 3 2 4 0 0 0 46
1 49 52 37 27 33 19 12 11 7 7 5 259
2 127 105 99 82 74 62 50 29 38 32 23 721
3 169 155 146 161 134 114 89 99 67 75 73 1282
4 190 180 201 167 220 175 155 181 138 149 109 1865
5 174 167 170 183 172 199 178 169 175 175 165 1927
6 99 101 114 133 149 155 160 172 167 175 181 1606
7 53 67 70 88 89 119 103 123 137 151 158 1158
8 19 27 32 34 45 69 54 67 64 108 119 638
9 7 7 13 16 14 21 31 44 51 44 57 305
10 2 4 3 8 6 9 14 18 24 26 25 139
11 0 0 2 2 1 2 2 7 5 12 9 42
12 0 0 1 1 0 1 1 0 1 3 1 9
13 0 0 0 0 0 0 0 0 1 1 0 2
14 0 0 0 0 0 0 0 0 0 1 0 1
Total 902 873 896 908 939 948 851 924 875 959 925 10000

that the player has a true batting average :25 or smaller. Put your answers in the table
below.
Observe 5 hits in 20 at-bats
Event Probability
Player’s true average is greater than :25
Player’s true average is :25 or smaller

(b) Repeat (a) assuming that the player gets 8 hits.

Observe 8 hits in 20 at-bats

Event Probability
Player’s true average is greater than :25
Player’s true average is :25 or smaller

(c) Compare your answers to parts (a) and (b). In which case (observing five hits or
observing eight hits) did you learn more about the player’s true batting ability?
Why?
7.8. Suppose Josh Donaldson has 50 at-bats at some point in the 2014 season and has 14 hits.
(a) Find Donaldson’s current batting average (his average over the 50 at bats).
(b) Find a 95% probability interval for Donaldson’s true batting average for the 2003
season.
(c) Is it possible that Donaldson has a true batting average of :320? Is it likely? Why?

7.9. Suppose you are following Paul Goldschmidt’s batting average during the 2015 season.
(a) Suppose that Goldschmidt has 40 hits after 100 at-bats. Find a 95% probability
interval for Goldschmidt’s true 2003 batting average.
(b) Suppose that Goldschmidt has 80 hits after 200 at-bats. Find a 95% probability
interval for Goldschmidt’s batting average.
(c) Compare the two probability intervals you computed in (a) and (b) with respect to
the center and length of the interval.
7.10. In 2014, Victor Martinez had 188 hits in 561 at-bats, and Michael Brantley had 200 hits
in 611 at-bats.
(a) Find the 2014 AVGs for Martinez and Brantley.
(b) Find 95% probability intervals for Martinez and Brantley’s true batting averages.
(c) Are you confident that Martinez really was a better hitter than Brantley during the
2014 season? (Use the probability intervals you calculated in (b) to answer the
question.)

Further Reading
Albert and Bennett (2003), Chapter 3, introduce statistical inference in the context of baseball. A
spinner is used to model a player’s hitting ability, and they use a simulation, such as described in
Case Study 7.3, to learn about a player’s ability based on his batting performance. Introductory
inference is described from a Bayesian perspective in Berry (1996) and Albert and Rossman
(2001). Basic inferential methods for one proportion are contained in Devore and Peck (2011)
and Moore, McCabe and Craig (2012).

What’s On-Deck?
In this chapter, we focus on two interesting statistical inferential topics related to baseball the
interpretation of situational data and the search for true streakiness in baseball data. Today
baseball hitting and pitching data is recorded in very fine detail, and it is popular to report the
performance of hitters and pitchers in a large number of situations. For example, we record
how a player hits in home and away games, against left and right-handed pitchers, during
different months of the season, during different pitch counts, and against different teams. The
reporting of this situational data raises an interesting question: how much of the variation in this
data corresponds to real effects and how much of the variation is attributed to luck or chance
variation? In Case Study 8.1, we look at the situational hitting data that is reported for a single
player. When we look at the situational hitting data for the home vs. away situation for a number
of players (Case Study 8.2), we will see some interesting effects. Some players will hit for a
much higher average during home games and other players perform much better during away
games. But when we graph the situational effects for a group of players during two consecutive
years, we will see that there is no association. In other words, players don’t appear to possess
an ability to perform unusually well (or poorly) during home games.
In Case Studies 8.3 and 8.4, we describe some useful statistical models for situational data.
One can represent the hitting abilities of players by means of a normal or bell-shaped curve.
Many situations are in the “no effect” scenario here the player has the same probability of getting
a hit in either situation. Other situations are so-called biases the situation, such as playing at
home will add a constant number to every player’s hitting probability. The most interesting
situation can be regarded as an “ability effect” where the particular situation is to one player’s
advantage, and to another player’s disadvantage. Generally speaking, most of the variation in
the reported situational data is essentially noise or chance variation, and it is difficult to pick up
real situational effects in baseball data.
In the last two case studies, we look at the general topic of streakiness in baseball hitting
data. In Case Study 8.5, we look at a player, Michael Brantley, and discuss ways of measuring
streakiness in his day-to-day hitting data. By looking at some of these streaky statistics, there
may be some evidence that Brantley is genuinely a streaky hitter. But Case Study 8.6 shows
that these patterns of streakiness are also common in results in tossing a coin many times.
The conclusion from this brief study is that genuine streakiness in hitting data is difficult to
detect.

175
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
176 Topics in Statistical Inference

8.1 Situational Hitting Statistics for Mike Trout

Topics Covered: Introduction to situational hitting data.
In this chapter, we first focus on situational statistics. These are currently very popular
among baseball fans. The fangraphs website contains a remarkable collection of hitting and
pitching situational stats and it’s fun reading.
When you watch a baseball game, you’ll hear the announcer say something like “Francisco
Lindor has a .421 batting average when he is facing Cole Hamels at Progressive Field.” You are
supposed to be surprised by this statement. If you are watching Lindor bat against Hamels in
Progressive Field, you might expect Lindor to get a hit.
But is this the right interpretation? How can we make sense of all of these “interesting”
situational stats that we hear in the media?
Let’s focus on Mike Trout, the Angels hitter, who had a great 2014 season. Here are Trout’s
basic hitting stats that we are familiar with.

Avg PA OBP SLG

Total .287 705 .377 .561

Next, we see how Trout did against left-handed pitchers and right handers. Generally a
hitter bats better against a pitcher who throws with an arm opposite from which he takes his
batting stance. Managers believe in this effect and make substitutions based on this belief.
It is surprising that, Trout, a right-handed hitter, had a better AVG and SLG against right-
handers.

Handedness PA AVG OBP SLG

vs Left 176 .275 .386 .523
vs Right 529 .291 .374 .574

Next, we see how Trout batted at home and on the road. Generally it is believed that
ballplayers perform better at home (more comfortable surroundings, loving fans, home cooking,
etc.) Trout had a higher slugging percentage at home, although his batting average was smaller
at home.

PA AVG OBP SLG

Home 349 .282 .378 .578
Away 356 .292 .376 .545

The next situation refers to the runners on base. We see that Trout was a better hitter
(actually performed better) with men on base compared with the bases empty.

Runners on Base PA AVG OBP SLG

Bases Empty 408 .275 .360 .544
Men on Base 297 .306 .401 .587

Next we see how Trout hit each month of the baseball season. There is quite some variation
here. He was hot in June and cold in August. How do we interpret these numbers? Was Trout a
much better hitter in March/April than in May?

Month PA AVG OBP SLG

Mar/Apr 124 .321 .403 .596
May 118 .263 .356 .495
Jun 102 .361 .471 .759
Jul 126 .265 .341 .504
Aug 132 .254 .318 .492
Sept/Oct 103 .274 .398 .571
The next situations relate to the notion of leverage which is a measure of the importance of
the particular game situation (inning, outs, and runners on base) towards the goal of winning the
game. The batting outcome in a “low leverage” situation has little effect on the game outcome.
In contrast, a hit in a “high leverage situation” may have a large influence on the game outcome.
We see that Trout did especially well in high leverage situations.

PA AVG OBP SLG

Low Leverage 387 .276 .357 .501
Medium Leverage 250 .288 .380 .611
High Leverage 68 .358 .485 .755
Next, we see how Trout hit plate apperances that passed through different pitch counts.
Generally players bat better when they have a pitch advantage (like 1-0, 2-0 or 3-0), and bat
much worse when there are two strikes (the pitcher has the advantage). Trout’s averages are
consistent with this pattern.

Pitch Count PA AVG OBP SLG

Through 3-0 46 .375 .761 .563
Through 3-1 93 .280 .602 .560
Through 3-2 142 .290 .493 .630
Through 2-0 103 .385 .602 .600
Through 1-0 308 .312 .435 .611
Through 2-1 181 .350 .497 .686
Through 1-1 360 .273 .367 .562
Through 0-1 370 .268 .335 .539
Through 2-2 227 .260 .348 .505
Through 1-2 249 .211 .265 .404
Through 0-2 131 .216 .244 .440
Next, we see how Trout performed when he hit grounders, fly balls (flies), or line drives
(liners). Trout had a very high slugging percentage when he hit a fly ball or a line drive.

Batted Ball Type PA AVG OBP SLG

Grounders 145 .352 .352 .400
Flies 202 .330 .317 1.021
Liners 81 .734 .716 1.038

Last, we see how Trout did on balls hit to the left, center, and right portions of the baseball
field. We see that Trout hit better when the ball was hit to left and center—this is a typical
pattern for a right-handed hitter.

Handedness PA AVG OBP SLG

as R to Left 158 .494 .487 .853
as R to Center 152 .473 .461 1.068
as R to Right 118 .228 .220 .412

To sum up, although Trout overall hit for a .287 batting average in the 2014 season, he
appears to bat much better (or much worse) in particular situations. Specifically, we see that
Trout

batted 31 points better with men on base

batted 10 points worse at home games,
batted for a high average in June and March/April and a low average in August.
hit for a .358 average in situations with high leverage

The main question we will address in the following case studies is:

Do these observed situational effects correspond to real effects?

For example, can we say that there is really an advantage to hitting in situations of high leverage?
Maybe Trout has the same ability to get a base hit in low and high leverage situations and, by
luck or chance variation, he happened to hit better in high leverage situations this year. Or maybe
hitters have a general advantage in situations of high leverage. We will see in this chapter that
much of the variation that we see in situational data such as Trout’s batting averages can be
explained by chance, and it is relatively difficult to pick out real situational effects.

8.2 Observed Situational Effects for Many Players

Topics Covered: Stemplot, scatterplot, relationship between two variables, distinction between
ability and performance.
To start to get an understanding of the patterns of situational data, we look at Table 8.1
where we see the HOME on-base percentage and AWAY on-base percentage for a group of
20 hitters for the years 2013 and 2014. Let’s focus first on the 2013 data. For each player, we
compute the difference

DIFF D HOME OBP in 2013 AWAY OBP in 2013.

For example, Robinson Cano’s OBP was .401 at home games and .366 at away games for
a difference of DIFF D :401 :366 D :035: Figure 8.1 displays a dotplot of the differences of
the home and way OBPs in the 2013 season. What can we see in this dotplot?

We can compute the median difference; it is :003. This means, on average, that a hitter bats
3 points lower during HOME games than during AWAY games. This is a little surprising since
we thought players hit better at home, on average.

Table 8.1. Home and away on-base percentages for a group of 20 hitters for the 2013
and 2014 seasons
2013 2014
Name Home Away Diff Home Away Diff
Pedro Alvarez .282 .310 :028 .315 .309 .006
Adrian Beltre .375 .367 .008 .415 .364 .051
Carlos Beltran .338 .340 :002 .319 .280 .039
Michael Bourn .314 .317 :003 .340 .294 .046
Robinson Cano .401 .366 .035 .377 .386 :009
Matt Carpenter .432 .353 .079 .353 .396 :043
Alejandro De Aza .338 .308 .030 .320 .308 .012
Ian Desmond .361 .301 .060 .318 .308 .010
Josh Donaldson .367 .400 :033 .322 .361 :039
Alcides Escobar .252 .266 :014 .324 .310 .014
Jedd Gyorko .301 .301 0 .294 .264 .030
Eric Hosmer .352 .355 :003 .260 .370 :110
Juan Lagares .266 .297 :031 .314 .327 :013
James Loney .290 .404 :114 .352 .321 .031
Justin Morneau .336 .312 .024 .363 .364 :001
Buster Posey .366 .377 :011 .307 .417 :110
Alex Rios .309 .338 :029 .319 .304 .015
Kyle Seager .316 .360 :044 .370 .301 .069
Jean Segura .314 .344 :030 .305 .274 .031
Ben Zobrist .368 .341 .027 .379 .329 .050

-0.10 -0.05 0.00 0.05

Home OBP - Away OBP

Figure 8.1. Dotplot of differences in home and away on-base percentages for 20 players in the 2013
season.

However, there is a tremendous range in these difference values. One hitter had a :114
difference—this player batted 114 points better on away games. In contrast, another hitter had
a difference of .079. This player batted 79 points higher at home. Generally, situational data
is interesting since one will typically see a great range of differences there will be many large
positive values and large negative values.

Since we see many “extreme” home/away effects, it is tempting to try to explain why a
particular player is doing so well (or so poorly) at home. But do players really have different
abilities to use the home field advantage? Is it possible that one player uses the home field to
great advantage in his hitting, while a second player has the same batting ability at home and

away games? We can learn about the presence of home/away batting abilities by looking at data
for many players.

Demonstration that Players Have Different Batting Abilities

It is obvious to most baseball fans that players have different batting abilities. But how can we
demonstrate this fact from hitting data? Let’s look at the home hitting data in Table 8.1. For each
player, we collect his 2013 home on-base percentage and his 2014 home on-base percentage.
Figure 8.2 displays a scatterplot of the 2013 and 2014 home averages for the 20 players. Looking
at the graph, we see that the points drift from the lower left to the upper right regions this means
that the 2013 and 2014 batting averages are positively associated. (The correlation between the
two variables is .34.) This means that players who hit well at home in 2013 tend also to hit
well at home the following year. Likewise, a weak hitting season by a player in 2013 tends to
be associated with a weak-hitting performance the following season. The explanation for the
positive association in the graph is that players have different batting abilities—the better hitters
correspond to points in the upper right section of the graph and the weak hitters correspond to
points in the lower left section.
0.40
2014 HOME OBP
0.35
0.30

0.25 0.30 0.35 0.40

2013 HOME OBP

Figure 8.2. Scatterplot of 2013 and 2014 home on-base percentages for 20 players.

Do Players Have Different Home/Away Abilities?

We use the above strategy to search for situational abilities. Let’s consider a particular ballplayer
who bats 50 points better at home than away games. Is this real? In other words, does this player
really have an extra ability to hit well during home games? We can see if this is a real effect
by looking at next year’s performance. If this player continues to hit much better at home, then
we would be more confident that he has some extra home batting ability. If we have home/away
batting data for a group of players for two consecutive years, we look for a general relationship
between a player’s home/away effect one year with the corresponding effect the next year.
In Table 8.1, we have computed the difference between the home and away on-base percent-
ages (DIFF) for both the 2013 and 2014 baseball seasons for our 20 players. To see if there is a

general relationship between a player’s 2013 DIFF and his 2014 DIFF, we construct a scatterplot
in Figure 8.3.

0.05
2014 Difference
0.00
-0.05
-0.10

-0.10 -0.05 0.00 0.05

2013 Difference

Figure 8.3. Scatterplot of differences in home and away on-base percentages for 20 players in 2013 and
2014 seasons.

In contrast to the pattern in Figure 8.2, we don’t see a strong trend in this graph. (The
correlation between the two variables in this case is actually the negative value :16:) There
does not appear to exist a tendency to show the same type of home/away effect for two consecutive
years. This suggests that the home/away effect is not an ability characteristic. This means that a
player who hits unusually well at home one particular year generally will not hit unusually well
at home the next year likewise poor home hitters one year will not be poor hitters at home the
following year. Remember that there will be a positive association between a player’s batting
average (or any other batting measure) for two consecutive years—this means that players have
intrinsic batting abilities. But there is no general tendency for players to hit better (or worse) in
the home/away situation for two consecutive years.

8.3 Modeling On-Base Percentages for Many Players

Topics Covered: Probabilities of hitting for many players, a random effects model, assessing
the goodness of fit of a model, normal distribution.
In our earlier case studies:

We distinguished between the Ability of a hitter from his Performance.

We observe a hitter’s performance during the season; we don’t know a batter’s ability, but
we’ll learn about it from the hitting data.
We model a hitter’s ability by a spinner. The spinner has two areas, “On-Base” and “Not
On-Base”, and the area of the On-Base region is given by p. This is our measure of ability.

Figure 8.4. Stemplot of the on-base percentages of all players with at least 300 at-bats in the 2014 season.

Now instead of one player, we consider all 238 players who were “regulars” (had at least 300
at-bats) in the 2014 MLB season. Figure 8.4 displays a stemplot of the on-base percentages.
We see that the on-base percentages are bell-shaped about the mean value of .326. The
highest batting average was Troy Tulowitzki at .432 and the lowest was John Schierholtz at .240.
Can we construct a model for the true on-base percentages (the abilities) for these 238
hitters?
The simplest model I can think of is what we call the “One Spinner” model. Maybe all of
the 238 players have the same batting ability. Each player is using the same spinner with on-base
probability p D :326 (the average of all players). If you believe this, then the variation in the
season on-base percentages that we observe in the stemplot above are simply due to chance
variation. Players all have the same ability, but some like Tulowitzki are lucky and have high
season averages; others, like Schierholtz, were unlucky and had low averages.
Now, if you know anything about baseball, you should be thinking that this is a crazy
model—players do have different batting abilities. (In fact, we illustrated this fact in the last
case study where we looked at batting averages of some players for two consecutive seasons.) I
agree, but I want to demonstrate how a statistician checks the suitability of a probability model.

Simulating Hitting Data Using a “One Spinner” Model

To see if the “One Spinner” model is reasonable, we simulate hitting data from the model.
We imagine 238 spinners, each with on-base probability p D :326. We spin each spinner
for a whole season using the same numbers of at-bats as the 2014 players and obtain 238
season on-base percentages. We did this one time. In the back to back stemplot display in

Figure 8.5, we have placed the simulated batting averages on the left and the actual 2014 batting
averages on the right.

