Basic Statistics With R

Basic Statistics With R

Reaching Decisions With Data

Stephen C. Loftus
Division of Science, Technology, Engineering and Math
Sweet Briar College
Sweet Briar, VA, United States
Academic Press is an imprint of Elsevier
125 London Wall, London EC2Y 5AS, United Kingdom
525 B Street, Suite 1650, San Diego, CA 92101, United States
50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom
Copyright © 2022 Elsevier Inc. All rights reserved.
No part of this publication may be reproduced or transmitted in any form or by any means,
electronic or mechanical, including photocopying, recording, or any information storage and
retrieval system, without permission in writing from the publisher. Details on how to seek
permission, further information about the Publisher’s permissions policies and our arrangements
with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency,
can be found at our website: www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the
Publisher (other than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and
experience broaden our understanding, changes in research methods, professional practices, or
medical treatment may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in
evaluating and using any information, methods, compounds, or experiments described herein. In
using such information or methods they should be mindful of their own safety and the safety of
others, including parties for whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors,
assume any liability for any injury and/or damage to persons or property as a matter of products
liability, negligence or otherwise, or from any use or operation of any methods, products,
instructions, or ideas contained in the material herein.

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library

ISBN: 978-0-12-820788-8

For information on all Academic Press publications

visit our website at https://www.elsevier.com/books-and-journals

Publisher: Katey Birtcher

Editorial Project Manager: Alice Grant
Production Project Manager: Beula Christopher
Designer: Patrick C. Ferguson
Typeset by VTeX
To Michelle, to whom I should have listened sooner.
Contents

Biography xv
Preface xvii
Acknowledgments xix

Part I
An introduction to statistics and R
1. What is statistics and why is it important?
1.1 Introduction 3
1.2 So what is statistics? 4
1.2.1 The process of statistics 4
1.2.2 Hypothesis/questions 4
1.2.3 Data collection 5
1.2.4 Data description 5
1.2.5 Statistical inference 5
1.2.6 Theories/decisions 6
1.3 Computation and statistics 6

2. An introduction to R
2.1 Installation 7
2.2 Classes of data 7
2.3 Mathematical operations in R 8
2.4 Variables 9
2.5 Vectors 11
2.6 Data frames 12
2.7 Practice problems 13
2.8 Conclusion 14

Part II
Collecting data and loading it into R
3. Data collection: methods and concerns
3.1 Introduction 17

vii
viii Contents

3.2 Components of data collection 17

3.3 Observational studies 18
3.3.1 Biases in survey sampling 19
3.3.2 Practice problems 21
3.4 Designed experiments 21
3.4.1 Practice problems 23
3.5 Observational studies and experiments: which to use? 23
3.5.1 Practice problems 24
3.6 Conclusion 25

4. R tutorial: subsetting data, random numbers, and

selecting a random sample
4.1 Introduction 27
4.2 Subsetting vectors 27
4.3 Subsetting data frames 29
4.4 Random numbers in R 31
4.5 Select a random sample 32
4.6 Getting help in R 33
4.7 Practice problems 33
4.8 Conclusion 35

5. R tutorial: libraries and loading data into R

5.1 Introduction 37
5.2 Libraries in R 37
5.3 Loading datasets stored in libraries 42
5.4 Loading csv files into R 42
5.5 Practice problems 43
5.6 Conclusion 43

Part III
Exploring and describing data
6. Exploratory data analyses: describing our data
6.1 Introduction 47
6.2 Parameters and statistics 47
6.3 Parameters, statistics, and EDA for categorical variables 48
6.3.1 Practice problems 50
6.4 Parameters, statistics, and EDA for a single quantitative variable 51
6.4.1 Statistics for the center of a variable 51
6.4.2 Practice problems 53
6.4.3 Statistics for the spread of a variable 54
6.4.4 Practice problems 56
6.5 Visual summaries for a single quantitative variables 57
6.6 Identifying outliers 59
 Contents ix

6.6.1 Practice problems 61

6.7 Exploring relationships between variables 61
6.8 Exploring association between categorical predictor and
quantitative response 62
6.8.1 Practice problems 65
6.9 Exploring association between two quantitative variables 65
6.9.1 Practice problems 71
6.10 Conclusion 72

7. R tutorial: EDA in R
7.1 Introduction 73
7.2 Frequency and contingency tables in R 73
7.3 Numerical exploratory analyses in R 74
7.3.1 Summaries for the center of a variable 74
7.3.2 Summaries for the spread of a variable 75
7.3.3 Summaries for the association between two quantitative
variables 76
7.4 Missing data 77
7.5 Practice problems 78
7.6 Graphical exploratory analyses in R 78
7.6.1 Scatterplots 78
7.6.2 Histograms 80
7.7 Boxplots 82
7.8 Practice problems 84
7.9 Conclusion 85

Part IV
Mechanisms of inference
8. An incredibly brief introduction to probability
8.1 Introduction 89
8.2 Random phenomena, probability, and the Law of Large
Numbers 90
8.3 What is the role of probability in inference? 91
8.4 Calculating probability and the axioms of probability 92
8.5 Random variables and probability distributions 94
8.6 The binomial distribution 95
8.7 The normal distribution 96
8.8 Practice problems 98
8.9 Conclusion 99

9. Sampling distributions, or why exploratory analyses

are not enough
9.1 Introduction 101
x Contents

9.2 Sampling distributions 101

9.3 Properties of sampling distributions and the central limit
theorem 105
9.4 Practice problems 107
9.5 Conclusion 107

10. The idea behind testing hypotheses

10.1 Introduction 109
10.2 A lady tasting tea 109
10.3 Hypothesis testing 110
10.3.1 What are we testing? 110
10.3.2 How rare is our data? 112
10.3.3 What is our level of doubt? 113
10.4 Practice problems 115
10.5 Conclusion 115

11. Making hypothesis testing work with the central limit

theorem
11.1 Introduction 117
11.2 Recap of the normal distribution 117
11.3 Getting probabilities from the normal distributions 118
11.3.1 Practice problems 119
11.4 Connecting data to p-values 119
11.4.1 Practice problems 124
11.5 Conclusion 125

12. The idea of interval estimates

12.1 Introduction 127
12.2 Point and interval estimates 127
12.3 When intervals are “right” 128
12.4 Confidence intervals 128
12.5 Creating confidence intervals 129
12.6 Interpreting confidence intervals 132
12.7 Practice problems 133
12.8 Conclusion 133

Part V
Statistical inference
13. Hypothesis tests for a single parameter
13.1 Introduction 137
13.2 One-sample test for proportions 138
13.2.1 State hypotheses 138
 Contents xi

13.2.2 Set significance level 138

13.2.3 Collect and summarize data 139
13.2.4 Calculate test statistic 139
13.2.5 Calculate p-values 140
13.2.6 Conclude 141
13.2.7 Practice problems 142
13.3 One-sample t-test for means 143
13.3.1 State hypotheses 143
13.3.2 Set significance level 144
13.3.3 Collect and summarize data 144
13.3.4 Calculate test statistic 144
13.3.5 Calculate p-values 145
13.3.6 A brief interlude: the t distribution 146
13.3.7 Conclude 148
13.3.8 Practice problems 150
13.4 Conclusion 150

14. Confidence intervals for a single parameter

14.1 Introduction 151
14.2 Confidence interval for p 151
14.2.1 Practice problems 153
14.3 Confidence interval for μ 153
14.3.1 Practice problems 155
14.4 Other uses of confidence intervals 156
14.4.1 Confidence intervals for p and sample size calculations 156
14.4.2 Practice problems 159
14.4.3 Confidence intervals for μ and hypothesis testing 159
14.4.4 Practice problems 161
14.5 Conclusion 162

15. Hypothesis tests for two parameters

15.1 Introduction 163
15.2 Two-sample test for proportions 164
15.2.1 State hypotheses 164
15.2.2 Set significance level 165
15.2.3 Collect and summarize data 165
15.2.4 Calculate the test statistic 166
15.2.5 Calculate p-values 168
15.2.6 Conclude 169
15.2.7 Practice problems 170
15.3 Two-sample t-test for means 171
15.3.1 State hypotheses 171
15.3.2 Set significance level 172
15.3.3 Collect and summarize data 172
15.3.4 Calculate the test statistic 173
15.3.5 Calculate p-values 175
xii Contents

15.3.6 Conclusion 177

15.3.7 Practice problems 178
15.4 Paired t-test for means 179
15.4.1 State hypotheses 180
15.4.2 Set significance level 181
15.4.3 Collect and summarize data 181
15.4.4 Calculate the test statistic 182
15.4.5 Calculating p-values 182
15.4.6 Conclude 183
15.4.7 Practice problems 184
15.5 Conclusion 185

16. Confidence intervals for two parameters

16.1 Introduction 187
16.2 Confidence interval for p1 − p2 187
16.2.1 Practice problems 191
16.3 Confidence interval for μ1 − μ2 191
16.3.1 Equal variances 192
16.3.2 Unequal variances 193
16.3.3 Interpretation and example 195
16.3.4 Practice problems 196
16.4 Confidence intervals for μD 197
16.4.1 Practice problems 198
16.5 Confidence intervals for μ1 − μ2 , μD , and hypothesis testing 199
16.5.1 Practice problems 201
16.6 Conclusion 201

17. R tutorial: statistical inference in R

17.1 Introduction 203
17.2 Choosing the right test 203
17.3 Inference for proportions 204
17.3.1 Inference for a single proportion 205
17.3.2 Inference for two proportions 207
17.3.3 Practice problems 208
17.4 Inference for means 209
17.4.1 Inference for a single mean 209
17.4.2 Inference for two means 210
17.4.3 Paired inference for means 212
17.4.4 Practice problems 213
17.5 Conclusion 213

18. Inference for two quantitative variables

18.1 Introduction 215
18.2 Test for correlations 216
18.2.1 State hypotheses 216
 Contents xiii

18.2.2 Set significance level 217

18.2.3 Collect and summarize data 217
18.2.4 Calculate the test statistic 217
18.2.5 Calculate p-values 218
18.2.6 Conclusion 219
18.2.7 Practice problems 221
18.3 Confidence intervals for correlations 222
18.4 Test for correlations in R 223
18.5 Confidence intervals for correlations 224
18.6 Practice problems 224
18.7 Conclusion 225

19. Simple linear regression

19.1 Introduction 227
19.2 Basic of lines 228
19.3 The simple linear regression model 229
19.4 Estimating the regression model 230
19.5 Regression in R 232
19.6 Practice problems 234
19.7 Using regression to create predictions 234
19.8 Practice problems 235
19.9 The assumptions of regression 236
19.9.1 Assumption 1: linearity 236
19.9.2 Assumption 2: independence 239
19.9.3 Assumption 3: zero mean 239
19.9.4 Assumption 4: homoskedasticity 239
19.10 Inference for regression 240
19.10.1 State hypotheses 240
19.10.2 Set significance level 241
19.10.3 Collect and summarize data 241
19.10.4 Calculate our test statistic 241
19.10.5 Calculate p-values 242
19.10.6 Conclusion 243
19.10.7 Inference for regression in R 243
19.11 How good is our regression? 244
19.12 Practice problems 247
19.13 Conclusion 247

20. Statistics: the world beyond this book

20.1 Questions beyond the techniques of this book 249
20.2 The answers statistics gives 252
20.3 Where does this leave us? 254

A. Solutions to practice problems

xiv Contents

B. List of R datasets

References 271
Index 277

Online Resources
Please visit the student companion site for access to Data Sets referenced in
the text: https://www.elsevier.com/books-and-journals/book-companion/
9780128207888
Biography

Dr. Stephen Loftus is a Visiting Professor of Mathematics at Sweet Briar Col-

lege. Prior to this, he received a Ph.D. in Statistics from Virginia Tech and
worked as an Analyst in Baseball Research and Development for the Tampa
Bay Rays. His research focuses on Bayesian applications in large datasets and
sabermetrics.

xv
Preface

Over the past 15 years—ever since I took AP Statistics as a high school junior—
I have heard many horror stories from friends, family, and complete strangers
about their experiences with statistics. “It was too difficult” or “It just did not
click for me” were common refrains. As such, the thought of doing statistics or
working with data fills them with fear.
This is something I consider very unfortunate, for two reasons. The first is
that statistics is a subject with which I am personally fascinated. The ability to
draw information out of and tell stories with data is something that has drawn
me to the subject. Second, and much more importantly, the ability to work with
data is becoming an essential skill in the 21st century. As the amount of data
increases in every field, employees of all types are expected to take a larger part
in the process of drawing decisions from this data. This cannot happen when a
person’s sole experience with statistics is negative.
What follows is an attempt to try to provide a minimal-stress introduction
to statistics for some or a gentle reintroduction to the subject for others. In this
book, we will be looking at many of the foundations of statistics; from how
data is collected, to exploratory analysis, to basic statistical inference. In doing
so, we will have many opportunities to practice the techniques learned through
examples and practice problems.
Additionally, many of the problems posed by modern statistics require the
use of software to find a solution. As such, the statistical methods taught in this
book are accompanied by instruction in the statistical programming language R.
While not intended to be a comprehensive introduction to either coding or R, it
should provide a good starting point for individuals working in this language for
the first time.
With all this in mind, let us go forward into the process of statistics.

Stephen C. Loftus

xvii
Acknowledgments

I would like to thank everyone who helped to make this book come to fruition, to
include Raina Robeva, for putting me in touch with the right people to start this
project, and everyone at Elsevier who had a hand in this process, particularly
Katey Birtcher, Alice Grant, Andreh Akeh, and Beula Christopher. The initial
three reviewers provided helpful suggestions relating to content that could aug-
ment one’s understanding. Additionally, I would like to thank all of my various
math and statistics teachers, professors, and mentors who trained me to be the
statistician I am today. Furthermore, I would like to thank the students at Sweet
Briar who gave me cause to create this text. Finally, I must especially thank my
wife, Michelle, who first suggested I turn my class notes into a textbook a full
year prior to my contact with Elsevier, and also to my family for their support.

xix
Chapter 1

What is statistics and why is it

important?
Contents
1.1 Introduction 3 1.2.4 Data description 5
1.2 So what is statistics? 4 1.2.5 Statistical inference 5
1.2.1 The process of statistics 4 1.2.6 Theories/decisions 6
1.2.2 Hypothesis/questions 4 1.3 Computation and statistics 6
1.2.3 Data collection 5

1.1 Introduction
On August 5, 2009, the Technology section of The New York Times ran the
following headline: “For Today’s Graduate, Just One Word: Statistics.” The ar-
ticle’s author, Steve Lohr, used the moment to argue that the most important
skill for college graduates entering the workforce was—and is still today—
statistics [1].
Statistics, as an academic discipline, is relatively young. Aristotle wrote on
chemistry in the 4th century BCE, and calculus was discovered by Newton and
Leibnitz in the late 17th century. The first major papers on statistics were pub-
lished in the late 19th century, but many of the foundational statistical techniques
still used today were developed in the 1920s.
What changed from that time to today? Before the past decade, the field of
statistics was generally regarded with skepticism, dismissed with pithy quips
that are still quoted today—for a few examples, consider Benjamin Disraeli’s
famous salvo, “There are three types of lies: lies, damned lies, and statistics,” or
George Canning’s, “I can prove anything by statistics except the truth.” How-
ever, the days of the statistician punch line seem to be passing. Now, being a
statistician or data scientist is consistently regarded by many outlets—including
U.S. News and World Report [2]—as one of the best jobs of the year.
What changed was the importance of data. Nearly every business field or
academic discipline now recognizes the importance of data in decision-making.
Data supports or disproves theories and pushes the scientific process forward in
biology and chemistry. Data fuels policymaking in economics and political sci-
ence. Data pervades all fields at all times, to the point that the World Economic

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00010-9 3

Copyright © 2022 Elsevier Inc. All rights reserved.
4 Basic Statistics With R

Forum declared data a new class of economic asset, akin to currency or precious
metals [3].

1.2 So what is statistics?

If data is what fuels decision-making, statistics is what converts that potential
energy into meaningful action. However, it is not generally thought of in this
way. When asked, many people think of statistics as an obscure bit of trivia,
something akin to Cal Ripken’s batting average in 1983—it was .318 by the
way. Thought of this way, statistics becomes a useless subject and irrelevant
to developing decisions. Or perhaps, a little closer to correct, you think that
statistics is just a number calculated of data. When thought of this way, statistics
becomes a field of formulas, merely something to be memorized and repeated
at an instant. Neither of these conceptions give statistics its due.
Statistics is ultimately the field of study that translates data into decisions,
theories, and knowledge. Put another way, it is the art or science of learning
from data. As such, statistics extends well beyond calculating results from for-
mulas. It reaches into all parts of the scientific decision-making process. It is an
essential part of any field that intends on trying to use data to reach any form of
conclusion.
For example, baseball has long been a sport that was rich in numbers. For
years, few people attempted to harness the power available in this data. In 2001,
Oakland Athletics General Manager, Billy Beane, transitioned toward a more
data-driven decision-making process based on sabermetrics, the study of sports
statistics. That season, during which the Athletics won over 100 games, was
chronicled by Michael Lewis in his 2003 best seller Moneyball [4]. In the years
since, every organization in Major League Baseball has moved toward using
the vast amounts of data to help identify undervalued players and new strategic
insights.

1.2.1 The process of statistics

The process of statistical decision-making comes in five steps that mirror the
scientific process: (1) Hypothesis/Questions, (2) Data Collection, (3) Data De-
scription, (4) Statistical Inference, and (5) Theories/Decisions. (See Fig. 1.1.)
Over the course of this book, we will go through each of these steps in
various levels of detail, focusing mostly on Data Description and Statistical In-
ference. For the moment, we will give a few words of description to each of
these steps.

1.2.2 Hypothesis/questions
We always want to try to answer questions about the world. We see an event,
some sort of phenomenon, and we want to try and understand it. With that in
mind, we ask questions or develop hypothesis to explain what we have seen.
 What is statistics and why is it important? Chapter | 1 5

FIGURE 1.1 The Process of Statistics. Notice that this process is arranged as a circle as, like the
scientific process, the process of Statistics is a constant procedure, continuously looking for better
solutions and better decisions as time goes on.

For example, medical professionals may want to try to see if the use of exter-
nal warming, such as blankets or forced-air warming blankets, helps decrease
instances of hypothermia in the operating room. Agricultural researchers may
want to know what types of fertilizer give the best yields. Political pollsters may
want to gauge public opinion of a law to be enacted.

1.2.3 Data collection

All of these questions need answering, and in order to obtain these answers we
need data. Definitive answers would require measuring values under all possi-
ble conditions or surveying every member of the population. In practice, this is
impossible, either due to the cost or time involved in considering every person
or possibility.
So, we need a smaller group of people or objects, called a sample, from
which to get data in order to answer our questions. Usually this data is obtained
through one of two methods: designed experiments and observational studies.
The way we collect our data ultimately affects the questions we can answer, as
well as possibly affecting the results of our study.

1.2.4 Data description

Once we have our data, we of course want to use it to answer our questions.
However, the datasets that we collect are often very large, as some datasets can
include hundreds of thousands of observations or variables. We cannot answer
questions with such unwieldy information, so we must summarize the data in
some way to make our data more manageable. These summaries can be visual,
such as scatterplots or histograms, or numerical, such as means or medians. All
of these summaries are generally referred to as descriptive statistics.

1.2.5 Statistical inference

While summarizing the data through descriptive statistics helps make the data
easier to deal with, it cannot be used to make decisions. This is because our
6 Basic Statistics With R

descriptive statistics are based on our sample, and no two samples are the same.
Ideally, we should make the same decision regardless of our data, at least most
of the time. Statistical inference helps us to make decisions by quantifying how
unlikely our data is assuming a specific state of the world. If our collected data
is rare enough, we conclude that our assumed state of the world is incorrect.

1.2.6 Theories/decisions
Once we complete our statistical inference, the subject matter experts—the
economists, the scientists, etc.—have to use the results accordingly. If you are
the expert in the field, it is important to understand what the results state. Other-
wise, you must be able to effectively communicate the results of your statistical
inference to other individuals, a skill that comes only with practice.

1.3 Computation and statistics

While these five steps are the basics of statistical thinking, there is one final
aspect that runs throughout the whole statistical process: computing. All of the
calculations that do go into statistics can be very tedious, or in some cases im-
possible, especially as the size of datasets increase. In order to overcome this,
we will need some form of statistical software in order to find probabilities,
collect data, and calculate summaries of our data.
A variety of statistical software programs exist, but in this book we will con-
centrate on using R [5], an open-source software freely available for download
online at the Comprehensive R Archive Network, https://cran.r-project.org/.
Throughout this book, we will go through the code required to accomplish many
of the statistical methods we discuss, along with several supporting examples
and opportunities to practice the code.
Chapter 2

An introduction to R
Contents
2.1 Installation 7 2.5 Vectors 11
2.2 Classes of data 7 2.6 Data frames 12
2.3 Mathematical operations in R 8 2.7 Practice problems 13
2.4 Variables 9 2.8 Conclusion 14

2.1 Installation
R is an open-source programming language that is used for data analysis and
graphics creation. Across a wide variety of academic fields, R is the standard
statistical software and boasts an extensive online community. As such, solu-
tions to problems one might encounter in working with R can easily be found
through a quick web search. Because of these reasons, we will be using R to
conduct our analyses in this book.
In order to install R, you have to download the program—be sure to choose
the correct version for your Windows or Macintosh machine—from the Com-
prehensive R Archive Network at https://cran.r-project.org/. The installation
wizard is simple to follow, and thus will not be explained in detail.
The R console is fairly simple for both Windows and Mac operating sys-
tems. As a coding language, there are minimal point-and-click aspects to the
system. All code for calculations, statistical methods, and creating graphics is
typed into the console and executed by hitting the <Enter> key. The output
for calculations and statistical methods will be seen in the same console where
the code is typed. When creating graphics, other windows will open in the pro-
gram containing the various plots that we create. Unless otherwise noted, there
is no difference in the code to complete a task between Windows and Macintosh
machines. (See Fig. 2.1.)

2.2 Classes of data

In R, there are many classes of data, but we will only focus on two: numeric and
character. Numeric data is exactly what it sounds like. It is data that is made up
of numbers, whether whole numbers or decimals. Numeric data can be acted on
by standard mathematical operators, which we will discuss shortly. Character
data is made up of a string of letters and numbers. It cannot be acted on by

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00011-0 7

Copyright © 2022 Elsevier Inc. All rights reserved.
8 Basic Statistics With R

FIGURE 2.1 The R console for Windows (Top) and Mac (Bottom) operating systems.

standard mathematical operators, but it can be acted on by logical operators. In

order for R to recognize a character string, it has to be entered using quotes. We
will come back to this a little later on.

2.3 Mathematical operations in R

Adding, subtracting, multiplying, dividing, and using exponents in R is a very
simple process. In fact, it uses the same conventions as many other programs,
including Microsoft Excel, to perform these operations. Unsurprisingly, the four
 An introduction to R Chapter | 2 9

arithmetic operations are done by using +, −, ∗, and /. Exponential operations

are done using the caret symbol (ˆ) or two asterisks ∗∗. (See Fig. 2.2.)

FIGURE 2.2 Basic mathematical operations in R.

To do the calculations, you type the formula into the R console and hit
<Enter>. Be sure to be careful about your order of operations and, if needed,
your parentheses, as R will follow the standard order of operations learned in
math class.
R has particular conventions about parentheses in particular. Where the
mathematical formula 2(1 + 2) is completely valid, R does not know how to
read this, and will return an error. In order to enter this in a way that R will
understand, you will have to say

2*(1+2)

which will return the answer 6. In addition, for every parenthesis that you open
in an equation, a closing parenthesis is necessary. If an open parenthesis does not
have a matching close parenthesis, R will not return your answer, but will return
an empty line with a + sign at the left. To exit out of the line, hit <Escape> and
reenter your formula. (See Fig. 2.3.)

FIGURE 2.3 A parenthesis error in R. Exit the line by hitting the <Escape> key.

2.4 Variables
Variables are the crux of most of what we will be using in R. These variables
are saved data of either numeric or character values. The simplest variable is a
variable with a single saved value. In general, to save a variable with a specific
value in R, you type

variable name = value

10 Basic Statistics With R

and then hit <Enter> to save the variable. From that point onward that variable
will be associated with that value, until you either exit R or overwrite the value.
For example, if you wanted to assign the value 71 (inches) to a variable named
height, you would enter

height = 71

and then hit <Enter>. From that point on, height will be associated with the
value 71, and can be acted on by mathematical operators. For example, say you
wanted to find the value of height in feet. You would want to divide the height—
71 inches—by 12. To do this in R, you would type

height/12

and hit <Enter> to get the height in feet value of 5.9167. If you wanted to
keep that value around, it could be stored in a new variable by using that exact
formula, for example,

heightfeet=height/12

At which point the variable heightfeet will be associated with the value
71/12=5.9167. It is important to note that for variable names in R, you can
use any combination of letters, numbers, and select special characters such as
periods or underscores. Spaces and other special characters—such as the excla-
mation point or commas—are not allowed in variable names. If a variable name
breaks a naming rule—a disallowed character, for example—R will return an
error, saying “Error: unexpected input” or “Error: unexpected symbol.” (See
Fig. 2.4.)

FIGURE 2.4 Working with numeric, single-valued variables in R.

One other important aspect of variable names is that they are case sensi-
tive, meaning that capitalization matters to variable names. To R, the variable
named height—what we defined to be equal to 71—and the variable named
Height—something we never created in R so it does not exist—are two differ-
ent variables. If you enter Height into the console, R will not return the value 71
but will tell you that the variable “Height” is not found.
 An introduction to R Chapter | 2 11

Character variables have a similar entering process in R, with one key dis-
tinction. Character values must be entered using quotes, otherwise R will return
an error. The entry process is again

variable name = "value"

and then hit <Enter> to save the variable. The variable is then associated with
that value until exiting R or until it is overwritten. For example,

name = "Stephen Loftus"

and the variable name will be associated with the value “Stephen Loftus” from
that point onward. Note that if you do not put the value of your variable in
quotes, R will return an error. All of the naming conventions for character vari-
able names and numeric variable names are identical. (See Fig. 2.5.)

FIGURE 2.5 Working with character, single-valued variables in R.

2.5 Vectors
Variables can be more than just single values. Commonly in our datasets, vari-
ables will have one value for every participant in our study. These multiple
values are stored in vectors in R, which will be stored in data frames (more
on them in a minute). In R, vectors are defined using the notation

vector name = c( value 1 , value 2 , ... , last value )

So, for example, if you recorded heights for five people and stored them in a
vector named height, it would be

height = c(75, 74, 67, 83, 75)

and then hit <Enter>. The variable height will now be associated with those five
values until exiting R or overwriting. Vectors can be acted on using arithmetic
operations similar to single-value variables. For example, to find the height in
feet of these five individuals, you would divide the whole height vector by 12,
using

height/12
12 Basic Statistics With R

This would return 6.25, 6.17, 5.58, 6.92, and 6.25, the heights of each of these
five people in feet. These values can be stored in a vector of their own if desired.
In addition, two vectors of the same length can be added together, with the
first elements of both vectors being added together, the second elements added
together, etc. (See Fig. 2.6.)