Figure 8.5. Back-to-back stemplots of simulated on-base percentages from one-spinner model (left) and
2014 MLB on-base percentages (right).

How does the simulated data compare with the actual data? There is a substantial difference:
the simulated batting averages appear to have less variation or spread than the real averages.
We can measure spread of a dataset by the standard deviation s. For the dataset of real
batting averages, we compute s D :0329 and for the simulated data, s D :0217. This confirms
what we just said the simulated averages aren’t as spread out as the real averages.
To see if this always happens, we did the simulation from the “One Spinner” model many
times. In four additional simulation runs, we obtained

s D :0212; s D :0212; s D :0201; s D :0208:

Each time, the standard deviation is smaller than the standard deviation (.0329) for the real
data. So this model does not seem to generate data that is similar to the 2014 dataset of on-base
percentages.
We conclude the “One Spinner” model isn’t appropriate and consequently that batters have
different abilities.

A Many Spinners Model

What is an alternative probability model that can represent hitting data in baseball? Here is a
“Many Spinners” model that seems to better represent what is really going on.

We first represent the hitting abilities of the 238 players by a normal curve shown in Figure
8.6 with an average of .326 and standard deviation of .025. In other words, the on-base
probabilities of the players are variable according to a normal curve with mean .326.

15
10
5
0

0.1 0.2 0.3 0.4 0.5

On-Base Probability

Figure 8.6. Bell-shaped curve to represent the on-base probabilities of the MLB players.

We sample at random 238 on-base probabilities from this normal curve. Then we represent
the abilities of the players using a set of spinners, where the “On-Base” areas are different
for different players. One player might have a on-base area of p D :310, another may have a
on-base area of p D :330, and so on.

We tried simulating data using this “Many Spinners” model. We first simulated a set of
random hitting probabilities and then simulated hitting data using these probabilities. Figure 8.7

Figure 8.7. Back-to-back stemplots of simulated batting averages from many-spinners model (left) and
2014 MLB on-base percentages (right).

shows back-to-back stemplots of the simulated season on-base percentages and the actual 2014
on-base percentages.
Comparing the two stemplots, the distribution of the simulated data from the “Many Spin-
ners” data does resemble the distribution of the actual 2014 on-base percentages. In particular,
the standard deviations match up—the standard deviation of the simulated data is .0320, which
is close to the standard deviation of the observed data .0329. So the “Many Spinners” model
seems to be a good representation for the hitting abilities of many players. We will use this
probability model in the following case study when we model situational hitting data.

8.4 Models for Situational Effects

Topics Covered: Probability models for situational hitting data, bias model, model with ability
effects.
What are good models for situational data? We describe three basic models here that
seem to describe the pattern in hitting data for all of the situations that are displayed in the
fangraphs.com site.
Recall our basic probability model for batting data for a group of players described in Case
Study 8.3. Players have different abilities (values of p) selected from a normal curve (with a
mean of .326 and a standard deviation of .025), and batting data is found by spinning a bunch
of spinners, where each spinner has a different hitting probability.

The “No Effect” Model

To describe the simplest situational hitting model, suppose we have five spinners with on-
base probabilities of p D :2; :3; :4; :5; :6. These spinners represent the abilities for five baseball
players. Suppose that we spin the spinners in the dark and the light.
Now, it is reasonable to think that the lightness of the room has no effect on the probability
of getting a hit on a particular spinner. So if the first spinner has a hitting probability of p D :2
in the light, then this spinner will also have a hitting probability of p D :2 in the dark. Likewise,
the .3 spinner will have a hitting probability of .3 both in the light and the dark, the .4 spinner
will have the same hitting probability in the dark and in the light, etc. In this case there is no
true situational effect so we call this the “no effect” model.
Now suppose we spin the p D :2 spinner 100 times in the dark and 100 times in the light.
It is certainly possible that we’ll get 22 on-base events in the dark and only 18 on-base events in
the light for an observed situational effect of

Observed Situational Effect D 22=100 18=100 D :04

But this effect is just due to chance there is no true situational effect. An example of this “no
effect” situation is before and after the All-Star Game. Generally, players appear to have the
same batting abilities before and after the All-Star Game. Certainly, we will observe some
players one season who bat better before the All-Star Game and we’ll see other players who are
“hot” after the All-Star Game. But most of the variation in the differences in batting averages
before and after the All-Star Game is due to chance variation.

The “Bias” Model

The bias model is a different way of representing the hitting abilities of players in two situations.
Here the situation changes the hitting probability by the same amount for all hitters.

Let’s consider our dark and light example again. Let’s suppose that it is hard to see the
spinner in the dark, and the light increases the on-base probability by .05 for all hitters. So if
one hitter has a dark on-base probability of p D :3, it will be p D :3 C :05 D :35 in the light.
Another hitter with a dark on-base probability of p D :4 will have an on-base probability of
p D :4 C :05 D :45 in the light.
The bias model seems appropriate for the home vs away hitting data. A particular ballpark
has a positive or negative impact on a player’s batting average. Coors Field is an obvious example
of a ballpark that helps a player’s batting ability, and Dodger Stadium is an example that hurts a
player’s batting ability. But the effect of the ballpark is to add a constant number to each player’s
batting probability. So Coors Field may add a positive number, say .10 to each player’s true
batting average. There is a situational effect here, but this effect is the same for all players.

A Simulation
I used a statistics computing package to illustrate what situational data looks like if there is a
bias effect. I assumed

there are 100 players,

each player has 300 at-bats at HOME and 300 at-bats AWAY,
there is a bias situational effect of twenty points due to playing at home so each player’s
HOME on-base probability is .020 higher than his AWAY on-base probability.

I had the program first simulate 100 “Away” true on-base percentages from a normal curve
with mean .326 and standard deviation .025—these hitting probabilities are put in the “p.AWAY”
column. We compute “Home” true batting averages in the “p.HOME” column by adding .020
to each probability in the “p.AWAY” column. We then simulated hits for 300 at-bats at the home
games using the Home hitting probabilities—these numbers are put in the “h.HOME” column.
Similarly, we simulated hits for 300 away at-bats using the Away hitting probability—these hit
numbers are in the “h.AWAY” column. We compute on-base percentages at home and away
(these OPB’s are in the “OBP.HOME” and “OBP.AWAY” columns), and compute observed
differences

OBP(home) OBP (away).

A part of the output of this simulation is shown in Table 8.2.

Table 8.2. Some situational hitting data assuming that the home ballpark
adds .020 to the probability of getting on base for all players
p.AWAY p.HOME OBP.AWAY OBP.HOME DIFF
1 .32 .34 .30 .30 .00
2 .36 .38 .37 .43 .06
3 .30 .32 .32 .34 .02
4 .35 .37 .37 .35 .02
5 .35 .37 .36 .37 .01
6 .37 .39 .37 .37 .00

Figure 8.8 displays a dotplot of the observed situational effects for the 100 players contained
in the DIFF column. The median value of DIFF is equal to .020 and the extreme values are
:083 and :093.

-0.05 0.00 0.05 0.10

OBP(Home) - OBP(Away)

Figure 8.8. Dotplot of observed situational effects from simulated data where there is a bias situational
effect.

Here we see that when there this is a bias situational effect, the observed effects can look
interesting. One player in the simulation hit for 93 points better at home games and another hit
for 83 points higher at away. This high variation is misunderstood by baseball fans. People think
that some players (like the one that hit 93 points higher during home games) have a special
ability to play better at home, when really the home ballpark has the same positive effect on all
players.

The “Ability” model

The last model, the so-called “ability” model, is the most complicated description of situational
data. This probability model says that there are real situational effects, and they are different
among players. I’ll illustrate this scenario using the following “Joe Cool” and “Harry Hyper”
example—two players who react differently to a home crowd.
Joe Cool plays the same way everywhere he plays. So if Joe hits for a on-base probability
of .350 at home, he’ll also hit for a on-base probability of .350 on the road. Harry, in contrast,
is very emotional (like Pete Rose) and feeds off of the enthusiasm of the home crowd. If Harry
hits for a on-base probability of .325 on away games, he might hit 50 points higher, or .375, at
home.
There are relatively few situations (among the ones that are discussed) that are ability
effects. The one exception is the batting performance of the hitter under different pitch counts.
Suppose one looks at two batting averages:

the batting average when there are two strikes on the batter,
the batting average when the batter is ahead in the count (that is, the pitch count is 3-0, 3-1,
2-0, 2-1, or 1-0).

If one looks at the situational hitting data for all players in this situation, there is evidence that
pitch count is an ability effect.
This means that players have different abilities under various pitch counts. A big slugger
(Ryan Howard comes to mind) is an ineffective hitter when he is behind in the count. When the
count is 0 balls and 2 strikes, he’s likely to strike out. In contrast, the good contact hitters (like

Ichiro Suzuki) are effective hitters even when they are behind in the count. Suppose we define
the pitch count situational effect to be

DIFF D OBP(ahead in the count) OBP(behind in the count).

There is evidence to suggest that big sluggers (like Howard) have large values of this
situational effect, and other hitters (the contact-hitter like Suzuki) have relatively small pitch
count effects. How does one find a group ability effect in situational data? Recall our discussion
about looking for abilities in hitting data in Case Study 8.2. Suppose you look at a group of
players and compute their situational effect (say, on-base percentage ahead in the count minus
on-base percentage behind in the count) for two consecutive years, say 2013 and 2014. Construct
a scatterplot of the two years of situational effects. You have found a group ability situational
effect if you find a positive association in the scatterplot between the 2013 effect and the 2014
effect. This means that players with high situational effects in 2013 will tend to have high
situational effects in 2014. Also players who have low situational effects one year will tend to
have low effects the second year. Remember that most situational effects in baseball hitting are
“no effects” or “biases”—the pitch count is one of the few situations where players generally
appear to have different abilities to take advantage of the situation.

8.5 Is Michael Brantley Streaky?

Topics Covered: Distinction between streaky performance and streaky ability, moving averages,
runs.
In this case study we talk about streakiness or the hot hand. Do players really get hot and
cold? Here I’m not talking about the hot and cold streaks that we observe in baseball data. I’m
instead talking about a player’s ability to “get into a groove” in hitting or pitching. Maybe during
a certain week, the player’s batting stroke feels just right or he sees the ball particularly well.
Another week, he feels different maybe his batting stroke is out of sync or he has an injury that
makes hitting more difficult.
Is there evidence for a player to be truly streaky, that is, have an ability to be hot and cold?
We focus on the Indians’ player Michael Brantley. To look for streakiness in Brantley’s
hitting data, we focus on his game-to-game hitting data for the 2014 season.

Moving Average
One way of looking for streaky behavior is to compute a moving average—this is a short-term
batting average using a window of a particular number of games.
Suppose that we wish to compute moving averages using a window of five games.

The first moving average is the batting average for games 1 through 5.
The second moving average is the batting average of games 2 through 6.
The third moving average is the batting average of games 3 through 7.

Essentially we’re just computing averages over short time intervals.

In Table 8.3, I compute moving averages for the 2014 Brantley data using a window of 9
games.

Table 8.3. Computation of moving averages for day-by-day batting data

for Michael Brantley
Game AB H 9-Game AB 9-Game H Moving AVG
1 4 2
2 3 1
3 4 1
4 4 1
5 4 1 34 10 .294
6 5 3 34 9 .265
7 4 0 35 9 .257
8 3 1 34 9 .265
9 3 0 33 9 .273
10 4 1 32 8 .250
11 4 1 32 7 .219
12 3 1 33 8 .242
13 3 1 33 7 .212
14 3 0 33 9 .273
15 5 2 33 9 .273

1. In games 1–9, Brantley had 34 AB and 10 H for an average of 10=34 D :294. This moving
average of .294 is put in the table across the average game number 5 (5 is the average of
games 1; 2; : : : ; 9).
2. In the table, we see a moving average of .250 across game 10. We look at the nine games
centered about game 10 these are games 6 through 14. In this period, Brantley had 8 hits in
32 at-bats for a batting average of 8=32 D :25.

If we graph the game numbers against the corresponding moving averages, the graph shown
in Figure 8.9 is obtained.
0.5
0.4
Average
0.3
0.2

0 50 100 150
Game

Figure 8.9. Moving average plot of Mike Brantley’s batting average using a window of nine games.

This graph is a picture of the hot and cold periods of Brantley’s hitting for the 2014 season.
We see

a hot period about game 70—during one nine game period, Brantley batted over .500,
this hot period was surrounded by two cold periods where he batted close to .250,
Brantley finished the season with a very cold period where his average was under .100 followed
by a hot period where he batted over .4. So one indication of streakiness is the extreme hills
and valleys that we see in the moving average plot.

Runs
Another way of measuring streakiness is based on runs of good and bad days. (Not to be
confused with runs scored in a baseball game.)
Suppose we classify each day of Brantley’s hitting as being either “Hot” or “Cold”—we
say he is Hot if his day batting average is over .327 (his season avg); otherwise he is classified
as Cold. Table 8.4 shows Brantley’s batting results for the first 13 games and the hot and cold
classification.

Table 8.4. Batting results for Michael Brantley

for the first 13 games of the 2014 season and
classification of the result in hot and cold states
Game AB H Hot or Cold?
1 4 2 Hot
2 3 1 Hot
3 4 1 Cold
4 4 1 Cold
5 4 1 Cold
6 5 3 Hot
7 4 0 Cold
8 3 1 Hot
9 3 0 Cold
10 4 1 Cold
11 4 1 Cold
12 3 1 Hot
13 3 1 Hot

In Game 1, Brantley was 2=4 D :500; this is larger than .327, so we call it a Hot day. In
Game 2, he was 1=4 D :250 which is under .300, so he is Cold that day.
Now we look for Runs in this sequence. A run is simply a consecutive sequence of Hot’s or
a sequence of Cold’s. (This is different from the usual meaning of run in baseball.) Specifically,
we count

1. the total number of runs ,

2. the length of the longest run (either Hot or Cold).

In the above sequence of 13 games, we see that

total number of runs D 7,

longest run is (Cold, Cold, Cold), so length of longest run D 3.

If the player is truly streaky, we expect to see

a small number of runs,

long runs of Hot’s or Cold’s, so the length of the longest run would be large.

This makes sense. If a player is streaky, then he will have long runs of hot games and long runs
of cold games, so the number of runs in the sequence will be small.
Here we have discussed ways of detecting the observed streakiness that we see in baseball
data. However, that does not mean that the player (or team) is truly streaky. In the next case
study, we’ll clearly make a distinction between a player’s streaky ability (or lack of streaky
ability) and his streaky performance during a baseball season. We will see that a fair die, an
object with consistent ability, can exhibit very streaky performance.

8.6 A Streaky Die

Topics Covered: Moving averages, runs, simulation of dice tossing, patterns in dice tossing
data.
In the previous case study, we looked for streaky behavior in the day-to-day hitting data for
Michael Brantley’s 2014 baseball season. Specifically, we looked at

moving averages using a window of nine games we were looking for unusually high or low
moving averages to say that Brantley is streaky,
runs of good and bad hitting days here we looked at the total number of runs, and the length
of the longest run—a small total number of runs and/or a long run (of hot or cold) games
would indicate streakiness.

Using moving averages and runs, we saw some interesting features in Brantley hitting data:

In one 9-game stretch, Brantley hit .552; in another 9-game stretch, he only hit .118.
Brantley had a total of 78 runs (either runs of Hot or runs of Cold). His longest run had length
9. But are these really interesting? Do they mean that Brantley is truly a streaky hitter? In
other words, can we tell from these data that Brantley has an ability to be streaky? Maybe he
isn’t really streaky and by luck or chance variation, we just happened to see some interesting
streaky behavior.

Mr. Consistent
Let’s consider this last question more carefully. Suppose that Brantley is truly not streaky. What
would that mean?
The opposite of a streaky hitter is a consistent hitter. This type of hitter always gets a hit
with the same probability no matter what the time or situation. This consistent hitter will get a
hit with the same probability p in every at-bat in the season.
We can represent hitting of a consistent hitter by use of a die. Consider a 10-sided die with
the sides 0, 1, 2, 3, 4, 5, 6, 7, 8, 9. Suppose we roll the die and a “hit” is recorded if the die
rolls 1, 2, 3; otherwise, the hitter is out. Then the probability of a hit will be 3=10 D :3. More

importantly, the chance of a hit is always .3 and the outcome of the die won’t depend on what
rolls occurred in the past.
We can use this die to simulate Brantley’s hit outcomes if he were really a consistent hitter.
We will simulate hitting data for all 162 games by working in groups—group 1 will simulate
Brantley’s hitting for games 1–18, group 2 will simulate Brantley’s hitting for games 19–36,
etc.
When we complete this activity we will see if this simulated data looks streaky. In particular,
we will look at

moving averages using a window width of 9 games,

the number of runs and the length of the longest run.

What I think we’ll discover is that this simulated data can look pretty streaky. Even when a
hitter is truly consistent, there can be interesting patterns in the moving average graph that can
be interpreted as streaky behavior.
Figure 8.10 shows the moving average graph (using a width of nine games) for the data that
a statistics class generated using a 10-sided die.
0.15 0.20 0.25 0.30 0.35 0.40 0.45
Average

0 50 100 150
Game

Figure 8.10. Moving average plot of simulated hitting data for a truly consistent hitter where the proba-
bility of a hit is equal to .3.

We see a lot of interesting patterns. Our hitter had an early slump around game 20 and had
one significant hot streak where he batted in the .500 range. But remember this hitter is truly a
consistent batter what we are observing is the streakiness that is inherent in chance variation.
To show you that this is not a fluke occurrence, I did our simulation five times on a computer.
I’m using the same assumptions as above.

1. Our player is a truly consistent hitter with probability p D :3 of getting a hit on a single
at-bat.
2. This player has exactly 4 at-bats in each game.

Figure 8.11 shows the moving average plots for my six simulations (I’m still using a window
width of nine games).

Simulation 1 Simulation 2 Simulation 3

0.5

0.4

0.3

0.2
Average

0.1

Simulation 4 Simulation 5 Simulation 6

0.5

0.4

0.3

0.2

0.1

0 50 100 150 0 50 100 150 0 50 100 150

Game

Figure 8.11. Moving average plots for six simulated datasets for a consistent hitter with a constant
probability of .3 of getting a hit.

You should notice a lot of up-and-down behavior in the five moving average plots. So even
if Brantley were a truly consistent hitter, he would likely have several interesting hot and cold
streaks.
The moral of the story is that you should be cautious about interpreting streaky behavior of
baseball players. Dice have a clearly consistent ability—just the opposite from true streakiness.
But we’ve seen that dice can look streaky!

8.7 Exercises
8.0. Table 8.5 gives some situational OBPs for Rickey Henderson in the 1999 baseball season.
This table shows how Henderson did for day and night games, for games home and away,
for games played on grass and turf fields, for games played in domed and open ballparks,
and against left- and right-handed pitchers.
(a) For each breakdown, compute the number of (approximate) plate appearances (PA)
by adding at-bats (AB) to walks (BB).
(b) Let pday denote Henderson’s on-base probability when he is playing a day game and
let pnight denote his on-base probability when he’s playing at night. Using the data in

Table 8.5. Situational on-base percentages of

Rickey Henderson for the 1990 season
Breakdown AB BB PA OBP
Day 144 21 .400
Night 294 61 .433
Home 200 34 .376
Away 238 48 .462
Grass 364 60 .408
Turf 74 22 .490
Dome 43 10 .472
Open 395 72 .417
vs. Right 333 54 .403
vs. Left 105 28 .481

the table, construct 90% probability intervals for pday and pnight . Comparing the two
intervals, can you conclude that Henderson really has a higher on-base probability at
night games?
(c) Using the same method as in (b), compare the home/away OBPs, the grass/turf OBPs,
the dome/open OBPs, and the right/left OBPs.
(d) Based on your work in (c) and (d), rank the five situations with respect to the most
significant effect to the least significant effect.
8.1. Table 8.6 gives batting statistics for 14 randomly selected players in the 2013 season. The
first three columns give the number of at-bats (AB), hits (H), and batting average (AVG)

Table 8.6. Home and away batting averages for 14 randomly selected players from
the 2013 season
Home Away
Name AB H AVG AB H AVG Home—Away
David Ortiz 272 83 .305 246 77 .313
Daniel Murphy 320 84 .262 338 104 .308
Evan Longoria 291 75 .258 323 90 .279
Khris Davis 72 22 .306 64 16 .250
Christian Yelich 121 31 .256 119 38 .319
Marlon Byrd 261 67 .257 271 88 .325
Xander Bogaerts 22 5 .227 22 6 .273
Marcell Ozuna 127 32 .252 148 41 .277
Todd Frazier 264 70 .265 267 54 .202
Matt Carpenter 311 112 .360 315 87 .276
Andrelton Simmons 286 72 .252 320 78 .244
Zack Cozart 289 68 .235 278 76 .273
Martin Prado 292 85 .291 317 87 .274
Jacoby Ellsbury 284 84 .296 293 88 .300

for all games played at Home, and the next three columns give the same statistics for all
games played Away from home.
(a) For each player, compute the difference in batting averages

Difference D AVG (Home) AVG (Away)

and put the differences in the “Home Away” column of Table 8.6.
(b) Draw a stemplot of the batting average differences that you computed in part (a).
(c) Looking at the stemplot constructed in part (b), what is the median value of the
difference? Also find the smallest and largest differences, and find the players that
had these extreme values.
(d) From your work, can you say that players generally hit better at home? If so, by how
much on the average?
8.2. Table 8.7 gives 2014 batting statistics for 15 players before the All-Star break (months
April, May, and June) and after the All-Star break (months July, August, and September).
(a) Repeat parts (a) – (c) of Exercise 8.1 for this dataset. Here the PRE-POST column of
Table 8.7 will contain the differences in batting average

AVG(Pre All-Star Game) AVG(Post All-Star Game).

(b) From your work, can you say that players generally play better after or before the
All-Star break? If so, how much on the average?

Table 8.7. Batting statistics for 15 players before and after the 2014
All-Star break
Pre AS Game Post AS Game
Name AB H AVG AB H AVG Pre—Post
David Ortiz 296 74 .250 222 62 .279
Daniel Murphy 340 103 .303 256 69 .270
Jose Abreu 272 76 .279 284 100 .352
Evan Longoria 332 87 .262 292 71 .243
Khris Davis 290 75 .259 211 47 .223
Christian Yelich 255 66 .259 327 99 .303
Marlon Byrd 310 83 .268 281 73 .260
Xander Bogaerts 290 72 .248 248 57 .230
Marcell Ozuna 291 77 .265 274 75 .274
Todd Frazier 310 89 .287 287 74 .258
Matt Carpenter 317 89 .281 278 73 .263
Andrelton Simmons 291 72 .247 249 60 .241
Zack Cozart 276 63 .228 230 49 .213
Martin Prado 314 84 .268 222 67 .302
Jacoby Ellsbury 302 87 .288 273 69 .253

8.3. Table 8.8 displays 2014 batting statistics for 15 players when the batter had an initial ball
(a 1-0 count) and when the batter had an initial strike (a 0-1 count).

Table 8.8. Batting statistics for 15 players with an initial ball (1-0 count) and an
initial strike (0-1 count)
Initial Ball Initial Strike
Name AB H AVG AB H AVG Initial B—Initial S
David Ortiz 255 64 .251 207 53 .256
Daniel Murphy 304 76 .250 206 67 .325
Jose Abreu 263 78 .297 208 67 .322
Evan Longoria 324 70 .216 205 56 .273
Khris Davis 244 49 .201 199 53 .266
Christian Yelich 306 64 .209 223 79 .354
Marlon Byrd 315 78 .248 201 48 .239
Xander Bogaerts 301 65 .216 194 50 .258
Marcell Ozuna 298 74 .248 195 57 .292
Todd Frazier 293 67 .229 226 70 .310
Matt Carpenter 341 93 .273 226 60 .265
Andrelton Simmons 244 63 .258 198 46 .232
Zack Cozart 261 58 .222 184 39 .212
Martin Prado 307 89 .290 199 52 .261
Jacoby Ellsbury 264 64 .242 226 67 .296

(a) Repeat parts (a)–(c) of Exercise 8.1 for this dataset. Here the “Initial B Initial S”
column of Table 8.8 will contain the differences in batting average

AVG(Initial Ball) AVG(Initial Strike).

(b) From your work, can you say that players generally hit better when they have an
initial ball as opposed to an initial strike? If so, how much on the average?
8.4. Suppose that one is interested in how players hit on odd-numbered days (like April 7,
May 13, June 9) as opposed to even-numbered days (like April 10, May 14, July 20). Here
we are pretty sure that there is no true situational effect for any player. (Why would any
player actually be a better or worse hitter on odd-numbered days?) We simulate the type
of hitting data that one might see in this scenario by means of the following experiment.
For each of 15 players, we first simulate their abilities (their hitting probabilities) using a
normal curve with mean .276 and standard deviation .021. Here are the twenty simulated
abilities:

.262 .282 .291 .293 .321 .227 .279 .248

.288 .309 .264 .238 .295 .309 .287

Then we simulate the situational data as follows. For each player, we simulate the results
of 300 at-bats for the odd-numbered days using the above hitting probabilities, and then

simulate the results of 300 at-bats for the even-numbered days using the same set of hitting
probabilities. We obtain the data in Table 8.9.

Table 8.9. Simulated batting statistics for 15 players on even-numbered

and odd-numbered days
Even # Days Odd # Days
Name AB H AVG AB H AVG Even—Odd
Jim 300 82 .273 300 78 .260
Pat 300 80 .267 300 87 .290
Edsel 300 87 .290 300 94 .313
Arjun 300 84 .280 300 72 .240
Rich 300 97 .323 300 94 .313
Pete 300 67 .223 300 59 .197
Sam 300 89 .297 300 80 .267
Dave 300 91 .303 300 61 .203
Curt 300 87 .290 300 84 .280
Ben 300 85 .283 300 89 .297
Adam 300 80 .267 300 93 .310
Dale 300 67 .223 300 72 .240
Joe 300 72 .240 300 80 .267
Dick 300 75 .250 300 92 .307
Brad 300 90 .300 300 77 .257

(a) Compute all of the batting average differences

Difference D AVG(Even-Numbered Days) AVG(Odd-Numbered Days)

and put the differences in the “Even Odd” column of Table 8.9.
(b) Graph the batting average differences using a stemplot.
(c) Find the average difference, the low and high differences, and find the batters who
had these extreme values.
(d) Do these simulated data resemble any of the situational data described in the chapter?
Explain.
8.5. Suppose that it is found out that some ballparks are using a special baseball that travels
further when hit. Moreover, it is known that this special baseball will add 30 points to
every player’s hitting probability. Table 8.10 generates some simulated hitting data using
this scenario.
To simulate these data, we first simulate hitting probabilities for the 15 players using a
normal curve with mean .261 and standard deviation .021. These probabilities correspond
to the abilities of the hitters playing with the Usual ball.

.290 .279 .267 .252 .249 .264 .244

.276 .287 .244 .294 .295 .279 .242 .288

To obtain the hitting probabilities for the players with the Special ball, we add 30 points
to each of the Usual hitting probabilities.

.320 .309 .297 .282 .279 .294 .274

.306 .317 .274 .324 .325 .309 .272 .318

Then we simulate hitting data for 300 at-bats using the Usual probabilities and for 300
at-bats using the Special probabilities—we show the data in Table 8.10.

Table 8.10. Simulated hitting data for 15 players using a special baseball and
the usual baseball when there is a true situational bias
Usual Baseball Special Baseball
Name AB H AVG AB H AVG Special—Usual
Larry 300 89 .297 300 81 .270
Moe 300 90 .300 300 98 .327
Curly 300 85 .283 300 79 .263
Harold 300 95 .317 300 74 .247
Harvey 300 72 .240 300 85 .283
Lee 300 74 .247 300 97 .323
Charles 300 72 .240 300 81 .270
Bill 300 77 .257 300 105 .350
Bob 300 83 .277 300 101 .337
Rick 300 81 .270 300 78 .260
Britt 300 89 .297 300 94 .313
Clark 300 86 .287 300 99 .330
Randy 300 85 .283 300 91 .303
Tom 300 86 .287 300 75 .250
Gene 300 87 .290 300 102 .340

(a) Compute all of the batting average differences

Difference D AVG(Special Ball) AVG(Usual Ball)

and put the differences in the “Special Usual” column of Table 8.10.
(b) Graph the batting average differences using a stemplot.
(c) Find the average difference, the low and high differences, and find the batters who
had these extreme values.
(d) Do these simulated data resemble any of the situational data described in the chapter?
Explain.
8.6. (Exercise 8.5 continued.) Suppose that players have different abilities to use the “Special
Ball” that was discussed in Exercise 8.5. For the first five players listed in Table 8.11 (the
first group), the special ball adds 50 points to the “Usual Ball” hitting probability. For the
next five players in the table (the second group), the special ball adds 30 points to the
player’s hitting probability, and for the final five players (the third group), the special ball

only adds ten points to the hitting probability. Below we give the hitting probabilities for
the 15 players. Table 8.11 shows hitting data simulated from this model.

Hitting Probabilities Hitting Probabilities

Players for Usual Ball for Special Ball
First .282, .248, .291 Add 50 .332, .298, .341
Group .310, .261 Points .360, .311
Second .243, .246, .288 Add 30 .273, .276, .318
Group .303, .290 Points .333, .320
Third .251, .276, .273 Add 10 .261, .286, .283
Group .268, .290 Points .278, .300

(a) Compute the differences in batting average (Special Ball Usual Ball) and put the
differences in the last column of Table 8.11.
(b) Construct a stemplot of the differences.
(c) Compute the average batting average difference, and find the smallest and largest
differences.
(d) Compare the batting average differences with the differences found in Exercise 8.5.
Can you offer any explanation for the different patterns that you see in stemplots
from Exercise 8.5 and this exercise?

Table 8.11. Simulated hitting data for 15 players using a special baseball and
the usual baseball when there are ability situational effects
Usual Baseball Special Baseball
AB H AVG AB H AVG Special—Usual
First Larry 300 90 .300 300 91 .303
group Moe 300 77 .257 300 91 .303
Curly 300 99 .330 300 95 .317
Harold 300 88 .293 300 95 .317
Harvey 300 82 .273 300 106 .353
Second Gene 300 67 .223 300 73 .243
group Bill 300 86 .287 300 82 .273
Bob 300 93 .310 300 105 .350
Rick 300 92 .307 300 115 .383
Lynn 300 74 .247 300 95 .317
Third Jean 300 86 .287 300 82 .273
group Joe 300 93 .310 300 73 .243
Bret 300 84 .280 300 79 .263
Britt 300 79 .263 300 73 .243
Clark 300 86 .287 300 81 .270

8.7. In Exercise 8.1, we looked at home and away batting averages for 14 players during the
2013 season. Table 8.12 gives home and away averages for the same 14 players, but for
the following (2014) season.

Table 8.12. Home and away batting averages for 14 randomly selected players
from the 2014 season
Home Away
Name AB H AVG AB H AVG Home—Away
David Ortiz 250 67 .268 268 69 .257
Daniel Murphy 276 69 .250 320 103 .322
Evan Longoria 303 79 .261 321 79 .246
Khris Davis 243 57 .235 258 65 .252
Christian Yelich 273 81 .297 309 84 .272
Marlon Byrd 299 78 .261 292 78 .267
Xander Bogaerts 255 66 .259 283 63 .223
Marcell Ozuna 278 80 .288 287 72 .251
Todd Frazier 299 85 .284 298 78 .262
Matt Carpenter 287 68 .237 308 94 .305
Andrelton Simmons 261 69 .264 279 63 .226
Zack Cozart 255 60 .235 251 52 .207
Martin Prado 261 73 .280 275 78 .284
Jacoby Ellsbury 264 80 .303 311 76 .244

(a) Compute the differences in batting averages AVG(home) AVG(away) and place
your work in Table 8.12.
(b) Construct a stemplot of the batting average differences. Based on this graph, would
you say that there is a general tendency for players to bat better at home? Why or
why not?
(c) Suppose that some batters hit unusually well at home, and other batters actually hit
better on the road. If players did have different abilities to bat at home games versus
away games, then one would expect to see a relationship between the batting average
difference (Home Away) for 2013 and the batting average difference for 2013.
Using the data from Table 8.6 and Table 8.12, construct a scatterplot of the 2013 and
2014 differences.
(d) Based on your scatterplot drawn in (c), do you see a relationship between a player’s
batting average difference for 2013 and 2014? Interpret what this means in terms of
one’s ability to hit well at home versus away games.
8.8. Situational data are also recorded for pitchers. Table 8.13 displays pitching statistics for
20 pitchers for the three-year period 1997–1999. For each pitcher, the table gives
Throws: the throwing hand of the pitcher,
Batting Average—Left: the batting average of left-handed hitters against the pitcher,
Batting Average—Right: the batting average of right-handed hitters against the pitcher,
Batting Average—Pitch 1–15: the batting average of the hitters on pitches 1–15 during
a game,
Batting Average—Pitch 46–60: the batting average of the hitters on pitches 46–60
during a game,
Batting Average—Pitch 91–105: the batting average of the hitters on pitches 91–105
during a game.

(a) If a pitcher is right-handed, then who will be a more effective hitter a left-hander or
a right-hander? Why?
(b) If a pitcher is left-handed, then who will be a more effective hitter a left-hander or a
right-hander?
(c) Define an opposite-side hitter as one who bats at the opposite side from the arm of
the pitcher. (Likewise, define a same-side hitter as a hitter who bats from the same
side as the arm of the pitcher.) For each pitcher, compute the difference

DIFF1 D AVG(opposite-side hitter) AVG(same-side hitter).

Put your batting average differences in the “DIFF1” column of Table 8.13.
(d) Graph the differences and find the average value. Conclude what you have learned
about the effectiveness of opposite-side hitters compared to same-side hitters.
8.9. (Exercise 8.8 continued.) Table 8.13 also shows the batting average of hitters during
different pitch counts of each pitcher.
(a) How do you expect a starting pitcher to perform at the beginning of a game? Do
you think pitchers need to warm-up and don’t reach full effectiveness until they have
pitched a few innings?

Table 8.13. Situational statistics for 20 pitchers for the three-year period 1997–1999
Batting Batting Avg.
Average DIFF1 Pitch
Pitcher Throws Left Right Opp—Same 1–15 46–60 91–105 DIFF2
John Smoltz Right .252 .227 .293 .216 .260
Alan Ashby Right .271 .250 .266 .279 .284
Alex Fernandez Right .284 .203 .283 .212 .208
Andy Pettite Right .295 .266 .255 .287 .302
Bartolo Colon Right .267 .248 .248 .282 .219
Chuck Finley Left .236 .249 .226 .294 .254
Charles Nagy Right .291 .291 .294 .298 .272
David Cone Right .237 .219 .212 .260 .273
Doug Drabek Left .282 .292 .260 .261 .254
Darryl Kile Right .284 .251 .299 .296 .267
Greg Maddux Right .244 .254 .284 .245 .216
Kevin Brown Right .249 .216 .259 .264 .221
Randy Johnson Left .178 .213 .216 .221 .258
Curt Schilling Right .244 .220 .250 .210 .232
Tom Glavine Left .250 .251 .283 .256 .224
Terry Mulholland Left .273 .270 .253 .269 .266
Pedro Martinez Right .210 .193 .171 .204 .221
Todd Stottlemyre Right .279 .216 .259 .242 .300
Wilson Alvarez Right .258 .238 .291 .244 .224
Mike Mussina Left .244 .251 .257 .246 .238

(b) How do you expect a starting pitcher to perform at the end of the game after he has
thrown 90 pitches? Is fatigue a factor for a starting pitcher?
(c) If a starting pitcher gets tired, what effect would that have on the batting average of
opposing hitters?
(d) To see how a pitcher performs in the middle of the game as opposed to the beginning,
we can compute the difference

DIFF2 D AVG(pitches 46–60) AVG(pitchers 1–15).