FIGURE 2.6 Working with numeric vectors in R.

Saving vectors of character values is identical to numeric, with the exception

of entering the values within quotes. It is important to note that you cannot
mix numeric and character values in the same vector. If you attempt to, R will
automatically change all numeric value to characters. For example, if we wanted
to save the names of five noted athletes in a vector called name, we would use
the code

name = c("Alex Ovechkin", "Mike Trout", "Lionel Messi",

"Giannis Antetokounmpo", "Patrick Mahomes")

2.6 Data frames

The most common way that we will receive our data in R is through data frames.
In a data frame, each row gives you all the variable values for one observation in
your study, while each column gives you all the values for a variable. To create
a data frame with vectors that we already created in R, we use

data frame name = data.frame(vector 1 name , ... ,

last vector name)

and hit <Enter>. Naming conventions are identical for data frames and vari-
ables. It is important to note that in creating data frames, we need to ensure the
ordering for each vector is the same. In other words, the first values in vector 1
and vector 2 both belong to observation 1, the second values from observation 2,
and so on. So, for example, to create a data frame name athlete with the athlete’s
names and heights from the previous section, we would use

athlete=data.frame(name, height)
 An introduction to R Chapter | 2 13

Printing this out in R, we can see that Alex Ovechkin is 75 inches tall, Mike
Trout is 74 inches tall, etc. Data frames are able to have both character vectors
and numeric vectors in the same data frame. (See Fig. 2.7.)

FIGURE 2.7 Working with data frames in R.

2.7 Practice problems

Consider the following set of attributes about the American Film Institute’s top-
five movies ever from their 2007 list.
1. What code would you use to create a vector named Movie with the values
Citizen Kane, The Godfather, Casablanca, Raging Bull, and Singing in the
Rain?
2. What code would you use to create a vector—giving the year that the movies
in Problem 1 were made—named Year with the values 1941, 1972, 1942,
1980, and 1952?
3. What code would you use to create a vector—giving the run times in minutes
of the movies in Problem 1—named RunTime with the values 119, 177, 102,
129, and 103?
4. What code would you use to find the run times of the movies in hours and
save them in a vector called RunTimeHours?
5. What code would you use to create a data frame named MovieInfo containing
the vectors created in Problem 1, Problem 2, and Problem 3?
Consider the following set of attributes about a series of LucasArts—an
early video game company under the umbrella of George Lucas’s Lucasfilm
company—video games.
6. What code would you use to create a vector named Title with the values
The Secret of Monkey Island, Indiana Jones, and the Fate of Atlantis, Day
of the Tentacle, and Grim Fandango?
7. What code would you use to create a vector—giving the year that the games
in Problem 6 were released—named Release with the values 1990, 1992,
1993, and 1998?
8. LucasArts was founded in 1982. What code would you use to calculate
how many years after the founding of the company was the game released?
14 Basic Statistics With R

9. Each of these games fall under the genre of adventure games. In 2011,
adventuregamers.com created a ranking of the top 100 adventure games of
all time. Create a vector—containing the rankings of the aforementioned
games—named Rank with the values 14, 11, 6, and 1 [6].
10. What code would you use to create a data frame called AdventureGames
containing the vectors contained in Problem 6, Problem 7, and Problem 9?

2.8 Conclusion
R provides a flexible coding framework in which we can store entered data—
both numbers and character strings—in vectors and data frames alike. Addi-
tionally, it can act as a calculator, doing standard mathematical operations on
numbers or inputted data as a whole. Our next goal will be to select individ-
ual values from a vector, variables from a data frame, and rows—representing
observations—from a dataset.
Chapter 3

Data collection: methods and

concerns
Contents
3.1 Introduction 17 3.4.1 Practice problems 23
3.2 Components of data collection 17 3.5 Observational studies and
3.3 Observational studies 18 experiments: which to use? 23
3.3.1 Biases in survey sampling 19 3.5.1 Practice problems 24
3.3.2 Practice problems 21 3.6 Conclusion 25
3.4 Designed experiments 21

3.1 Introduction
The first step in the process of statistics is the development of a hypothesis or
question to test. While careful thought does need to be taken when choosing the
question you want answered, we will assume that you have a clear definition of
the question you want answered. With this in mind, we move on to the second
step in the process: collecting data.
Data collection is one of the more overlooked roles of statistics in science,
but is crucial nonetheless. If data is not carefully collected, it is very possible
that we will influence the results of our study or possibly make it so that we
are unable to answer our question at all. In this chapter, we will look at two
of the most common methods of data collection—observational studies and de-
signed experiments—and see what each of these methods bring to the process
of statistics.

3.2 Components of data collection

As we mentioned in the Introduction, the goal of statistics is to answer some
question, specifically about a population that we are interested in. A population
is the entire set of subjects of interest, while a subject is an individual or object
that we will measure. Again, it is generally impossible for us to gather data on
the entire population, so in order to answer these questions we try and collect
data on a subset of the population. This subset that we collect data on or have
data on is called a sample.
As we collect our data, we are generally recording variables of interest from
our subjects. Broadly speaking, a variable is any recordable aspect of our sub-

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00013-4 17

Copyright © 2022 Elsevier Inc. All rights reserved.
18 Basic Statistics With R

ject, regardless of what it is. There are the many types of variables that we
can record, but we will concentrate on two main types. Quantitative variables,
sometimes referred to as numeric variables, are any variable where the recorded
value is a number and that number indicates some sort of magnitude. Exam-
ples of quantitative variables could include height, weight, or age. Qualitative
variables, more commonly referred to as categorical variables, are any variable
where the recorded value is a category. Sex, political affiliation, job, or academic
class are all examples of categorical variables.
As an example, consider a situation where we collect 5-digit zip codes from
our subjects. Would a zip code be considered a quantitative variable or a cat-
egorical variable? Zip codes are comprised of numbers, which might imply
quantitative. However, the numbers of zip codes do not indicate any magni-
tude. A zip code 34815 does not imply anything better or worse than a zip code
of 26970. As such, zip codes are categorical variables.
In order to answer our questions using data, we of course must gather sam-
ple. Generally speaking, there are two ways we can get our data: observational
studies and designed experiments. Both methods have their strengths and weak-
nesses, as well as scenarios when one method is better than the other. To begin,
let us look at observational studies, how they are conducted, and questions we
need to be concerned about in order to properly conduct such a study.

3.3 Observational studies

Observational studies are studies where researchers observe and collect vari-
ables about a subject without affecting the subject. This is a passive method of
data collection, as the researcher does not manipulate the conditions that the
subjects experience.
The most common form of observational study is the survey. In surveys, a
researcher selects a sample of subjects from the population of interest and gath-
ers information from them. Therefore, in order to conduct a survey, we would
need a list of every member of the population. This complete list is referred to as
the sampling frame. In practice, it can be impossible and at best very difficult
to get the full list of all members of the population. However, there are methods
of sampling outside the scope of this book that do not necessarily require such
a list. For the sake of simplicity, however, we will assume going forward that a
sampling frame is available.
Once we have the sampling frame, we have to choose our sample. This is
not a trivial step, as all statistical inference from observational studies hinges on
two assumptions about our sample.
1. We assume that our sample is representative of our population.
2. We assume that all members of the population have an equal chance of being
chosen in our sample.
If our sample does not reflect our population of interest, we cannot trust any
of the results that we come to over the course of the statistical process.
 Data collection: methods and concerns Chapter | 3 19

In order to ensure that these assumptions are met, we have to use a ran-
dom sample of subjects from our population. There are many different ways to
choose a random sample, but for this chapter we will concentrate on the simplest
form of a random sample: the Simple Random Sample (SRS).
In an SRS, we start—as with all forms of random sampling—with our sam-
pling frame. Each subject in the sampling frame is assigned a number or some
form of identifier, and then we select n numbers at random to create our sam-
ple. In this way, each member of the population has an equal chance of being
selected, and on the average our sample will reflect the population.
An important aspect of the SRS is the selection of random numbers and
how we generate them. We cannot just choose n numbers that come to mind to
make up our random numbers. They must be generated through some random
process—such as drawing numbers out of a hat—or through computational soft-
ware such as R. This is because, as amazing as the human mind is, it is incapable
of replicating true randomness. We will look into how we choose random num-
bers for our sample using R in the next chapter.
While there are many more complicated—and arguably better—sampling
methods, the SRS illustrates one of the foundations of statistical inference: ran-
domness. Further, by looking only at the SRS, we are able to see possible biases
existent in sampling more clearly, a key consideration in collecting a sample.

3.3.1 Biases in survey sampling

True random sampling can be difficult in practice. Whether due to the difficulties
of getting a truly comprehensive sampling frame or generating random numbers,
we often have to make compromises in our sampling. However, when we do
make these compromises, we have to be sure that we avoid introducing bias
into our data. Bias is any time that a sample’s responses favor some part of the
population disproportionally.
There are many examples of bias that are commonly seen in survey samples.
Sampling bias is introduced through the way that we collect our data. Sampling
bias can be seen in leading questions, or questions that are intended to provoke
a specific response. The question “Do you think that good parents read to their
children?” is designed to elicit a different response than the more measured “Do
you think that children who are read to experience benefits?”
An additional form of sampling bias comes from collecting data using con-
venience or volunteer sampling. This is when data is collected on individuals
who are readily available or who volunteer to take a survey. Convenience and
volunteer sampling can lead to particularly poor conclusions, as people who are
convenient for sampling or readily volunteer to participate in a sample often
share common traits, thus weighting their group’s opinions more heavily in the
results.
The most famous example of sampling bias occurred in the polls of the
1948 presidential election. In this election, pollsters and pundits alike expected
20 Basic Statistics With R

a landslide victory for candidate Thomas Dewey over incumbent Harry Truman.
Despite the seeming certainty of the result, President Truman was reelected in a
landslide, winning the popular vote by 4.5%.
How were the polls wrong, leading to one of the most famous pictures in
American political history? It turns out that the polls leading up to the elec-
tion were tainted twice by sampling bias. First, several major polls stopped
collecting data two weeks before election day, not allowing them to track the
ever-shifting public opinion toward Truman. In fact, some—admittedly less
scientific—polls that showed a Truman lead were discontinued due to their “ir-
regular” results. Even more importantly, several major polls showed particular
bias toward Dewey because they were conducted via telephone. At the time
telephones were more commonly found in wealthy homes, and wealthy individ-
uals were more likely to vote for Dewey, resulting in poll results showing clear
Dewey leads right up until election day.
A second type of bias common in surveys is nonresponse bias. Nonresponse
bias is introduced because people did not respond to a survey, possibly for a
common reason. This extends beyond the common complaint “Why do tele-
marketers or survey people always call around dinner time?” Nonresponse bias
can be a symptom of some common trait among the nonresponders. For exam-
ple, if a researcher fails to reach individuals via a phone survey in the evening,
it may be that the nonrespondents are, say, forced to work a second job in the
evenings due to lower incomes.
In Lahaut et al. (2002) [7], researchers sent out surveys to 310 individu-
als trying to asses rates of alcohol consumption across various socioeconomic
conditions. Of the 133 responses to their survey, they found that 27.2% of
respondents abstained from alcohol use. At a later date, they were able to con-
tact 80 nonrespondents to their original survey and asked them the same set
of questions. They found that 52.5% of this nonrespondent group abstained
from alcohol use, a drastic difference from their original survey results. Had the
nonresponse bias been ignored, they likely would have come to very different
conclusions for their study.
A final type of common bias is response bias. Response bias occurs when
respondents give incorrect responses to the surveyor, whether intentionally—
i.e., lying—or unintentionally—i.e., failing to remember the proper response for
them. Response bias is harder for researchers to control, as it entirely depends
on the respondent’s honesty and accuracy.
One instance where response bias partially influenced results comes from
the 2016 presidential election. That year, Hillary Clinton was widely expected to
defeat Donald Trump by a landslide; however, Trump won the Electoral College
and a higher share of the popular vote than expected. Many news outlets [8]
pointed to “Shy Trump Voters,” voters who did not say that they would vote
Trump due to social desirability biases, as a major reason why the polls were so
inaccurate. (See Fig. 3.1.)
 Data collection: methods and concerns Chapter | 3 21

FIGURE 3.1 County by county results from 2016 presidential election. Some analysts believed
that “Shy Trump Voters,” an example of response bias, swung the election.

When creating a survey, you can only directly control the sampling bias of
your survey. Response and nonresponse bias entirely come from the respon-
dents, and are out of a surveyor’s control. With this in mind, while there are
some methods to correct these latter two biases—such a demographic weight-
ing for nonresponse bias—the majority of efforts to correct biases are directed
toward avoiding any form of sampling bias.

3.3.2 Practice problems

1. A researcher conducts a pair of surveys, one set in person and one via e-mail,
to determine the rate of depression among college students. In the in-person
interviews, only 10% of respondents admitted to having dealt with depres-
sion in the past year, while in the e-mail survey 34% admitted to depression.
Why might this be an example of a response bias in the survey?
2. A surveyor conducts a phone survey calling the home phones of his sample
between the hours of 9 AM and 5 PM on weekdays. What type of bias could
this introduce into their data?

3.4 Designed experiments

The design of experiments represents an entire field within the subject of statis-
tics. However, we will only give a brief overview of the topic, concentrating on
the basic principles of designed experiments through the simplest experimen-
tal design. An experiment is a study in which the researcher assigns subjects
to specific conditions and observes the outcomes. In contrast to observational
22 Basic Statistics With R

studies, experiments are an active method of collecting data, as the researcher

manipulates the conditions that the subjects experience.
The goal in an experiment is to see what explanatory variables affect a given
response. An explanatory variable is any variable that helps us understand or
predict the response, while the response is the variable of interest that we want
to measure. In order to do this effectively, researchers compare treatments, or
the combinations of levels of the explanatory variables.
In experimental design, there are two principles that we are primarily con-
cerned with in this course: randomization and replication. Randomization is
achieved in experiments through the random assignment of subjects in our sam-
ple to treatments. Just as random sampling was essential to survey sampling,
randomization is essential for designed experiments. Replication is the process
of assigning multiple subjects to the same treatment in order to ensure more
accurate results. We will discuss these principles through the simplest type of
experiment, the Completely Randomized Design (CRD).
In a CRD, each subject in the sample is randomly assigned to a treatment,
generally with each treatment getting an identical number of subjects. We will
discuss how to do this in R in the next chapter. In assigning our treatments in
this way, we ensure that the experiment is both randomized and replicated.
The emphasis on replication makes sense, as it’s intuitively understood that
the more subjects in a study, the more reliable the results. However, why is ran-
domization so important? There are many reasons for this, all which are crucial
to reliable and trustworthy research. The first reason for randomization is that
randomization helps us to avoid bias in our experiment. If we were to design
our experiment in such a way that one treatment was more likely to be better or
worse, we could not trust the results.
The second reason why randomization is important is that randomization
helps us to avoid accusations of bias. Having randomization helps other re-
searchers to trust that we did not affect the results of our studies to our benefit
through bias, which lends credibility to the outcome of our experiment.
The third reason for randomization is that randomization ensures objective
treatment assignment. In medical studies, a researcher does not want to assign
a patient to the placebo group of an experiment. Generally, placebos do not
result in improved condition, and researchers want their subjects to ultimately
have an improved medical condition. However, without the baseline provided
by placebos, researchers would not know the effectiveness of treatments.
Finally, randomness helps us control for confounding variables. A con-
founding variable is a variable that is related to both the explanatory variable
and response, which falsely makes it look like the explanatory variable predicts
the response.
A notable example of confounding variables was found in Tonks (1999) [9].
This initial observational study found that myopia in children was related to the
use of nightlights at ages 2 and under. Despite this, later designed experiments
found no such relationship. It was later discovered that myopic parents were
 Data collection: methods and concerns Chapter | 3 23

more likely to use nightlights to make it easier to check on their children at

night, and myopic parents were naturally more likely to have myopic children.
The confounding variable of myopia in parents made it look like the use of
nightlights—common for myopic parents—was related to myopia in children—
also common for myopic parents. Where observational studies were not able to
pick this up, designed experiments were able to weed out the influence of the
confounder.

3.4.1 Practice problems

3. A drug company is interested in testing whether their blood pressure medi-
cation is more effective at lowering blood pressure than the current industry
standard. To test this, they look at two variables: the blood pressure medica-
tion (Drug A and Drug B) and the exercise level for the subject (1 hour per
week, 5 hours per week). What are the response, explanatory variables, and
treatments in this experiment?
4. An educator wants to know if team-based learning is more effective than
the traditional model. To test this, students were assigned to one of the two
teaching methods (team-based learning or traditional learning), and their
improvement was measured by taking the difference in pre-class and post-
class examinations for each student. What is the explanatory variable, the
response, and the treatments for this experiment?

3.5 Observational studies and experiments: which to use?

So, between observational studies and designed experiments, which method of
data collection should you use for your study? This depends on many factors,
but we will focus on three: budget, ethical concerns, and the goals of the study.
When it comes to budget, observational studies are generally preferred to de-
signed experiments. In terms of both time involved in the study and the financial
cost of gathering the data, observational studies are less expensive than experi-
ments.
Observational studies also generally raise fewer ethical concerns than exper-
iments. Because designed experiments involve actively manipulating the condi-
tions that subjects will endure, there are many instances where an experimenter
cannot assign subjects to a treatment without causing harm to the subject. While
it is a generally understood fact that smoking causes cancer, it would be un-
ethical to run an experiment to show the link between smoking and cancer in
humans. This is because such an experiment would involve forcing an individ-
ual to smoke, an ethical issue to be sure.
If observational studies are preferred in matters of budget and ethics, then
why do we conduct experiments? The reason is that experiments have a major
advantage in the questions that are able to be answered by a study. Experiments
are able to establish what is called causality, that an explanatory variable causes
24 Basic Statistics With R

Observational studies Designed experiments

• Passive data collection • Active data collection
• Less expensive • More expensive
• Fewer ethical concerns • More ethical concerns
• Observe explanatory • Manipulate explanatory
variables in subjects variables in subjects
• Randomly select • Randomly assign
subjects to observe subjects to treatments
• Cannot control • More control over
confounding variables confounding variables
• Cannot establish causality • Can establish causality

something in the response. Observational studies cannot establish that one vari-
able causes another, only that they are related in some way.
This is because experiments control for confounding factors through ran-
domization in a way that observational studies are unable to do. Remember the
example from earlier about confounding variables, in which it appeared that us-
ing nightlights earlier in life led to myopia in children. The observational study
did not consider the confounding variable of myopic parents, which ultimately
did explain the myopia in the children. The experiment assigned families to ei-
ther the nightlight or no nightlight group randomly. Some families with myopic
parents would have been forced to not use a nightlight, despite their possible
preference. This would have lessened the chance that the confounding variable
would have made it appear that myopia was related to nightlight use.
So, with these concerns in mind, we can choose between observational stud-
ies and experiments accordingly. If you want to establish causality between your
explanatory variable and response, you have to use an experiment. If you only
want to establish that the variables are related in some way, an observational
study will likely be sufficient and also less expensive and prone to ethical con-
cerns than experiments.

3.5.1 Practice problems

5. A researcher wants to look into the relationship between hours of exercise
weekly and weight. They survey 200 people, asking how many hours a week
they exercise and then measuring their weight. Based on the results of the
study, the researcher concludes that exercising more causes lower weights.
Why is the researcher unable to make this statement?
6. In the observational study in problem 5, what are the subjects, the explana-
tory variable, and the response?
7. For this observational study, what could some potential confounding vari-
ables that could invalidate the results?
 Data collection: methods and concerns Chapter | 3 25

8. What would be a more effective way to investigate the causal relationship

between exercise and weight?

3.6 Conclusion
While not often thought of as statistics proper, data collection is an essential
portion of the process of statistics. As such, it needs to also be considered with
a statistical mindset. Thus, we have a myriad of considerations, including cost,
ethics, the types of data collected, the method of collection, and the ways that
the sample could have been biased. Each of these considerations can greatly
affect the questions that can be answered using our data, particularly focusing
on whether or not we are able to make causal claims about our data. Ignoring
these aspects of statistics can lead to lost time, money, and the inability to answer
desired questions.
Once we have data, our next concern will be gaining an understanding of
our data. This is done through looking at our variables, summaries both numeric
and graphical, through a process called Exploratory Data Analysis, or EDA.
Chapter 4

R tutorial: subsetting data,

random numbers, and selecting
a random sample
Contents
4.1 Introduction 27 4.5 Select a random sample 32
4.2 Subsetting vectors 27 4.6 Getting help in R 33
4.3 Subsetting data frames 29 4.7 Practice problems 33
4.4 Random numbers in R 31 4.8 Conclusion 35

4.1 Introduction
In this chapter, we will concentrate on how to choose a random sample from
a sampling frame. In order to do this, we will need to build up a few skills
first. Initially, we will talk about subsetting vectors and data frames, followed
by a little information about how R generates random numbers. Finally, we will
discuss the sample function to select a sample of numbers, and then combine
all the information to select a random sample from a sampling frame. In all of
these examples, it is assumed that you have a sampling frame loaded into R (we
will discuss how to do this in Chapter 5).

4.2 Subsetting vectors

In many instances when working with data in R, we do not want to work with
the entire vector or data frame, but merely a portion of it. In other words, we
want to work with a subset of the data. To begin, we will look at how to select a
subset of a vector. Say that we have a vector hours defined below

hours=c(8.84, 3.26, 2.81, 0.64, 0.60, 0.53, 0.37, 0.35, 0.31, 0.24)

This data is the average daily hours spent on the ten most popular daily activities
in 2019, according to the Bureau of Labor Statistics [95]. R subsets data using
the square brackets [] after the vector. You can place any vector indices or logical
statements to select specific elements of the vector. For example, if you wanted
to know the first element of the vector age, you would type

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00014-6 27

Copyright © 2022 Elsevier Inc. All rights reserved.
28 Basic Statistics With R

hours[1]

and hit <Enter>. R will return the value 8.84, or the average number of hours
spent sleeping—the activity with the most amount of time spent on it. With
that in mind, how would you look up the eighth entry of the hours vector, or
the average number of hours spent socializing or communicating in 2019—the
eighth-most popular daily activity? You would type hours[8] and hit <Enter>.
It is possible to look up more than one entry in the vector in a single line of
code. In addition to having a single number inside the square brackets, you can
have a vector with all the entries of interest. For example, if you want to select
the hours spent on the most popular, third-most popular, and ninth-most popular
activities—sleeping, watching television, and participating in sports, exercise,
or recreation—you would use the code
hours[c(1,3,9)]

which would return 8.84, 2.81, and 0.31. What code would you use to look up
the age of father and son John Adams (2nd president) and John Quincy Adams
(6th) at their inaugurations? The code would be
age[c(2,6)]

Finally, you can look up specific values in a vector based on a logical state-
ment. That is, you could look up specifically the entries in the vector that are
greater than some value (or less than, or equal to, etc.). For example, if you
wanted to look up the average hours spent for activities that were specifically
than an hour, you would use the code
hours[hours>1]

which will return 8.84, 3.26, 2.81, as these are the only three ages in the vector
that are greater than 1. To use logical statements, we need to know the logical
operators that R recognizes.

Symbol Logical statement

> Greater Than
>= Greater Than or Equal To
< Less Than
<= Less Than or Equal To
== Equal To
!= Not Equal To
& And
| Or

The last two statements help us make multiple statements to subset the data.
For example, if you wanted the average hours spent in the vector that fall be-
tween 30 and 45 minutes—that is, greater than or equal to 0.5 hours and less
than or equal to 0.75 hours—we would use
 R tutorial: subsetting data Chapter | 4 29

hours[hours>=0.5&hours<=0.75]

which would return 0.64, 0.6, and 0.53. As another example, if we wanted to
select the hours that are either less than 15 minutes—0.25 hours—or longer
than 4 hours (not including 0.25 and 4), we would use

hours[hours<0.25|hours>4]

which would return 8.84 and 0.24. So, with this in mind, what code would
you use to find hours that fell between 15 and 30 minutes (including both end-
points)? How about hours that are equal to 0.6 or greater than 2 (not including
2)? For the first, it would be hours[hours>=0.25&hours<=0.5], and the second
is hours[hours==0.6|hours>2].

4.3 Subsetting data frames

Subsetting data frames is a similar process as vectors, with one difference.
Vectors are a one-dimensional entity, while data frames have two dimensions
defined by the rows and columns. So, instead of using [index] that we used for
vectors, we will have to use [row index, column index]. So we have to subset
by either the rows, columns, or both. We will start with both, rows only, then
finally the columns only (referred to by their column names). To start, assume
that we flesh out the vector that we saw earlier to create a data frame named
Activities that has the activity name, average hours spent, and category of the 10
most popular daily activities.

Name AverageHours Category

Sleeping 8.84 Personal Care
Working 3.26 Work-Related
Watching Television 2.81 Leisure
Socializing 0.64 Leisure
Food Preparation 0.60 Household
Housework 0.53 Household
Childcare 0.37 Caring for Household
Consumer Goods Purchase 0.35 Purchasing
Participating in Recreation 0.31 Leisure
Attending Class 0.24 Education

Now, the data frame is basically just a matrix, so we can select an element of
the data frame by defining a specific row-column combination. So, if we wanted
to choose the category of food preparation, we would want to choose the fifth
row and third column of the data frame, using
30 Basic Statistics With R

Activities[5,3]

which returns the value “Household.” So, say, what code would we use if we
wanted the name of the ninth-most popular activity? The code would be

Activities[9,1]

In addition to choosing specific elements of the data frame, we can choose

specific rows of the data frame. We do this by leaving the column index blank
after the comma. This would give the row or rows that you list in the row index.
So, for example, if you wanted the information on Attending Class—the tenth
row—we would use the code

Activities[10,]

which would return the vector with

‘‘Attending Class’’ 0.24 ‘‘Education’’

as its values. We can choose more than one row as well by using vectors in the
row index. For example, if you wanted the information on Sleeping (Row 1),
Socializing (Row 4), and Childcare (Row 7) the code would be

Activities[c(1,4,7)]

which would return

Name AverageHours Category

Sleeping 8.84 Personal Care
Socializing 0.64 Leisure
Childcare 0.37 Caring for Household

So, with this in mind, what code would you use to look up the information
for Watching Television (Row 3) and Housework (Row 6)? The code would be
Activities[c(3,6),].
Thus far, we have concentrated on subsetting data frames using the rows.
However, we can work with the columns in data frames as well. We saw earlier
that data frames are composed of the named vectors we created. The Activities
data frame, for example, is composed of three vectors named Name, Aver-
ageHours, and Category. We may be interested in looking at only individual
vectors. For example, given this dataset, we may only want to look at the cate-
gories of these activities. In R, we can look at columns within a data frame using
the $ operator, specifically

data frame name$variable name

So, to look at the categories of these popular daily activities, we would use
 R tutorial: subsetting data Chapter | 4 31

Activities$Category

which returns Personal, Work, Leisure, etc. How would we look at the average
hours spent on these activities? Activities$AverageHours.
We use this $ variable notation to help us choose rows based on logical state-
ments. For example, we may want to look at the information for all activities that
receive more than an hour daily, using
Activities[Activities$AverageHours>1,]

which would return

Name AverageHours Category

Sleeping 8.84 Personal Care
Working 3.26 Work-Related
Watching Television 2.81 Leisure

How would you look at the information for all activities not classified as
Leisure? Activities[Activities$Category!=Leisure”,]. Note that this answer in-
cludes having Leisure in quotes, as it is a character value.