For each pitcher, compute this difference and put the values in the “DIFF2” column
of Table 8.13.
(e) Graph the differences you computed in part (d) and compute an average value. Can
you detect a general tendency for pitchers to tire out between the beginning and
middle parts of the game?
8.10. (Exercise 8.8 continued.)
If one thought that pitchers generally tire out during a game, then one would expect
the opposing batting average to be larger for pitch count 91–105 than for pitch count
46–60.
(a) In Table 8.13, count the number of pitchers who pitch better on pitches 91–105
(compared to pitches 46–60), and count the number of pitchers who do better on
pitches 46–60 put the counts in the table below. Next, find the proportion in each
group and put the values in the “Proportion” column

Count Proportion
Pitchers who do better on pitches 91–105
Pitchers who do better on pitches 46–60

(b) What have you learned from the proportions you computed in (a)?
(c) Suppose that you now look at how batters perform during different pitch counts.
If batters generally have a smaller AVG for pitches 91–105, does that mean that
pitchers tend to get stronger during a baseball game? Explain why this might be a
wrong conclusion.
8.11. (Exercise 5-1 continued.) Table 8.14 shows Ichiro Suzuki’s daily batting record for the
first 20 games in the 2004 baseball season. To see how Suzuki’s performs in short periods,
one can compute moving averages of five games. For example, in games 1 through
5, Table 8.14 shows that Suzuki had seven hits in 22 at-bats for a batting average of
7=22 D :318. In the next group of five games (2 through 6), Suzuki was 8 for 23 for a
batting average of 8=23 D :348.
(a) For each group of five games, find the number of at-bats, hits, and five-game moving
batting average. Put the answers in the table.
(b) Graph the moving averages against the middle-game number on the grid below.
(For example, you would graph the first moving average .318 against the mid-
game number 3, the moving average .348 against the mid-game number 4, and
so on.)

Table 8.14. Computation of moving averages for Ichiro Suzuki’s 2004 daily batting record
Game AB H 5 Games 5-Game AB 5-Game H Moving AVG
1 4 1
2 5 1
3 4 0 1 through 5 22 7 7=22 D :318
4 5 3 2 through 6 23 8 8=23 D :348
5 4 2 3 through 7
6 5 2 4 through 8
7 5 2 5 through 9
8 5 2 6 through 10
9 5 0 7 through 11
10 5 0 8 through 12
11 4 1 9 through 13
12 4 3 10 through 14
13 6 2 11 through 15
14 3 0 12 through 16
15 3 1 13 through 17
16 4 0 14 through 18
17 4 2 15 through 19
18 4 1 16 through 20
19 5 1
20 4 1

(c) Describe the pattern that you see in the moving average plot. Are there any periods
that Suzuki appears to be hot or cold?
0.5
0.4
Moving Average
0.3
0.2
0.1
0.0

0 5 10 15 20
Game

8.12. (Exercise 8.11 continued.) Table 8.15 shows Suzuki’s day-to-day batting performance for
the first 20 games of the 2004 season. Suzuki’s batting average for the whole season

Table 8.15. Ichiro Suzuki’s 2004 daily batting record for the first 20 games
Hot or Hot or
Game AB H Game AVG Cold Game AB H Game AVG Cold
1 4 1 1=4 D :250 Cold 11 4 1
2 5 1 1=5 D :200 Cold 12 4 3
3 4 0 13 6 2
4 5 3 14 3 0
5 4 2 15 3 1
6 5 2 16 4 0
7 5 2 17 4 2
8 5 2 18 4 1
9 5 0 19 5 1
10 5 0 20 4 1

was .372. We say that Suzuki is hot for a particular game if his game batting average
is over .372; otherwise we say that Suzuki is cold. For example, on game 1, his game
average was 1=4 D :250. Since this is smaler than .300, we say that he was hot in
game 1.
(a) Complete Table 8.15. For each game, compute the game batting average and classify
the game as hot or cold.
(b) From the sequence of Hot and Cold games, find
i. the longest run of Hot games,
ii. the longest run of Cold games,
iii. the total number of runs.
8.13. (Exercise 8.11 continued.) For the game-to-game batting data for Ichiro Suzuki, count the
number of games that Suzuki had 0 hits, 1 hit, etc. Put your counts in the table below.

Number of hits 0 1 2 3
Count

8.14. Table 8.16 shows the game results for the New York Yankees and the Atlanta Braves for
the first thirty games of the 2015 season.

Table 8.16. Game results for New York Yankees and Atlanta Braves for first thirty
games of the 2015 season
Yankees
Games 1–15: L W L
L L W W L L W W W L W W
Games 16–30: W W L
W W W L W W W L W L W W
Braves
Games 1–15: W W W W W L W L L W L W L L L
Games 16–30: L W L W L L L W L W L W W L L

(a) Compute all moving fractions of wins using a width of 9 games for both teams. That
is, find the proportion of wins in games 1–9, in games 2–10, . . . , in games 22–30.
Put your results in the table below

Yankees Braves Yankees Braves

Mid Fraction Fraction Mid Fraction Fraction
Games game of wins of wins Games game of wins of wns
1–9 5 12–20 16
2–10 6 13–21 17
3–11 7 14–22 18
4–12 8 15–23 19
5–13 9 16–24 20
6–14 10 17–25 21
7–15 11 18–26 22
8–16 12 19–27 23
9–17 13 20–28 24
10–18 14 21–29 25
11–19 15 22–30 26

(b) Graph the moving fractions against the mid-game for the two teams on the axes
below.
1.0
0.8
Fraction of Wins
0.6
0.4
0.2
0.0

5 10 15 20 25
Mid-Game

(c) Comment on the pattern of moving-fractions that you see in the graph. Did the
Yankees or Braves seem unusually hot or cold during these first 30 games?

8.15. (Exercise 8.14 continued.) For each of the win/loss sequences of the Yankees and Braves
shown in Table 8.16, find (1) the longest run of wins or losses, and (2) the total number
of runs. Put your answers in the table below.

Team Longest run of wins or losses Number of runs

Yankees
Braves

8.16. (Exercise 8.12 continued.) Suppose Ichiro Suzuki is truly a consistent hitter. The chance
that he gets a hit on a single at-bat is .372 (his true batting average), and the results
of different at-bats are independent. Suppose we simulate data from this model using
Suzuki’s actual number of at-bats for the 20 games. The results of this simulation are
placed in Table 8.17.

Table 8.17. Simulated data for 20 games assuming Suzuki is a consistent hitter with
hit probability of .372.
Hot Hot
Game AB H Game AVG or Cold Game AB H Game AVG or Cold
1 4 2 2=4 D :500 Hot 11 4 2
2 5 1 12 4 0
3 4 1 13 6 1
4 5 2 14 3 1
5 4 2 15 3 1
6 5 2 16 4 2
7 5 1 17 4 1
8 5 1 18 4 2
9 5 3 19 5 1
10 5 3 20 4 3

(a) As in Exercise 8.12, classify the hitting of each game as either hot or cold, depending
if the game batting average is larger or smaller than .372. Put your results in the table.
(b) Compute
i. the longest run of Hot games,
ii. the longest run of Cold games,
iii. the total number of runs.
(c) Compare the results of these simulated data (the longest run of Hot, the longest run
of Cold, and the total number of runs) with the results using Suzuki’s actual data in
Table 8.15. Is there any evidence that Suzuki is a streaky hitter?
8.17. (Exercise 8.16 continued.) In Exercise 8.16, we assumed that Suzuki was really a consistent
hitter with probability of .372 of getting a hit, and simulated 20 games of hitting assuming
this consistent model. Suppose we repeat this experiment 20 times. For each experiment,
we simulate 20 games of hitting (assuming Suzuki is really a consistent hitter), and then
classify each game as either hot or cold. The results are shown in Table 8.18.
(a) For each experiment, find the longest run of either Hot or Cold. Put the longest run
in the LONGEST RUN column of the table.

Table 8.18. Simulated data for twenty experiments, where one experiment consists of 20
games assuming Suzuki is a consistent hitter with hit probability of .372. Suzuki’s
performance in each game is classified as hot (H) or cold (C)
Game Longest
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 run
C H H H H H H C H H H H C H H H H H C H
C H H C C H C H C C C H H C H H C C C H
H H C C C H H H C H C C C C C H H H C H
H H H H C H C C C C H C C C C H C C H C
H C H H H C C H H C H C C H C C C H H H
C C H C C H C H C H H H H C C H C H H H
C H C C H C H C C H H C C H C C H H H H
H C C H C C C H H C C C H C H C H C H H
C H H C H H H H C C H C H C C H H C H C
H H H C C H H C H C H C H C C H H H H H
C H C H C C H H C H H C H C C C C H C H
H C C H H C H H H H C C C C C C H C H C
C H H H H H H H H C C H C C H H H C H C
H H C H H H C C H H H C C C C H H H C H
H C C H H C H H C C C H H C H H H H C C
H C H C H C H H H H C C H H C C H C C C
C C H H C H C H C H C H C C C H H C H H
C C C H H H C H C H H H H C C C C C H C
H H C H H C C C H H H H C C C H H H C C
H H H C H H H H H C H C H C H C H H H C

(b) Construct a dotplot of the longest run values using the number line below.

+---+---+---+---+---+---+---+---+---+---+--
0 1 2 3 4 5 6 7 8 9 10 LONGEST RUN

(c) For Suzuki’s data in Table 8.15, the longest run (of either Hot or Cold) was equal to
. Does this value seem unusually small or large if Suzuki was really a consistent
hitter? (Compare Suzuki’s value with the longest run values plotted in part (b).)
8.18. Was Chris Davis a streaky home run hitter in 2013? Using Davis’ home run log from Case
Study 5-1, one can construct a moving average plot of his home run rate. The moving
averages are shown in Figure 8.12 using a window of ten games. We see that Davis showed
some streaky behavior—his 10-game home run rate was over .15 about games 50 and 70
and his home run rate dropped to 0 about game 100.
How would Davis perform if he were truly a consistent home run hitter? In 2013,
Davis hit 53 home runs in 673 plate appearances for a season rate of 53=673 D :0789.
Suppose that the probability that Davis hits a home run in a single plate appearance
is p D :0789, and the results of different plate appearances are independent. Using
this consistent model, four seasons of home run hitting were simulated using the same

0.15
0.10
Average
0.05
0.00

0 50 100 150
Game

Figure 8.12. Moving average plot of 2013 Chris Davis’ home run rates using a window of ten games.

game-to-game plate appearances as Davis. Moving average plots of the home run rates
for the four simulations are shown in Figure 8.13.
Discuss any unusual features of each of these four simulations. Do you think that
Davis’ moving average plot is different, with respect to streakiness, from the plots of the
four simulated consistent hitters? Do you think that Davis was a truly streaky home run
hitter?

Simulation 1 Simulation 2

0.20

0.15

0.10

0.05
Average

0.00
Simulation 3 Simulation 4

0.20

0.15

0.10

0.05

0.00
0 50 100 150 0 50 100 150
Game

Figure 8.13. Moving average plot of home run rates of simulated data using a consistent model with
home run probability .0789.

8.19. On the baseball-reference.com web site, one can explore the patterns of hot streaks
and slumps of any Major League baseball team in history. (Look for the streaks link on
the web site.) The Oakland Athletics had an interesting pattern of wins and losses during
the 2002 season. Figure 8.14 displays a moving average of the winning proportion using
a window of 20 games.
(a) Looking at Figure 8.14, describe the slumps and hot streaks of Oakland during this
season.
(b) Oakland concluded this season with a 103-59 win/loss record with a winning per-
centage of 63.6%. Simulate sequences of wins and losses for an entire 162 game
season assuming Oakland is a truly consistent team with probability of winning each
game equal to .636. By comparing the simulated sequences with Figure 8.14, can
you say that Oakland was truly a streaky team in the 2002 season?
1.0
0.8
Winning Fraction
0.6
0.4

20 40 60 80 100 120 140

Game

Figure 8.14. Moving average plot of winning fraction of the 2002 Oakland Athletics using a window of
20 games.

(c) Choose another team in history that had a reputation for being unusually streaky
during a particular season. By using the simulation method described in part (b),
decide if the pattern of wins and losses for this team was different from that of a truly
consistent team.

Further Reading
Situational baseball data for batters and pitchers is presented in many internet sites such as
FanGraphs. Chapter 4 of Albert and Bennett (2003) describes the different probability models
that can be used for situational hitting data. Based on an analysis of situational data for the 1998
season, they classify the different true situational effects as “no effects”, biases, or ability effects.
Gilovich et al (1985) describe the tendency of people to misinterpret the inherent streaky nature
of random sequences. Albert and Bennett (2003), Chapter 5 discuss hitters who are genuinely
consistent or streaky, and describe how one can perform inference about streakiness.

What’s On Deck?
In this chapter, we introduce a special probability model, called a Markov Chain, to represent
the sequence of plays in a baseball game. The state of an inning is defined by the number of
outs and the runners on the three bases. A half-inning of baseball can be regarded as a sequence
of states until there are three outs. One can represent the random movement between states by
a Markov Chain, where one moves from one state to another state with a given probability. In
Case Study 9.1, we introduce a Markov Chain using an example of a person traveling between
cities, and in Case Study 9.2 we extend the basic structure of a Markov Chain to a baseball
game. To specify a Markov Chain, one needs only to specify the probabilities that control the
movement between states, and we use actual baseball game data to estimate these probabilities.
The remainder of the chapter shows the usefulness of this probability model to answer questions
about baseball. Using the model, one can estimate the number of players that come to bat during
the inning, and compute the probability that a team will score at least one run in an inning.
One of the most useful computations is the expected number of runs scored from a particular
state in an inning. Using these expected numbers of runs scored, one can assess the value of
a particular hit, such as a home run. In addition, one can use these expected numbers of runs
scored to judge the value of a particular batting play. Baseball fans are interested in the value of
particular baseball strategies, such as a sacrifice bunt and a steal, and this probability model is
useful in seeing if a particular strategy is helpful towards the general goal of scoring runs.

9.1 Introduction to a Markov Chain

Topics Covered: Transition probabilities, absorbing states, expected number of visits to different
states, matrices, matrix multiplication and inversion.
In this chapter, we will show how a baseball game can be modeled using a special probability
model called a Markov Chain. In this first case study, we give a gentle introduction to Markov
Chains and the remainder of the case studies will show the application of this idea to baseball.
Suppose a baseball fan is planning an interesting (one might call it bizarre) trip to New
York City. He will start from San Francisco (SF). The next day, he will either stay in SF another
night or he will fly to St. Louis (STL); he is equally likely to stay or fly to STL. If he does
go to STL, then he will stay in STL another night with probability .4, or fly back to SF with
probability .3, or fly to Chicago (CHI) with probability .3. If he is in CHI, the next day he is

211
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
212 Modeling Baseball Using a Markov Chain

equally likely to stay another night, fly to SF, fly to STL, or fly to New York City (NY). Once
he arrives at New York City, he will stay there—his trip is over.
Obviously the exact trip of this fan is random. All he knows for sure is that he will eventually
arrive in New York City. But what is the probability that he will arrive in NY in exactly two
days? In three days? How long will it take this fan, on average, to get to NY? And how many
days can this fan expect to stay in SF, STL, and CHI?
The map in Figure 9.1 shows the possible city-to-city connections of our traveler. An arrow
from one city to another city indicates the particular trip is possible and the number label is the
probability of this trip. An arrow from a city back to the same city corresponds to a night where
the traveler stays over.

.25 New York

.25
.25 1
.5 Chicago
.5 .25
.3
San Francisco St. Louis

.3 .4

Figure 9.1. City-to-city connections of a traveler in trip from San Francisco to New York City.

A Markov Chain describes movement between a number of locations, called states. Here
there are four states for our traveler, SF, STL, CHI, and NY. Given that the traveler is in a
particular state (city), then we will move to another state the next day with specific probabilities.
The probability that he will move to another city depends only on his current location and not
on the previous cities he visited. (This is a special property of a Markov Chain.) Given the
information above, we know if the person is currently in SF, then he will travel to

SF STL CHI NY

with respective probabilities

.5 .5 0 0.

If he is currently in STL, then he travels to these four cities the next day with respective
probabilities

.3 .4 .3 0.

These probabilities are called transition probabilities—they describe the likelihoods of moving
between various states in the Markov Chain. We summarize all of these transition probabilities
by means of a transition matrix P shown in Table 9.1.

Table 9.1. Matrix of transition probabilities P for the traveler example

SF STL CHI NY
SF 0.50 0.50 0.00 0.00
STL 0.30 0.40 0.30 0.00
CHI 0.25 0.25 0.25 0.25
NY 0.00 0.00 0.00 1.00

The first row of this matrix P gives the transition probabilities of the traveler starting from
SF, the second row gives the probabilities starting from STL, and so on.
One special feature of this particular Markov Chain is that once the traveler arrives at New
York City, he will remain there. We call the state NY an absorbing state—as indicated in the
transition matrix, the probability of remaining in an absorbing state is one.
There are a number of nice results about Markov Chains that will make it easy to answer
the questions posed above.

What is the Probability of Reaching States After a Specific Number of Moves?

The matrix P gives the probabilities of reaching various states in exactly one move. We can
obtain the probabilities of reaching states in two moves by squaring the matrix P :
2 3 2 3
:5 :5 0 0 :5 :5 0 0
6 :3 :4 :3 0 7 6 :3 :4 :3 0 7
P DP P D6
4 :25
76 7
:25 :25 :25 5 4 :25 :25 :25 :25 5
0 0 0 1 0 0 0 1
2 3
:4 :45 :15 0
6 :345 :385 :195 :075 7
D6
4 :2625
7
:2875 :1375 ::3125 5
0 0 0 1

The first row of this matrix, Œ:4; :45; :15; 0, gives the probability that, starting at SF, we will
be in the respective cities SF, STL, CHI, NY in exactly two days. So he will be in Chicago in
two days with probability .15. Note that the probability that he’ll be in NY in two days is 0.
If we multiply the matrix P additional times, we obtain the probabilities of being in the
states after more than two days. If we multiply

P 6 D P P P P P P;

we obtain the six-step transition probabilities

2 3
:323 :361 :154 :163
6 :293 :328 :140 :240 7
P6 D 6
4 :219
7
:244 :104 :433 5
0 0 0 1

The first row gives the probabilities of being in the four states, starting in SF, in exactly six days.
We see that if we start in SF, the chance that we will be in NY in six days is .163.

How Long Will One Stay in the States?