4.4 Random numbers in R

Now, we are moving toward selecting a random sample, so in order to do so we
will need to create random numbers in R. Now, there is something important to
note about most any random number generator you come across, whether in R
or some other program; the numbers it generates are not truly random. They are
created using some algorithm (R uses something called the Mersenne Twister)
that takes the previous value inputted and outputs a new “random” number.
How can we trust these pseudo-random numbers? The point of all these ran-
dom number generators is to not repeat too often. If a random number generator
repeats too often, it cannot be trusted to be roughly “random.” The Mersenne
Twister, the algorithm that R uses, takes approximately 219937 − 1 runs to re-
peat, which is many more runs than any reasonable sample size that you might
need for a study.
If these algorithms use the previous “random” number to generate the next,
it stands to reason that it needs an initial number to start. This initial number is
referred to as the seed. This can be set by the user or R will default to a seed tied
to the clock time. If you needed to set a seed, the code in R would be
set.seed(Number)

where you supply the number for the seed. If you supply a seed, you will al-
ways get the exact same series of random numbers with that particular seed. For
example, in the R screenshot below, we set the seed to be 8 and then generate
10 “random” numbers between 0 and 1, reset the seed to 8 and generate 10 new
“random” numbers. You will notice that both sets are identical. (See Fig. 4.1.)
32 Basic Statistics With R

FIGURE 4.1 Setting the seed and generating random numbers in R.

4.5 Select a random sample

Now that we roughly know how random numbers are generated, we can generate
a random sample. R has a specific function that will choose the row numbers
from your sampling frame, the function sample.int. In R, functions have specific
inputs that will then return specific output(s). In sample.int, the two inputs are
the number of subjects in the sampling frame and the number of subjects in your
sample. The general code is

sample.int(n=number of subjects in sampling frame,

size=desired number of subjects in sample)

which will return a vector of row numbers that will be selected in the sample.
For example, one of the most important datasets in baseball analysis is the
PitchF/x dataset. This dataset contains a large number of variables—including
pitch velocity, revolutions of the pitch, pitch type, where the pitch crossed the
plate, among others—about ever pitch thrown in a regular season Major League
Baseball game since 2007, 8,812,107 after the conclusion of the 2019 season.
Trying to analyze all these pitches would be difficult at best, so we could choose
to analyze a subset—say of 1000 pitches—of this data. If we wanted to choose
which 1000 pitches to analyze (or in other words, which rows of the data frame),
we would use

sample.int(n=8812107, size=1000)

Note that R does not allow the use of commas to break up thousands or mil-
lions. This code would only tell us which rows to use. In practice, we would want
to look up—or possibly save—those specific rows from the sampling frame. We
would do this by subsetting the rows of the data frame using the output of the
sample.int function. Assuming that the entire PitchF/x dataset is stored in a data
frame called pitch, we would specifically use

pitch[sample.int(n=8812107, size=1000),]

to print out the 1000 rows that we selected. Now, generally we want to save those
chosen rows in a data frame, so we would save that like we would generally save
anything in R. The code could look something like
 R tutorial: subsetting data Chapter | 4 33

sample=pitch[sample.int(n=8812107, size=1000),]

which would save the selected rows to a data frame called sample. Say that we
wanted to select a sample of 5000 pitches from the pitch data frame (same size
as before) and save it to a data frame called pitchsample, What would the code
be? pitchsample=pitch[sample.int(n=7821149, size=5000),].

4.6 Getting help in R

Throughout any analyses, we will occasionally run into errors with code, forget
the arguments that a function takes, or come across a new function that we want
to use. In those instances, we will need help in navigating these functions. If
you know what you are attempting to accomplish in R without knowing the
exact function you need, the simplest way to get help is to Google what you are
attempting to do. This is also the case if you get an unknown error. The R online
community is very numerous and helpful, with websites such as Stack Overflow
having many solutions to unknown errors or help for new functions.
If you are having difficulty with a particular function in R, there is a simple
method to get information about the arguments and outputs of the functions.
Namely, we use the help function or its shortcut ?. The help function takes in
several arguments, but the key one is the function that you need help with. For
example, if you needed help with the sample.int function, you could use either
of the following lines of code.
help(sample.int)

?sample.int

R will then open up a window that provides a description of the function, its
inputs, and its outputs. Even further, this window will provide sample code that
should hopefully make it easier to accomplish what your analysis intends to.

4.7 Practice problems

Suppose we have the following data frame named Colleges:
College Employees TopSalary MedianSalary
William and Mary 2104 425000 56496
Christopher Newport 922 381486 47895
George Mason 4043 536714 63029
James Madison 2833 428400 53080
Longwood 746 328268 52000
Norfolk State 919 295000 49605
Old Dominion 2369 448272 54416
Radford 1273 312080 51000
continued on next page
34 Basic Statistics With R

College Employees TopSalary MedianSalary

Mary Washington 721 449865 53045
Virginia 7431 561099 60048
Virginia Commonwealth 5825 503154 55000
Virginia Military Institute 550 364269 44999
Virginia Tech 7303 500000 51656
Virginia State 761 356524 55925

1. What code would you use to select the first, third, tenth, and twelfth entries
in the TopSalary vector from the Colleges data frame?
2. What code would you use to select the elements of the MedianSalary vector
where the TopSalary is greater than $400,000?
3. What code would you use to select the rows of the data frame for colleges
with less than or equal to 1000 employees?
4. What code would you use to select a sample of 5 colleges from this data
frame (there are 14 rows)?

Suppose we have the following data frame named Countries:

Nation Region Population PctIncrease GDPcapita

China Asia 1409517397 0.40 8582
India Asia 1339180127 1.10 1852
United States North America 324459463 0.70 57467
Indonesia Asia 263991379 1.10 3895
Brazil South America 209288278 0.80 10309
Pakistan Asia 197015955 2.00 1629
Nigeria Africa 190886311 2.60 2640
Bangladesh Asia 164669751 1.10 1524
Russia Europe 143989754 0.00 10248
Mexico North America 129163276 1.30 8562

5. What could would you use to select the rows of the data frame that have
GDP per capita less than 10000 and are not in the Asia region?
6. What code would you use to select a sample of three nations from this data
frame (There are 10 rows)?
7. What code would you use to select which nations saw a population percent
increase greater that 1.5%?

Suppose we have the following data frame named Olympics:

 R tutorial: subsetting data Chapter | 4 35

Year Type Host Competitors Events Na- Leader

tions
1992 Summer Spain 9356 257 169 Unified
Team
1992 Winter France 1801 57 64 Ger-
many
1994 Winter Norway 1737 61 67 Russia
1996 Summer United 10318 271 197 United
States States
1998 Winter Japan 2176 68 72 Ger-
many
2000 Summer Australia 10651 300 199 United
States
2002 Winter United 2399 78 78 Nor-
States way
2004 Summer Greece 10625 301 201 United
States
2006 Winter Italy 2508 84 80 Ger-
many
2008 Summer China 10942 302 204 China
2010 Winter Canada 2566 86 82 Canada
2012 Summer United 10768 302 204 United
Kingdom States
2014 Winter Russia 2873 98 88 Russia
2016 Summer Brazil 11238 306 207 United
States
2018 Winter South 2922 102 92 Nor-
Korea way

8. What code would you use to select the rows of the data frame where the
host nation was also the medal leader?
9. What code would you use to select the rows of the data frame where the
number of competitors per event is greater than 35?
10. What code would you use to select the rows of the data frame where the
number of competing nations in the Winter Olympics is at least 80?

4.8 Conclusion
Where we previously had the ability to input data into R, we now add the ability
to pick off individual data values, columns, and rows within a dataset. This will
allow us to eventually compare groups in order to see if and how they differ.
However, at this time we have still have to enter our data ourselves. We will
next endeavor to load pre-entered datasets into R, either through csv files or
libraries designed for R.
Chapter 5

R tutorial: libraries and loading

data into R
Contents
5.1 Introduction 37 5.4 Loading csv files into R 42
5.2 Libraries in R 37 5.5 Practice problems 43
5.3 Loading datasets stored in libraries 42 5.6 Conclusion 43

5.1 Introduction
Once we have collected data, our next step in the process of statistics is to begin
analyzing our data through exploratory data analyses. In very few cases, we can
analyze data by hand because large datasets make calculation by hand tedious if
not impossible. As such, we need to be able to load data into R to take advantage
of the programs functions and computing ability. This chapter will focus on
libraries in R—which hold data as well as specific functions that help make
analysis easier—and loading data into R—either from libraries or csv files.

5.2 Libraries in R
Because R is entirely open-source software, anyone can create packages of
user-defined functions to help make analysis easier. These packages are called
libraries and are mostly stored on the CRAN. These libraries can be downloaded
and installed in R and then called into use at any time.
The first step to installing the libraries is downloading and installing them,
specifically demonstrating this through installing the MASS package [10] into R.
This process of downloading and installation differs for Windows machines and
Macs (the process is much easier for Macs), so we will detail both here, begin-
ning with the Windows machine.
In order to install packages in R, you must run R in administrator mode on
your computer. We do this by right-clicking on the R icon and selecting “Run as
Administrator” from the menu. (See Fig. 5.1.)
After this, once R opens up, we click on the “Packages” tab on the top menu
and select “Install Packages” from the drop-down menu. (See Fig. 5.2.)

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00015-8 37

Copyright © 2022 Elsevier Inc. All rights reserved.
38 Basic Statistics With R

FIGURE 5.1 Running R as an administrator.

FIGURE 5.2 Installing Packages for Windows.

After this, we need to select a CRAN mirror. Basically, all this does is define
which saved version of the CRAN, which are consistently updated. The choice
of CRAN mirrorusually has no effect on installing the packages. (See Fig. 5.3.)
After this, we are given a full alphabetical list of libraries available in R.
Scroll down to select MASS and click okay. The package will install and can be
called upon whenever the user needs it. (See Fig. 5.4.)
To call a library—and its functions and datasets—into use in R, we use the
library command. The code for this is

library(Library Name)
 R tutorial: libraries and loading data into R Chapter | 5 39

FIGURE 5.3 Selecting the CRAN Mirror for Windows.

And then hit <Enter>. Unless there is an error in your code, R will just show
a blank line and you then can use any of the functions or datasets stored in the
library. So, to call on the functions in the MASS library, we would load the MASS
library using

library(MASS)

For Mac operating systems, the process is considerably simpler. To begin, open
the R application regularly, and click the Packages & Data tab at the top. From
the dropdown menu, select Package Installer. (See Fig. 5.5.)
40 Basic Statistics With R

FIGURE 5.4 Selecting a library from the list.

In the search box, type the name of the package (in this case MASS) and
click Get List. This will return all packages with names including the string of
characters that you searched for. Click on the package you want to install, and
before clicking Install Selected, check the box that says Install Dependencies.
Some R libraries depend on other libraries to work, and this last step ensures
that all the other dependent packages are installed as well. (See Fig. 5.6.)
At this point, the procedure is identical to the Windows listed above. To
call a library—and its functions and datasets—into use in R, we use the library
command. The code for this is
 R tutorial: libraries and loading data into R Chapter | 5 41

FIGURE 5.5 The Package Installer for Macs.

FIGURE 5.6 Searching for libraries on Macs.

library(Library Name)

And then hit <Enter>. Unless there is an error in your code, R will just show
a blank line and you then can use any of the functions or datasets stored
in the library. So, to call on the functions in the MASS library, we would
use

library(MASS)
42 Basic Statistics With R

5.3 Loading datasets stored in libraries

As was mentioned, the libraries of R contain various functions and datasets for
analysis. Loading these datasets into the R workspace is very simple given the
name of the dataset. The data function in R calls the datasets into the workspace,
using the code

data(Dataset Name)

and then hitting <Enter>. As an example, consider the nlschools dataset in the
MASS library [10,11]. This dataset is the test scores for over 2000 8th grade
students in the Netherlands. To load this, we would use the code

data(nlschools)

and then hit <Enter>. There is now a data frame in the workspace named
nlschools that can be used for analysis.

5.4 Loading csv files into R

In addition to datasets stored in libraries, we can also load datasets stored in
csv files into R. The function to be able to do this is the read.csv function. The
function is executed using the code

read.csv(File Location)

The file location can occasionally be difficult to find, and even when known
there are opportunities to make a mistake typing in the file pathway into the
function. With that in mind, we will use the file.choose function inside read.csv.
The code could look something like

data=read.csv(file.choose())

This will cause R to open a window where you are able to navigate to your csv
file and select it to be loaded into R. It is important to note that you have to
save the loaded data frame as something, otherwise the read.csv function will
just print out the dataset, making it unusable it for analysis. So, in the code
above, the data frame that is read into R is saved under the name data, but in
practice you can choose any name you want for the dataset assuming it follows
the naming conventions we laid out earlier.
 R tutorial: libraries and loading data into R Chapter | 5 43

5.5 Practice problems

1. Say that we previously installed the Ecdat [12] library into R and wanted to
call the library to access datasets from it. What code would we use to call the
library?
2. Say that we then wanted to call the dataset Diamond [12,13] from the Ecdat
library. What code would we use to load this dataset into R?

5.6 Conclusion
Up until this point, any data that we wanted to work with in R would have to
be entered by ourselves. This would be a very tedious process, especially as
datasets get larger. Through the use of libraries and the read.csv function, we
are able to easily load datasets into the R workspace. I might add that R is very
flexible in the forms of data that it can work with. With the right functions and
knowledge, R can get data from PDF files, text files, and web pages.
Now that we have ways to load data into R and subset the data into groups,
we can now move into the portion of statistics that comes to mind when people
discuss the subject: analyzing data. Doing so in R will open up a wide array of
functions and graphs to us, allowing us to more express our data in understand-
able terms.
Chapter 6

Exploratory data analyses:

describing our data
Contents
6.1 Introduction 47 6.5 Visual summaries for a single
6.2 Parameters and statistics 47 quantitative variables 57
6.3 Parameters, statistics, and EDA 6.6 Identifying outliers 59
for categorical variables 48 6.6.1 Practice problems 61
6.3.1 Practice problems 50 6.7 Exploring relationships
between variables 61
6.4 Parameters, statistics, and EDA
6.8 Exploring association between
for a single quantitative variable 51
categorical predictor and
6.4.1 Statistics for the center of quantitative response 62
a variable 51 6.8.1 Practice problems 65
6.4.2 Practice problems 53 6.9 Exploring association between
6.4.3 Statistics for the spread two quantitative variables 65
of a variable 54 6.9.1 Practice problems 71
6.4.4 Practice problems 56 6.10 Conclusion 72

6.1 Introduction
Once we have collected data, the next step in the statistical process is begin-
ning to explore the data. Now, we often cannot explore the entire dataset, as the
number of subjects in our study can be quite large. As such, we need to work
with summaries of our data, both graphical summaries and numeric summaries.
These summaries, referred to as descriptive statistics, are not the end of the sta-
tistical process—which does not really have an end as it is continually repeating
itself. Further, we cannot—or should not—make definitive decisions from de-
scriptive statistics. Rather, descriptive statistics can give us a general idea what
is going on in the dataset and point us in the direction of interesting relationships
between variables in the dataset.

6.2 Parameters and statistics

Recall that the goal of the process of statistics is to help us answer questions
from data. Really, we are not answering a question about the population as a
whole as we are answering questions about a single number of the population.
When asking, say, if exercise helps to lower blood pressure, we are really ask-

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00017-1 47

Copyright © 2022 Elsevier Inc. All rights reserved.
48 Basic Statistics With R

ing, “On the average, does the population of individuals who engage in exercise
experience a decrease in blood pressure level?” This single number that de-
scribes the entire population is called a parameter or population parameter.
It is considered fixed and unknown, and in order to go about answering ques-
tions through the process of statistics we need to estimate the value(s) of the
parameter(s) of interest. (See Fig. 6.1.)

FIGURE 6.1 Parameters and statistics.

In order to estimate these parameters, we need our sample, whether from

a designed experiment or observational study. But just like we cannot use the
entire sample to answer a question, we cannot use the entire sample to estimate
a population parameter. Rather, we will use an appropriately chosen summary
number of the sample to estimate our parameters. This is a statistic, or any
number calculated from our sample.
Each type of variable (quantitative and categorical) has different population
parameters and statistics that we have to work with. As there are fewer things to
be concerned with for categorical variables, we will begin there.

6.3 Parameters, statistics, and EDA for categorical variables

Because categorical variables are by nature categories, the parameters and statis-
tics that we are interested in are the percentage of population or sample, respec-
tively, that falls into each category. When we are looking at the population, the
main parameter of interest is the population proportion (notated p), or the per-
centage of the population that falls into some category of interest. This, like all
parameters, is an unknown and fixed quantity that we need to estimate. For the
population proportion, we use the sample proportion (notated p̂), or the per-
centage of the sample that falls into some category of interest. Mathematically,
this would be
Number of Subjects in Category
p̂ = .
Sample size
Assuming that our sample is a good representation of our population, an
assurance usually made good by taking a random sample, the sample proportion
should do a good job estimating our population proportion.
Because we are working with categorical variables and are ultimately inter-
ested in the percentage of subjects who fall into particular categories, the most
 Exploratory data analyses: describing our data Chapter | 6 49

common summary is the frequency table. A frequency table gives the counts
for the number of subjects that fall in each category, and occasionally the total
number of subjects in the study. From the frequency table, we can easily get the
sample proportions for each category in our categorical response.
For example, in 2016, the Pew Research Group conducted a survey [14]
of 1488 people to see through what forms Americans consume books. They
found that 395 people did not read books, 577 people only read print books, 425
people read both digital and print books, and 91 people only read digital books.
The frequency table would be

No books Print only Digital only Print and digital

395 577 91 425

What would the sample proportion be for the people who read only print
books? We would divide the number of people who read print-only books by
the total sample size, so the sample proportion for print-only readers would be
577
p̂ = = 0.378.
1488
Oftentimes, we are interested in how two categorical variables are related.
With this in mind, a frequency table, which counts the number of subjects in
each category for a single variable only, is an inadequate summary. When we are
interested in the relationship between two variables, the best descriptive statistic
is a contingency table. A contingency table is a m × p table where the counts
in each cell of the table give the number of subjects who fall in the specific
combination of categories defined by the row-column combination. In addition,
contingency tables give the marginal totals for each row and column, which
gives the subjects who fall into a row or column category ignoring the other
dimension of the table, essentially creating a frequency table for either rows or
columns.
In 2017, the Pew Research Group conducted a survey [15] of 3701 people to
see what causes Americans to be active consumers of science news. Among the
variables they looked at were ethnic group, including black, white, and Hispanic.
These subjects were then asked if they were an active, casual, or uninterested
consumer of science news. The contingency table for the data was

Active Casual Uninterested No answer Row totals

White 487 916 1431 28 2862
Black 59 98 227 8 392
Hispanic 89 152 183 23 447
Column totals 635 1166 1841 59 3701

What would the frequency table be for the columns (Defined by the question
“Are you an active, casual, or uninterested consumer of science news?”), and
50 Basic Statistics With R

what would the sample proportion be for subjects who are Hispanic and who
are active science consumers? If look at the column totals, we would find that
the frequency table for the columns is

Active Casual Uninterested No answer

635 1166 1841 59

Additionally, by finding the Hispanic row and Active column in the table,
we would find that the sample proportion of individuals who are Hispanic and
89
active science consumers is p̂ = = 0.024.
3701
Calculating frequency and contingency tables by hand is a tedious task. For-
tunately, R has a function that creates both of these tables for us. The table
function in R takes in one or multiple categorical variables and outputs the de-
sired frequency or contingency table. The code used for this is

table(variable1, variable2)

where variable1 is the first variable desired in the table and variable2 is the
optional second variable, used if you are creating a contingency table. As with
all the functions discussed in this chapter, more details and examples will be
included in Chapter 7.
Visual summaries for categorical variables are very popular in the media, es-
pecially pie charts and bar charts. We will not discuss these charts because they
often cause multiple problems in interpretation, particularly pie charts. Edward
Tufte, a statistician noted in the field of data visualization, once stated that “Pie
chart users deserve the same suspicion and skepticism as those who mix up its
and it’s or there and their.” In short, the best descriptive statistics for categorical
data remain sample proportions, frequency tables, and contingency tables.

6.3.1 Practice problems

A survey [16] of adults in the United States was taken about whether the pop-
ulation believes technology has had a positive effect on society. The subjects
were asked “The effect of technology on our society has been...” with possible
responses Mostly Positive, Equally Positive and Negative, and Mostly Negative.
The frequency table for the survey was

Positive Equal positive and negative Negative No answer

2458 1796 378 94

1. What is the population of interest in this survey?

2. What is the sample size of this survey?
3. What is the sample proportion of people who believe technology has had a
Mostly Positive effect on society?
 Exploratory data analyses: describing our data Chapter | 6 51

In 2018, the Pew Research Group conducted a survey looking at people’s

opinions on genetically modified organisms. Two of the questions were “How
much of the food you consume is organic?” and “Do you have internet access?”
The survey answers to this are given below [81].

No Internet Internet
None of it 48 334
Not much 66 990
Some of it 62 848
Most of it 11 166

4. What is the sample size of this survey?

5. What is the sample proportion of people who do not have internet access?
6. What would the frequency table be for the question “How much of the food
you consume is organic?”

6.4 Parameters, statistics, and EDA for a single quantitative

variable
Quantitative variables have a wider variety of descriptive statistics and ex-
ploratory analyses to consider because quantitative variables have more aspects
than their categorical counterparts. For categorical variables, all we ultimately
have to consider is the proportion of individuals in the categories for the re-
sponse to get an understanding of the data. For a single quantitative variable, we
have to consider all aspects of the distribution of values, including the center of
the values, the spread of the values, and the shape of the distribution. We will
discuss various statistics that can be used to describe each of these aspects.

6.4.1 Statistics for the center of a variable

The most important aspect to consider for any quantitative variable is where the
center of that variable lies. Whether we are considering the population or
the sample, most statistical techniques involve considering the center of the
distribution of values for some variable of interest. Generally speaking, when
considering the population as a whole, there are two population parameters that
describe the center: the population mean (notated μ) and the population median.
The population mean is the average value for every member of the population,
while the population median is the midpoint of the population, where 50% of
the population lies below and above the value. As with any parameter, these
values are considered fixed and unknown.
Because the true values of these parameters are generally unknown, they
need to be estimated through sample statistics, specifically the sample mean and
sample median. To define these, we first need to talk about a little notation for
52 Basic Statistics With R

our samples. Say that we have a sample of size n. We assign the values of a
variable that we collect to be

x1 , x2 , ..., xn

where x1 is the value of the variable for the first subject, x2 is the value of the
variable for the second subject, xn is the value of the last subject, etc. In addition
to ordering the values of our variable by subject number, we could alternatively
order the variable values in our sample from smallest to largest. We notate this
by
x (1) , x (2) , ..., x (n)
where x (1) is the smallest value of the variable in the sample, x (2) is the second-
smallest value in the sample, x (n) is the largest value in the sample, etc.
Now that we have defined our sample, we can talk about the sample statistics.
The sample mean (notated x̄), our best estimate for the population mean, is
merely the average value for all values in the sample, so
 
1
n
1
x̄ = xi = x1 + x2 + ... + xn
n n
i=1

The sample median (notated x̃), our best estimate for the population me-
dian, is the median value for the variable in the sample, so there are two possible
formulas for the sample median. If the sample size is odd, the sample median is
 n+1 
x̃ = x 2
n−1
or the points in the sample where half of the data— data points—are
2
greater than and less than that value. If the sample size is even, the sample me-
dian is
n n 
x 2 + x 2 +1
x̃ =
2
or the average between the two values in the sample closest to the middle.
For example, Nigel Richards is generally considered one of the best Scrabble
players in the world. At a 2014 tournament, he put up scores of

550, 526, 458, 440, 440, 389, 432, 508, 433, 433, 548, 497, 531, 484.
 
1
The sample mean for this data would be x̄ = 550 + 526 + ... + 484 =
14
476.36. The sample median would require first ordering the scores for smallest
to largest from 432 to 550. Then, as the sample size is even at n = 14, we find
 Exploratory data analyses: describing our data Chapter | 6 53

the two values closest to the middle of 458 and 484, and average them to get the
median of x̃ = 471.
349, 432, 433, 433, 440, 440, 458, 484, 497, 500, 508, 526, 531, 548
In R, the mean and median are calculated using the mean and median func-
tions respectively. The code to calculate the average and median for some data—
named variable below—would be
mean(variable)

median(variable)

Now, because there are two summaries for the center of the data, a reason-
able question would be, which is better? Both the sample median and mean have
properties that make them the better option. The sample mean is not robust,
which means it is highly sensitive to outliers. An outlier is any strange point in
the dataset far away from other values. We will discuss them more later on. If a
single value in a sample is changed to become an outlier, the sample mean will
drastically change. On the other hand, the sample median is robust, which im-
plies that the sample median is not highly affected by outliers. Generally speak-
ing, we like our statistics to be robust, so this is a point in the median’s favor.
Despite this, generally speaking we use the sample mean to summarize the
center of a variable. This is because the mean has several theoretical properties
that make it very convenient to do statistical inference, much more so than the
sample median. We will not discuss these properties now, but will return to them
in later chapters. It suffices to say that unless otherwise specified, we will use
the sample mean to describe the center of a variable.

6.4.2 Practice problems

William Gosset, a noted statistician, produced major analyses on a dataset that
compared the effect of two sleeping drugs on hours of sleep gained [17]. Two
variables were collected, the hours of sleep gained due to drug X, and the hours
of sleep gained due to drug Y .
Patient Drug X Drug Y
1 0.7 1.9
2 −1.6 0.8
3 −0.2 1.1
4 −1.2 0.1
5 −0.1 −0.1
6 3.4 4.4
7 3.7 5.5
8 0.8 1.6
9 0.0 4.6
10 2.0 3.4
54 Basic Statistics With R

7. What is x1 , y8 , x (2) , y (6) ?

8. What is ȳ?
9. What is x̃?
A group of 15 laboratories was looking at the amount of the pesticide DDT
in kale (in parts per million). The 15 ordered measurements are given below [49]

2.79, 2.93, 3.06, 3.07, 3.08, 3.18, 3.22, 3.22,

3.33, 3.34, 3.34, 3.38, 3.56, 3.78, 4.64

10. What would be DDT (3) and DDT (14) ?

¯ ?
11. What is the average of DDT, DDT
˜
12. What is the median of DDT, DDT ?