If we remove the row and column corresponding to the single absorbing state from the matrix
P , we have the matrix Q of transition probabilities
2 3
0:5 0:5 :0
4
Q D 0:3 0:4 0:3 5
0:25 0:25 0:25

To find the expected number of times that the process visits the states before being absorbed,
we simply compute the matrix E D .I Q/1 , where I is the 3 by 3 identity matrix. Here
0 2 311 2 3
0:5 0:5 :0 10 10 4
@ 4
E D I 0:3 0:4 0:3 5 A 4
D 8 10 45
0:25 0:25 0:25 6 6:67 4

Let’s interpret this matrix. The first row, Œ 10 10 4 , gives the expected number of visits to
the cities SF, STL, and CHI if we start at San Francisco. So in the trip, the person will be in
SF, on average, ten days (including the starting day), STL for ten days, and CHI for four days.
Likewise, the second and third rows of the matrix E give the expected number of visits if we
start in St Louis and Chicago, respectively.
The matrix E tells us how the traveler will do on average. But actually there is a lot of
variation in the length of the trip. To see this, I had the computer simulate 1000 trips originating
from San Francisco using the matrix of transition probabilities. Figure 9.2 is a histogram of the
lengths (in days) of the 1000 trips. The distribution is right-skewed where the length ranged
from three to 142 days. The mean length of a trip was 23.31 days. This is reasonable—looking
at the first row of the matrix E, we see that we will spend a total of 10 C 10 C 4 D 24 days in
our journey.
300
250
200

Mean number of days = 23.31

Count
150
100
50
0

0 50 100 150
Number of Days

Figure 9.2. Histogram of lengths of 1000 trips from San Francisco simulated from the Markov Chain.

How Many Frequent Flyer Miles?

Suppose our traveler plans on flying and he is interested in how many frequent flyer miles he
will accumulate in his trip. First we record in a table the distances between all of these cities as
displayed in Table 9.2.

Table 9.2. Distances between all cities in the traveler example

SF STL CHI NY
SF 0 2062 2133 2908
STL 2062 0 294 953
CHI 2133 294 0 790
NY 2908 953 790 0

Then we find the expected length of a one-day trip from each city. If we start in SF, then we
will travel 0, 2062, 2133, 2908 miles with respective probabilities 0.5, 0.5, 0, 0. So the expected
length of a one-day trip from SF is
0 0:5 C 2062 0:5 C 2133 0 C 2908 0 D 1031:
Similarly, we find the expected length of a one-day trip from STL and CHI to be 709.8 and
804.3 miles, respectively.
Last, using a standard result for Markov Chains, we can compute the expected length of
the trip from each starting city by multiplying our matrix E by the column of expected length
of one-day trips.
2 3 2 3 2 3
10 10 4 1031 20; 625
4 8 10 4 5 4 709:8 5 D 4 18; 563 5
6 6:67 4 804:3 14; 135
The product vector gives the expected length in miles of the total trip starting from each of the
three cities. So if we start at San Francisco, we can expect to log 20,625 frequent flyer miles
before getting to New York City. We will travel, on average, 18,563 miles if we start our trip at
St. Louis.

9.2 A Half-inning of Baseball as a Markov Chain

Topics Covered: State of an inning in a baseball game, transitions between states, batting plays,
runs scored.
In the first case study, we described the basic structure of a Markov Chain. How can this
model be applied to baseball?
We focus on the run production of a baseball team during its half-inning at-bat. A state will
describe the current runners on bases situation and the number of outs. Each base (first, second,
and third) can either be occupied or not. There are a total of eight possible base situations that
will be graphically represented in Table 9.3 using symbols on a diamond shape.

Table 9.3. Possible runner situations in a baseball game

Empty 1st 2nd 3rd 1st, 2nd 1st, 3rd 2nd, 3rd Loaded

When a batter comes up, the number of outs at any one time during an inning can either
be 0, 1, or 2. So when a batter comes to bat, there are eight possible base situations and three
different out situations, and so there are 8 3 D 24 different base/out situations. If we add the
final situation, three outs, to this list, we have a total of 24 + 1 =25 possible states. In Table 9.4
all of the states are presented by use of a table classified by number of outs and the base
situation.

Table 9.4. Diagram of all 25 possible states of an inning defined by the runners on base
situation and the number of outs
Bases Situation

0 outs
1 out
2 outs
3 Outs

A player comes to bat when the inning is in a particular state, say runners on 1st and 2nd
with one out. There will be a batting play, such as a hit, out, or walk, that will change the inning
state. For example, suppose a batter comes up with a runner on 1st and one out

( , 1 out)

He singles into right field, moving the runner from 1st to 3rd. The new inning state is

( , 1 out)

We can view a half-inning of baseball as a sequence of changes in state, starting with (no runners
on base, 0 out) and ending with (3 outs). (To simplify the following discussion, we will ignore
non-batting plays, such as steals and balks, that can also change the state of an inning. A more
realistic model would include these non-batting plays in the analysis.)
A Markov Chain can be used to model the change in states during an inning. We assume
that the probability of moving to a particular (bases, outs) state only depends on the current state
and not on any earlier state. We let pij denote the probability of moving from the i th state to the
j th state. We let P denote the matrix of transition probabilities with 25 rows and 25 columns.
Note that the (3 outs) situation is an absorbing state in the Markov Chain since one cannot leave
this state once it is entered. In other words, the inning is over when there are three outs.
When there is a change in the (bases, outs) state, runs can score. Suppose that there are
Rbefore runners on base and Obefore outs before the batting play, and Rafter runners and Oafter outs
after the play. Then the number of runs scored in this transition would be

Runs Scored D .Rbefore C Obefore C 1/ .Rafter C Oafter /:

Let’s illustrate this computation for one batting play. Suppose there are runners on 1st and
2nd with one out. The batter doubles, scoring both runners, leaving a runner on 2nd with one
out. Table 9.5 verifies that the Runs Scored formula gives that two runs scored on this play.

Table 9.5. Illustration of the number of runs scored for one batting play
Situation Situation Runs
before Play .Rbefore ; Obefore / Play after Play .Rafter ; Oafter / Scored
, 1 out (2, 1) Double , 1 out (1, 1) 2

9.3 Useful Markov Chain Calculations

Topics Covered: Computation of probabilities by use of frequencies, matrix computations,
expected number of visits to states, expected runs scored in a half-inning, computation of event
probabilities by simulation.
To use this Markov Chain model, we need to estimate the matrix of transition probabilities
P . We use play-by-play data for all the Major League teams in the 2014 season to estimate this
matrix.
Let’s illustrate computing the transition probabilities starting from the bases empty, no outs
state. All half-innings begin with this particular state, although this state may happen more than
once in a particular inning. In the 2014 season, there were 45,312 instances where this state
occurred. There are five possible transitions from this state, depending on the batting play:

The batter can hit a home run, and the state remains at bases empty and no outs.
The batter can get out, and the state changes to bases empty and one out.
The batter can get to first base by a single, walk, hit-by-pitch, or an error and the state changes
to runner on first and no outs.
The batter can hit a double, and the state changes to runner on second with no outs.
The batter can hit a triple, and the state changes to runner on third base with no outs.

Table 9.6 below shows the five possible transitions from (bases empty, no outs) and the
frequency with which each transition occurred. Note that the most common transition is an out
this event happened 31,059 times out of a total of 45,312 transitions and so the probability of
this transition is estimated to be 31;059=45;312 D 0:6854. In a similar fashion, we can estimate
the probability of the four other possible transitions.

Table 9.6. All possible transitions from the no-runners, no outs state. For each state, the
number of runs scored, the number of times this transition occurred, and the corresponding
probability are given
Batting Play End State Runs Scored Count Probability
Home run , 0 out 1 1149 .0254
Out , 1 out 0 31059 .6854
Single or walk , 0 out 0 10663 .2353
Double , 0 out 0 2205 .0487
Triple , 0 out 0 236 .0052
Total 45312 1.000

Suppose instead that the current state is two outs with a runner on first. In this case, there
are more possible transitions types depending on the advancement of the runner from first base.
Table 9.7 shows all of the possible transitions from (runner on first, two outs), the frequency
of each transition, and the corresponding probability. We see that the most likely play is an out
that results in the third out the probability of this transition is 7988=11; 630 D 0:6868. Other
possible transitions, ordered in terms of their likelihood of occurring, are

a single, walk, or hit-by-pitch that result in runners on 1st and 2nd,

a single where the runner on first advances to 3rd base,
a home run that clears the bases,
a double that scores the baserunner on 1st,
a double that advances the baserunner to 3rd base,
a triple that scores the baserunner on 1st,
a single that scores the baserunner on 1st.

We see that the single scoring the runner on first is quite an unusual play—it happened only
10 times that particular season.

Table 9.7. All possible transitions from the (runner on first, two outs) state. For each state, the
number of runs scored, the number of times this transition occurred, and the corresponding
probability are given
Batting Play End State Runs Scored Count Probability
Home run , 2 outs 2 282 .0242
Triple , 2 outs 1 78 .0067
Double , 2 outs 1 255 .0219
Double , 2 outs 0 248 .0213
Single , 2 outs 1 10 .0009
Single , 2 outs 0 593 .0510
Single or Walk , 2 outs 0 2176 .1871
Out 3 outs 0 7988 .6868
Total 11630 1.000

Suppose that we compute the probabilities of all possible transitions starting from each
of the 24 possible initial states. We can then construct the matrix P (of dimension 25 by 25)
that contains the transition probabilities for all (runners on base, number of outs) states. Given
this transition matrix, we can perform several matrix calculations, such as done in the first case
study, to find some quantities of interest.

Reaching Various States After a Given Number of Batters

Suppose three players come to bat at the beginning of an inning. What will be the state of the
inning after these three plate appearances? What’s the chance that there will be at least one
runner in scoring position?

Remember the matrix P gives the probabilities of one-step transitions. To obtain the
probabilities of three-step transitions, we simply multiply the matrix by itself three times:
P 3 D P P P:
Table 9.8 gives the first row of the matrix, displaying it in a familiar two-way format where
rows correspond to the number of outs and the columns represent the runners situation.

Table 9.8. Transition probabilities of being in different states after three plays starting from the
(no outs, no runners) state
Bases Situation

0 outs .0022 .0030 .0021 .0005 .0074 .0041 .0030 .0109

1 out .0159 .0229 .0160 .0064 .0825 .0283 .0190 .0000
2 outs .0411 .2441 .0789 .0300 .0000 .0000 .0000 .0000
3 outs .3818

These probabilities represent the chances of being in different inning states after three
batters starting from the (no outs, no runners) state. We see that the most likely state after three
at-bats is (3 outs) with a probability of .3818. The next most likely state is (2 outs, runner on
first) with a probability of .2441. Several three-batter movements are impossible. For example,
we see from the table that the probability of getting to (2 outs, runners on first and second) from
three batters has a probability of 0.
What is the chance of having a runner in scoring position (that is, a runner on 2nd or 3rd base)
after three at-bats? We look at the above table and sum the probabilities over the , , , ,
, states where one runner or more are in scoring position. So the probability of runners
in scoring position after 3 batters is equal to :0021 C :0064 C C :0000 C :0000 D :2891.

How Many Batters?

As before, let Q denote the submatrix of P found by deleting the one row and one column
corresponding to the absorbing (3 outs) state. The matrix E D .I Q/1 contains the expected
times that the inning will be in each state starting with each of the 24 possible beginning states.
Suppose that we are currently at the (no runners on, no outs) state that begins the inning. The first
row of the matrix E will give the expected number of visits to all the states (before absorption)
given that one starts in this (no runners on, no outs) state. This first row of E is displayed in a
convenient table format in Table 9.9.
Table 9.9. The expected number of visits to each state starting from the (no runners on, no
outs) state
Bases Situation

0 outs 1.037 0.249 0.059 0.008 0.065 0.021 0.013 0.016

1 out 0.745 0.302 0.104 0.030 0.113 0.047 0.031 0.039
2 outs 0.583 0.311 0.127 0.051 0.141 0.066 0.033 0.045
SUM 4.236

Of course, since we are starting at the (no runners on, no outs) state, we’ll visit this state at
least once this table tells us that we will visit it, on the average, 1.037 times. Also, we will visit
the (runner on first, no outs) state an average of .249 times, the (runner on second, no outs) state
an average of .059 times, and so on.
If we sum all of these expected state counts, we will obtain the expected number of state
visits before absorption. In baseball lingo, this sum will be the expected number of batters
before the inning is over. Here the sum is 4.236, which means that, on average, there will be
4.236 batters in the remainder of the inning starting with the (no runners on, no outs) state.
In a similar fashion, one can use the matrix E to compute the expected number of batters
in the remainder of the inning starting from each of the 24 possible states. Table 9.10 shows the
“expected number of batters” matrix.

Table 9.10. Expected number of batters in the remainder of the inning starting from each
possible state
Bases Situation

0 outs 4.236 4.022 4.316 4.317 3.994 4.105 4.370 4.010

1 out 2.844 2.683 2.910 2.982 2.648 2.687 3.007 2.662
2 outs 1.446 1.447 1.497 1.513 1.407 1.424 1.494 1.389

A manager can use this matrix in strategic decisions during a game. For example, suppose
there are bases loaded with one out. Looking at this table, we see that, on average, there will be
2.662 batters in the remainder of this inning. A manager can use this to plan his batting lineup;
for example, it might help him make a decision regarding the use of a pinch-hitter.

Expected Runs in the Remainder of the Inning

The expected number of visits matrix E can be used to compute the expected number of runs
in the remainder of the inning starting from each state. This computation is similar to the
computation for the expected number of frequent flyer miles for our first example. Let Rone step
denote the column vector that contains the expected number of runs that will be scored in a
single batting play starting from each of the 24 states. Let’s illustrate the computation of the first
element of Rone step . Suppose that one starts in the (no runners on, no outs) state. In one batting
play, only 0 and 1 runs can score. The probability of scoring one run is the chance of hitting a
home run that is estimated to be .0254, and so the chance of scoring no runs is estimated to be
1 :0254 D :9746. The expected number of runs scored in one batting play is therefore

1 :0254 C 0 :946 D :0254:

If we do this computation for each of the 24 states, we obtain the vector Rone step that is displayed
in matrix form in Table 9.11.
The vector of expected number of runs scored, denoted by R, is found by multiplying the
number of visits matrix E by Rone step :

R D E Rone step :

Table 9.11. The expected number of runs scored in a single batting play starting at each
possible state. The vector contains these values
Bases Situation

0 outs .025 .072 .149 .491 .201 .713 .603 .843

1 out .021 .076 .170 .476 .248 .543 .639 .814
2 outs .022 .078 .175 .223 .245 .271 .317 .505

Note that R is a column vector with 24 entries, where the entries correspond to the expected
number of runs in the remainder of the inning starting from each of the 24 states. This vector R
is displayed, in matrix form, in Table 9.12.

Table 9.12. The expected number of runs scored in the remainder of the inning starting at each
possible state
Bases Situation

0 outs 0.44 0.81 1.06 1.29 1.36 1.73 1.90 2.21

1 out 0.22 0.46 0.61 0.90 0.83 1.06 1.33 1.49
2 outs 0.08 0.19 0.28 0.33 0.38 0.40 0.47 0.63

This fundamental matrix, often called the “run expectancy matrix”, is useful for many
purposes in baseball research. Starting an inning in the (no runners, no outs) state, we see from
Table 9.12 that a team will score, on average, .44 runs. In contrast, when there are bases loaded
with one out, there is a good potential to score runs; using the table, we see that, on average,
1.49 runs will score from this state. This run expectancy matrix gives the potential for a team
scoring runs starting from each (bases, outs) situation.

Computing Other Event Probabilities

Once the Markov Chain model is defined by means of the transition probability matrix P , one
can compute the probability of any event of interest by simulating the chain many times. For
example, suppose a manager is interested in the probability that the team will score at least one
run if there is a runner on 3rd base with one out. On a computer, one can simulate the batting
plays in the remainder of the inning by using the transition probability matrix, starting with the
row of the matrix corresponding to the (3rd base, one out) state, and stopping when the (3 outs)
state has been reached. One records if one or more runs were scored in this simulated inning.
Then one repeats this simulation process for a large number of innings, each time recording if
one or more runs occurred. Then the probability of scoring at least one run, Prob(at least one
run), can be approximated by
# of innings where at least one run was scored
:
# of innings simulated
We actually simulated the Markov Chain starting from each of the 24 possible inning states.
For each starting state, we simulated the remainder of the inning 10,000 times. From each state,

we computed the probability that at least one run is scored in the remainder of the innings. The
estimated probabilities are displayed in Table 9.13.

Table 9.13. The probability of scoring at least one run in the remainder of the inning starting
from each possible state
Bases Situation

0 outs .246 .391 .595 .827 .589 .865 .852 .842

1 out .138 .236 .380 .635 .386 .605 .674 .643
2 outs .059 .101 .196 .244 .206 .241 .217 .273

When the inning starts (no outs, no runners on), the probability the team will score at least
one run is .246. This chance of scoring drops to .059 when there are two outs in the inning.
A team is most likely to score when the bases are loaded with no outs—the chance of scoring
is .842.

9.4 The Value of Different On-base Events

Topics Covered: Value of a batting event defined by expected runs, mean value of a particular
type of hit.
Using the run expectancy matrix R defined in the previous case study, we can define the
value of a batting event. Suppose a batter comes to bat in a particular state (runners on base and
number of outs) where the expected number of runs scored in the remainder of the inning is
Rbefore . After the batter event, there is a new (runners on base, outs) situation with an expected
number of runs scored called Rafter . Then the value of this batting event is
VALUE Rafter Rbefore C (runs scored on play).

Values of a Terrible Play and a Great Play

Let us illustrate this formula for two extreme cases that correspond to the least and most valuable
batting plays. Suppose that the bases are loaded with no outs. The batter hits a sharp grounder
to third; the third baseman touches third base, throws it quickly to the second baseman who
touches second base and throws it to first, completing an (unusual) triple play. Clearly this is a
bad play for the hitter (and the team)—the question is how bad?
When the batter came to the plate, the run potential (bases loaded, no outs) is 2.21 runs.
After the play, there are three outs and a run potential of zero. No runs scored on this play. The
value of this plate appearance is
VALUE D .0 2:21/ C 0 D 2:21:
So this bad hitting play essentially cost the team about two runs. This is the worst possible play
where the worth is defined in terms of this value measure.
Let’s contrast this with a great batting play. The bases are loaded with two outs and the
batter hits a deep fly that goes over the center field fence it’s a grand slam! When this batter
came to bat, the run potential of (bases loaded, two outs) is .63 runs. After the play, the bases are
empty (with 2 outs) and the run potential is .08 runs. Four runs scored on this play. The value

of this plate appearance is

VALUE D .:08 :63/ C 4 D 3:45:

One might think the value of this home run would be four runs after all, four runs scored on
this play. But, by clearing the bases, the batter has decreased the run potential in the remainder
of the inning, and the value measure adjusts for this decrease.