6.4.3 Statistics for the spread of a variable

The second-most important aspect of a quantitative variable is the spread of the
variable. While two variables may have identical centers, if they have differ-
ent amounts of spread they will have drastically different values. Consider two
variables X and Y with values x1 = 99, x2 = 100, x3 = 101, y1 = 1, y2 = 100,
y3 = 199. Both x̄ = ȳ = 100, but they have very different spreads and very dif-
ferent values. There are many different statistics to describe the spread of the
data, beginning with the range.
The range of a dataset is the maximum value of a variable in the sample
minus the minimum value of a variable in the sample, or x (n) − x (1) . This is
the simplest summary of the spread, and as such it does have some deficiencies.
Namely, is the range robust, or resistant, to outliers? No, and the reason is fairly
clear. Since the range is the maximum minus the minimum, if one data point
was changed to be an outlier in the dataset, the range would increase greatly.
The range can be calculated in R a variety of ways. One could get the min-
imum and maximum value individually and then compute the range, but the
simpler way is to use the range function. The range function returns both the
minimum and maximum, making our range calculations easier. Similar to table,
mean, and median, we only have to feed the variable into the function, using
code
range(variable)

In another direction, a way we could describe the spread of individual values

in a sample are the deviations (notated di ), or the difference of each observation
and the sample mean. Mathematically speaking,

di = xi − x̄

Now, while deviations are useful for individual data points, we want to have
a single number to summarize the spread for the variable in the entire sample.
 Exploratory data analyses: describing our data Chapter | 6 55

We can use some summary of the deviations to create a summary of spread. We

could use the average deviation to describe spread, but there is a slight prob-
lem with that. Specifically, the average deviation will always be 0, making it
inappropriate to describe the spread. We need to find a different summary of the
deviations with values that will reflect the spread in the data.
The variance is just such a summary to describe the spread of a variable.
Variance is calculated as the average of the squared deviations, either for the
population or the sample. The population variance (notated σ 2 ) is the average
squared deviation for the entire population, and as with all parameters is fixed
and unknown. The sample variance (notated s 2 ) estimates the population vari-
ance and is the average squared deviation in the sample. Mathematically, this is
1  1  
s2 = (xi − x̄)2 = (x1 − x̄)2 + ... + (xn − x̄)2
n−1 n−1
i

You will notice that this is not an average in the strict sense, as the sum of
the squared deviations is divided by n − 1 and not n. This is to make the sample
variance an unbiased estimator of the population variance. Directly related to
the variance is the standard deviation. The standard deviation is the square
root of the variance, and also comes as an unknown√population parameter—√
notated σ —and sample statistic—notated s. So, σ = σ 2 , and s = s 2 . If the
standard deviation is known, the variance is known and vice versa.
R has two simple functions to calculate the variance and standard deviation,
the var and sd functions, respectively. All that is required is the data that we
want the mean and variance for, using code
var(variable)

sd(variable)

Because the sample variances and sample standard deviations are a func-
tion of the squared deviations, they will always be greater than or equal to zero.
A sample variance or sample standard deviation of 0 implies that every obser-
vation in the sample has the same value. As the sample variance or standard
deviation increases, the spread of the data is larger.
For example, The Sweet Briar soccer team allowed the following number
of goals (from fewest to most) in the 15 games of the 2017 season: 1, 1, 2,
2, 4, 5, 6, 6, 7, 7, 10, 10, 10, 10, 11. If we were to find the variance—and
standard deviation of the number of goals allowed—we first need to find the
sample average, which is x̄ = 6.133. Our sample variance will then be
1
s2 =
15 − 1
 
× (1 − 6.133)2 + (1 − 6.133)2 + ... + (10 − 6.133)2 + (11 − 6.133)2

= 12.69524.
56 Basic Statistics With R

Our√ sample standard deviation will be the square root of our sample variance,
so s = 12.69524 = 3.563038. While the variance and standard deviation are
the most commonly-used measures of spread in the data, neither are robust to
outliers. If we changed our largest number of goals allowed from 11 to 50, the
variance and standard deviation would drastically change. This brings up the
question, are there any robust measures of spread that provide alternatives to the
variance or standard deviation?
The Interquartile Range (IQR) is just such a robust measure of spread in
the dataset. The IQR is defined to be the distance between the 75th and 25th
percentile in the sample. The 75th percentile is the point in the dataset where
75% of the data is less than that value and the 25th percentile is the point in the
dataset where 25% of the data is less than that value. When calculating the 75th
percentile, we essentially find the median of the upper half of the dataset, while
the 25th percentile is the median of the lower half of the dataset.
So let us look at our Sweet Briar goals allowed again. As a reminder, the
team allowed the following number of goals in 2017: 1, 1, 2, 2, 4, 5, 6, 6, 7, 7,
10, 10, 10, 10, 11. To find the 75th and 25th percentiles, we need to first find the
median in order to split the dataset into the upper and lower halves. Since there
are 15 observations, the median will be x (8) = 6. This splits the dataset into two
halves:

Upper Half: 7, 7, 10, 10, 10, 10, 11

Lower Half: 1, 1, 2, 2, 4, 5, 6

The 75th percentile is the median of the upper half, or x (12) = 10. The 25th
percentile is the median of the lower half, or x (4) = 2. So the IQR will be 10-
2=8. R uses the function IQR to calculate the IQR.

IQR(variable)

It is important to note that R uses a different algorithm to calculate the IQR

than above, so in small sample sizes the IQR calculated by hand will often differ
from R.

6.4.4 Practice problems

New York Yankees relief pitcher Aroldis Chapman is generally considered the
hardest thrower in baseball, being the first known pitcher to break 105 MPH. In
a September 2017 appearance, he threw the following 12 pitches, with the speed
saved as the variable x.

100.7, 102.9, 103.7, 100.6, 99.7, 99.8,

100.7, 100.2, 101.1, 101.9, 102.9, 102.9

13. What is x̄?

 Exploratory data analyses: describing our data Chapter | 6 57

14. What is x̃?

15. What is s 2 ?
16. What is s?
17. What is the IQR?
The myocardial oxygen consumption is a measure of how much oxygen is
required for the heart to function optimally. Seven dogs went to the vet and their
myocardial oxygen consumption was measured to be the variable x [82].
78, 92, 116, 90, 106, 78, 99
18. What is x̄?
19. What is x̃?
20. What is s 2 ?
21. What is s?
22. What is the IQR?

6.5 Visual summaries for a single quantitative variables

The last aspect of a quantitative variable is the shape of the distribution of values
for a variable. While there are numeric statistics that can be calculated to de-
scribe the shape, we mostly investigate this through visual summaries. For our
purposes, we will focus on one of the most common summaries: histograms.
Histograms help us to visually display the shape of the distribution of nu-
meric values by taking the numeric values and placing them in equally-sized
intervals referred to as bins. The histogram then graphs the frequencies of each
bin using bars.
In the 2017 Gubernatorial election [18], Amherst County had 11 precincts
to report in. The turnout in each of these districts was

Precinct Turnout
1 0.409
2 0.427
3 0.439
4 0.477
5 0.461
6 0.475
7 0.524
8 0.52
9 0.445
10 0.363
11 0.399

To create the histogram, we need to decide on the size of the bins. For this
data, we can use the intervals from 0.35 to 0.4, 0.4 to 0.45, 0.45 to 0.5, and 0.5
58 Basic Statistics With R

to 0.55. There are two observations in the first bin, four in the second bin, three
in the third bin, and two in the fourth. Graphing this yields the histogram seen
in Fig. 6.2.

FIGURE 6.2 Histogram of Amherst County voter turnout.

R uses the hist function to create histograms. There are multiple extra options
that we will discuss in Chapter 7, but the most basic histogram can be created
by solely inputting the data.

hist(variable)

When we plot a histogram, there are several things to look for that tell us
information about the data. The first is the shape of the histogram, either sym-
metric or skewed. Symmetric histograms are what the name implies, that the
data evenly split and identically shaped on either side of the middle value. How-
ever, note that symmetric histograms do not have to be perfectly symmetric, only
roughly speaking. Skewed histograms are nonsymmetric histograms and can be
either skewed right or left, with the direction of the skew being whichever side
has the long tail of data.
The second thing to look for is the number of modes, or local high points,
in the histogram. The most common numbers of modes are one (unimodal) and
two (bimodal). On very rare occasions, you see a trimodal histogram. Modes
that are very close together may be in fact one mode depending on how the
values are binned together. (See Fig. 6.4.)
The last three things to look for are where the center of the data is located,
how spread out the data is, and if there are any outliers. Generally, we use the
sample mean and sample variance to describe the center and spread, but how do
we identify outliers? We have mentioned several times that various statistics and
summaries are sensitive to outliers without talking about how to identify poten-
tial outliers—stated this way because we do not necessarily know for certain if
a data point is an outlier.
 Exploratory data analyses: describing our data Chapter | 6 59

FIGURE 6.3 Various shapes of histograms (from left to right) Symmetric, Skewed Right, and
Skewed Left.

FIGURE 6.4 Number of modes of histograms (from left to right) Unimodal and Bimodal.

6.6 Identifying outliers

The first possibility to identify an outlier is visually through histograms. If there
is a clear data point that exists outside of the normal area of the data, it is entirely
possible that the data point is an outlier.
While this visual method can be useful, we generally want something more
rigorous and quantitative to determine if an observation is a potential outlier.
With this in mind, we can use what is called a z-score (notated zi ) to identify
potential outliers. The z-score for an observation is the difference between an
observed value and the sample average, scaled by the sample standard deviation.
That is,
xi − x̄
zi = .
s
The general idea is that if the observation is an outlier, it will be very far
from the average, and thus will have a very large positive or very small negative
60 Basic Statistics With R

FIGURE 6.5 A histogram with a potential outlier out near 10.

z-score. Generally speaking, if |zi | > 3—that is, the observation is more than
three standard deviations from the average—we will flag the observation as a
potential outlier.
For example, on the average, the Old Faithful geyser [10,19] has 72.31 min-
utes in between eruptions with a standard deviation of 13.89 minutes. If you
have been waiting for 2 hours for an eruption, is this observation possibly an
120 − 72.31
outlier? Yes, as a 2-hour wait has a z-score of = 3.43, which has
13.89
an absolute value greater than 3.
While z-scores can be useful for identifying outliers, they come with a major
caveat: They can only be used to identify outliers for bell-shaped data. Bell-
shaped data is a specific type of symmetric data that is ultimately very important
to statistics, and we will discuss it later on in the course. The symmetric his-
togram in Fig. 6.3 is an example of bell-shaped data, while Fig. 6.5 shows an
example of bell-shaped data with a single outlier.
If a dataset is not bell-shaped, we cannot use the z-score to identify outliers
because ultimately the calculation of the z-score requires bell-shaped data to be
valid. As many datasets are not going to be bell-shaped, we need another method
to identify potential outliers.
Such a method can be found using the Interquartile Range. Once we find the
75th percentile (also referred to as the third quartile or Q3) and 25th percentile
(also called the first quartile or Q1), we calculate the Interquartile Range as
I QR = Q3 − Q1. Then we can label potential any outliers if an observation is

xi > Q3 + 1.5 × I QR or xi < Q1 − 1.5 × I QR.

The general idea here is similar to z-scores. If an observation is particularly

far from where the majority of the data is—here defined as from the first quartile
to the third quartile—it is likely an outlier. The major difference between the z-
score method and the IQR method is that IQR does not require bell-shaped data
 Exploratory data analyses: describing our data Chapter | 6 61

and is reliant on the robust IQR rather than the nonrobust sample average and
sample standard deviation. But both methods do have the uses and instances
where one should be chosen.
If you made a histogram of the Old Faithful data, you would see that it
is highly nonbell shaped. With that in mind, we need to use the IQR to iden-
tify outliers. The 75th percentile (or third quartile) is at 82 and 25th percentile
(first quartile) is at 58, so the IQR is 82 − 58 =24. This means that any data
points greater than 82 + 24 × 1.5 = 118 or less than 58 − 24 × 1.5 = 22 are
outliers. So, if you have been waiting for two hours, this would still be an out-
lier.

6.6.1 Practice problems

Charles Darwin studied the differences in growth between self-fertilized and
cross-fertilized pairs of corn plants grown under identical conditions. The dif-
ference in heights—cross-fertilized minus self-fertilized—was measured, with
the following 15 data points recorded. [83]

6.125, −8.375, 1.000, 2.000, 0.750, 2.875, 3.500, 5.125,

1.750, 3.625, 7.000, 3.000, 9.375, 7.500, −6.000

23. Using z-scores, are there any outliers in the data?

24. Using the IQR, are the any outliers in the data?
Albert A. Michelson was a Polish physicist whose experiments helped esti-
mate the speed of light. The 20 observations below are the results of one of those
experiments, with the variable measured being the speed of light in kilometers
per second above 299,000 kilometers. [84]

850, 850, 1000, 810, 960, 800, 830, 830, 880, 720,
880, 840, 890, 770, 910, 720, 890, 810, 870, 940

25. Using z-scores, are there any outliers in the data?

26. Using IQR, are there any outliers in the data?
Describing the attributes of the histogram helps us to understand our single
quantitative variables.
27. Describe the shapes, modes, centers, outliers, etc. of the histograms below.
(See Fig. 6.6.)

6.7 Exploring relationships between variables

While we can do many analyses looking at a single variable, many questions
in statistics look at the relationship between two variables. We could look at
how gender affects pay, how yearly salary affects political affiliation, etc. In
62 Basic Statistics With R

FIGURE 6.6 Practice histograms.

these cases we are looking for an association between an explanatory variable

and a response. A response is the variable of interest, what we ultimately want
to predict or compare. The explanatory variable (sometimes referred to as a
predictor) is the variable that we hope explains or predicts our response. Finally,
an association between two variables implies that there is a relationship between
two variables where the value of your explanatory variable makes certain values
of your response more likely.
For example, say a researcher is looking into what diet yields the greatest
increase in weight for young chickens. They assign each chicken randomly to
one of four diets, and measures their weight before the experiment and at the
end of 60 days on the diet and calculates the difference. What is the explanatory
variable and response for this experiment? The response is the change in weight
and the explanatory variable is the diet.
We have looked into the association between two categorical variables using
contingency tables, but from here on we will look at exploratory analyses for
a categorical predictor and quantitative response, followed by the association
between two quantitative variables.

6.8 Exploring association between categorical predictor and

quantitative response
When we are looking at the relationship between a categorical predictor and
quantitative response, there are many routes that we can possibly go. We can
 Exploratory data analyses: describing our data Chapter | 6 63

simply begin by comparing sample averages or sample medians across the lev-
els of a categorical predictor to look for major differences. This does not take
variability of the response into account, so we need to bring that into the pic-
ture.
While there are many methods to explore these relationships, we are go-
ing to focus on a visual comparison, specifically boxplots. Boxplots allow us
to visually see if there is overlap for a response among multiple groups, but
they can be used to visualize single variables as well. The boxplot is basically a
visual display of the five number summary (5NS) of a variable. The five num-
ber summary is a simple summary of the values of a variable, consisting of
the minimum, maximum, median, first quartile (Q1, 25th percentile), and third
quartile (Q3, 75th percentile). Again, a boxplot is just a visual plotting of this
five number summary, and the steps are quite simple.
1. Calculate the five number summary.
2. Determine if there are any outliers using the interquartile range (IQR).
3. On a number line, draw three vertical lines at the first quartile (Q1), median,
and third quartile (Q3). Form a box with these three lines.
4. Draw a horizontal line from Q1 down to the smallest nonoutlier value.
5. Draw a horizontal line from Q3 up to the largest nonoutlier value.
6. Mark all outliers with a symbol.
R uses the function summary to calculate the five number summary. This
function also returns the sample average as well as the number of missing data
points if any exist.
summary(variable)

Consider the following dataset giving the number of pages in a sample of

soft-back textbooks: 203, 336, 378, 385, 417, 436, 469, 585. To begin, we
have to create the five number summary, which would be Min=103, Max=585,
Median=401, Q1=357, and Q3=452.5. Next, we need to identify any out-
liers using the IQR. The IQR is 452.5 − 357 = 95.5. This means that any
data points that are greater than 452.5 + 1.5 × 95.5 = 595.75 or less than
357 − 1.5 × 95 = 213.75. So we do have one outlier at 203. Next, we place
three vertical lines at Q1, median, and Q3 and create a box with the lines.
The next two steps have us drawing us to the largest and smallest nonoutlier
data values. Finally, we mark the outlier point with a symbol of your choos-
ing.
In addition to creating individual boxplots, we can place boxplots side-by-
side to compare two or more categories of a categorical explanatory variable.
If there is considerable overlap between the two boxplots, it is likely that there
is no association between the explanatory variable and response. If there is no
overlap, there is possibly some association. (See Fig. 6.7.)
For example, Fig. 6.8 shows the home attendance for the 2019 Baltimore
Orioles’ day and night games. In looking at the boxplot, there is a little overlap
between the two groups, suggesting that there might be an association be-
64 Basic Statistics With R

FIGURE 6.7 Step-by-step construction of a boxplot.

tween Day/Night games and attendance. This makes some intuitive sense, as
day games either tend to be held on weekends, holidays, or special occasions—
such as Opening Day—when people are more likely to be able to attend
games.
The boxplot function in R allows us to create both individual and side-by-
side boxplots. We’ll discuss the particulars of this function in Chapter 7, but the
basic boxplot can be created using

boxplot(variable)
 Exploratory data analyses: describing our data Chapter | 6 65

FIGURE 6.8 Side-by-side boxplot for home attendance for the 2019 Orioles day and night games.

6.8.1 Practice problems

28. Given the following data (Barry Bonds’s intentional walks in a season),
create the boxplot:
2, 3, 3, 9, 14, 15, 18, 22, 22, 22, 25, 29, 30, 32, 34, 35, 38, 43, 43, 61, 68,
120
29. Edgar Anderson’s iris dataset [20] is one of the most significant historical
statistics datasets. Given the following five number summaries (no outliers)
lf the sepal length for two species of iris, create the boxplot.

Species Min Q1 Median Q3 Max

Setosa 4.3 4.8 5 5.2 5.8
Versicolor 4.9 5.6 5.9 6.3 7.0

30. Does there appear to be an association between species and sepal length?

6.9 Exploring association between two quantitative variables

In addition to exploring the association between a categorical predictors and
quantitative response, we can investigate the association between two quantita-
tive variables. We still have explanatory variables and responses, and our goal
is still to find associations between the two. However, as our predictor is now
quantitative, the methods of investigating this association are going to change.
The initial way we investigate the association between two quantitative vari-
ables is to plot them using a scatterplot. A scatterplot is a plot where every pair
of explanatory variable (X) and response (Y ) is plotted as an (x, y) pair on the
Cartesian plane. (See Fig. 6.9.)
The plot function allows us to create scatterplots for two variables. I will
add that using plot in certain other scenarios will create graphics other than
66 Basic Statistics With R

FIGURE 6.9 A scatterplot of cat heart weight (Y) versus cat body weight (X) [10,21].

scatterplots, but using the code below, we can create a simple scatterplot where
variable1 will be the x-coordinate and variable2 will be the y-coordinate.

plot(variable1, variable2)

Just like when using a histogram to consider the shape of a variable, there
are specific things we look for in a scatterplot. The first thing we look for is
some sort of a pattern in the data. This could be linear, quadratic, or any other
pattern we can think of. (See Fig. 6.10.)
Second, we look to see if there is an increasing or decreasing relationship
between our predictor and response. Increasing implies that as our predictor
gets larger, so does our response. On the other hand, decreasing implies that as
our predictor gets larger, our response gets smaller. (See Fig. 6.11.)
The third thing we look for is any anomalies, specifically outliers, in our pre-
dictor or response. Existence of these outliers can drastically change or possibly
invalidate our results. Other anomalies may include clusters in our predictor,
response, or both. (See Fig. 6.12.)
Finally, assuming the relationship between the predictor and response is lin-
ear, we look for how strong the association is. Stronger associations between
predictor and response imply the scatterplot looks much more like a line, while
a weak association between predictor and response implies that the scatterplot
looks like random scatter. (See Fig. 6.13.)
Now, using “strong” and “weak” to denote the strength of the association
between a predictor and the response can be very subjective and vague. We
need a better way to describe this association, specifically with a statistic that
can be calculated from our sample.
This better way to describe associations between variables is correlation.
It exists as a population parameter (ρ, the fixed and unknown true correlation
between two quantitative variables for the entire population) and as a sample
 Exploratory data analyses: describing our data Chapter | 6 67

FIGURE 6.10 Four possible scatterplot patterns (clockwise from top left): Linear, Quadratic, Ex-
ponential, No Pattern.

FIGURE 6.11 Scatterplot relationships (from left to right): Increasing relationship, No relation-
ship, Decreasing relationship.

statistic (r, the calculated correlation between two quantitative variables in our
sample). Correlations are able to summarize both the strength of the associa-
tion between two variables as well the direction (Increasing/Decreasing) of the
relationship between the predictor and response.
68 Basic Statistics With R

FIGURE 6.12 Scatterplot anomalies (from left to right): Outlier in the predictor, Outlier in the
response, Clustered Data.

FIGURE 6.13 Scatterplot strength of association (from left to right): Weak association, Strong
association.

The correlation is ultimately the average of the product of z-scores for our
two variables in question. That is,
n   
1  xi − x̄ yi − ȳ
r=
n−1 sx sy
i=1

where x̄ and ȳ are the sample means for x and y, while sx and sy as the sample
standard deviations for x and y. So, if we are calculating the correlation by hand,
we follow the following steps:
1. Calculate the sample means for x and y
2. Calculate the sample standard deviations for x and y
3. Calculate the z-score for each observation of x
4. Calculate the z-score for each observation of y
 Exploratory data analyses: describing our data Chapter | 6 69

5. Multiply the z-score for x and the z-score for y together for each observation
6. Add all the multiplied z-scores together and divide by the sample size n
minus 1.
An agricultural researcher in the 1980s was looking into the effect of over-
seeding on crop yield [22,23]. He planted the fields with various seed rates and
recorded the resulting yield.

Seed rate Grains Per Barley head

50 21.2
75 19.9
100 19.2
125 18.4
150 17.9

To find the correlation, we first need to find the sample averages for seed rate
(100) and yield weight (19.32). Next, we need the sample standard deviations
for seed rate (39.53) and yield weight (1.29). Then we need to calculate z-scores
and multiply them together

Observation Seed rate z-score Yield weight z-score Multiplied z-scores

1 −1.26 1.45 −1.83
2 −0.63 0.45 −0.28
3 0.00 −0.09 0.00
4 0.63 −1.71 −0.45
5 1.26 −1.09 −1.38

Finally, we add all the values in the third column together and divide by
5 − 1 = 4 to get −3.94/4 = −0.985 for our correlation.
As is quite obvious, this process is very tedious to do by hand. Thus, we often
turn to the cor function in R. Given two variables that we want to correlate, this
function will return the correlation as calculated above.

cor(variable1, variable2)

The sample correlation r tells us first the direction of the relationship, where
if r is positive this implies that there is a increasing relationship between the two
variables. r is bounded between −1 and 1, and can also tell us the strength of
the association between our two variables. If |r| is close to 1, this implies a very
strong relationship between the two variables, while if |r| is close to 0, there is
a weak relationship between the variables. (See Fig. 6.14.)
There are two main caveats when dealing with correlations. The first is that
correlations are only valid and interpretable for linear relationships between two
variables. If the two variables we are interested in have any other relationship—
quadratic, cubic, exponential, etc.—we cannot interpret our correlations and our
70 Basic Statistics With R

FIGURE 6.14 Plots of various correlations (clockwise from top left): r = 0, r = 0.99.r = −0.25,
r = −0.75.

calculated value is not valid. In addition, the correlation is not a robust statistic,
so outliers in either variable will strongly affect the correlation value.
The second caveat—and I cannot stress this enough—is that if two variables
are associated (or have a high correlation) this does not imply that one variable
causes the other. Put another way, Correlation does not imply causation. The
only thing that can establish causation is a designed experiment like the ones
we discussed earlier. For proof, I would suggest visiting the website Spurious
Correlations [24], which is filled with many instances of pairs of highly corre-
lated variables that very likely have no causal relationship. For example, while
the marriage rate in Virginia is incredibly highly correlated with the per capita
consumption of high fructose corn syrup (r = 0.982) [25], we would in no way
expect high marriage rates in Virginia to cause the per capita consumption of
corn syrup to increase. For another example, while the number of lawyers in
Virginia is incredibly highly correlated with the per capita consumption of mar-
garine (r = −0.946) [26], we would in no way expect high numbers of lawyers
in Virginia to cause the per capita consumption of margarine to decrease. The
lesson here is, again, correlation does not imply causation. (See Fig. 6.15.)
Ultimately, when looking at associations between two quantitative variables,
we need to use both scatterplots and correlations to get at understanding the
relationship between predictor and response. Further than that, we need to use
 Exploratory data analyses: describing our data Chapter | 6 71

FIGURE 6.15 Plots of spurious correlations (from left to right): Marriage rates in Virginia versus
per capita consumption of corn syrup (r = 0.9816), number of lawyers in Virginia versus per capita
consumption of margarine (r = −0.9461).

a fair bit of common sense to ensure that we do not try to claim more than the
data allows us to claim.

6.9.1 Practice problems

31. Describe the scatterplots given below in terms of pattern, direction, etc.
72 Basic Statistics With R

32. Given the following data, calculate the correlation between the two vari-
ables [27].

Total Runs in the World Series Visitors (Millions) to Disneyland Tokyo

39 13.906
39 14.293
59 13.646
41 14.452
68 13.996
22 14.847

33. Given the following data describing the amount of wear on each shoe in a
pair of shoes, calculate the correlation between the two variables [85].

Shoe A Shoe B
13.2 14.0
8.2 8.8
10.9 11.2
14.3 14.2
10.7 11.8
6.6 6.4
9.5 9.8
10.8 11.3
8.8 9.3
13.3 13.6

6.10 Conclusion
Exploratory data analysis takes many forms, from numerical summaries of a
single variable to graphical summaries of multiple variables. Each of these sum-
maries has its uses, and they all share a goal of trying to make our data more
understandable. However, in choosing a summary we need to be sure that it is
appropriate, as well as considering how the summary will be affected by the
data as a whole.
No matter how useful our exploratory analyses may be, they cannot be the
foundation of decision-making on their own. This is because these summaries
are calculated from random samples or randomized experiments, and as such
will have a certain amount of variability ascribed to them. We need to account
for this variability, which is the basis of statistical inference.
Chapter 7

R tutorial: EDA in R
Contents
7.1 Introduction 73 7.4 Missing data 77
7.2 Frequency and contingency 7.5 Practice problems 78
tables in R 73 7.6 Graphical exploratory analyses
7.3 Numerical exploratory analyses in R 78
in R 74 7.6.1 Scatterplots 78
7.3.1 Summaries for the center 7.6.2 Histograms 80
of a variable 74 7.7 Boxplots 82
7.3.2 Summaries for the spread 7.8 Practice problems 84
of a variable 75 7.9 Conclusion 85
7.3.3 Summaries for the
association between two
quantitative variables 76

7.1 Introduction
In our previous R tutorial, we talked about how we load data into our workspace,
using either the libraries available for use with R or loading in csv files use the
read.csv function. In this tutorial, we will talk about calculating the summaries
discussed in Chapter 6 as well as generating various plots used in visualizing
data.

7.2 Frequency and contingency tables in R

When working with categorical variable(s), our main form of exploratory anal-
yses is the frequency (for one variable) or contingency (for two variables) table.
From these, it is trivial to calculate sample proportions. In R, generating both
of these types of tables is done with a single function, the table function. For a
frequency table, the code for the table function is
table(Variable)

and then hit <Enter>. The “Variable” mentioned can be either then name of an
entered vector or a variable selected from a larger data frame using the $ nota-
tion. R then returns the frequency table of the variable you specified in the table
statement. Consider the quine dataset [10,28] in the MASS library. This data
was an investigation into Australian student absences, with data collected on
each student including ethnicity (aboriginal and nonaboriginal), sex (male and

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00018-3 73

Copyright © 2022 Elsevier Inc. All rights reserved.
74 Basic Statistics With R

female), learning ability (average learner and slow learner), etc. If we wanted
to know the ethnic breakdown of the subjects in the survey-defined by the Eth
variable in the quine data frame—we would use the code given below to see that
there were 69 aboriginal students and 77 nonaboriginal students.
table(quine$Eth)

For contingency tables, the code is almost identical, except instead of one vari-
able in the table we have two, so the code will change to
table(Variable1, Variable2)

Again, these variables can be entered vectors or variables selected from a larger
data frame. Using the same quine dataset, we could look at the contingency
table for student ethnicity—the Eth variable in the quine data frame—and their
learning ability—the Lrn variable in the quine data frame. Our code would be
table(quine$Eth, quine$Lrn)

to get the contingency table given below.

Average learner Slow learner

Aboriginal 40 29
Nonaboriginal 43 34

Notice that R does not give the marginal totals for the rows or columns,
nor the table total. These can be determined by adding the appropriate rows
and columns as desired or taking advantage of other functions in R such as
addmargins.

7.3 Numerical exploratory analyses in R

In this section, we will look at calculating summaries of quantitative variable
in R. Thus, we will focus on getting summaries for the center and spread of a
single quantitative variable, as well as the correlation between two quantitative
variables. Visual summaries will be covered in a later section.

7.3.1 Summaries for the center of a variable

To begin, we recall the two main summaries for the center of a variable that
we learned before: the sample mean and the sample median. In R, we can easily
calculate these quantities for a variable stored in a vector. To calculate the mean,
our code would be
mean(Variable)

and the code for the median would be

 R tutorial: EDA in R Chapter | 7 75

median(Variable)

In both cases, the “Variable” can either be saved in an inputted vector or saved
within a data frame. So, for example, let us look at the Housing dataset [12,29]
in the Ecdat library. This dataset looks at the characteristics and sales price of
houses in the city of Windsor. If we wanted to find the mean of the sales price
of the houses—the price variable—we would use
mean(Housing$price)

and for the median

median(Housing$price)

To find that the mean sales price was 68,121.6 and the median was 62,000. We
can also use the mean and median functions in conjunction with subsetting vec-
tors to compare the means and medians for multiple groups. For example, say
that we wanted to compare the average sales price of houses based on whether
they were in a preferred area—found in the variable prefarea with values of
“yes” and “no”—we would use
mean(Housing$price[Housing$prefarea=="yes"])
mean(Housing$price[Housing$prefarea=="no"])

to see that the average sales price for preferred areas is 83,986.37 versus
63,263.49 for nonpreferred areas. The code would be similar for comparing
medians, using the median function in place of the mean function.

7.3.2 Summaries for the spread of a variable

In looking at the spread of a variable, we focused in on four summaries: range,
variance, standard deviation, and interquartile range (IQR). Each of these sum-
maries can be calculated in R with one single line of code apiece. We will begin
with the range and the interquartile range. As a reminder, the range is the maxi-
mum value minus the minimum value, while the interquartile range is the third
quartile (Q3) minus the first quartile (Q1). These four values, plus the median,
make up the five number summary for a variable. So, if we can get the five
number summary we can get both of these summaries. In R, we can get the five
number summary for a variable—either in an inputted vector or saved in a data
frame—with the code
summary(Variable)

This will give you the minimum, first quartile, median, third quartile, and
maximum, plus the added mean for a variable. From this, we can calculate the
range as well as the IQR. Alternatively, we could use the range function to get
the minimum and maximum values of a variable or IQR function to calculate
the IQR directly.
76 Basic Statistics With R

range(Variable)
IQR(Variable)

While the summary function does not include variance or standard deviation,
they both are simple to calculate using R. Our variance can be found using the
var function with the code
var(Variable)

and the standard deviation using the function sd with the code
sd(Variable)

To illustrate all these options, we can look at the Pima.te dataset [10,30] in
the MASS library. This data looks at possible indicators of diabetes for females
in the Pima Native American tribe. If, say, we wanted to find the range and
IQR for the plasma glucose level (the glu variable), we would first find the five
number summary by using
summary(Pima.te$glu)

to find that the five number summary for the variable is

Min Q1 Med Q3 Max

65 96 112 136.2 197

To get that the range is 197 − 65 = 132 and the IQR is 136.2 − 96 = 40.2.
The range and IQR can be confirmed using the code
range(Pima.te$glu)
IQR(Pima.te$glu)

To get the variance, we would use

var(Pima.te$glu)

to find that s 2 = 930.32, and to get standard deviation we use

sd(Pima.te$glu)

to get s = 30.5. Notice that s is the square root of s 2 , as is part of the definition
of variance and standard deviation.

7.3.3 Summaries for the association between two quantitative

variables
Our one numeric summary of the association between two quantitative variables
is correlation. By hand, correlation is a complicated multistep procedure with
many possible pitfalls. In R, the correlation is calculated using the cor function
using the code
 R tutorial: EDA in R Chapter | 7 77

cor(Variable1, Variable2)

where both variables could either be an inputted vector or stored in a data frame.
As an example, let us look at the cars dataset [31] stored in the base R environ-
ment. This dataset looks at the relationship between the speed of a car and the
resultant stopping distance—specifically for cars in the 1920s. To calculate the
correlation between the speed—stored in the variable speed—and the stopping
distance—stored in dist—we would use
cor(cars$speed, cars$dist)

to find the correlation of r = 0.8069, indicating the intuitive strong positive

relationship between speed and stopping distance. Note that the order of the
variables could be switched and the correlation would remain the same. Addi-
tionally, recall that correlation is only interpretable if the relationship between
the two variables is linear, something we can confirm by visually through a scat-
terplot. In the next section, we will talk about the various plots and visualizations
we can use to glean information from our data.

7.4 Missing data

One thing you may notice depending on the dataset that you are using is that
your mean, median, or other numeric summary may return a value of NA. This
is because there is some value in that variable that is missing for some reason.
It may be because the subject never answered a question, data was lost, or any
number of other possibilities. When done by hand, we would likely just calculate
our desired summary with all the other data values, but R does not do this on its
own. Rather, we have to tell it to ignore the missing values.
Most functions in R—including many of the ones above—allow us to use
the na.rm option in the function. This is a TRUE/FALSE option that just tells R
to ignore any NA values in its calculations. So, for example, if you were trying
to get the average of a variable that you know has NA values in it, you would
use the code
mean(variable, na.rm=TRUE)

This process is similar for many other functions, including median, var, and
sd. One notable difference is the cor function. As there are two variables in-
volved in this function, we only want to use observations that have no missing
values, or complete observations. In this case, we would add use=“complete”
to our cor function, with the result looking something like
cor(variable1, variable2, use="complete")

It should be noted that the na.rm and use options can be included in any
instance of mean or cor, but if there are no missing observations these options
will have no effect on the resulting calculations.
78 Basic Statistics With R

7.5 Practice problems

Load the leuk dataset from the MASS library [10,32]. This dataset is the survival
times (time), white blood cell count (wbc), and the presence of a morphologic
characteristic of white blood cells (ag).
1. Generate the frequency table for the presence of the morphologic character-
istic.
2. Find the median and mean for survival time.
3. Find the range, IQR, variance, and standard deviation for white blood cell
count.
4. Find the correlation between white blood cell count and survival time.
Load the survey dataset from the MASS library [10,86]. This dataset contains
the survey responses of a class of college students.
5. Create the contingency table of whether or not the student smoked (Smoke)
and the student’s exercise regimen (Exer).
6. Find the mean and median of the student’s heart rate (Pulse).
7. Find the range, IQR, variance, and standard deviation for student age (Age).
8. Find the correlation between the span of the student’s writing hand (Wr.Hnd)
and nonwriting hand (NW.Hnd).
Load the Housing dataset from the Ecdat library [12,29]. This dataset looks
at the variables that affect the sales price of houses.
9. Create the contingency table of whether or not the house has a recreation
room (recroom) and whether or not the house had a full basement (full-
base).
10. Find the mean and median of the house’s lot size (lotsize).
11. Find the range, IQR, variance, and standard deviation for the sales price
(price).
12. Find the correlation between the sales price of the house (price) and the
number of bedrooms (bedrooms).

7.6 Graphical exploratory analyses in R

One of the simplest and most common ways to begin exploratory analyses in
Statistics is through graphs of some form. We talked about three of these in the
last chapter: the scatterplot, the boxplot, and the histogram. In R, all three of
these plots are simple to generate, and their functions have several options to
modify the plots as necessary.

7.6.1 Scatterplots
Scatterplots are used to determine the relationship between two quantitative
variables. This is important for many reasons, including identifying the afore-
mentioned linear relationship required for correlation interpretability. The base
code that we use to generate a scatterplot for two variables is
 R tutorial: EDA in R Chapter | 7 79

plot(X-variable, Y-variable)

where “X-variable” is the variable that you want to plot along the x-axis, while
“Y-variable” is the variable that you want to plot along the y-axis. So, say that
we wanted to confirm the linearity of the relationship between speed and stop-
ping distance in the cars dataset. The code we would use is
plot(cars$speed, cars$dist)

to generate the left-most plot in Fig. 7.1. As we can see, it is a very spartan plot
and somewhat difficult to read. However, plot has a number of options that can
be included after the variables to improve the look and readability of the plot.

FIGURE 7.1 Left-to-right. A base scatterplot for speed versus stopping distance and an enhanced
scatterplot of the same data using several of the described options in plot.

• xlab, ylab: The axis title for the x-axis and y-axis, respectively. These are
changed through xlab=“Title” and ylab=“Title”. The default axis titles are
the variable names entered in plot.
• main: The main title for the plot. This is changed through main=“Title”. The
default is no title.
• pch: The plotting character, or what the (x, y) points look like when plotted.
The standard options are numbers 1 through 25 with each number being a
different plotting character—shown below in Fig. 7.2. The default value is
pch=1.
• col: The color of the plotted points. This value can be entered as a string—
col=“red”—a number—col=4 results in a blue point—or even a HTML
hex code—col=“#8B1F41” for Chicago Maroon. The default value is
col=“black”.
• cex, cex.lab, cex.axis: The magnification factor of the plotted points—
cex—the axis titles—cex.lab—and the axis values—cex.axis. These are all
numeric values, with a baseline default of 1. Larger numbers increase the
size of points or text, smaller numbers shrink the size.
80 Basic Statistics With R

FIGURE 7.2 Different plotting character—pch—options with their corresponding number value.

• xlim, ylim: The range of the plot window for the x-axis and y-axis, respec-
tively. This is entered as a vector of values—for example, xlim=c(minimum,
maximum). By default, R sets the xlim and ylim to be the range of the X and
Y variables, respectively.
Using all these options, let us improve the speed and stopping distance plot
with the result being seen in the right-most plot in Fig. 7.1. As we can see, the
linearity assumption between the speed and stopping distance seems reasonable
given the data.

plot(cars$speed, cars$dist, pch=16,

xlab="Speed", ylab="Stopping Distance",
main="Speed vs. Stopping Distance",
cex=1.25, cex.lab=1.25, cex.axis=1.25)

7.6.2 Histograms
Histograms are used to help us describe the shape of our data. In R, histograms
are created through the hist function, with the baseline code being

hist(Variable)

For an example, we can look at the geyser dataset [10,19] in the MASS pack-
age in R. This dataset is the Old Faithful eruption wait and duration that we
discussed last chapter. To create the histogram for the wait time between erup-
tions (the wait variable in the geyser data frame), we would use

hist(geyser$wait)

The wait time variable seems to be bimodal, symmetric, and centered around
70. Like scatterplots, histograms have several options that can affect the look or
readability of the histogram.
 R tutorial: EDA in R Chapter | 7 81

FIGURE 7.3 Left-to-right: A base histogram for the wait time between eruptions and an enhanced
histogram of the same data using several of the described options.

• breaks: The number of bins for the data. This can be entered as either the raw
number—breaks=10—or a vector defining the cutoff points of the bins—
breaks=c(10,15,20,25,30). The effect of the breaks variable can be seen in
Fig. 7.4.
• xlab, ylab: The axis title for the x-axis and y-axis, respectively. These are
changed through xlab=“Title” and ylab=“Title”. The default axis titles are
the variable names entered in plot.
• main: The main title for the plot. This is changed through main=“Title”. The
default is no title.
• col: The color of the bars. This value can be entered as a string—col=“red”—
a number—col=4 results in a blue point—or even a HTML hex code—
col=“#8B1F41” for Chicago Maroon. The default value is col=“black”.
• cex.lab, cex.axis: The magnification factor of the axis titles—cex.lab—and
the axis values—cex.axis. These are all numeric values, with a baseline de-
fault of 1. Larger numbers increase the size of points or text, smaller numbers
shrink the size.
• xlim, ylim: The range of the plot window for the x-axis and y-axis, respec-
tively. This is entered as a vector of values—for example, xlim=c(minimum,
maximum). By default, R sets the xlim and ylim to be the range of the X and
Y variables, respectively.
We can use these histogram options to improve the base histogram in Fig. 7.3
with the following code, resulting in the right-most histogram in that same fig-
ure.

hist(geyser$wait, breaks=20,
main="Histogram of Time Between Eruptions",
xlab="Wait Time", cex.axis=1.25, cex.lab=1.25, cex.main=1.5)
82 Basic Statistics With R

FIGURE 7.4 The effect of the number of breaks on the resulting histogram. From left to right, 20
breaks and 50 breaks.

7.7 Boxplots
In using boxplots, we can either make a boxplot for a single variable to make
side-by-side boxplots exploring the association between a categorical predictor
and a quantitative response. Both of these goals can be accomplished through
a single function: the boxplot function. The base code to create a boxplot for a
single variable is

boxplot(Variable)

As an example, we can look at the birthwt dataset [10,33] in the MASS pack-
age of R. This dataset is looking at several possible indicators of low birth
weight in newborns. If we just wanted to look at the birth weights of these
subjects (The variable bwt in the birthwt data frame), our code would be

boxplot(birthwt$bwt)

Like the previous plots, the boxplot has multiple options to improve the plot.
• pch: The plotting character for outlier points. The standard options are num-
bers 1 through 25 with each number being a different plotting character—
shown in Fig. 7.2. The default value is pch=1.
• xlab, ylab: The axis title for the x-axis and y-axis, respectively. These are
changed through xlab=“Title” and ylab=“Title”. The default axis titles are
the variable names entered in plot.
• main: The main title for the plot. This is changed through main=“Title”. The
default is no title.
• cex.lab, cex.axis: The magnification factor of the axis titles—cex.lab—and
the axis values—cex.axis. These are all numeric values, with a baseline de-
fault of 1. Larger numbers increase the size of points or text, smaller numbers
shrink the size.
 R tutorial: EDA in R Chapter | 7 83

FIGURE 7.5 From left to right: The base boxplot of the birth weights from the birthwt dataset and
the enhanced boxplot of the same data using some of the options.

• xlim, ylim: The range of the plot window for the x-axis and y-axis, respec-
tively. This is entered as a vector of values—for example, xlim=c(minimum,
maximum). By default, R sets the xlim and ylim to be the range of the X and
Y variables, respectively.
We can use these options to modify the base histogram in Fig. 7.5, using the
following code. The result will be the right-most boxplot in that same figure.
boxplot(birthwt$bwt, ylab="Birth Weight",
pch=16, cex.axis=1.25, cex.lab=1.25)

In order to assess the relationship between a quantitative response and cate-

gorical predictor, we are able to use the boxplot function with a small addition.
boxplot(Response~Predictor)

Where “Response” is the response variable and “Predictor” is the predictor

variable. So, for example, if we wanted to take a look at the effect of a mother
smoking—the variable smoke in the birthwt data frame—on the birth weight of
their child, our code would be
boxplot(birthwt$bwt~birthwt$smoke)

The base code of the boxplot can be modified with the same options for the
single-variable boxplot. Additionally, we can make changes to the plot with one
further option. (See Fig. 7.6.)
• names: The names of the levels for the predictor variable. This can be
changed with names=c(Value1, Value2,...). By default, the boxplot names
will be the variable levels. It is important that the values in the names argu-
ment are in the matching order as the values of the original variable.
We can use these options to modify the base boxplot in Fig. 7.5, resulting
in the right-most boxplot in the same figure. Notice that in the names argument,
84 Basic Statistics With R

FIGURE 7.6 From left to right: The base boxplot of the birth weights from the birthwt dataset and
the enhanced boxplot of the same data using some of the options.

we order the values “Nonsmoker” and “Smoker” to correspond to the values 0

and 1 in the smoke variable.

boxplot(birthwt$bwt~birthwt$smoke,
names=c(Nonsmoker, Smoker),
ylab="Birth Weight", xlab="Mother’s Smoking Status",
pch=16, cex.axis=1.25, cex.lab=1.25)

7.8 Practice problems

Load the Star dataset from the Ecdat library [12,34]. This dataset looks at the
affect on class sizes on student learning.
13. Generate the scatterplot of the student’s math score tmathssk and reading
score treadssk.
14. Generate the histogram of the years of teaching experience totexpk.
15. Create a new variable in the Star dataset called totalscore that is the sum
of the student’s math score tmathssk and reading score treadssk.
16. Generate a boxplot of the student’s total score totalscore split out by the
class size type classk.
Load the survey dataset from the MASS library [10,86]. This dataset contains
the survey responses of a class of college students.
17. Generate the scatterplot of the student’s height Height and writing hand
span Wr.Hnd.
18. Generate the histogram of student age Age.
19. Generate a boxplot of the student’s heart rate Pulse split out by the student’s
exercise regimen Exer.
 R tutorial: EDA in R Chapter | 7 85

7.9 Conclusion
With the exploratory analyses learned in this chapter, we suddenly have a lot
of tools at our disposal to help us understand and visualize our data. These are
not the only options available in statistics or in R, but they do represent some
of the most common. The next time we return to R, we will have the tools of
statistical inference at our disposal, and thus will be focused on implementing
these techniques in R.
Chapter 8

An incredibly brief introduction

to probability
Contents
8.1 Introduction 89 8.5 Random variables and
8.2 Random phenomena, probability distributions 94
probability, and the Law of 8.6 The binomial distribution 95
Large Numbers 90 8.7 The normal distribution 96
8.3 What is the role of probability in 8.8 Practice problems 98
inference? 91 8.9 Conclusion 99
8.4 Calculating probability and the
axioms of probability 92

8.1 Introduction
Now that we are able to summarize our data, we want to be able to take these
summaries and answer questions using statistical inference. Ultimately, for us
to use statistical inference, we need a basic understanding of probability. This is
because much of statistical inference comes down to the question, “What is the
chance that we observe more extreme data if a certain hypothesis is true?” To
quantify that chance, we need probability.
Unlike statistics, probability as a field of study has been around since the
1600s. Originally, many of the rules and theorems were designed to solve prob-
lems related to gambling, but the early field is full of many of the most signifi-
cant names in mathematics, including Pascal, Fermat, Bernoulli, and Gauss.
We have occasional encounters with probability in our daily lives. We see
the probability that it will rain or snow in weather reports. Sports talks about
the probability of teams winning before and during the games. Then there is of
course gambling, the father of probability, where we have odds in craps, poker,
roulette, and blackjack. Despite this, there are often misunderstandings or mis-
conceptions about probability and the general idea behind it. (See Fig. 8.1.)
This chapter will introduce us to the basics of probability, and the general
ideas that we will return to in statistical inference. We will begin with the ideas
behind random phenomena, random variables, the role of probability, and the
rules of probability.

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00020-1 89

Copyright © 2022 Elsevier Inc. All rights reserved.
90 Basic Statistics With R

FIGURE 8.1 A win probability plot for Game 5 of the 2017 World Series. Plotted is the Houston
Astros’s probability of winning the game at the time.

8.2 Random phenomena, probability, and the Law of Large

Numbers
To begin, we need to discuss why we need probability and what it does. In short,
probability helps us to explain random phenomena. A random phenomenon is
any event that we do not know the outcome of in advance. This can be anything
from a coin flip to the height of an individual. As long as we cannot know
the outcome of the event, it qualifies. And probability is a methodology that
quantifies the likelihood of outcomes for random phenomena.
There are several interpretations of probability, including subjective proba-
bility, classical probability, and frequentist probability. Subjective probability
bases the likelihood of outcomes on individual belief in how likely the outcome
is. In this formulation, probability is—of course—subjective and varies from
individual to individual. While there are many interesting consequences of this
interpretation, we will not focus on it in this text.
The second interpretation of probability we will discuss is classical prob-
ability. In classical probability, the probability of any event is the proportion
of possible outcomes that fall into that event. While classical probability has
its uses in simple situations—we will talk about this later when we discuss the
rules of probabilities—it is limited by the assumption that each possible out-
come is equally likely. This is rarely true, leading us to one final interpretation
of probability that has no such assumption.
The most common interpretation of probability is the frequentist proba-
bility model. The general idea behind this interpretation is that if we repeat a
random phenomenon a large number of times, it will get closer and closer to the
true probability of the event occurring. This is also sometimes referred to as the
long-run proportion view of probability.
 An incredibly brief introduction to probability Chapter | 8 91

Say we are flipping a coin and want to know the probability of flipping a
heads. Assuming the coin is fair, the chances of flipping a head should be 0.5.
If we flip a coin over and over, the proportion of heads should get closer and
closer to 0.5, as can be seen in Fig. 8.2.

FIGURE 8.2 The frequentist proportion view of probability as applied to flipping a fair coin. As
the number of flips gets larger and larger, the proportion of heads gets closer and closer to 0.5.

Why does this long-run view of probability work? The reason is a statistical
theorem called the Law of Large Numbers. This law states that as our sample
size gets larger—or n → ∞—the sample mean x̄ of any sample we take will
converge toward the true value of the mean μ. Similarly, as the sample size
gets larger, the sample proportion p̂ will converge toward the true value of the
proportion p. Mathematically, we can write this as

lim x̄ = μ
n→∞
lim p̂ = p.
n→∞

As an example, the probability that you roll any number on a fair, six-sided
die is 1/6. So, the proportion of 3s will converge to 1/6 as the sample size goes
to infinity.

8.3 What is the role of probability in inference?

We now know what probability is and how it helps us in general. However,
what we are most interested in is defining the role of probability in statistical
inference. By explaining random phenomena, probability helps us measure un-
certainty. Put another way, probability can help us quantify how confident we
are in the conclusions that we draw from inference.
Think about the following scenario: You play a game against an opponent
where you each put up $5. Your opponent rolls a single die, and if a 6 comes up,
they win the bet. Otherwise, you win the bet. Generally speaking, this seems like
92 Basic Statistics With R

a really good game for you because you should win five times more often than
your opponent. Now, say that your opponent rolls five consecutive 6s, leading
him to win $25. What would you likely think? After two or maybe even three
6s, you might think that your opponent got really lucky. By the fifth consecutive
6, you would likely be thinking that your opponent either cheated, that the die
was not fair, or both. You would not be unwarranted in thinking this. If the die
was fair and your opponent did not cheat, the chances of rolling five consecutive
6s is 1/7776, a very unlikely event.
In a similar sort of question, say you were flipping a coin—that you assume
is fair—20 times and counting the number of times you see a heads. You would
expect to see 10 heads if the coin was fair. If you saw 12 heads in the 20 flips,
would you start to think that the coin was unfair? 14 heads? How many heads
would you need to see before you start to suspect that the coin is unfair? This
is the general idea behind statistical inference: Assume a certain state of the
world, collect data, and consider if the data makes you think your assumption is
incorrect. Probability will help us quantify this, and we will go into more detail
on this topic in later chapters.

8.4 Calculating probability and the axioms of probability

When we are discussing probability, we really are discussing the chance that
some outcome occurred for a random phenomenon. But to actually calculate that
chance, we have to know the full list of all possible outcomes for that random
phenomenon. This full list of all outcomes for a random phenomenon is called
the sample space, usually denoted S. For example, if we were flipping a coin
one single time, we would say that the sample space would be heads and tails,
or S = {H, T }. If we were to extend that to flipping a coin two times, our sample
space would be heads on both flips, Tails on both, heads then tails, and tails then
heads, or S = {H H, H T , T H, T T }. Similarly, if we were to roll two dice and
added the results together, the sample space would be S = {2, 3, 4, ...10, 11, 12}.
After we have listed out all the possibilities for a random phenomenon, we
can go about considering the probability of any event. An event is any subset
of the sample space. Often these events are noted with the capital Latin letter
(A, B, etc.). For example, for the random phenomenon of flipping a coin twice
(S = {H H, H T , T H, T T }), an event we could consider is the event where there
is exactly one head seen in the two flips (A = {H T , T H }), or the event where
we see at least one tail in the two flips (A = {T H, H T , T T }).
In classical probability, when we calculate the probability of some event A,
we assume that every individual outcome in the sample space S is equally likely.
If that is the case, then the probability of an event A occurring is the number of
outcomes in A divided by the number of outcomes in S, or

# of outcomes in A
P (A) = .
# of outcomes in S
 An incredibly brief introduction to probability Chapter | 8 93

Consider the example again where you roll two dice and add them together—
essentially the gambling game called craps. There are 36 total combinations of
first die and second die, defined below with die 1 on the rows, die 2 on the
columns.