The Value of a Home Run

Using the notion of value, we can measure the effectiveness of different types of batting plays,
such as hits, walks, sacrifice flies, and outs. Here we focus on the biggest hit, the home run.
There were a total of 4186 home runs hit in the MLB in 2014. But these home runs were
not equally valuable—certainly a home run hit with runners on base is more valuable than a
home run hit with the bases empty. In Table 9.14, we classify all of the home runs by the bases
situation and the number of outs. In each cell, the number on top is the value of the home run
in that particular situation and the number in parentheses is the number of home runs hit in that
situation.
Table 9.14. Value and the number of home runs that occur in all possible situations using 2014
data
Bases Situation

0 outs 1 1.63 1.38 1.08 2.07 1.67 1.56 2.10

(1142) (257) (69) (9) (51) (27) (6) (13)
1 out 1 1.76 1.62 1.29 2.40 2.14 1.88 2.73
(687) (306) (135) (37) (111) (39) (23) (34)
2 outs 1 1.89 1.78 1.74 2.70 2.67 2.58 3.45
(561) (280) (114) (54) (117) (43) (34) (37)

Note that most of the home runs were hit with the bases empty in fact, 2390 (57%) of the
home runs hit were solo shots and the value of each of these home runs is 1. In contrast, only
13% of the home runs were hit with two or more runners on base. The values of these “two
runners or more on base” home runs vary from 1.56 (runners on 2nd and 3rd and no outs) to
3.45 (bases loaded with 2 outs).
To measure the value of a home run, we average all of the values for the 4186 home runs
hit in the 2014 season. To find this average, we first total all of the values of the home runs in
Table 9.14, and divide this total by the number of home runs. In Table 9.15, we find the total
value in each situation, and show the sum of these totals in the lower right cell of the table. The
average value of a home run is then equal to
1142 1 C 257 1:63 C C 37 3:45
Avg value of home run D
4186
5861:98
D D 1:40:
4186
This average value of a home run may seem small, since we credit a home run with four
bases in the computation of a slugging percentage. But this average of 1.40 runs represents a

typical run value for this type of hit. In the exercises, we will investigate the run values for other
types of batting plays.

9.5 Answering Questions About Baseball Strategy

Topics Covered: Evaluating the worth of a play by use of expected runs.
The 2001 World Series between the New York Yankees and the Arizona Diamondbacks is
considered one of the most exciting World Series in history. We focus on Game 4 that involved
repeated use of a well-known baseball strategy. The question we want to address is whether this
strategy really helps in scoring runs.
In the top of the first inning, the lead off hitter for Arizona, Tony Womack, singled to center.
Then Craig Counsell, the second batter, was instructed to hit a sacrifice bunt. The bunt was
effective—Counsell was thrown out at first and Womack advanced to second base. The inning
finished without Womack scoring. In the top of the third inning, the same situation happened.
Womack opened the inning with a walk and Counsell again sacrificed with a bunt to move
Womack to second. (Again Womack didn’t score.) In the top of the 5th inning, Womack started
the inning with a double. Counsell again hit a sacrifice bunt, moving Womack to third. The next
hitter, Luiz Gonzalez, hit a fly ball, but Womack was thrown out at home plate, ending the inning.
It appears that the Arizona manager, Bob Brenly, likes to play the sacrifice bunt. Counsell
was instructed to sacrifice his at-bat (and get an out) in order to advance the runner one additional
base. Is the sacrifice bunt a smart play in baseball?
We have the tools to evaluate the effectiveness of this play in creating runs for a team. As
in evaluating the home run, we use the run potential matrix that gives the average number of
runs scored in each bases/outs situation.

Sacrifice Bunt to Move Runner from First to Second

We first look at the situation that occurred in the first and third innings of the World Series game.
Arizona has a runner on first with no outs. Looking at our run expectancy matrix, this situation
has a potential of .81 runs. If Counsell hits a successful sacrifice bunt, the batter is out, but the
runner moves to second base. The run potential of (runner on 2nd, one out) is .61 runs. So the
value of the sacrifice bunt in this situation is

VALUE D 0:61 0:81 D 0:20

So Arizona really has hurt themselves on the average the team has decreased their run production
by .2 runs.
But this calculation is assuming that we are primarily interested in scoring as many runs
as possible. Maybe the Arizona manager wants to score one run or more and thinks that he has
put the team in a better position to score (that is, get one run or more) using the sacrifice bunt.
Using the Markov Chain model, we showed earlier how we could simulate the probability of
scoring in the remainder of the inning from each of the 24 starting states. From Table 9.13, we
obtain

Prob(scoring with runner on 1st and no outs) D :391;

Prob(scoring with runner on 2nd and 1 out) D :380:

Even from this perspective, Arizona has decreased their chances of scoring slightly by using the
sacrifice bunt.
Let’s return to the situation where these sacrifice bunts occurred. Counsell was instructed
to sacrifice in the 1st and 3rd innings of the game. In these early innings, it would seem that the
goal of a team would be to score as many runs as possible. Also, Craig Counsell is one of the
better hitters on the Diamondbacks and he is likely to get a base hit that would greatly increase
Arizona’s run potential. So it would seem that the sacrifice hit was not an effective strategy in
the situations where it was used.

9.6 Exercises
9.0. Rickey Henderson is considered the greatest leadoff hitter, but how often did he actually
lead off? Let’s focus on Henderson’s 273 plate appearances at home games for the 1990
baseball season we record the state of the inning (outs and runners) for each plate appear-
ance. Table 9.15 classifies these 273 plate appearances with respect to the number of outs
(0, 1 and 2) and the eight possible runner situations.

Table 9.15. Count of number of plate appearances in different runner and out
situations for home games for the 1990 season
Bases Situation

0 outs 109 17 0 0 2 2 0 0
1 out 29 13 8 1 3 2 0 2
2 outs 37 18 11 4 6 5 2 2
Total 273

(a) What fraction of times did Henderson actually bat with the bases empty and no outs?
(These were essentially the times when he was a leadoff batter.)
(b) What fraction of times did Henderson bat with exactly one runner on base? With
exactly two runners on base? When the bases were loaded?
(c) Suppose that you were able to find a similar classification of plate appearances for
Barry Bonds for the 1990 season. Would you expect Bonds to have similar fractions
of plate appearances with the bases empty, one runner, two runners, and three runners
as Rickey Henderson? Explain.
9.1. A baseball fan has been celebrating the recent success of his team at location D. He takes
a random walk down the street hoping to arrive at his home (location H). From location
D, he is sure to go to location J in the next minute. From location J, he is equally likely in
the next minute to return to D or go ahead to location K. From location K, he is equally
likely the nesxt minute to go home (location H) or back to location J. Once he is home, he
will remain there with probability 1. Figure 9.3 shows the four locations, arrows to show
the possible transitions, and the transition probabilities. A Markov Chain with states D, J,
K, H and transition matrix P shown in Table 9.16 can represent this random walk. (Note
that H is an absorbing state in the chain.) The matrices P 2 , P 3 , and P 4 are also shown
below.

1 .5 .5

D J K H

.5 .5
1

Figure 9.3. States and possible moves in a random walk from location D to home (H).

Table 9.16. Transition probability matrix for the random walk

D J K H
D 0 1 0 0
P D J 0.5 0 0.5 0
K 0 0.5 0 0.5
H 0 0 0 1

2 3 2 3
:5 0 :5 0 0 :75 0 :25
6 0 :75 0 :25 7 6 :375 0 :375 :25 7
P2 D 6
4 :25
7P3 D 6 7
0 :25 :50 5 4 0 :375 0 :625 5
0 0 0 1 0 0 0 1

2 3
:375 0 :375 :25
6 0 :5625 0 :4375 7
P4 D 6
4 :1875
7
0 :1875 :6250 5
0 0 0 1
(a) One possible path of our baseball fan is DJDJKH. Find the probability of this path
using the transition probabilities.
(b) If the fan starts at D, is it possible for our fan to return to D after three minutes?
(Recall that each step takes one minute.) Why or why not?
(c) Using the given matrices, find the probability that the fan (starting at D) will arrive
home in four minutes.
(d) If the fan starts at J, find the probability that he returns to J in two minutes.
(e) If the fan starts at J, where is his most likely location in three minutes?
9.2. (Exercise 9.1 continued.) Let Q denote the matrix obtained by deleting the last row
and last column corresponding to the absorbing state from the transition matrix P . The
fundamental matrix is displayed below.
2 3
3 4 2
E D .I Q/1 D 4 2 4 2 5
1 2 2
(a) If the fan starts at D, how many minutes does the typical fan expect to spend at
location J before he arrives home?
(b) If the fan starts at J, how many minutes will the typical fan spend, on average, at
location D before he gets home?

(c) Starting from D, how many minutes will the fan take, on average, before arriving
home?
(d) Is it possible for the fan to take 20 minutes to arrive home (starting from D)? Why
or why not?
9.3. Consider the following simplified version of baseball. When a player comes to bat, he gets
out with probability .6, he hits a double with probability .3, and he hits a home run with
probability .1. No other batting events besides outs, doubles, and home runs are possible.
There are only seven possible (outs, bases) situations in this game given by

(0 outs, ), (0 outs, ), (1 out, ),

(1 out, ), (2 outs, ), (2 outs, ), (3 outs)

(a) Fill in the transition probability matrix P below. The impossible transitions are
indicated by zeros in the matrix.

Transition matrix P
0 outs 0 outs 1 out 1 out 2 outs 2 outs 3 outs

0 outs, 0 0 0 0
0 outs, 0 0 0 0
1 out, 0 0 0 0
1 out, 0 0 0 0
2 outs, 0 0 0 0
2 outs, 0 0 0 0
3 outs 0 0 0 0 0 0

(b) For each of the possible transitions, find the number of runs scored on the play and
place in the table below. The impossible transitions are crossed out in the table.

Runs matrix
0 outs 0 outs 1 out 1 out 2 outs 2 outs 3 outs

0 outs, xxx xxx xxx xxx

0 outs, xxx xxx xxx xxx
1 out, xxx xxx xxx xxx
1 out, xxx xxx xxx xxx
2 outs, xxx xxx xxx xxx
2 outs, xxx xxx xxx xxx
3 outs xxx xxx xxx xxx xxx xxx

9.4. (Exercise 9.3 continued.) Consider a Markov Chain model for the simplified game of
baseball with transition matrix found in Exercise 9.3.
(a) Find the two-step probability matrix. Using this matrix, find the probability that there
will be a runner on 2nd and one out after two players have batted in the inning.
(b) Compute the expected number of visits matrix E. Using this matrix, find the average
number of batters in a half-inning of baseball.
(c) If the current state of the inning is one out with a runner on 2nd, find the expected
number of batters in the remainder of the inning.
(d) Find the run potential vector . This vector gives the expected number of runs in the
remainder of the inning starting at each possible state.
(e) Suppose that a player comes to bat with a runner on 2nd with one out. He hits a
double. Using the run potential vector R, find the value of this play.
9.5. Consider again the run potential matrix shown in Table 9.17 that gives the expected runs
in the remainder of the inning for each of the 24 possible outs/bases situations.

Table 9.17. The expected number of runs scored in the remainder of the inning
starting at each possible state
Bases Situation

0 outs 0.44 0.81 1.06 1.29 1.36 1.73 1.90 2.21

1 out 0.22 0.46 0.61 0.90 0.83 1.06 1.33 1.49
2 outs 0.08 0.19 0.28 0.33 0.38 0.40 0.47 0.63

Using this matrix, find the value of the following batting plays.
(a) There are runners on the corners (first and third) with one out. The batter hits a
double, scoring both runners.
(b) There is a runner on 1st with no outs. The batter hits a grounder, which is converted
to a double play, getting both the runner and the batter out.
(c) The bases are loaded with no outs. The batter hits a grand slam home run. Compare
with the value of the grand slam with two outs.
9.6. What is the value of a single when there are runners on 1st and 2nd with no outs?
The value of this hit depends on the advancement of the runners. In the 1987 National
League, a single occurred in this situation (runners on 1st and 2nd with no outs) a total
of 136 times. Table 9.18 shows the three types of run advancement and the count of each
type.

Table 9.18. Three types of runner advancement and the count of each type when
there is a single with runners on 1st and 2nd and no outs
Starting state Final state Count Runs scored Value
, no outs 43
, no outs , no outs 35
, no outs 58

(a) For each final state, compute the number of runs scored on the play. Put these values
in the “Runs scored” column of the table.
(b) Using the run potential matrix in Table 9.12 and the runs scored from (a), find the
value of each transition. Put the values in the “Value” column of the table.
(c) Use the previous calculations to find the mean value of a single when runners are on
1st and 2nd with no outs.
9.7. Suppose that the batter hits a single with the bases loaded and one out. How important is
this play? The value depends on the advancement of the runners. This play occurred 88
times in the 1987 National League. Table 9.19 below shows the possible advancement of
the runners and the number of times each type of advancement occurred.

Table 9.19. Five types of runner advancement and the count of each type when
there is a single with the bases loaded with one out
Starting state Final state Count Runs scored Value
, 1 out 2
, 1 out 34
, 1 out , 1 out 14
, 1 out 5
, 1 out 33

(a) For each final state, compute the number of runs scored on the play. Put these values
in the “Runs scored” column of the table.
(b) Using the run potential matrix and the runs scored from (a), find the value of each
transition. Put the values in the “Value” column of the table.
(c) Use the previous calculations to find the mean value of a single when the bases are
loaded with one out.
9.8. How valuable is a walk? The value of this play depends on the beginning (runners, outs)
state. In Table 9.20, the value of the walk is shown for each of the 24 possible states
and the number of times that the state occurred (using 2014 MLB data) is displayed in
parentheses. For example, we see that there was a walk with the bases empty and no outs

Table 9.20. Values and the number of walks that occur in all possible situations
Bases Situation

0 outs 0.37 0.56 0.32 0.40 0.97 0.57 0.47 1.00

(3025) (566) (268) (46) (136) (50) (78) (32)
1 out 0.24 0.36 0.22 0.16 0.67 0.41 0.15 1.00
(2214) (683) (657) (247) (289) (119) (287) (84)
2 outs 0.11 0.19 0.08 0.07 0.25 0.24 0.14 1.00
(1973) (791) (1022) (402) (401) (200) (312) (138)

3025 times. The value of this particular play (moving from , 0 outs to , 0 outs) is .37
runs.
(a) In what situation(s) is a walk most valuable? When is a walk least valuable?
(b) Explain why a walk has a value of one run when the beginning state is bases loaded .
(c) In what situation is a walk most likely to occur? When is a walk least likely to
occur?
(d) Find the average value of a walk over all situations.
9.9. In Game 7 of the 2014 World Series, Hunter Pence and Brandon Belt were starters for
the San Francisco Giants. The tables below show how the players did in all of their plate
appearances in that particular game. Table 9.21 gives the inning in which the player batted,
the before and after game states and the batting play.

Table 9.21. Results of all plate appearances of Hunter Pence and Brandon Belt
in Game 7 of the 2014 World Series
Hunter Pence
Inning Before state Play After state Value
2nd , 0 outs Single , 0 outs
4th , 0 outs Single , 0 outs
6th , 0 outs Groundout double play , 2 outs
8th , 2 outs Groundout 3 outs
Brandon Belt
Inning Before state Play After state Value
2nd , 0 outs Single , 0 outs
4th , 0 outs Flyball , 1 out
6th , 2 outs Single , 2 outs
8th , 0 outs Groundout , 1 out

(a) For each batting play of each player, find the value using the run potential matrix in
Table 9.12. Record the values in the “Value” column of the tables.
(b) What was the most valuable batting play for each hitter?
(c) Which player had the better batting performance in this particular game? Explain.
9.10. (Value of a sacrifice bunt.) Suppose that there is a runner on 2nd base with no outs. Is it a
good play to have the batter hit a sacrifice bunt to move the runner from 2nd to 3rd? (Use
the run potential matrix in Table 9.12 in your explanation. Alternately, you can use the
matrix that gives the probability of scoring at least one run in all situations.)
9.11. (Value of a steal.) Suppose the first batter in the inning gets a walk. He is thinking about
stealing 2nd base.
(a) Suppose that the runner attempts to steal 2nd base and is successful. Find the value
of this play. (It should be positive.)

(b) Suppose that the runner attempts to steal 2nd base and is unsuccessful. (The catcher
throws him out.) Find the value of this play. (This value should be negative.)
(c) Suppose that this particular runner is successful in stealing 2nd base 70% of the
time. Does the team benefit (in terms of runs scored) by having this player steal?
Explain.
(d) Let the probability that the runner is successful in stealing 2nd base be equal to p.
Find the value of p such that the value of the stealing play is equal to zero. (So if the
runner has a success rate that is larger than p, it would benefit the team to have him
attempt the steal of 2nd base.)
9.12. Table 9.22 displays data about the changes in pitch count using data from the 2014 season.
The column “End” refers to the end of the plate appearance by a ball put in play, a walk,
or a strikeout. The row “0-0”, this table shows there were 93,466 occurrences of “adding
a strike” (to count 0-1), 74,098 occurrences of “adding a ball” (to count 1-0), and 22,438
occurrences to “End.” For two-strike outs, like 0-2, 1-2, 2-2, and 3-2, it is possible to
transition to the same count. For example, we see there were 8643 transitions from a
count of 0-2 to 0-2.