1 2 3 4 5 6
1 2 3 4 5 6 7
2 3 4 5 6 7 8
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

Each one of these 36 outcomes is equally likely, assuming the dice are fair.
This makes the game an opportunity to exercise classical probability. Say that
we are interested in the event A where we roll a 7 or 11. In the grid, there are
eight 7s or 11s, so the probability that event A occurs is P (A) = 8/36 or 0.2222.
Now, say that we want to know the probability of event B where we roll a 2, 3,
or 12. There are four 2s, 3s, or 12s, leading to P (B) = 4/36 = 0.1111.
On occasion, we are not interested in some event A occurring so much as
the chance that is does not occur. This event, A not occurring, is referred to as
the complement of A, or AC . The complement of A is made up of all elements
in the sample space S that are not in A.
Say you run a simple lottery, with a drawing of two balls with possible num-
bers being 1 through 3. The sample space will be S = {11, 12, 13, 21, 22, 23, 31,
32, 33}. If the event A is where the winning combo includes at least a single 2,
what will AC be? AC = {11, 13, 31, 33}
Notice that if you combine A and AC , you will get the entire sample space
S. Once we notice that, we are able to define our basic rules for probability, re-
ferred to as the Axioms of Probability or Kolmogorov Axioms—after Russian
probabilist Andrey Kolmogorov, who set down the axioms in 1933.
1. P (A) ≥ 0
2. P (S) = 1
3. If events A and B are disjoint sets—in other words, have no overlap—then
P (A or B) = P (A) + P (B).
These axioms hold regardless of the paradigm of probability being consid-
ered. Additionally, further properties can be derived from these consequences.
• P (AC ) = 1 − P (A), known as the complement rule.
• If event A contains all of event B, then P (A) ≥ P (B).
• If A is the empty set—an event with no items—then P (A) = 0.
94 Basic Statistics With R

8.5 Random variables and probability distributions

We have talked about the basics of probability, but now we want to talk
about one more key concept for probability. Right now, we have been talking
about probability in terms of events and random phenomena. These phenom-
ena in many instances are defined by categorical values, such as flipping a coin
(Heads/Tails) or the outcome of a game (Win/Loss). In general, mathematics
likes to deal with numeric values, so we need another definition to really gener-
alize random phenomena, regardless of categorical or numeric outcomes.
This is where random variables come in. A random variable is a numeric
measure or summary for a random phenomenon. For random phenomenon with
numeric outcomes, such as rolling a dice, the random variable could be the
same as the random phenomenon. However, for categorical random phenom-
ena, things will have to be adjusted slightly. Instead of looking at the outcome
of a two coin flips, with a sample space of S = {H H, H T , T H, T T }, we could
create the numeric measure of the number of heads seen in two coin flips. This
converts the categorical phenomenon to a numeric summary, making it easier to
work with.
Generally speaking, we notate the random variable that we define with a
capital X, or occasionally Y or Z, and any observed value of the random variable
is notated using lower case x, y, or z.
Say that we are playing roulette, which involves spinning a wheel with 38
options numbered 0, 00, and 1 through 36. Say we are interested in the number
of evens that come up in five spins. Our random variable would be X=# of
evens in five spins. Now, if we observed a five-spin run of 5, 11, 6, 21, 32, our
observed value of the random variable would be x = 2.
There are two types of random variables: discrete and continuous. A discrete
random variable is one that has a countable number of outcomes. Simply, this
means that you can list out every single possible outcome for that random vari-
able, it is a discrete random variable. Our previous example of the number of
heads in two coin flips is discrete, since the possible outcomes are 0, 1, or 2
heads in the two flips. Another example of a discrete random variable is the
number of times we have to flip a coin until we get our first head. It could take
one, two, or even—theoretically—an infinite number of flips. But we can list
them all out: 1, 2, 3, ....
A continuous random variable is a random variable with an uncountable
number of outcomes, meaning that we cannot list out every outcome. For ex-
ample, say that your random variable was the amount of time it took to run a
mile. If you were able to measure the exact time, this would be any number
between—theoretically—0 and infinity. It is impossible to write down each of
those numbers, so this would make this a continuous random variable.
We are able to calculate probabilities for these random variables, just as we
did for events. However, because random variables are more universally defined
across categorical and numeric phenomena, it is easier to get a summary of the
 An incredibly brief introduction to probability Chapter | 8 95

probabilities, specifically through probability distributions. A probability dis-

tribution is a summary of a random variable that gives all possible values of the
random variable along with their probability of occurring. These probabilities
can be defined by a function, or can be specified by the researcher. Often, when
the probability distribution is defined by a function, the probability distribution
has what are called parameters. These parameters are values that affect what
the probability distribution looks like, and thus what the probability of each out-
come will be.
There are an infinite number of possible probability distributions, but we are
going to talk briefly about two of the more common probability distributions:
the binomial distribution and the Normal distribution.

8.6 The binomial distribution

Say that you were to flip a coin once. In this case, there are two possible out-
comes: heads and tails. We know that there is a certain probability that you
observe a heads in an individual flip—which we will call p—assumed to be
0.5 if the coin is fair. Now, say that you repeat this process a total of 10 times.
A common random variable in this scenario would be how many heads you
observe in these 10 flips.
The binomial distribution is a probability distribution designed for this par-
ticular scenario. This discrete distribution describes the number of successes—
however success is defined—X observed in n independent trials. Because we
are looking at the number of successes in a limited number of trials, the possible
values for X would be 0, 1, 2, ..., n, making it a probability distribution for a
discrete random variable.
The binomial distribution has its probabilities defined by a function, which
means that there are parameters that describe the shape and values of the func-
tion. For a binomial distribution there are two parameters: the number of re-
peated trials n and the probability of success in an individual trial p. So, our
example at the start of this section described a scenario where a coin was flipped
10 times with a probability of success being 0.5. This means that the number of
heads observed—heads being considered a success in this case—X will follow
a binomial distribution with 10 trials and p = 0.5, notated X ∼ Bin(10, 0.5).
The parameters of the binomial distribution can also give us information
about how many successes we can expect to see in our n trials. In general, if
X ∼ Bin(n, p), then the expected—or average—number of successes in those n
trials will be n × p. For example, consider our initial example where the number
of heads X in 10 coin flips followed X ∼ Bin(10, 0.5). This means that the
expected or average number of heads we would expect to see in these 10 flips
would be 10×0.5 = 5. This knowledge can help us to understand if the observed
result was higher or lower than what should be expected.
96 Basic Statistics With R

8.7 The normal distribution

Normal distributions are one of the most if not the most important probability
distributions in statistics. We will talk about Normal distributions in more detail
later, but we want to at least introduce them now. (See Fig. 8.3.)

FIGURE 8.3 The Normal distribution.

Normal distributions are a “bell-shaped” distribution, similar to the bell-

shaped data that we reference earlier. The Normal distribution is symmetric,
unimodal, and bell-shaped, and a random variable that is defined by a normal
distribution can take on values from −∞ to ∞, making it a probability distri-
bution for continuous random variables. The shape of Normal distributions are
defined by a specific function, and as such, the Normal distribution has param-
eters that define that function. In the Normal distribution case, there are two
parameters: the mean parameter μ and the variance parameter σ 2 .
The first parameter, the mean parameter μ, defines the center of the Normal
distribution. If a random variable is defined by, or follows, a Normal distribution
with mean parameter μ, the average value for that random variable will be μ. In
other words, if we could collect every single value from that normal distribution
and average them together, we would find that the average is μ.
The second parameter, the variance parameter σ 2 , defines the spread of the
Normal distribution. If a random variable is defined by, or follows, a Normal
distribution with variance parameter σ 2 , the variance for that random variable
will be σ 2 . Again, if we could collect every single value from that normal dis-
tribution and calculate their variance, we would find that the average is σ 2 .
We can see the effect of these parameters in Fig. 8.4. If the mean μ changes,
the probability distribution will be shifted to the left or right, but ultimately the
distribution will be centered around the value of μ and will all have the same
variance—i.e., spread. If the variance σ 2 is increased or decreased, the proba-
bility distribution will be stretched out or compressed in, while still retaining the
overall bell shape and still centered around the mean μ.
 An incredibly brief introduction to probability Chapter | 8 97

FIGURE 8.4 The effect of the mean and variance parameters on the Normal distribution. On the
left, the effect of changing the mean μ. On the right, the effect of changing the variance σ 2 .

Of the various Normal distributions, there is one particular distribution that

is the most important distribution for statistics: the Standard Normal distribu-
tion. This is the Normal distribution with mean parameter μ = 0 and variance
parameter σ 2 = 1. Why is this particular Normal so important? It turns out that
we can convert any Normal distribution, with any mean and variance values, to
a Standard Normal. If a random variable X follows a Normal distribution with
mean μ and variance σ 2 —notated X ∼ N (μ, σ 2 )—then if we create a random
variable Z such that
X−μ
Z=
σ
then Z will follow a Standard Normal distribution. As an example, the height
of the American male roughly follows a Normal distribution with a mean of
μ = 68 inches and a variance of σ 2 = 9. If we wanted to convert this to a random
variable Z so that Z follows a Standard Normal distribution, we would take

X − 68
Z= .
3

While this ability to convert any Normal distribution to a Standard Normal

seems like a trivial fact, it will become extremely important for us to test claims
when doing statistical inference. In addition to being able to convert Normal
distributions to a common standard Normal, Normal distributions can be com-
bined through addition and subtraction. Say that we have two random variables;
X following a Normal distribution with mean μX and variance σX2 and Y fol-
lowing a Normal distribution with mean μY and variance σY2 . The sum of those
two variables X + Y will also follow a Normal distribution with mean μX + μY
98 Basic Statistics With R

and variance σX2 + σY2 . That is,

   
X ∼ N μX , σX2 , Y ∼ N μY , σY2
 
X + Y ∼ N μX + μY , σX2 + σY2 .

So, for example, say X ∼ N (6, 16) and Y ∼ N (3, 4). If we were to add X
and Y together, then X + Y will follow a Normal distribution with the mean
being 6 + 3 = 9 and variance 16 + 4 = 20. That is, X + Y ∼ N (9, 20).
A similar property holds for subtracting two Normal distributions. If X fol-
lows a Normal distribution with mean μX and variance σX2 and Y follows a
Normal distribution with mean μY and variance σY2 , then the difference of those
two variables X − Y will also follow a Normal distribution with mean μX − μY
and variance σX2 + σY2 . That is,
   
X ∼ N μX , σX2 , Y ∼ N μY , σY2
 
X − Y ∼ N μX − μY , σX2 + σY2 .

Taking the same example from above—X ∼ N (6, 16) and Y ∼ N (3, 4).
Then the difference X − Y will follow a Normal distribution with the mean
being 6 − 3 = 3 and variance 16 + 4 = 20. That is, X + Y ∼ N (3, 20).

8.8 Practice problems

1. Consider you have a full deck of 52 cards. What would be the sample space
for the random phenomenon of drawing a single card?
2. Say your event A would be drawing a jack or queen or king (i.e., a face
card) from a standard deck of cards. What is P (A)?
3. Say your event A was drawing a spade from a standard deck of cards. What
is P (AC )?
4. Say that you roll a 100-sided die, with the smallest value being 1 and the
largest being 100. What is the sample space for a single roll?
5. Assuming all rolls are equally likely, what is the probability of rolling a
number divisible by 10 on a 100-sided die?
6. Say that you flip a coin three times and your random variable X is the
number of tails in those three flips. What values can X take on?
7. Say that you flip a coin three times and your random variable X is the
number of tails in those three flips. Assuming each outcome in the sample
space is equally likely, what is P (X = 2)?
8. In 2019, the Baltimore Orioles played the Boston Red Sox 19 times. Say
that in any individual game, the probability that the Orioles would win is
p = 0.4. If we say that X is the number of Orioles wins observed in these
19 games—and assumed that the results of each game was independent—
what distribution would X follow?
 An incredibly brief introduction to probability Chapter | 8 99

9. Say you were at a craps table, where the probability of winning is p =

0.492. If you were to play eight games of craps and X is the number of
games won in those eight games, what distribution would X follow?
10. In the craps table example above, where the probability of winning is p =
0.492 and eight games of craps were played, how many games would you
expect to win?
11. Say that you have a random variable that follows a Normal distribution
with mean −2 and variance of 7 (X ∼ N (−2, 7)). How would you define
Z so that Z follows a Standard Normal (Z ∼ N (0, 1))?
12. Say that the random variable X follow the probability distribution X ∼
N (3, 1). How would you define Z so that Z follows a Standard Normal
(Z ∼ N (0, 1))?
13. Say that you have two random variables that follow Normal distributions,
X ∼ N (−3, 1) and Y ∼ N (5, 4). What will be the distributions of X + Y
and X − Y ?
14. Say that the random variables X and Y follow identical standard Normal
distributions, or N (0, 1). What will be the distributions of X + Y and X −
Y?

8.9 Conclusion
Probability allow us to quantify uncertainty in a way that makes a wide variety
of events comparable. There are multiple ways to do this quantification, but
they all follow a similar set of rules that define our probabilities. These rules
and calculations come together in our probability distributions, which allow us
to describe our random variables in a way that use just a couple of numbers,
called our parameters. These distributions will allow us to ultimately say how
rare our observed data is in a way that is essential to statistical inference.
Chapter 9

Sampling distributions, or why

exploratory analyses are not
enough
Contents
9.1 Introduction 101 9.4 Practice problems 107
9.2 Sampling distributions 101 9.5 Conclusion 107
9.3 Properties of sampling
distributions and the central
limit theorem 105

9.1 Introduction
As we saw in previous chapters, exploratory data analyses can give us a fair bit
of information about variables in our dataset. These analyses can show us the
center and shape of numeric data, the proportions of categories, and if variables
appear to be associated with each other. However, exploratory analyses are not
enough to help us answer questions. In this chapter, we will explore the reasons
why, and thus move toward the ideas behind statistical inference.

9.2 Sampling distributions

To this point, we have been working on calculating statistics from our sam-
ple. However, remember that we do not want to merely calculate statistics. Our
real goal is to do inference about our parameters. So, let us quickly recall what
are parameters and statistics. Parameters are numbers that describe our entire
population of interest. Our statistics are numbers that are calculated from our
sample, and that generally speaking estimate our parameters. For example, the
sample mean x̄ estimates the population mean μ, the sample proportion p̂ es-
timates our population proportion p, and the sample variance s 2 estimates the
population variance σ 2 . The former in each of those pairs is a statistic calculated
from a sample, while the latter is a parameter.
When we calculate a sample statistic, we get an estimate of our population
parameter that we do not know the exact value of. In classical statistics, one of
our fundamental assumptions is that all our population parameters are fixed and

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00021-3 101

Copyright © 2022 Elsevier Inc. All rights reserved.
102 Basic Statistics With R

unknown quantities. If our parameters were known, there would be no reason to

go through the whole statistical process.
Even if we were able to know our population parameter, our sample statistic
would rarely equal exactly that parameter. This is because our sample statistic
is a random quantity, dependent on our sample, and thus would only happen on
rare occasions. As an example, let us consider a simple case where our popula-
tion is small and our population parameter is known.
Say that we are interested in estimating the average number of credit hours
taken in a semester for a population of five students, with their number of credit
hours given in the table below.

Student ID Credits
1 18
2 15
3 12
4 13
5 15

This population would have a population mean of μ = 14.6. Now let us

further say that for some reason we were only able to take a sample size of
n = 2, meaning that there are only 10 possible combinations of two students
selected from the population of five that we started with. These combinations
are given below, along with the different sample means that will result from
each sample.

Students Average Students Average

1,2 16.5 2,4 14
1,3 15 2,5 15
1,4 15.5 3,4 12.5
1,5 16.5 3,5 13.5
2,3 13.5 4,5 14

The true population mean is 14.6, but depending on the sample, we can
have a variety of sample averages. So, our sample mean—or any of our sam-
ple statistics—is a random variable because our sample is chosen randomly.
Recall that a random variable is a numeric summary of some event that we do
not know the outcome of in advance, in this case our sample. Since the mean is
a random variable, like every other random variable it will follow some proba-
bility distribution. The sampling distribution is the probability distribution for
a sample statistic for a sample of size n. Every sample statistic has a sampling
distribution, with some being easier to work with than others. In general, the
sampling distribution helps give us an idea if the data that we collected is un-
usual in any way. We can view the sample distribution of our example above,
visualized through a histogram. (See Fig. 9.1.)
 Sampling distributions Chapter | 9 103

FIGURE 9.1 Histogram describing the sampling distribution of the sample mean for our credit
hours example.

In addition to the distribution of our sample statistic, we also want to know

how much variability exists in our sample statistic. If there is a large amount
of variability, we could see a wide variety of values for the sample statistic. To
see how much variability was in a variable, we eventually settled on calculating
the variance and standard deviation. We can do the same for a sample statistic,
that is, calculate the variance or standard deviation of a sample statistic. The
standard deviation of this sample statistic is what is referred to as the standard
error, commonly notated as s.e.(Statistic).
Consider the same example above, where we calculate a sample mean of
credit hours for a sample size of two students from our population.

Students Average Students Average

1,2 16.5 2,4 14
1,3 15 2,5 15
1,4 15.5 3,4 12.5
1,5 16.5 3,5 13.5
2,3 13.5 4,5 14

We know the population mean—and the average of all the sample means—
is 14.6. So we can calculate the standard error of these sample means to be
s.e.(x̄) = 1.26
Now, we mentioned that the sampling distribution is the probability distribu-
tion for a sample statistic for a sample of size n. This implies that our sampling
distribution is in some way affected by the sample size, and if the distribution
is affected by the sample size, it is entirely reasonable to expect that the stan-
dard error of the statistic is also affected by the sample size. We know from the
Law of Large Numbers that as the sample size gets larger, our sample statistics
104 Basic Statistics With R

get closer and closer to their parameter’s true value. But what happens to the
standard error?
Consider the same example above, where we calculate a sample mean, but
this time we up our sample size to four. This will have five possible samples
with their sample means given below.

Students Average
1,2,3,4 14.5
1,2,3,5 15
1,2,4,5 15.25
1,3,4,5 14.5
2,3,4,5 13.75

The average of the sample means is still 14.6, so the standard error would
now become s.e.(x̄) = 0.515. So, our standard error decreases as our sample
size increases. This makes intuitive sense as well. If we were able to sample
the entire population, we would only calculate one sample mean—equal to the
population mean—and you would not have any variability in that one sample
mean. In general,
s.e.(Statistic) → 0 as n → ∞.
So, since the variability of the sample statistic decreases as the sample size
gets larger, this means that our sampling distribution will become more com-
pressed as the sample size gets larger. Consider the situation where you are
finding the average commute time for a population of 20 employees at a com-
pany. The individual commute times are given in Fig. 9.2.

FIGURE 9.2 Individual commute times for 20 employees.

If you knew the commute times for all 20 employees and considered all
possible samples of size n = 5, n = 10, and n = 15, you get the sampling dis-
 Sampling distributions Chapter | 9 105

FIGURE 9.3 Sampling distribution for sample average commute time for n = 5, n = 10, and
n = 15.

tributions below, which become more compressed as our sample size increases.
(See Fig. 9.3.)
To summarize, sample statistics are—by the nature of random samples—
random variables themselves. Thus, they have probability distributions—called
the sampling distribution—and variability—-called the standard error. These
distributions and variability are affected by the sample size, with the variability
decreasing and the distributions constricting as the sample size increases.

9.3 Properties of sampling distributions and the central limit

theorem
Now, all these sampling distributions as currently explained can be useful, ex-
cept for one slight problem: thus far, all our sampling distributions require us to
consider every possible combination of subjects in our sample. Knowing them
requires us to collect every possible sample, and as we know, collecting just a
single sample is difficult enough. Collecting all of them is next to impossible.
With this in mind, unless there are certain properties about the sampling dis-
tribution that are known in advance, the sampling distribution is near impossible
to get. Fortunately, there are several properties about the sampling distribution
that are known. To begin, it is known that the sampling distribution is centered
around the true value of the population parameter. That is, if we were able to
take the average of all the possible values in the sampling distribution, we will
get the value of the population parameter. Further, the spread of the sampling
distribution can be described by the standard error.
The final property deals with the sampling distributions for the sample pro-
portion and the sample average. If you look at the sampling distributions for
the commute time hypothetical, you will notice that they look particularly bell-
106 Basic Statistics With R

shaped. The same holds if you look at the sampling distribution of the sample
proportion.
Consider the scenario where you flip a coin 100 times and count the number
of heads in the 100 flips. Say you did this 10,000 times. The sampling distribu-
tion for the sample proportion of heads would look like Fig. 9.4.

FIGURE 9.4 Sampling distribution for the proportion of heads in 100 coin flips.

This bell-shaped nature of the sampling distribution for these two statistics
is not a coincidence. It turns out that as the sample size goes to infinity, the
sampling distribution for the sample mean becomes a Normal with parameters
μ equal to the true value of the population mean and variance being the standard
error of x̄ squared. That is,
 
x̄ ∼ N μ, s.e.(x̄)2 as n → ∞.

A similar property holds for the sample proportion. Namely, as the sample
size goes to infinity, the sampling distribution for the sample proportion will
become Normal with parameters μ being the true value of the population pro-
portion p and variance being the standard error of p̂ squared. That is,
 
p̂ ∼ N p, s.e.(p̂)2 as n → ∞.

This result—that the sampling distributions of the sample mean and propor-
tion becoming Normal—is called the central limit theorem. Thanks in part to
this result, it is possible for us to do statistical inference in the form of hypothesis
testing and confidence intervals, which will begin next chapter.
As an example of the central limit theorem, let us look at the example seen
in Fig. 9.4. In this example, we are flipping a fair coin—true probability of
flipping a heads is p = 0.5–100 times—and calculate the sample proportion p̂.
If this is the case, the standard error of p̂ would be s.e.(p̂) = 0.05, meaning that
our sample proportion would roughly follow a Normal distribution with mean
 Sampling distributions Chapter | 9 107

p = 0.5 and variance of s.e.(p̂)2 = 0.052 . Mathematically, we would write this

as
 
p̂ ∼ N 0.5, 0.052 .

9.4 Practice problems

1. Say you are flipping a fair coin (50/50 chance of heads/tails) 50 times and
calculating the proportion of tails in the flips. If you repeated this process a
large number of times, the sampling distribution for the proportion of tails
would become approximately Normal centered around what value?
2. Say that you are going to take a sample of size n = 100 from a population
and your colleague takes a sample of size n = 500 from the same popula-
tion. Will you or your colleague have smaller standard errors for the sample
statistics?
3. Say that you are calculating the sample average for a sample from a popu-
lation. As your sample size increases, what distribution (or class of distribu-
tions) will the sampling distribution for your sample average be?
4. Say you are taking a sample of people’s stride lengths and calculating the
sample average stride length. As you sample more individuals, what happens
to the standard error of the sample average stride length? What distribution
will the sampling distribution for the sample average approach?

9.5 Conclusion
As we had previously suggested, data summaries are not sufficient to be the
foundation for making decisions. This is because they are calculated from ran-
dom samples or randomized experiments and are therefore random variables
themselves. However, this means that they follow probability distributions—
called the sampling distribution—and in some cases these probability distri-
butions are known. Particularly in the case of the sample average or sample
proportion they follow Normal distributions, a result known as the central limit
theorem. This result will help us going forward to be able to understand how un-
likely our data is, providing us the eventual foundation for coming to a decision
based on our data.
Chapter 10

The idea behind testing

hypotheses
Contents
10.1 Introduction 109 10.3.3 What is our level of
10.2 A lady tasting tea 109 doubt? 113
10.3 Hypothesis testing 110 10.4 Practice problems 115
10.3.1 What are we testing? 110 10.5 Conclusion 115
10.3.2 How rare is our data? 112

10.1 Introduction
We have finally reached the most important step of the process of statistics:
statistical inference. Inferential statistics allow us to consider the data that we
have collected in the light of probability and assumptions we make about the
state of the world. In doing so, we consider if the data we collected is rare under
our assumed world state and whether we should revise our assumptions. This
chapter will focus on introducing us to the most common of inferential statistics:
the hypothesis test. To begin, let us look at how hypothesis testing came about,
with a tale that reminds us how scientific and statistical discoveries can come
about in the most unlikely ways.

10.2 A lady tasting tea

One day in 1920s England, a group of university professors and researchers de-
cided to go to lunch. Among the party was Dr. Muriel Bristol and R.A. Fisher,
a phycologist and statistician working at nearby Rothamstead Experimental Sta-
tion.
During the course of the meal, tea with milk and sugar was served, and
Dr. Bristol claimed that there was a difference in the taste of the cup of tea
depending on if the tea was poured into the milk or the milk was poured into
the tea. In addition, she claimed she could tell the difference between the two
possibilities. On first hearing this, the claim sounds ridiculous, but Fisher is
interested in pursuing the idea further.
According to Fisher’s line of thinking, if Dr. Bristol could not tell the
difference—as most of the group there believed—when presented a series of
cups of tea with various mixing methods, she would be expected to correctly

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00022-5 109

Copyright © 2022 Elsevier Inc. All rights reserved.
110 Basic Statistics With R

identify half of the cups, just by pure chance. However, if she identified many
more cups correctly than expected, then the group assumption that she could not
tell the difference was likely incorrect.
And that is exactly what they did. Fisher and his colleagues presented Dr.
Bristol with eight cups of tea—four where the tea was poured into the milk
and four where the milk was poured into the tea—in random order. They then
recorded Dr. Bristol’s guesses and tallied up the results. It turned out that Dr.
Bristol got all eight guesses correct.

10.3 Hypothesis testing

As strange as that tale is, it is true. Further, it was the motivation behind the
whole idea of hypothesis testing, a statistical technique that appears in academic
journals of every discipline. Hypothesis testing is a class of procedures that use
sample data to examine a claim about a population parameter. In essence, we
are placing a claim on trial, using the data we collected as evidence to evaluate
the claim.
In this way, hypothesis testing is something akin to a court case. In court
cases, a defendant is put on trial, and this defendant is assumed innocent until
their guilt is proven. The burden of proof is on the prosecution, who must prove
that the defendant is guilty beyond the shadow of a reasonable doubt. In hypoth-
esis testing, we place some claim on trial and assume that the claim is true until
proven otherwise. In this case, the burden of proof is on the data that we collect.
We have to see data that is unlikely to occur by chance if the claim on trial is
true.
Generally speaking, hypothesis testing is a very methodical procedure. Even
though there are several varied tests that we will learn about, they are all com-
prised of an overarching framework of steps that we will see over and over.
Because of this, we are going to introduce the steps and ideas behind hypothesis
testing, specifically using the example of testing if a coin is fair.

10.3.1 What are we testing?

In hypothesis testing, our first step is to “state the charges”—to continue the
court case metaphor—specifying exactly what claim we are testing. In hypoth-
esis testing, this comes in two parts, called the null hypothesis and alternative
hypothesis. The null hypothesis—notated H0 —expresses the status quo, or the
current state of the world. Essentially, it assumes nothing interesting is going on
in our dataset.
The alternative hypothesis—notated HA —is the hypothesis that we ulti-
mately want to find evidence for or the hypothesis that we want to prove to be
true. These alternative hypotheses usually tend to be more interesting, or at least
generate more interesting results than the null hypothesis.
In hypothesis testing, we assume that the null hypothesis is true until proven
otherwise. Essentially, we go into our hypothesis tests assuming that the current
 The idea behind testing hypotheses Chapter | 10 111

state of the world is in fact true, that there really is nothing strange going on
in the data. This mirrors court cases, the defendant—our null hypothesis—is
innocent until proven guilty.
To illustrate this, let us consider the case where you want to discover if the
coin is fair, which we want to determine through a hypothesis test. To begin this
hypothesis test, we need to both our null and alternative hypotheses. In most
cases, when writing out the null and alternative hypotheses, we want to express
our claims in terms of some population parameter. This allows us to more easily
test these hypotheses mathematically, and requires us to define our population
parameters in many cases.
In the case of our fair coin, that parameter is the true probability of flipping a
heads—or the true proportion of heads. Let us define p to be this true probability
of flipping a heads, or p = P r(H ). If the coin is fair, the probability of flipping
a heads would be p = 0.5. This is our assumed state of the world, our status quo,
a world where nothing strange is happening. Because of this, this claim—that
p = 0.5—is our null hypothesis. We write this out as

H0 : p = 0.5.

Once we have defined our null hypothesis, we need to define our alternative
hypothesis. This alternative is going to completely depend on what we want
to show. For example, we might think that the coin that we are flipping makes
heads more likely to occur. If this is the case, the population proportion of heads
will be greater than 0.5, or p > 0.5. In that case, our alternative hypothesis is

HA : p > 0.5.

Or you believe that the coin makes tails more likely to occur, or heads less
likely to occur. In this case, the population proportion of heads will be less than
0.5, or p < 0.5. For this, our alternative hypothesis is

HA : p < 0.5.

These types of alternative hypotheses are referred to as one-sided tests, be-

cause they are only concerned with whether our parameter is less than or greater
than the value listed in the hypotheses—0.5 in this case. The last option is re-
ferred to as a two-sided test, and is the case where we want to investigate if the
coin is unfair in either direction. We do not know if the coin makes heads more
likely or less likely, we just believe it differs from fair in some way. This would
result in the population proportion just being not equal to 0.5, or p = 0.5. For
this last case, the alternative hypothesis is

HA : p = 0.5.

When defining our hypotheses, we cannot have any overlap between the null
and alternative. This makes reasonable sense, as we cannot have both of these
112 Basic Statistics With R

hypotheses be true. For example, the pair of hypotheses listed below are not a
valid set of null and alternative hypotheses.

H0 : p = 0.5
HA : p ≥ 0.5.

10.3.2 How rare is our data?

Once we have set out what we want to test, we get our data. Say that you flip
the coin 10 times to test the fairness of the coin, and come up with the following
sequence:
T , T , H, H, T , H, H, H, H, T .
Added up, that is six heads and four tails. Now, this data acts as the evi-
dence for our trial on our hypotheses. Say, for the sake of argument, that our
hypotheses are

H0 : p = 0.5
HA : p > 0.5

so we would hope to see more heads than tails to support our alternative hypoth-
esis. The question for you is: Is six heads and four tails enough evidence that
the coin is unfair and makes heads more likely? Six heads and four tails is pretty
close to 50/50, so maybe it is not enough evidence for you. What about seven
heads and three tails? Is that enough evidence? What do you consider enough
evidence?
The idea behind hypothesis testing is that we go in assuming that our null
hypothesis is true, and if we see sufficiently rare data we conclude that the as-
sumption in the truth of the null hypothesis is wrong and we have statistically
significant results. The problem is how to determine if our data is “sufficiently
rare.” That is where the p-value comes in. A p-value is the probability that we
see data that is at least as extreme as what we observed, assuming of course that
the null hypothesis is true. For example, in the coin flip sequence above, our
p-value would be the probability that we see 6 or more heads in a series of 10
coin flips if the coin is indeed fair (A value of 0.377).
P-values can have many misinterpretations. The p-value is not...
• the probability that the sample statistic equals our observed value
• the probability that the null hypothesis is true.
The second misinterpretation is much more common as well as more dam-
aging. It is important to recall that our parameters—about which we make our
hypotheses—are fixed values, so the hypotheses will either be true or false, with
no probability one way or the other. The p-value only makes a statement of how
rare our data is assuming that the null hypothesis is true. There are several ways
 The idea behind testing hypotheses Chapter | 10 113

to calculate p-values, but we will leave the discussion of calculating p-values for
a later chapter.

10.3.3 What is our level of doubt?

Now that we have hypotheses, data, and evidence in the form of a p-value, we
want to make a decision. In hypothesis testing, there are two possible decisions.
Our first possible option is to decide that the data we collected is sufficiently
rare, and thus unlikely to occur if the null hypothesis is true. Therefore, our as-
sumption that the null hypothesis is incorrect and we reject the null hypothesis.
Our other possible option is to conclude that the data we collected is not
sufficiently rare or unlikely to occur, assuming the null hypothesis is true. Thus
we cannot say that the null hypothesis is wrong and we fail to reject the null
hypothesis. Even though we cannot say the null hypothesis is wrong, we also
cannot say that null hypothesis is right. At most, we can say that the null hypoth-
esis is plausible. As defendants in court cases are often ruled “not guilty,” null
hypotheses can only be rejected—i.e., wrong—or failed to reject—i.e., plausi-
ble.
But how do we make our decision? Which conclusion do we reach? It seems
reasonable that p-values would be involved, and this is in fact true. In hypothesis
testing, if we also assume that the null hypothesis is true and our p-values reach
a certain level, we can generally agree that we are observing something rare.
We just have to decide the level of rarity where we start thinking that the null
hypothesis is wrong. In essence, we need to set our “reasonable doubt.” We do
this by setting our significance level. The significance level—notated α—is a
pre-chosen boundary that sets our level of reasonable doubt with which we will
make our decision. If our p-value is less than our significance level α—i.e., our
evidence surpasses our level of reasonable doubt—we reject the null hypothesis.
On the other hand, if our p-value is greater than or equal to our significance level
α—or our evidence fails to surpass our reasonable doubt level—we fail to reject
the null hypothesis.
But how do we choose our significance level? We can go back to the court
case analogy to think about this. In sentencing, it is likely that we will get some
decisions incorrect based on our level of reasonable doubt. If we have a very
stringent interpretation of reasonable doubt—that is, it takes a lot of evidence
to convict someone—we will likely let several guilty individuals go free. If we
have a very loose definition of reasonable doubt—where it takes comparatively
little evidence to convict—we will likely convict several innocent individuals.
As a juror, you would have to balance these two competing concerns in order
make the best set of decisions.
The same problem exists for hypothesis testing. In setting α and making
these decisions, you can either make the correct choice—reject H0 when it is
indeed false in reality or fail to reject H0 when it is true in reality—or make an
error—reject H0 when it is true in reality or fail to reject H0 when it is false in
114 Basic Statistics With R

reality. In statistics, these particular errors have specific names. If we reject the
null hypothesis when it is in reality true, this is called a Type I Error. If you
fail to reject the null hypothesis when it is in reality false, this is called a Type
II Error.

Test result
Reject H0 Fail to reject H0
H0 True Type I Error Correct
Reality
H0 False Correct Type II Error

Ideally, we would like to avoid both errors, but this is impossible. If we make
it hard to reject the null hypothesis by setting α to be very low—and avoid Type
I errors—we make it likely that we will fail to reject the null hypothesis when
we should—i.e., we make more Type II errors. If we make is easy to reject the
null hypothesis by setting α higher—and avoiding Type II errors—we make it
more likely that we will reject the null hypothesis when we should not do so—
i.e., we make more Type I errors. This all becomes a large balancing act where
we have to consider the consequences of making a Type I or a Type II error.
Generally, it is much more costly to commit a Type I error. Our null hypoth-
esis represents the current state of the world, so to reject that null hypothesis
and change the state of the world will require money, time, and manpower. If
we undertake all this cost supporting something that is not true in reality, this all
becomes a sunk cost and is usually more detrimental than remaining with the
status quo. With this in mind, we will try and control how often we make Type
I errors in our hypothesis test, even as we make it more likely that we will make
a few more Type II errors.
The next question is: How do we control Type I errors? It ultimately all
comes down to α. Our significance level determines when we reject the null
hypothesis, so it will determine how likely we are to make a Type I error. In
fact, α equals the probability that we make a Type I error, or

α = P r(Type I Error).

So, we need to choose α based on how comfortable we are in making a Type

I error. Consider if you were testing a drug that had potentially side effects. Here,
the cost of a Type I error is those side effects, so you may need to make Type
I errors very unlikely by setting α to be very small. In addition to this concern,
certain academic fields may specify standards for statistical significance that you
may have to consider. Generally speaking, the most common significance levels
are α = 0.05, α = 0.1, and α = 0.01, with α = 0.05 being the general standard.
Significance levels must be set before you see the data or calculate p-values.
You cannot calculate a p-value and then artificially select an α value that guaran-
tees statistically significant results, as this is dishonest and biases your results.
It would be akin to changing your “reasonable doubt” level in a court case to
 The idea behind testing hypotheses Chapter | 10 115

suit the evidence. You cannot bias your results toward one result or another and
retain your integrity in research.

10.4 Practice problems

Consider the case where you roll a six-sided die.
1. Say you believe that the die is unfair, specifically making sixes more likely
than it should be. What would be your null and alternative hypotheses in this
case?
2. Is the hypothesis test you set up a one-sided or two-sided test?
3. Say that you set your significance level for the test to be α = 0.05. What is
the probability that you commit a Type I error and conclude that the die is
unfair when it is in fact fair?
4. Say you conclude that it is plausible that the die is fair, when it is in fact
unfair. What type of error have you committed?
5. Say that the data you collect had you rolling 25 sixes and 75 non-sixes. This
results in a p-value of 0.01. Based on the significance level set earlier, what
would your conclusion be (In terms of the null hypothesis)?
Say that a survey conducted a year ago found that half of all social media
users used social media to check on someone they used to date [103].
6. Say you believe this proportion has changed in some way, but not specifically
greater or less than. What would be your null and alternative hypotheses?
7. Is the hypothesis test you set up a one-sided or two-sided test?
8. Say that you set your significance level for the test to be α = 0.01. What is the
probability that you commit a Type II error and conclude that the proportion
is still 0.5 when it is in fact is not equal to 0.5?
9. Say that the data you collect results in a sample proportion of 0.53 in 100
people. This results in a p-value of 0.5485. Based on the significance level
set earlier, what would your conclusion be (in terms of the null hypothesis)?

10.5 Conclusion
The philosophy behind hypothesis testing is fairly simple: place a claim about
our population parameters on trial, set a reasonable doubt, and let the data speak
as evidence. If the data is sufficiently rare that there is no reasonable way it
could have happened by chance, we conclude that we should reject our initial
claim. Otherwise, it should remain in place as a plausible claim. This mindset is
the foundation of many statistical techniques ranging from the simple to com-
plex. With this mindset now in place, we can work toward conducting tests for
population means and proportions.
Chapter 11

Making hypothesis testing work

with the central limit theorem
Contents
11.1 Introduction 117 11.3.1Practice problems 119
11.2 Recap of the normal distribution 117 11.4 Connecting data to p-values 119
11.3 Getting probabilities from the 11.4.1Practice problems 124
normal distributions 118 11.5 Conclusion 125

11.1 Introduction
Now that we have introduced the idea of hypothesis testing, we are going to
begin delving into it a little more. Specifically, we need to talk about how to get
our p-values, the driving force behind our decision-making process. There are
many techniques to get p-values, but a majority of them can be tedious or require
more theoretical explanations beyond the scope of an introductory course. The
most common, and indeed one of the simpler, methods of getting p-values takes
advantage of the central limit theorem, allowing us to turn sample statistics into
probabilities about how rare our data is. But first, we need to talk a little more
about the Normal distribution and how to work with this very important class of
distributions.

11.2 Recap of the normal distribution

Before we begin, let us recall what we previously learned about the Normal dis-
tribution. The Normal distribution is really a class of distributions; all unimodal,
symmetric, and bell-shaped. As the Normal distribution is defined by a function,
it has two parameters that give us the shape of the function; the mean parame-
ter μ and the variance parameter σ 2 . If a random variable X follows a Normal
distribution with mean μ and variance σ 2 —X ∼ N (μ, σ 2 ), the average of all
possible values of X will be μ and the variance for all possible values of X will
be σ 2 .
Within Normal distributions, there is one special case of the utmost impor-
tance: the Standard Normal distribution. The standard Normal distribution is a
Normal distribution with mean μ = 0 and variance σ 2 = 1. This distribution is
particularly useful because we can convert any Normal distribution to the stan-
dard Normal with relative ease. If a variable X follows Normal distributions

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00023-7 117

Copyright © 2022 Elsevier Inc. All rights reserved.
118 Basic Statistics With R

with mean μ and variance σ 2 , we create a new variable Z by taking

X−μ
Z= .
σ
Then Z will follow a standard Normal distribution. This will become very im-
portant to us when we are trying to do any hypothesis testing. Our other key
property of Normal distributions was that they could be combined using ad-
dition and subtraction. Namely, if two random variables X and Y both follow
Normal distributions, the X + Y and X − Y will also follow Normal distribu-
tions. Namely,
   
X ∼ N μX , σX2 , Y ∼ N μY , σY2
 
X + Y ∼ N μX + μY , σX2 + σY2
 
X − Y ∼ N μX − μY , σX2 + σY2 .

This property will also be important for hypothesis testing, specifically when we
are comparing two populations to see if they are similar or differ in some way.

11.3 Getting probabilities from the normal distributions

In order for us to use Normal distributions, we must be able to obtain probabil-
ities. The function that defines the Normal distribution does not make it easy to
get probabilities from it. Because of this, we have to use technology—in other
words, R—to get these values. The various statistical software packages that
exist make this job much easier. In R, the function we use to get probabilities
from all normal distributions is the pnorm function. The pnorm function takes
four arguments that entirely tell it what probability we are looking for, with the
format of the function is as follows:

pnorm(q, mean, sd, lower.tail)

The four arguments are as follows:

• q: The quantile of interest from the Normal distribution.
• mean: The mean μ—default value of 0—of the Normal distribution in ques-
tion.
• sd: The standard deviation σ —default value of 1—of the Normal distribution
in question.
• lower.tail: A TRUE/FALSE value—default value of TRUE—that defines if
whether we are interested in the probability
 that the Normal distribution is
less or greater than the quantile.
 For P N (0, 1) < q , we use lower.tail =
T RU E. For P N (0, 1) > q , we use lower.tail = FALSE.
It is important to note that the pnorm function uses the standard deviation σ
while we have defined the Normal distribution using the variance σ 2 . However,
 Making hypothesis testing work with the central limit theorem Chapter | 11 119

√
the variance and standard deviation are directly connected— σ 2 = σ —so we
just have to do is be cognizant of what we
 are inputting into
 the function.
So, for example, say that we want P N (5, 4) > 6.46 . To get this in R, we
would use the code

pnorm(q=6.46, mean=5, sd=2, lower.tail=FALSE)

Again, note√that we had sd = 2 since the variance of our Normal distribution

is σ 2 = 4 and 4 = 2. R then outputs the probability 0.2326951.

11.3.1 Practice problems

Find the following probabilities using R:
 
1. P N (3, 6.25) < 8
2. P N (−2, 16) > 7 
3. P N (1.2, 15.4) < 10
4. P N (0, 12) < 6 
5. P N (−3, 1) > 0

11.4 Connecting data to p-values

Now that we have reviewed some properties of the Normal distribution and in-
troduced how we can get probabilities from said Normals, we need to return to
hypothesis testing. In introducing hypothesis testing in the last chapter, we dis-
cussed the ideas behind hypothesis testing, specifically how we assume a state
of the world and only reject this state if the data is sufficiently rare. This rarity
is measured through p-values, the probability that we see more extreme data
assuming that the null hypothesis is true. There is a key step we left out in this
procedure: How do we get from our data to p-values? How do we calculate how
rare our data is?
To get at this idea, we are going to go back to the example that we were
working with earlier. In this example, we were flipping a coin, say 100 times,
and calculating the proportion of heads, in this case say 60 heads, to see if the
coin is fair or not. We ultimately want to get our p-value, the probability that we
see more extreme data given the null hypothesis is true.
To get a probability, we need a probability distribution. Specifically, we need
the probability distribution of the data or some statistic calculated from the data
we collect. Fortunately, we have the distribution of our sample statistics: namely,
the sampling distribution. We have, for this coin flipping example, the sampling
distribution for the proportion of heads. Even further, we know what this distri-
bution is thanks to the central limit theorem: it is a Normal distribution centered
around the true probability that we will flip a heads (p) and with variance equal
to the standard error of the sample proportion squared. That is,
 
p̂ ∼ N p, s.e.(p̂)2 .
120 Basic Statistics With R

However, this leaves a few things missing. We need to know what the true
value of p is along with the standard error of the sampling proportion. We never
know what the value of p is, as it is a population parameter and, therefore,
unknown. However, in hypothesis testing, we assume that the null hypothesis is
true, which gives us some assumptions about that value of p. Our null hypothesis
is that the coin is fair, or p = 0.5. So, if the null hypothesis is true, then the
sampling distribution of the sample proportion of heads would be

 
p̂ ∼ N 0.5, s.e.(p̂)2 .

This just leaves the standard error of the sampling proportion s.e.(p̂). As
we move forward, we will give you formulas for what the standard error of our
various sample statistics is. However, for this chapter we will fill in the standard
errors for the sampling distributions.
Our goal is to calculate our p-value, or the probability that we see more
extreme data from this sampling distribution. Put another way, we need the prob-
ability that we see a more extreme sample statistic. We have a sample statistic
that follows a Normal distribution for which we now know how to find probabil-
ities. Based on this distribution and the knowledge we now have, we just need
to understand what we mean by more extreme data.
This definition of “more extreme” depends on the alternative hypothesis. Say
that our alternative hypothesis is HA : p > 0.5, or that the coin makes flipping
a heads more likely. “More extreme” data for this alternative hypothesis—data
that would support the alternative hypothesis—would result in seeing at least
as many heads or more than we actually did. So, our p-value would be the
probability that we see greater than or equal to 60 heads in a series of 100
flips—or, the probability that we see a sample proportion greater than or equal
to 60/100 = 0.6.
To calculate this out, we need the full sampling distribution of p̂. The stan-
dard error of p̂ for this particular scenario—testing if the coin is fair using 100
flips—is 0.05, so the sampling distribution of p̂ under the assumption that the
null hypothesis is true would be is N (0.5, 0.052 ). Seeing “more extreme” data
in support of the alternative hypothesis that p > 0.5 is equivalent to seeing a
sample proportion greater than or equal to 0.6, so our p-value will be

 
P N (0.5, 0.052 ) ≥ 0.6 .

Based on the properties of Normal distributions, we can convert this to a

standard Normal by subtracting off the mean and dividing by the standard de-
viation of the Normal distribution. Any action that is done to one side of the
inequality in the probability statement must be done to the other side as well,
 Making hypothesis testing work with the central limit theorem Chapter | 11 121

just as is the case in algebra. Do so would give us

 
P N (0.5, 0.052 ) ≥ 0.6
 
N (0.5, 0.052 ) − 0.5 0.6 − 0.5
=P ≥
0.05 0.05

= P N (0, 1) ≥ 2).

We can find this probability using R with the code

pnorm(q=2, mean=0, sd=1, lower.tail=FALSE)

The resulting probability returned by R is 0.0228. So, if we observe 60 heads

in 100 flips with an alternative hypothesis
 of p > 0.5, our p-value will be 0.0228.
This probability is equivalent to P N (0.5, 0.052 ) ≥ 0.6 , which we can confirm
in R using the code
pnorm(q=0.6, mean=0.5, sd=0.05, lower.tail=FALSE)

What if we change that alternative hypothesis? What if we wanted to show

that the coin makes tails more likely, or in other words, HA : p < 0.5? More
extreme data now would be observing as many or fewer heads than we did. So,
our p-value would then be the probability that we see less than or equal to 60
heads in a series of 100 flips—or the probability that we see a sample proportion
less than 0.6.
Calculating this is similar to before. The standard error of p̂ is still 0.05 in
this scenario, so the sampling distribution of p̂—if the null hypothesis is true—
remains the same; a N (0.5, 0.052 )  
For this case, our p-value will P N (0.5, 0.052 ) ≤ 0.6 , since “more ex-
treme” data is observing a sample proportion is less than or equal to 0.6. Again,
we can convert this Normal distribution to the standard Normal by subtracting
off the mean and dividing by the standard deviation.
 
P N (0.5, 0.052 ) ≤ 0.6
 
N (0.5, 0.052 ) − 0.5 0.6 − 0.5
=P ≤
0.05 0.05

= P N (0, 1) ≤ 2).

Again, we would use the pnorm function in R to get this probability

pnorm(q=2, mean=0, sd=1, lower.tail=TRUE)

with the returned probability of 0.9772, which becomes our p-value for the al-
ternative hypothesis HA : p < 0.5.
The final alternative hypothesis that we have to consider is when HA : p =
0.5. This is going to require a little more thought, because data that would sup-
port the alternative hypothesis would mean that we saw evidence that the coin
122 Basic Statistics With R

makes heads more likely or tails more likely, so extreme comes now in two
directions.
In this way, we could define “more extreme” as further away from the null
hypothesis value—or 0.5 in this case—than what we observed. In our example,
the observed value for the sample proportion is 0.6, which is 0.1 away from
0.5. On the other side, 0.4 is 0.1 away from 0.5. So, more extreme would be all
sample proportions that are greater than or equal to 0.6 and less than or equal to
0.4. Now, our sample proportion will still follow a N (0.5, 0.052 ) according to
the central limit theorem. So, our p-value will wind out being
   
P N (0.5, 0.052 ) ≤ 0.4 + P N (0.5, 0.052 ) ≥ 0.6 .

We are still able to convert each of these Normals to a standard Normal by

subtracting off the mean and dividing by the standard deviation
   
P N (0.5, 0.052 ) ≤ 0.4 + P N (0.5, 0.052 ) ≥ 0.6
 
N (0.5, 0.052 ) − 0.5 0.4 − 0.5
=P ≤
0.05 0.05
 
N (0.5, 0.05 ) − 0.5 0.6 − 0.5
2
+P ≥
0.05 0.05
   
= P N (0, 1) ≤ −2 + P N (0, 1) ≥ 2 .

Now, we still have to calculate two probabilities to get our p-value. How-
ever,
 let us look at these two
 probabilities
 individually. If we do, we will see that
P N (0, 1) ≤ −2 and P N (0, 1) ≥ 2 are both 0.02275. This is not coinciden-
tal. Normal distributions are perfectly symmetric, which means that they will
have the same tail probabilities, or in other words,
   
P N (0, 1) ≤ −a = P N (0, 1) ≥ a .

We can use this to our advantage, taking our two probabilities and converting
them to a single probability.
   
= P N (0, 1) ≤ −2 + P N (0, 1) ≥ 2
   
= P N (0, 1) ≥ 2 + P N (0, 1) ≥ 2
 
= 2 × P N (0, 1) ≥ 2 .

We can get this using the pnorm function in R to find that our p-value is
0.0455. We will use this conversion in all instances where our alternative hy-
pothesis is of the form HA : p = p0 . (See Fig. 11.1.)
In calculating all these p-values, regardless of the alternative hypothesis,
there is a common thread. In order to see this common thread, we need to look at
how these p-value were derived by hand. If you look back at the calculations of
 Making hypothesis testing work with the central limit theorem Chapter | 11 123

FIGURE 11.1 P-values for the three alternative hypotheses visualized.

all three p-values, you will notice they all begin in the same place. Namely, look-
ing at how our sampling distribution—a N (0.5, 0.052 )—is related to our sample
proportion—p̂ = 0.6. After employing one of the properties of Normal distribu-
0.6 − 0.5
tions, this changes to how a standard Normal is connected to t = .
0.05

HA P-value
 
p > 0.5 P N (0, 1) ≥ t
 
p < 0.5 P N (0, 1) ≤ t
 
p = 0.5 2 × P N (0, 1) ≥ |t|

This value t takes into account information we know about the sample
statistic—in this case p̂—the null hypothesis value p0 , and the standard error—
here, s.e.(p̂). Specifically, it is
Sample Statistic – Null Hypothesis Value
t=
Standard Error
Essentially, it completely encompasses all our information about the hypoth-
esis test. This quantity t is referred to as the test statistic, and it calculates how
far our sample statistic is from the null hypothesis value, scaled by the standard
error. In general, we can use this knowledge about the relationship between the
test statistic, alternative hypothesis, and p-values to generalize which p-value
we should use for which alternative hypothesis. If we define our test statistic as
Sample Statistic – Null Hypothesis Value
t= ,
Standard Error
then our p-values will be as defined in the table below, solely depending on the
alternative hypothesis.
124 Basic Statistics With R

HA P-value
 
HA : p < p0 P N (0, 1) ≤ t
 
HA : p > p0 P N (0, 1) ≥ t
 
HA : p = p0 2 × P N (0, 1) ≥ |t|

We will see tables defining our p-values very similar to this for each individ-
ual test as we go through them, but this general idea behind getting our p-values
will carry across a variety of parameters and tests.

11.4.1 Practice problems

In the casino game of roulette, there are three colors on the wheel: red, black,
and green. Say you are playing roulette at a table and you believe that the wheel
is rigged, specifically that it makes red less likely to come up.
6. Under normal circumstances, if the table is fair, the probability that red will
come up is 0.47. What would be the null and alternative hypotheses for this
scenario?
7. Say you played 50 rounds of roulette and red came up 20 times, black 28
times, and green twice. What is the sample proportion for red coming up?
8. If the standard error for the sampling error in this example is s.e.(p̂) =
0.07, what would the test statistic be for this data?
9. If the null hypothesis is true and the standard error is as given above, what
would the sampling distribution be for p̂?
10. For the alternative hypothesis you defined earlier, what will the p-value be
for the hypothesis test?
Video poker is generally considered one of the better casino games in
terms of chances to win. In what is called “jacks or better” the probability of
winning—in which a player has a hand of a pair of jacks or better—is theoreti-
cally 0.4546.
11. Say that you are interested in testing if a video poker machine is rigged
to make the probability of winning lower. What would be your null and
alternative hypotheses?
12. Say you played 100 hands of poker and won 42 times. What is the sample
proportion of winning?
13. If the standard error for the sampling error in this example is s.e.(p̂) =
0.05, what would the test statistic be for this data?
14. If the null hypothesis is true and the standard error is as given above, what
would the sampling distribution be for p̂?
15. For the alternative hypothesis you defined earlier, what will the p-value be
for the hypothesis test?
 Making hypothesis testing work with the central limit theorem Chapter | 11 125

11.5 Conclusion
P-values represent an important part of our hypothesis-testing procedure. Ulti-
mately, they are the evidence with which we decide whether or not to reject our
null hypothesis. Based on their definition, they could be difficult to calculate
exactly. However, thanks to the central limit theorem discussed earlier, we are
able to connect the sampling distribution to our p-values. With an assist from R
calculating these p-values becomes a simple task, relying only on our alternative
hypothesis and the observed data. As we will see in the following chapter, the
importance of the central limit theorem also extends to the other common form
of statistical inference.
Chapter 12

The idea of interval estimates

Contents
12.1 Introduction 127 12.5 Creating confidence intervals 129
12.2 Point and interval estimates 127 12.6 Interpreting confidence intervals 132
12.3 When intervals are “right” 128 12.7 Practice problems 133
12.4 Confidence intervals 128 12.8 Conclusion 133

12.1 Introduction
So far in our discussion of statistical inference, we have concentrated solely on
the hypothesis tests. Further, in calculating our statistics from our sample we
have concentrated on only calculating single values as estimates for our popula-
tion parameters. However, there is a collection of estimators for our parameters
that involve an interval of plausible values for that parameter. In this chapter, we
will talk about the spirit and interpretation of these interval estimates.