Table 9.22. Number of transitions in pitch count using data from the
2014 season
Count Add Strike Add Ball Same End
0-0 93466 74098 0 22438
0-1 37710 37850 0 17886
0-2 0 20546 8643 17164
1-0 35860 25000 0 13238
1-1 32424 24731 0 16555
1-2 0 24889 14394 28081
2-0 12257 8081 0 4662
2-1 17010 10459 0 9519
2-2 0 14273 16073 25826
3-0 4567 0 0 3514
3-1 7116 0 0 7910
3-2 0 0 9417 23186

(a) Explain why the transitions in pitch counts can be viewed as a Markov Chain.
Give the possible states, and write down several rows of the transition probability
matrix P .
(b) Find the probability that the count will be 0-2 after two pitches.
(c) Find the probability the count will be 1-2 after three pitches.
(d) Using matrix operations, find the average number of pitches in a plate appearance.

Further Reading
A good description of the discrete Markov Chain probability model is contained in Kemeny
and Snell (1976). Pankin (1987) and Bukiet and Palacios (1997) describe the use of Markov

Chains to model baseball. Lindsey (1963), using actual game data, found the distribution of runs
scored in the remainder of the inning starting from each possible bases occupied/outs situation.
Lindsey’s run production data is used to estimate the value of different types of base hits in
Albert and Bennett (2003), Chapter 7.

A.1 The Game of Baseball

Baseball is one of the most popular games in the United States; it is often called the national
pastime. The game evolved out of various ball-and-stick games played in many areas of the
world, including the Russian game of lapta and the English game of rounders. It became a
popular sport in the eastern United States in the mid-1800s. Professional baseball started near
the end of the 19th century; the National League was founded in 1876 and the American League
in 1900. Currently in the United States, there are 30 professional teams in the American and
National Leagues and millions of people watch games in ballparks and on television.
Baseball is a game between two teams of nine players each, played on an enclosed field.
A game consists of nine innings. Each inning is divided into two halves; in the top half of the
inning, one team plays defense in the field and the second team plays offense, and in the bottom
half, the teams reverse roles. The fielding positions for the nine players playing defense are a
catcher, first baseman, second baseman, shortstop, third baseman, left fielder, center fielder and
right fielder. Figure A.1 shows a diagram of a baseball field and shows the fielding positions
for the defensive players. The corners of the diamond shape are the locations of the four bases,
home base, first base, second base, and third bases, that play an important part in the game. This
figure also shows the location of one offensive player, the batter, at home base.
The team that is batting during a particular half-inning, the offensive team, is trying to score
runs. A player from the offensive team begins by batting at home base. A run is the score made
by this player who advances from batter to runner and touches first, second, third, and home
bases in that order. A team wins a game by scoring more runs than its opponent at the end of
nine innings. There are some exceptions to the nine-inning game. A game that is tied after nine
innings continues into extra innings until one team has won, and a game may be shortened due
to inclement weather.
A basic play in baseball consists of a player on the defensive team, called a pitcher, throwing
a spherical ball (called a pitch) toward the batter. The batter is attempting to strike or hit the pitch
using a smooth round stick called a bat. After a number of thrown pitches, the batter will either
be put out or become a runner on one of the bases. The batter may be put out in several ways:
He hits a fly ball (a ball in the air) that is caught by one of the fielders.
He hits a ball in fair territory (explained below) and first base is tagged before the batter
reaches first base.
A third strike (explained below) is caught by the catcher.

233
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
234 An Introduction to Baseball

Center Fielder
Right Fielder
Left Fielder

Shortstop 2nd Baseman

3rd Baseman Pitcher 1st Baseman

Batter
Catcher

Figure A.1. Diagram of baseball field and bases and location of nine defensive players and batter.

Fair territory is the part of the playing field between the line from home plate to first base
and the line from home plate to third base. Foul territory is the region of the field outside of
fair territory. A strike is a pitch that is struck at by the batter and missed, or is not struck by the
batter and passes through a region called the strike zone. A ball is a pitch that is not struck at by
the batter and does not enter the strike zone in flight.
A hitter can advance to a runner and reach base safely by:

Receiving four pitches that are balls. In this case, the batter receives a walk or base-on-balls
and can advance to first base.
Hitting a ball in fair territory that is not caught by a fielder or thrown to first base before the
runner reaches first base. There are different types of hits depending on the advancement of
the runner on the play. A single is a hit where the runner reaches first base, a double is a hit
where the runner reaches second base, a triple is a hit where a runner reaches third base, and
a home run is a big hit (usually over the outfield fence) where the runner advances around all
bases safely.

A.2 One Half-Inning of Baseball

In a half-inning of baseball, the nine players on the offensive team will come to bat in sequence.
The players will continue to bat in the inning until three outs are made. To get a flavor of how
baseball is played, let us revisit the last game played in the 2014 Major League Baseball season.
The Kansas City Royals and San Francisco Giants were playing the final game of the World
Series – the winner of this game would be declared the best team of the season. We focus on
the top half of the 4th inning of the game where the Giants were batting and the game score
was tied at 2-2. The pitcher for the Royals at the beginning of the inning was Jeremy Guthrie.

Figure A.2 displays diagrams of the runner and batter situations for each of the six players that
came to bat this particular half-inning

Play 1 Play 2 Play 3

Sandoval

Sandoval Pence

Sandoval Pence Belt

0 Out 0 Out 0 Out

Play 4 Play 5 Play 6

Sandoval Pence Pence Morse Pence Morse

Morse Crawford Perez

1 Out 1 Out 2 Outs

Figure A.2. Batter and runner diagrams for each of six San Francisco Giant players that came to bat in
the top of the fourth inning of the final game of the 2014 World Series.

Batter 1: The inning started with no runners on base and no outs. The first batter for the Giants
was Pablo Sandoval. Sandoval hit a grounder to second-base for a single. The Giants now have
a runner on first base with no outs.
Batter 2: The next Giants batter Hunter Pence singles to center field. Sandoval advances to
second base. The situation is now runners on first and second with no outs.
Batter 3: Brandon Belt batted next for the Giants. Belt hits a deep fly that is caught in left field
for an out. On the fly ball, Sandoval is able to advance to third base, so the Giants have runners
on first and third with one out.
Batter 4: Michael Morse, the next batter, hits a linear to right field for a single. Sandoval scores
from third base and Pence advances to third base. Moore is credited with a run batted in as one
run scored on the basis of his hit. Now the Giants again have runners on first and third base with
one out.
Batter 5: The next hitter, Brandon Crawford, strikes out, and the runners remain on 1st and 3rd
with two outs.
Batter 6: Juan Perez, the next batter, hits a grounder to shortstop who throws to first base in
time for the third out.
With the run scored in the top of the fourth inning, the Giants took a 3-2 lead that would
eventually be the final score of the game—the Giants became the 2014 MLB champions.

A.3 The Boxscore: A Statistical Record of a Baseball Game

One notable aspect of the game of baseball is the wealth of numerical information that is
recorded about the game. A boxscore is a statistical record of a particular game. Figure A.3

Figure A.3. Boxscore of Game 7 of the 2014 World Series between the Giants and the Royals.

displays a boxscore for the last game in the 2014 World Series. We will define particular baseball
events and the associated notation by using this boxscore.
When a player comes to bat during an inning, he is making a plate appearance (PA). What
can happen during this plate appearance? The batter can get a hit (H) and there are four possible
hits: a single (1B), a double (2B), a triple (3B), and a home run (HR). The batter may get a walk
(base-on-balls abbreviated BB) by receiving four pitched ball—she advances to first base. Also,
the batter can advance to first base when he is hit by a pitch (HBP). The player might create
an out. Some outs like a sacrifice bunt (SH) and a sacrifice bunt (SB) advance runners on base.
Finally, the player might reach base by an error by a fielder (E).
An official at-bat (AB) is a plate appearance excluding walks, hit-by-pitches, sacrifice flies
and sacrifice hits. The top half of the boxscore lists all of the batters for both teams. For each
player, the boxscore first gives his fielding position. For example, we see that Blanco was the
center fielder (CF) for the Giants in this game. Then the boxscore lists

AB: the number of at-bats of the player in the game,

R: the number of runs scored by the player,
H: the number of hits by the player,
RBI: the number of runs batted in by the player,
BB: the number of walks by the player,
SO: the number of strikeouts by the player,
# P: the number of pitches received by the player,
AVG: the World Series batting average of the player

OBP: the WS on-base percentage of the player,

SLG: the WS slugging percentage of the player.
Under the basic batting table, the boxscore lists some special events not included in the
table. Under BATTING, the boxscore lists the players who hit doubles (2B), triples (3B) and
home runs (HR). In this listing, the inning in which the hit occurred and the opposing pitcher
are recorded. So the listing
2B P. Sandoval (3, W. Davis)
means that Pedro Sandoval hit a double in the third inning against Wade Davis. In addition, this
section gives specific information about RBI, SF, GIDP (ground out in a double play), and RISP.
The Giants were 2-6 in RISP which indicates that two of the six runners in scoring position
(either 2nd or 3rd base) eventually scored.
Below the batting tables is a table of statistics for the pitchers in the game. For each pitcher,
the table gives the number of innings pitched (IP), the number of hits (H) and runs (R) allowed.
Next it shows the number of earned runs (ER) allowed—these are runs allowed by the pitcher
not due to errors by the fielders of the team. Next the table gives the number of walks (BB)
and strikeouts (SO) and home runs (HR) allowed by each pitcher. The table gives the number
of pitches (PC) and the number of strikes thrown (ST). The last number, the earned run average
(ERA), gives the average number of earned runs allowed by this pitcher for nine innings for all
games played in the series.
In addition to this batting and pitching information, the boxscore gives a line summary of
the runs and hits scored in the game.
--------------------------------------------------
Giants - 020 100 110 -- 3
Royals - 020 000 000 -- 2
--------------------------------------------------

Each column of numbers corresponds to the number of runs scored by the two teams in a
given inning. We see that the Giants and Royals scored two runs in the second inning, and the
Giants scored one run in the top of the fourth that become the winning run in the game. Last,
the boxscore gives some other information about the game (not shown in Figure A.3). This
includes the names and positions of the umpires, the elapsed time (T) of the game, the ballpark
attendance, and some data on the weather during the game.

[1] Albert, Jim (2008), “Streaky hitting in baseball,” Journal of Quantitative Analysis in Sports, 4, 1.
[2] (1998), “Sabermetrics.” In Encyclopedia of Statistical Sciences (edited by S. Kotz, C. B.
Read and D. L. Banks). New York: John Wiley.
[3] Albert, James and Rossman, Allan (2001), Workshop Statistics: Discovery with Data, A Bayesian
Approach, Emeryville, CA: Key College.
[4] Albert, Jim and Bennett, Jay (2003), Curve Ball: Baseball, Statistics, and the Role of Chance in the
Game, Springer, New York.
[5] Bennett, Jay (1998), “Baseball.” In Statistics in Sport (edited by Jay Bennett), London; New York:
Arnold Publishers.
[6] Berry, Donald A. (1996), Statistics: A Bayesian Perspective, Belmont, CA: Wadsworth Publishing.
[7] Bukiet, Bruce, Harold, Elliotte, and Palacios, Jose (1997), “A Markov Chain Approach to Baseball,”
Operations Research, Vol. 45, No. 1, pp. 14–23.
[8] Cover, Thomas M. and Keilers, Carroll, W. (1977), “An Offensive Earned Run Average for Baseball,”
Operations Research, 25, pp. 729–740.
[9] D’Esopo, D. A. and Lefkowitz, B. (1977), “The Distribution of Runs in the Game of Baseball.”
In Optimal Strategies in Sports (edited by S. P. Ladany and R. E. Machol), pp. 55–62. New York:
North-Holland.
[10] Devore, Jay and Peck, Roxy (2011), Statistics: The Exploration and Analysis of Data, Pacific Grove,
CA: Brooks/Cole.
[11] Gilovich, Thomas, Vallone, Robert, and Tversky, Amos (1985), “The hot hand in basketball: On the
misperception of random sequences,” Cognitive Psychology, 17, 295–314.
[12] James, Bill (1982), The Bill James Baseball Abstract, New York: Ballantine Books.
[13] (1997), The Bill James Guide to Baseball Managers, New York: Scribners.
[14] (2001), The New Bill James Historical Baseball Abstract, New York: The Free Press.
[15] Katz, Stanley M., “Study of ’The Count,” 1986 Baseball Research Journal (#15), pp. 67–72.
[16] Kemeny, John G. and Snell, J. Laurie (1976). Finite Markov Chains, New York: Springer-Verlag.
[17] Lindsey, George R. (1963), “An Investigation of Strategies in Baseball,” Operations Research,
vol. 11, pp. 477–501.
[18] Major League Handbook (2001), Lincolnwood, IL: Sports Team Analysis & Tracking Systems.
[19] Marchi, Max and Albert, Jim (2013), Analyzing Baseball Data with R, Chapman and Hall.
[20] Moore, David, McCabe, George and Craig, Bruce (2012), Introduction to the Practice of Statistics
(4th edition), New York: W. H. Freeman Company.
[21] Mosteller, Frederick (1952), “The World Series Competition,” Journal of the American Statistical
Association, 47, 355–380.
[22] Neft, David S., and Cohen, Richard M. (1997), The Sports Encyclopedia: Baseball, New York: St.
Martin’s Press.
[23] Pankin, Mark D. (1987), “Baseball as a Markov Chain” In The Great American Baseball Stat Book,
pp. 520–524.

239
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
240 Bibliography

[24] R Core Team (2015), R: A Language and Environment for Statistical Computing, Vienna, Austria,
https://www.R-project.org/.
[25] Rothman, Stanley (2012), Sandlot Stats: Learning Statistics with Baseball, John Hopkins University
Press.
[26] Schell, Michael (1999), Baseball’s All-Time Best Hitters: How Statistics Can Level the Playing Field,
Princeton, NJ: Princeton University Press.
[27] Scheaffer, Richard L. and Young, Linda J. (2009), Introduction to Probability and its Applications,
Duxbury Press.
[28] StatCrunch (2015), Pearson Education.
[29] Tabor, Josh and Franklin, Chris (2011), Statistical Reasoning in Sports, New York: W. H. Freeman
Company.
[30] Thorn, John and Palmer, Palmer (1985), The Hidden Game of Baseball, New York: Doubleday.
[31] (eds) (1989), Total Baseball. New York: Warner Books.

68-95-99.7 rule, 57 Brenly, Bob, 224

Brett, George, 59
ability Bumgarner, Madison, 126
career, 49
model, 187 Cabrera, Melky, 140
peak, 49 Cabrera, Miguel, 80
ability and performance, 156 career trajectory, 49
absorbing state, 213 Carew, Rod, 59
adjusted ERA statistic, 53 Carter, Joe, 50
All-Star Baseball, 5, 116 Cobb, Ty, 132
Altuve, Jose, 103 consistent hitter, 191
association correlation, 79
negative, 76 Counsell, Craig, 224
positive, 75, 82 Crawford, Brandon, 235
attendance, ballpark, 23, 42 Cruz, Nelson, 103

balk, 216 data distribution, 15

Baseball Archive, 11 average, 15
baseball attendance, 102 shape, 15
baseball card, 9 spread, 15
Baseball Reference, 10, 20 Davis, Chris, 112, 207
batted ball data, 71 Davis, Wade, 237
batter’s runs average, 81 derived baseball statistic, 13
batting average, 7, 59 designated hitter, 16, 27
true, 164 dice
Bayesian thinking, 160 roll of two, 115
bell-shaped, 182 die
Belt, Brandon, 230, 235 roll of 10 sided, 191
Bernoulli trials, 143 roll of 20 sided, 120
bias model, 185 roll of one, 114
Big League Baseball, 114 to simulate hitting data, 191
binomial distribution, 139 DiMaggio, Joe, 28
Bogg, Wade, 165 Donaldson, Josh, 172
Bonds, Barry, 54, 149 dotplot, 18, 26
boxplot, 48 Cleveland-style, 25
parallel, 56 parallel, 27, 56
boxscore, 235
Brantley, Michael, 173, 188 earned run average, 8

241
Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)
Copyright no copyright American Mathematical Society. Duplication prohibited. Please report unauthorized use to [email protected].
Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.
242 Index

empirical rule, 57 mean, 57

equally likely outcomes, 115 mean square error, 83
ERA median, 17, 46
adjusted, 53 model
statistic, 8, 51 ability, 187
expected number bias, 185
of hits, 141 many spinners, 183
of visits to a state, 214, 219 no effect, 185
runs in an inning, 221 one spinner, 182
Morse, Michael, 235
Fangraphs, 11 moving average, 188
fastball speeds, 70 multiple linear regression, 86
Feller, Bob, 32 multiplication rule, 123
five-number summary, 17, 46, 56
Ford, Whitey, 51 negative association, 94
fundamental matrix, 221 negative binomial distribution, 143
no effect model, 185
geometric distribution, 151 nonlinear equation, 89
Glavine, Tom, 66 normal curve, 183
Goldschmidt, Paul, 173
goodness of fit, 82 official at-bat, 236
Gould, Stephen Jay, 64 on-base percentage, 8, 46
Greinke, Zack, 43, 70 true, 168
Gwynn, Tony, 59, 64, 134, 165 on-base profile, 139
one spinner model, 182
Hall of Fame pitchers, 38 OPS statistic, 9, 46, 81
Henderson, Rickey, 1, 28, 61, 94, 124, 148, 168, outcomes
193, 225 of one die roll, 114
histogram, 24 of two die rolls, 115
home run rate, 55, 76 outliers, 48
Howard, Ryan, 187
Pagan, Angel, 125
James, Bill, 1, 8 peak ability, 49
Jeter, Derek, 17 Pearson residual, 144
Johnson, Randy, 20, 134 Pedroia, Dustin, 169
Johnson, Walter, 101 Pence, Hunter, 230, 235
Perez, Juan, 235
Keeler, Willie, 64 performance, 156
Kershaw, Clayton, 70 Pesky, Johnny, 8
Koufax, Sandy, 31 pitch count, 177
pitches, number of, 101
lapta, 233 PITCHf/x system, 11
least-squares fit, 82, 87, 90, 92 plate appearance, 236
length of baseball game, 36 Poisson distribution, 149
leverage, 177 positive association, 75
probability
Maddux, Greg, 66, 120 of set of outcomes, 114
Mantle, Mickey, 132, 157 of sum of two dice, 121
many spinners model, 183 relative frequency interpretation, 112
Maris, Roger, 9, 51, 54, 157 subjective interpretation, 162
Markov Chain, 211 probability interval, 164
Martinez, Victor, 173 Pujols, Albert, 44, 45, 66, 71, 107
McGwire, Mark, 54, 120 Pythagorean method, 88

quartile, 17, 46 spinner

lower, 17 calculation of probability regions, 117
upper, 17 to model hitting, 116, 181
spinner model, 157
Raines, Tim, 4, 61 standard deviation, 57
Ramirez, Manny, 45, 66 standard normal density, 164
regression effect, 91 standardized score, 59
relative standing, 15 states of Markov Chain, 212
residual, 82, 91 statistics, meaning of, 2
Pearson, 144 stemplot, 14, 21, 22, 57, 59
Retrosheet, 11 back to back, 15, 23, 46, 51
Reynolds, Mark, 125 step
right skewed, 23 for determining outliers, 48
Roberts, Robin, 50 stolen base rate, 99
Rodriguez, Alex, 162 Strat-O-Matic Baseball, 119
root mean square error, 83 streaky
rounders, 233 ability, 188
run expectancy matrix, 221 performance, 188
run potential matrix, 224 strikeout rate, 21, 126
run value, of batting event, 222 Stuart, Dick, 8
runs created, 8, 81 subjective probability, 162
runs of hot and cold performance, 190 Suzuki, Ichiro, 71, 124, 188, 202, 206
Ruth, Babe, 30, 54, 116 symmetric, 24, 57

sabermetrics, 8 time of game, 101

SABR, 8 time series plot, 18, 22, 53
sacrifice bunt, 26 total average, 81
salaries, 40 transition probability matrix, 212
sample space, 114 tree diagram, 122
Sandoval, Pablo, 235 triple rate, 76
Sandoval, Pedro, 237 Trout, Mike, 80, 158, 176
scatterplot, 4, 19, 74, 82 two-way table, 171
labeled, 75
scatterplot matrix, 79 value
Schmidt, Mike, 117 of batting event, 222
simulation of home run, 223
of binomial distribution, 141 of sacrifice bunt, 224
of data from consistent hitter, 192
of dice rolls, 158 weights of batting events, 86
of hitting data, 167 Whiz Kids, 50
of situational hitting, 186 Williams, Ted, 59
one spinner model, 183 win/loss ratio, 88
run scoring, 152 winning percentage, team, 98
to learn about hitting ability, 160 Womack, Tony, 224
situational World Series
ability, 180 1950, 50
statistics, 176 1980, 50
skewed right, 23 1993, 50
slugging percentage, 8, 46, 57 2001, 224
smoothing curve, 19 2008, 50
Society of American Baseball Research, 8, 2014, 24, 234
80
sophomore slump, 91 z-score, 60