12.2 Point and interval estimates

In statistics, we calculate sample statistics in order to estimate our population
parameters. What we have seen so far are point estimates, or a single numeric
value used to estimate the corresponding population parameter. The sample av-
erage x̄ is the point estimate for the population average μ. The sample variance
s 2 is the point estimate for the population variance σ 2 . The sample proportion
p̂ is the point estimate for the population proportion p. All of these estimates
represent our “best guess” at the value of the population parameter.
Our point estimates or sample statistics give us a fair bit of information, but
as we know, sample statistics will vary from sample to sample. If we collected
several samples, we would have a bunch of plausible values for our popula-
tion parameter. It seems reasonable that under certain circumstances, we would
want a range of plausible values as an estimate, rather than a singular esti-
mate. This type of estimate is exactly what interval estimates are able to give
us. An interval estimate is a interval of possible values that estimates our un-
known population parameters. Interval estimates are probably most familiar to
us through their use in political polling, as approval ratings and election leads
are often stated as an interval. In these settings, an interval estimate is commonly
given in the form of a point estimate plus-or-minus a margin of error. The mar-

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00024-9 127

Copyright © 2022 Elsevier Inc. All rights reserved.
128 Basic Statistics With R

gin of error gives us an idea of how precise the point estimate is, expressing the
amount of variability that may exist in our estimate due to uncontrollable error.
For example, prior to the 2017 Virginia gubernatorial election, Quinnipiac
University conducted a poll and found that the Democratic candidate Ralph
Northam led the Republican, Ed Gillespie, by 9 points plus-or-minus 3.7 points
[35]. In this case, the point estimate for Northam’s lead is 9 points and the
margin of error is 3.7 points. This implied that in reality it would be plausi-
ble that Northam could lead the race by any number between 9 − 3.7 = 5.3 and
9 + 4.7 = 12.7 points.
Under ideal circumstances, our margin of error is able to give us an idea
of how confident we can be in the estimate of our population parameter. Say
we have two interval estimates where our point estimates are the same but our
margins of error differ: 6 ± 2 and 6 ± 4. Our first interval estimate implies that
the plausible values of our parameter are from 4 to 8, while the second implies
a range of 2 to 10. Because the range of plausible values is smaller for the
first interval estimate, we would be inclined to trust the results more. However,
there are many things that go into our margin of error: the amount of variability
inherent in our data, the sample size, and most importantly the probability that
our interval is “right.”

12.3 When intervals are “right”

When we define our interval estimates, we want them to have some sort of prob-
abilistic interpretation or some sort of probability statement attached to them.
Ideally, this probability will be connected to how often our interval is “right.”
However, what do we even mean by an interval being “right?” Can we even
know if we are “right?”
Let us backtrack to what we know about our parameters and intervals. In the
Frequentist paradigm of statistics—the paradigm with which we are working—
a population parameter is a fixed, unknown value. There will not be plausible
values for the parameter in reality, only a single value. On the other hand, our
interval estimates are calculated from our data, which implies that they will be
influenced by our sample and thus random.
We will say that our interval is “right” when the true value of our parameter
is covered by—or contained in—our interval. It is impossible to know whether
or not this is true—since the true value of any parameter is unknown—so we will
need to couch any interval estimate in terms of the probability that the interval is
right. Put another way, we need to know how confident we are in the correctness
of our interval estimate.

12.4 Confidence intervals

Now that we have defined what we mean by an interval being right, we need
to create an interval estimate that is directly connected the to probability of be-
ing “right.” Confidence intervals are an interval estimate that gives a range of
 The idea of interval estimates Chapter | 12 129

plausible values for our parameter calculated from our data whose probabilistic
interpretation is directly connected to if the interval covers the true value of the
parameter. A confidence interval that is calculated from a sample will cover the
true value of the population parameter a predetermined portion of the time; an
80% confidence interval will cover the true value of a parameter 80% of the
time. A 95% confidence interval will cover the true value of the parameter 95%.
In general, if we were able to take an infinite number of samples and calculate
(1 − α)100% confidence intervals from each of them, then (1 − α)100% of these
intervals would cover the true value of the parameter. This (1 − α)100% prede-
termined coverage proportion—the probability that the interval is “right”—is
called the confidence level.
Let us look at an example of this where we know the true value of our pa-
rameter. Say that we flipped a fair coin 100 times and created a 95% confidence
interval for the true probability of flipping a heads p—more details on this to
come. Again, this is a fair coin, so the true value of p = 0.5. Thus, we will know
whether or not our confidence interval covers the true value of p. Now, we re-
peat this process of collecting data and creating a confidence interval 99 more
times. Assuming that our 95% confidence interval is “right” 95% of the time, we
should see that approximately 95 out of the 100 intervals covers the true value
of p = 0.5. In Fig. 12.1, we can see this phenomenon, with exactly 95 out of the
100 intervals covering p = 0.5.

FIGURE 12.1 95% Confidence intervals for the proportion of heads p with a sample size of 100.
Approximately 95% of the confidence intervals should cover the true value of p, given by the vertical
dotted line at p = 0.5.

12.5 Creating confidence intervals

Generally speaking, our confidence intervals for our parameters will be of the
form
130 Basic Statistics With R

Point Estimate ± Margin of Error

similar to the interval estimates we talked about above. While the exact form of
our confidence intervals will vary depending on what parameter we are doing
confidence intervals for, we can see the process of making confidence intervals
and how they are connected to the probability of being “right.” Specifically, we
will look at this through calculating our confidence interval for the population
proportion p.
For our confidence interval to be “right” (1 − α)100% of the time, we need
the interval to have a lower bound and upper bound calculated from our sample
such that
 
P Lower < p < Upper = 1 − α.
So, let us start with a distribution that we have worked with before: the stan-
dard Normal distribution. We could find a value z∗ so that the probability that a
standard Normal is between −z∗ and z∗ is 1 − α, or
 
P − z∗ < N (0, 1) < z∗ = 1 − α.

We will call this z∗ the critical value. Now, if we can find something that
connects the population proportion p to the standard Normal distribution, we
will be able to create an interval. If we recall, there is an important theorem
that connects our sample proportion, population proportion, and the standard
Normal: the central limit theorem. This states that the sample proportion p̂ will
follow a Normal distribution assuming the sample size n is large enough, specif-
ically
 
p̂ ∼ N p, s.e.(p̂ 2 ) .
As we saw in hypothesis testing, we can translate this to a standard Normal
by subtracting off the mean p and dividing by the standard deviation s.e.(p̂)

p̂ − p  
∼ N 0, 1 .
s.e.(p̂)

So let us go back to our standard Normal probability, P (−z∗ < N (0, 1) <
p̂ − p p̂ − p  
z∗ ) = 1 − α. We can sub in for the N (0, 1), since ∼ N 0, 1 .
s.e.(p̂) s.e.(p̂)
Our standard Normal probability then becomes
 
∗ p̂ − p ∗
P −z < < z = 1 − α.
s.e.(p̂)

Now, we have our population proportion p inside a probability statement di-

rectly connected to our 1 − α probability. To find our lower and upper bounds—
 The idea of interval estimates Chapter | 12 131

and, therefore, our interval—we need to solve for p inside the probability state-
ment:
 
∗ p̂ − p ∗
P −z < <z =1−α
s.e.(p̂)
 
P − z∗ s.e.(p̂) < p̂ − p < z∗ s.e.(p̂) = 1 − α
 
P − p̂ − z∗ s.e.(p̂) < −p < −p̂ + z∗ s.e.(p̂) = 1 − α
 
P p̂ − z∗ s.e.(p̂) < p < p̂ + z∗ s.e.(p̂) = 1 − α.

So, we now have an interval with a lower bound of p̂ − z∗ s.e.(p̂) and an

upper bound of p̂ + z∗ s.e.(p̂). This interval is random, having been calcu-
lated from our data. Additionally, it will cover the true value of p—and thus
be “right”—(1 − α)100% of the time, which is the probability that we want.
With this in mind, our (1 − α)100% confidence interval for p will be
 
p̂ − z∗ s.e.(p̂), p̂ + z∗ s.e.(p̂) .

This leaves three components to calculating our confidence interval: our

sample proportion p̂, our standard error s.e.(p̂), and our critical value z∗ . We
calculate our sample proportion from our sample data as previously described.
For the moment, the standard error will be given. In later chapters, we will see
the formula for the standard errors of our various sample statistics. This just
leaves our critical value z∗ . In deriving our confidence interval, we said that we
selected z∗ based on how confident we wanted to be in our interval by choosing
z∗ such that
 
P − z∗ < N (0, 1) < z∗ = 1 − α.
Finding z∗ based on this definition can be difficult, as it requires us to find
two probabilities to get our answer. However, there is an easier way to find z∗
based
 on the fact that Normal
 distributions are symmetric. It turns out that if
P − z∗ < N (0, 1) < z∗ = 1 − α, then
 
P N (0, 1) < z∗ = 1 − α/2.

This will allow us to look up only one probability to find our z∗ . Let us see
this in practice. Say we wanted to calculate a 95% confidence interval for p.
To find our value of α, we solve (1 − α)100%=95% so that our α = 0.05. In
finding z∗ , this implies that
 
P N (0, 1) < z∗ = 1 − 0.05/2 = 0.975.

In order to get this critical value we turn to technology, specifically R. To find

the quantile of a Normal distribution associated with a specific probability, we
turn to the qnorm function. The qnorm function takes in three arguments: the
mean of the Normal distribution mean, the standard deviation of the Normal
132 Basic Statistics With R

distribution sd, and the probability associated with that quantile p. The code to
use this function is
qnorm(p, mean, sd)
 
And R will return the quantile z∗ such that P N (mean, sd) < z∗ = p. For
our example—finding the z∗ associated with a 95% confidence interval—we
would use the code
qnorm(p=0.975, mean=0, sd=1)

And R will return a z∗ value of 1.959964, which we can use to get our 95%
confidence interval. Let us take a look at this in practice. In 1996, the United
States General Social Survey took a survey of 96 people asking if they were
satisfied with their job [36,37]. They found that 79 of them were satisfied, so
the sample proportion would be p̂ = 79/96 = 0.8229. Say we wanted to create
a 90% confidence interval (α = 0.1) for the true proportion of people satisfied
with their job. The first thing is to find z∗ , so that
P (N(0, 1) < z∗ ) = 1 − 0.1/2 = 0.95.

Using R, we input the following code to get our value of z∗ = 1.644854:

qnorm(p=0.95, mean=0, sd=1)

Our final part of the confidence interval is the standard error of p̂, which in
this case is s.e.(p̂) = 0.039. Taking all this information, our 90% confidence
interval for p will be
   
0.8229 − 1.645 × 0.039, 0.8229 + 1.645 × 0.039 = 0.7585, 0.8871 .

12.6 Interpreting confidence intervals

A key portion of confidence intervals is correctly interpreting their meaning.
We talked about this previously when we defined what it meant for an interval
to be “right.” In general, if we calculate a (1 − α)100% confidence interval for
a parameter, this means that we are (1 − α)100% confident that the calculated
interval covers the true value of that parameter—in other words, will be “right.”
Let us see this in practice. From our previous example, we saw that the 90% con-
fidence interval for American job satisfaction was (0.7585, 0.8871). We would
interpret this saying, “We are 90% confident that the interval (0.7585, 0.8871)
covers the true proportion p of American satisfied with their jobs.”
Similar to p-values, our confidence interval have many incorrect interpreta-
tions that are commonly seen. With this in mind, it is worthwhile to mention
these incorrect interpretations and clarify why they are wrong.
• A (1 − α)100% confidence interval does not mean that the probability that
the parameter of interest lives in the set bounds of a confidence interval is 1 −
α. This interpretation implies that our parameter is a random variable rather
 The idea of interval estimates Chapter | 12 133

than a fixed value. This is false, as our parameter is fixed while the bounds
of our confidence interval are calculated from a sample and are therefore
random.
• A (1 − α)100% confidence interval does not mean that we are (1 − α)100%
confident that the parameter equals the interval defined by the given confi-
dence interval. Again, our parameter is a fixed, unknown, single value and
will not be equal to an interval.
• A (1 − α)100% confidence interval does not mean that (1 − α)100% of
the population falls within the bounds of a confidence interval. Confidence
intervals make statements about population parameters, not about individuals
within the population.
• A (1 − α)100% confidence interval does not mean that a calculated sample
statistic will fall in the confidence interval (1 − α)100% of the time. Again,
confidence intervals make statements about population parameters and not
about the individuals within the population.

12.7 Practice problems

In 2016, the Pew Research Group polled Americans to see their opinion of the
political party opposing their own [38]. Forty-five percent of Americans in the
survey had a “very unfavorable” view of the opposing party, with a standard
error of s.e.(p̂) = 0.007.
1. Find and interpret the 90% confidence interval for p.
2. Find and interpret the 95% confidence interval for p.
3. Find and interpret the 99% confidence interval for p
4. Find and interpret the 99.5% confidence interval for p.
In 2019, the Pew Research Group polled Americans to see what percentage
believed that focusing on renewable energy is a more important priority than
expanding fossil fuels [87]. They found that 77% of Americans held this view,
with a standard error of s.e.(p̂) = 0.007.
5. Find and interpret the 80% confidence interval for p.
6. Find and interpret the 90% confidence interval for p.
7. Find and interpret the 95% confidence interval for p

12.8 Conclusion
Confidence intervals represent another important component of statistical infer-
ence. Rather than testing a claim, they allow us to see the range of values that our
parameter may take. We will see how we can connect this directly to hypothesis
testing, but the two methods of inference are already connected via the central
limit theorem. Both techniques rely on this theorem to connect our data to our
parameter, allowing us to understand the values for these important summaries
of our population. Going forward, we can now fill in the specifics of several key
inferential situations, focusing on population means and proportions.
Chapter 13

Hypothesis tests for a single

parameter
Contents
13.1 Introduction 137 13.3.2 Set significance level 144
13.2 One-sample test for proportions 138 13.3.3 Collect and summarize
13.2.1 State hypotheses 138 data 144
13.2.2 Set significance level 138 13.3.4 Calculate test statistic 144
13.2.3 Collect and summarize 13.3.5 Calculate p-values 145
data 139
13.3.6 A brief interlude: the t
13.2.4 Calculate test statistic 139
distribution 146
13.2.5 Calculate p-values 140
13.2.6 Conclude 141 Back to p-values 146
13.2.7 Practice problems 142 13.3.7 Conclude 148
13.3 One-sample t-test for means 143 13.3.8 Practice problems 150
13.3.1 State hypotheses 143 13.4 Conclusion 150

13.1 Introduction
In the past few chapters, we have been laying the general foundations for sta-
tistical inference through hypothesis testing and confidence intervals. In this
chapter, we will start really going through the process of hypothesis testing, fill-
ing in the missing gaps from earlier chapters. We will specifically be looking at
one-sample tests in this chapter. One-sample hypothesis tests are concentrated
on the questions related to a single parameter, for example, is a coin fair—i.e.,
is p = 0.5?
While statistics is a process that is many times as much art as science, hy-
pothesis testing is a very methodical procedure that in all its forms consists of
a series of steps. These steps are given below and will be referred to as we go
through each of the tests in this chapter and future chapters:
1. State hypotheses
2. Set significance level
3. Collect and summarize data
4. Calculate test statistic
5. Calculate p-value
6. Conclude

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00026-2 137

Copyright © 2022 Elsevier Inc. All rights reserved.
138 Basic Statistics With R

In this chapter, our one-sample tests will include a one-sample test about our
population proportion p and a one-sample test about our population mean μ. We
will begin with the more familiar of these two tests, looking at our one-sample
test for the population proportion p.

13.2 One-sample test for proportions

The one-sample test for proportions is a hypothesis test used to see if the
population proportion is equal to or differs from some value. Because this test
specifically deals with proportions, we use this test when we have a categorical
response that we are interested in. This can help us answer a variety of questions,
such as is a die fair, do more people than not approve of a law, etc.
As we go through our six steps of this hypothesis test, let us illustrate them
with an example. Legacy students—students whose family went to the same
college—are an important component of college admissions. Nationally, 14% of
college students are legacies. Say we were interested in investigating if Sweet
Briar College has a different proportion of legacy students than the national
average. As our response—whether or not the student is a legacy—is categorical
and only deals with a single group or sample, testing the proportion using the
one-sample test for proportions is an appropriate step to take.

13.2.1 State hypotheses

We have previously discussed this first step. We have to state the null hypothesis
and alternative hypothesis in terms of some parameter and that parameters rela-
tionship to some value. The null hypothesis—representing our state of the world
or the assumption that nothing interesting is occurring in our data—is usually
of the form H0 : p = p0 . Our alternative hypothesis can take one of three forms:
either one-tailed—HA : p > p0 or HA : p < p0 —or two-tailed—HA : p = p0
while not overlapping with the null hypothesis.
In our example, we want to test if the true proportion of Sweet Briar legacy
students is equal to or differs from 0.14. This implies that our hypotheses will
be

H0 : p = 0.14
HA : p = 0.14.

13.2.2 Set significance level

Recall, our significance level α will be half of determining the conclusion of our
hypothesis test. In addition, it essentially sets our level of reasonable doubt. If
we set α to be low, we will have a harder time rejecting the null hypothesis, and a
high α makes it easier to reject the null hypothesis. All this is important because
it turns out that the probability of rejecting the null hypothesis wrongly—and
making a Type I error—is α. Individual fields can have differing standard about
 Hypothesis tests for a single parameter Chapter | 13 139

appropriate α levels, but the most commonly accepted significance level is α =

0.05.
It is important to set our significance level α at this point in the process
rather than later. If we set α later in the process—say, after we have collected
our data—we are going to bias our results or be tempted to go searching for val-
ues that yield significant results. This is not ethical in research, as it introduces
large amounts of bias and skepticism to the results. As such, we need to set our
significance level before we even collect the data.
For our example testing the proportion of Sweet Briar legacy students, there
are no particular consequences to Type I errors. With this in mind, we will use
the default value of α = 0.05.

13.2.3 Collect and summarize data

While the actual mechanics of collecting data is outside the realm of this class,
the summary of data definitely falls within the scope of what we have discussed.
Each test has specific data summaries that will be important. For the one-sample
test for proportions, we first and foremost need the sample proportion, or p̂. This
gives us an idea of how far the observed value is from the null hypothesis value
p0 . The other data summary that we need for this test is the sample size n. This
is because the standard error of our sample proportion—an essential part of our
test statistic—depends on our sample size. Once we have these two summaries
of our data, we can move on to calculating the test statistic.
In this example, we collected data from 79 Sweet Briar students and found
that 12 of them were legacy students. This implies that our sample size was
12
n = 79 with a sample proportion of p̂ = = 0.1519.
79

13.2.4 Calculate test statistic

As we discussed last chapter, our test statistic gives us an idea of how far our
observed data is from our null value, scaled be the standard error. We defined
this test statistic t to be
Sample Statistic – Null Hypothesis Value
t= .
Standard Error
For the one-sample test for proportions, the sample statistic that we are work-
ing with is always the sample proportion p̂. We defined the null hypothesis value
to be p0 . So our test statistic currently looks like

p̂ − p0
t= .
s.e.(p̂)

All that remains is our standard error s.e.(p̂). This can be a little more
difficult to calculate, but it turns out that the standard error of our sampling
140 Basic Statistics With R

proportion is

p(1 − p)
s.e.(p̂) =
n
where p is the true value of our population proportion and n is our sample size.
This is problematic because it relies on knowing the true value of p, which—as
a parameter—is unknown. If we knew the value of p, there would be no reason
to go through this procedure. This would seem to be a dead end in calculating
our standard error. Fortunately, one of our key assumptions in hypothesis testing
provides an answer. In hypothesis testing, we always assume that the null hy-
pothesis is true until proven otherwise. In this case, this would mean that p = p0
until we prove it false. We can substitute this value of p = p0 into our standard
error which becomes

p0 (1 − p0 )
s.e.(p̂) = .
n
Putting all this together, our test statistic for this particular test is

p̂ − p0
t= .
p0 (1 − p0 )
n
For our Sweet Briar example, our null hypothesis value will be p0 = 0.14,
our sample proportion is p̂ = 0.1519, and our sample size is n = 79. We can fill
this into our test statistic formula above to get

0.1519 − 0.14
t= = 0.3048.
0.14(1 − 0.14)
79

13.2.5 Calculate p-values

Once we have a test statistic, we need to calculate a p-value. The p-value is
entirely dependent on our alternative hypothesis, as it defines what we mean be
“more extreme” data. We discussed in a previous chapter how we connected our
p-values to our standard Normal distribution based on our alternative hypothesis,
so we will jump to the result here. If we have a test statistic t calculated as
defined in Step 4, then our p-values—dependent on HA —will be

HA P-value
 
HA : p < p0 P N (0, 1) ≤ t
 
HA : p > p0 P N (0, 1) ≥ t
 
HA : p = p0 2 × P N (0, 1) ≥ |t|
 Hypothesis tests for a single parameter Chapter | 13 141

We are able to use the standard Normal distribution because, as we discussed

earlier, the central limit theorem states that for large enough sample sizes, the
sample proportion p̂ will follow a Normal distribution. However, what sample
size is large enough? Generally speaking, the central limit theorem for the sam-
ple proportion will hold if (1) the sample size n is at least 30, and (2) we observe
at least 10 successes and failures in our sample. If these two properties hold, we
can proceed with the p-values given above with little worry about the results that
they give.
These p-values can be obtained using the pnorm function in R. It is important
to recall that the pnorm function returns P N (0, 1) < t by default. In order to
get our other probabilities, we will have to get our result by manipulating the
output and or changing our code in R.
For the Sweet Briar legacy example, our alternative hypothesis is HA : p =
0.14 and our previously calculated
  is t = 0.3048. This means that
test statistic
our p-value will be 2 × P N (0, 1) > |0.3048| . To get our p-value we will need
that initial probability, which we can get using the R code

pnorm(q=0.3048, mean=0, sd=1, lower.tail=FALSE).

 
This will return P N (0, 1) > |0.3048| = 0.3802592, meaning that all we
have to do to get our p-value is double this probability to get p-value = 2 ×
0.3802592 = 0.7605184.

13.2.6 Conclude
In general, our conclusions come in three parts—justification, result, and con-
text. We want to be sure to mention the result of our hypothesis test—we
either reject or fail to reject the null hypothesis—to go with our justification
for reaching that result—because our p-value was less than α or greater than
α, respectively. Finally, we want to be sure to put this conclusion in context, in
order for us to understand what this result means for our study. If we reject the
null hypothesis, we will conclude in favor of HA . However, if we fail to reject
the null hypothesis, we can only conclude that the null hypothesis is plausible—
remember that the null hypothesis can never be concluded true, only plausible.
In our Sweet Briar legacy example, our p-value = 0.7605184 is greater than
our significance level of α = 0.05. Since this is the case, we will fail to reject
the null hypothesis and conclude that it is plausible that the true proportion of
legacy students at Sweet Briar is equal to 0.14.
Now, let us see all our whole hypothesis testing procedure at once in another
example. In March 2016, a Pew Research poll found that 57% of Americans
were “frustrated” with the federal government [39]. Pew repeated the poll in
December 2017 and found that 826 of 1503 respondents said that they were
“frustrated” with the federal government. Say we wanted to test to see if the true
proportion of American frustrated with the government had decreased since the
142 Basic Statistics With R

2016 poll. Starting out, our hypotheses would be

H0 : p = 0.57
HA : p < 0.57.

We will set our significance level to be α = 0.05 for this data. The summaries
826
we need for this data is the sample proportion p̂ = = 0.549 and the sample
1503
size n = 1503. Assuming that the null hypothesis is true, our test statistic is
0.549 − 0.57
t= = −1.644.
0.57 × 0.43
1503
Because
 our alternative
 hypothesis is HA : p < 0.57, our p-value will
be P N (0, 1) < −1.64 . Using the R code given below, we find that this
probability—which is our p-value—will be p-value = 0.0500881.
pnorm(q=-1.644, mean=0, sd=1, lower.tail=TRUE).

So, because our p-value is greater than α = 0.05—even if only ever so

slightly—we fail to reject the null hypothesis and conclude that it is plausible
that the true value of the proportion of Americans who are frustrated with the
federal government is equal to 0.57.
All of this can be done in R using the prop.test function. The details of this
function will be discussed in Chapter 17, but the function takes in the sample
size n, the number of successes x, the null hypothesis value p, and the alternative
hypothesis and outputs our test statistic and p-value.

13.2.7 Practice problems

In 2014, 68% of Americans believed that vaccines should be required for all
children according to a Pew Research study [40]. In 2017, the group wanted to
see if this proportion had increased due to more awareness of the issue [41].
1. What would the null and alternative hypotheses be to test this question?
2. In the survey to investigate this question, Pew found that 1270 respondents
out of 1549 said that children should be required to be vaccinated. What is
the sample proportion for this data?
3. Based on the null hypothesis and data collected, what would your test statis-
tic be?
4. Based on your alternative hypothesis and test statistic, what will your p-value
be?
5. Assuming that α = 0.05, what conclusions will you draw from this test?
It is generally thought that approximately 50% of all email is spam email.
Say that you want to test if this percentage has shifted from this number, either
increasing or decreasing.
 Hypothesis tests for a single parameter Chapter | 13 143

6. What would the null and alternative hypotheses be to test this question?
7. In surveying a set of 276 e-mails, 122 were determined to be spam. What
is the sample proportion for this data?
8. Based on the null hypothesis and data collected, what would your test
statistic be?
9. Based on your alternative hypothesis and test statistic, what will your p-
value be?
10. Assuming that α = 0.1, what conclusions will you draw from this test?
In a 2011 survey, 79% of Americans said they had read a book in the past
year [42]. A 2019 survey found that out of 1502 respondents, 1082 had read a
book in the previous year [43].
11. Test if the population proportion of Americans who have read a book in the
past year has decreased from the 2012 level of p = 0.79 at the α = 0.05
level.

13.3 One-sample t-test for means

Now that we have discussed tests for a single proportion, let us move on to a
test for a single mean. The one-sample t-test for means is the hypothesis test
used to see if that population mean differs from some value. Because this test
deals with means, this implies that this test is used when we have a quantitative
response. In general, this test follows the same general procedure for hypothesis
tests, with the same six steps we saw in our test for proportions.
To walk through our steps, let us consider the following example. Consider
a video slot machine with seven possible outcomes—blank, bar, double bar,
triple bar, double diamond, cherries, and 7. Theoretically, the true average pay-
out should be