Alan Dundes Into The Endzone
100% (3)
Alan Dundes Into The Endzone
15 pages
Elementary Number Theory 1st Edition James S. Kraft PDF Download
100% (1)
Elementary Number Theory 1st Edition James S. Kraft PDF Download
61 pages
Math in Our World
100% (1)
Math in Our World
2,069 pages
Jbdlkdckldms
100% (1)
Jbdlkdckldms
366 pages
Mathematical Logic Through Python
No ratings yet
Mathematical Logic Through Python
284 pages
A Radical Approach To Real Analysis 2E - David M Bressoud
100% (1)
A Radical Approach To Real Analysis 2E - David M Bressoud
453 pages
Number Systems An Introduction To Algebra and Analysis (Sergei Ovchinnikov Etc.) (Z-Library)
No ratings yet
Number Systems An Introduction To Algebra and Analysis (Sergei Ovchinnikov Etc.) (Z-Library)
154 pages
Math
100% (2)
Math
156 pages
Linear Methods
100% (2)
Linear Methods
430 pages
Probability and Stochastic Processes 3rd Edition Student Solutions Manual
100% (6)
Probability and Stochastic Processes 3rd Edition Student Solutions Manual
400 pages
Mathematical Adventures
100% (2)
Mathematical Adventures
304 pages
113819770X Sanet - CD
No ratings yet
113819770X Sanet - CD
352 pages
Index To Mathematical Problems 1975-1979 PDF
100% (4)
Index To Mathematical Problems 1975-1979 PDF
520 pages
OceanofPDF - Com Game of Edges - Bruce Schoenfeld
No ratings yet
OceanofPDF - Com Game of Edges - Bruce Schoenfeld
224 pages
A Matlab (R) Companion To Comple - A. David Wunsch
100% (1)
A Matlab (R) Companion To Comple - A. David Wunsch
353 pages
TheBestofSJPerelman 10780295
No ratings yet
TheBestofSJPerelman 10780295
313 pages
Sandor Lehoczky, Richard Rusczyk - The Art of Problem Solving, Vol. 1 - The Basics (2006, AoPS Incorporated) - Compressed PDF
100% (15)
Sandor Lehoczky, Richard Rusczyk - The Art of Problem Solving, Vol. 1 - The Basics (2006, AoPS Incorporated) - Compressed PDF
288 pages
The Handbook of Essential Mathematics
98% (84)
The Handbook of Essential Mathematics
205 pages
Markov Chains and Boardgames Like Monopoly With Mathematica 4-5-2008
No ratings yet
Markov Chains and Boardgames Like Monopoly With Mathematica 4-5-2008
39 pages
An Introduction To Classical Real Analysis
100% (1)
An Introduction To Classical Real Analysis
593 pages
A Radical Approach To Real Analysis PDF
100% (1)
A Radical Approach To Real Analysis PDF
341 pages
#Introductory Mathematics - Applications and Methods - Gordon S. Marshall - Springer (1998) .1
100% (1)
#Introductory Mathematics - Applications and Methods - Gordon S. Marshall - Springer (1998) .1
233 pages
A Decade of The Berkeley Math Circle The American Experience 2 0821846833 9780821846834 Compress
100% (2)
A Decade of The Berkeley Math Circle The American Experience 2 0821846833 9780821846834 Compress
372 pages
Text 049
No ratings yet
Text 049
250 pages
Problems and Snapshots From The World of Probability
100% (1)
Problems and Snapshots From The World of Probability
252 pages
Gary N Felder Kenny M Felder - Mathematical Methods in Engineering and Physics-Wiley 2015
No ratings yet
Gary N Felder Kenny M Felder - Mathematical Methods in Engineering and Physics-Wiley 2015
822 pages
Mathematical Ideas 14th Edition PDF
No ratings yet
Mathematical Ideas 14th Edition PDF
22 pages
A Radical Approach To Real Analysis Second Edition Classroom Resource
91% (11)
A Radical Approach To Real Analysis Second Edition Classroom Resource
341 pages
Interactive Se FM
100% (1)
Interactive Se FM
49 pages
Linear Weights
No ratings yet
Linear Weights
27 pages
CP Baseball Program
No ratings yet
CP Baseball Program
24 pages
Game of Spreads How To Build Better Trading Algos With Game Theory
No ratings yet
Game of Spreads How To Build Better Trading Algos With Game Theory
3 pages
Calculus For The Life Sciences
No ratings yet
Calculus For The Life Sciences
2 pages
Full Issue
No ratings yet
Full Issue
152 pages
Mathematical Methods in Science
No ratings yet
Mathematical Methods in Science
247 pages
NTE82
No ratings yet
NTE82
154 pages
Index To Mathematical Problems 1975 1979
No ratings yet
Index To Mathematical Problems 1975 1979
520 pages
Introductory Mathematics. Applications and Methods
No ratings yet
Introductory Mathematics. Applications and Methods
232 pages
Text010 Endmatter
No ratings yet
Text010 Endmatter
70 pages
Geometry PDF
95% (22)
Geometry PDF
1,129 pages
Decembrie2014 PDF
100% (1)
Decembrie2014 PDF
100 pages
Dol037 Endmatter
No ratings yet
Dol037 Endmatter
19 pages
Probability and Stochastic Processes 2nd Roy D Yates and David J Goodman 2
0% (1)
Probability and Stochastic Processes 2nd Roy D Yates and David J Goodman 2
433 pages
Mansfield Linear Algebra PDF
No ratings yet
Mansfield Linear Algebra PDF
257 pages
MicroLeague Baseball
67% (3)
MicroLeague Baseball
60 pages
Upswing - (Fiction Story)
No ratings yet
Upswing - (Fiction Story)
18 pages
WMLL Announcer'S Script & Instructions: PREPARATION - Laptop Setup
No ratings yet
WMLL Announcer'S Script & Instructions: PREPARATION - Laptop Setup
2 pages
5 - Sample - Math References
No ratings yet
5 - Sample - Math References
1 page
4 - Sample - Math References
No ratings yet
4 - Sample - Math References
1 page
Amm Vol121n07
No ratings yet
Amm Vol121n07
104 pages
Problem Books in Mathematics: R. Halmos
No ratings yet
Problem Books in Mathematics: R. Halmos
5 pages
(SUGGESTED BOOKS) Mathematics - Complete Set
No ratings yet
(SUGGESTED BOOKS) Mathematics - Complete Set
5 pages
Grade 8 Physical Education 2nd Quarter
80% (25)
Grade 8 Physical Education 2nd Quarter
43 pages
Speechless by Adam P. Schmitt Chapter Sampler
No ratings yet
Speechless by Adam P. Schmitt Chapter Sampler
46 pages
Adult Softball League Rules: Nacogdoches Parks & Recreation
No ratings yet
Adult Softball League Rules: Nacogdoches Parks & Recreation
8 pages
American Mathematical Monthly - 1988-7
No ratings yet
American Mathematical Monthly - 1988-7
94 pages
Deadball Quick Start
No ratings yet
Deadball Quick Start
2 pages
101 Problems in Algebra
97% (29)
101 Problems in Algebra
160 pages
Money Ball
No ratings yet
Money Ball
143 pages
Rules of Baseball 2015
No ratings yet
Rules of Baseball 2015
282 pages
Major League Baseball: Special "Baseball Lingo" Section With Definitions Starting On
No ratings yet
Major League Baseball: Special "Baseball Lingo" Section With Definitions Starting On
53 pages
SUNYAC SpringSports Operational Plan21
No ratings yet
SUNYAC SpringSports Operational Plan21
91 pages
Undergraduate Texts in Mathematics: Axler F.W. Gehring Ribet
No ratings yet
Undergraduate Texts in Mathematics: Axler F.W. Gehring Ribet
9 pages
Undergraduate Texts in Mathematics: S. Axler
No ratings yet
Undergraduate Texts in Mathematics: S. Axler
11 pages
Umpirebible 01
No ratings yet
Umpirebible 01
14 pages
2010-11 Winona State Baseball Media Guide
No ratings yet
2010-11 Winona State Baseball Media Guide
56 pages
Fundamental Counting Principle and Permutation
No ratings yet
Fundamental Counting Principle and Permutation
12 pages
Winona State Softball 2011 Media Guide
No ratings yet
Winona State Softball 2011 Media Guide
55 pages
All-State Softball
No ratings yet
All-State Softball
1 page
Haines: Youngest Manager in Franchise History
No ratings yet
Haines: Youngest Manager in Franchise History
52 pages
BH2045 BasesLoaded Reference v5 Prev
No ratings yet
BH2045 BasesLoaded Reference v5 Prev
12 pages
Game Script
No ratings yet
Game Script
16 pages
Mini Scorecard
No ratings yet
Mini Scorecard
3 pages
Dolphin Tiki Extends Win Streak To Ten2
No ratings yet
Dolphin Tiki Extends Win Streak To Ten2
6 pages
Puzzles
No ratings yet
Puzzles
20 pages
Hidden Markov Model Run Scoring Baseball
No ratings yet
Hidden Markov Model Run Scoring Baseball
14 pages
All Stars - Spring 2011
No ratings yet
All Stars - Spring 2011
16 pages
BS Quiz 2017
No ratings yet
BS Quiz 2017
4 pages
Mills Pond Park Softball - Rules
No ratings yet
Mills Pond Park Softball - Rules
14 pages
The Baseball Once-Upon-A Times.: Robinson Nears 500th Home Run, Jackson Shines As O's Tip Yankees
No ratings yet
The Baseball Once-Upon-A Times.: Robinson Nears 500th Home Run, Jackson Shines As O's Tip Yankees
2 pages
The Republican's 87th Sandlot Baseball Tournament Rules and Entry Form
No ratings yet
The Republican's 87th Sandlot Baseball Tournament Rules and Entry Form
3 pages
Shape: The Hidden Geometry of Information, Biology, Strategy, Democracy, and EverythingElse
From Everand
Shape: The Hidden Geometry of Information, Biology, Strategy, Democracy, and EverythingElse
Jordan Ellenberg
4/5 (32)
Against Flow: Video Games and the Flowing Subject
From Everand
Against Flow: Video Games and the Flowing Subject
Braxton Soderman
No ratings yet
Resonant Games: Design Principles for Learning Games that Connect Hearts, Minds, and the Everyday
From Everand
Resonant Games: Design Principles for Learning Games that Connect Hearts, Minds, and the Everyday
Eric Klopfer
No ratings yet
Math Is Wonderful
From Everand
Math Is Wonderful
Fred Hardwicke
No ratings yet
The Everything Everyday Math Book: From Tipping to Taxes, All the Real-World, Everyday Math Skills You Need
From Everand
The Everything Everyday Math Book: From Tipping to Taxes, All the Real-World, Everyday Math Skills You Need
Christopher Monahan
5/5 (2)
The Universe And The Teacup: The Mathematics of Truth and Beauty
From Everand
The Universe And The Teacup: The Mathematics of Truth and Beauty
K. C. Cole
No ratings yet
Data Feminism
From Everand
Data Feminism
Catherine D'Ignazio
4.5/5 (10)
Fascinating Mathematical People: Interviews and Memoirs
From Everand
Fascinating Mathematical People: Interviews and Memoirs
Donald J. Albers
No ratings yet
You’Re Fired, God Hired: Turning a Negative into a Positive
From Everand
You’Re Fired, God Hired: Turning a Negative into a Positive
Dr. Beverly Lewis-Johnson
No ratings yet
The Handy Math Answer Book
From Everand
The Handy Math Answer Book
Patricia Barnes-Svarney
4/5 (6)
Math Games for Middle School: Challenges and Skill-Builders for Students at Every Level
From Everand
Math Games for Middle School: Challenges and Skill-Builders for Students at Every Level
Mario Salvadori
4.5/5 (6)
Introduction to Matrices and Vectors
From Everand
Introduction to Matrices and Vectors
Jacob T. Schwartz
No ratings yet
AP Q&A Statistics:With 600 Questions and Answers
From Everand
AP Q&A Statistics:With 600 Questions and Answers
Barron's Educational Series
No ratings yet

Text 034

Uploaded by

Text 034

Uploaded by

Purchased from American Mathematical Society for the exclusive use of Isabella Burgo (brisxi)

Published and Distributed by

MAA Service Center

Preface to the First Edition ix

Preface to the Second Edition xi

1 An Introduction to Baseball Statistics 1

2 Exploring a Single Batch of Baseball Data 13

3 Comparing Batches and Standardization 45

4 Relationships Between Measurement Variables 73

5 Introduction to Probability Using Tabletop Games 111

6 Probability Distributions and Baseball 139

7 Introduction to Statistical Inference 155

8 Topics in Statistical Inference 175

9 Modeling Baseball Using a Markov Chain 211

A An Introduction to Baseball 233

the greatest base stealer of all time,

Table 1.1. Batting statistics for Rickey Henderson’s career

How should the data be collected to make it useful in drawing conclusions?

Exploring a Single Batch of Baseball Data (Chapter 2)

0.338 0.420 0.408 0.398 0.414 0.399 0.419 0.358 0.423

0.32 0.34 0.36 0.38 0.40 0.42 0.44

Figure 1.1. Dotplot of season on-base percentages for Rickey Henderson

Comparing Batches and Standardization (Chapter 3)

0.391 0.353 0.393 0.393 0.405 0.413 0.429 0.350 0.395

Relationships between Measurement Variables (Chapter 4)

Introduction to Probability Using Tabletop Games (Chapter 5)

Probability Distributions and Baseball (Chapter 6)

Introduction to Statistical Inference (Chapter 7)

Topics in Statistical Inference (Chapter 8)

Modeling Baseball Using a Markov Chain (Chapter 9)

Table 1.2. Batting results for Rickey Henderson’s leadoff

Some Basic Measures of Baseball Performance

Measures for Batters

Measures for Pitchers

Better Measures of Hitting Ability

Career statistics for a player

Player statistics for a given season

Team statistics for a given season

Statistics through the years

Collecting baseball data

2.1 Looking at Teams’ Offensive Statistics

Home Run Totals

What observations can we make from this data distribution?

What do we see in these displays?

2.2 A Tribute to Derek Jeter

Jeter’s Home Runs

2000 2005 2010

2000 2005 2010

The equation of this line is

HR D 883:68  0:4337 Year

Jeter’s Season OPS Values

Figure 2.8. Stemplot of Jeter’s season OPS values.

2000 2005 2010

Figure 2.9. Time series plot of Jeter’s OPS numbers

2.3 A Tribute to Randy Johnson

Table 2.4. Strikeout rates for Randy Johnson

Figure 2.10. Stemplot of season strikeout rates for Randy Johnson.

1990 1995 2000 2005

Figure 2.11. Time series plot of Johnson’s season strikeout rates.

Johnson’s Win/Loss Record in 1995

2.4 Analyzing Baseball Attendance

15000 20000 25000 30000 35000 40000 45000 50000

20000 25000 30000 35000 40000 45000

2.5 Manager Statistics: the Use of Sacrifice Bunts

Table 2.7. Seasonal slugging percentages for Rickey Henderson

(a) Construct a stemplot of the slugging percentages.

Table 2.8. Career Statistics for Joe DiMaggio

(a) Find a five-number summary of the HR numbers.

Table 2.9. Career Statistics for Babe Ruth

(a) Compute a five-number summary.

Table 2.10. Pitching Statistics for Sandy Koufax

Table 2.11. Pitching statistics for Bob Feller

0.0 0.2 0.4 0.6 0.8 1.0

Figure 2.19. Histogram of OBPs of 2014 MLB players.

0.25 0.30 0.35 0.40

0.00 0.02 0.04 0.06 0.08

HR D 883:68 0:4337 Year

STEP D 1:5 .QU QL /:

STEP D 1:5 .1:031 0:953/ D 0:117

QL STEP D 0:953 0:117 D 0:836: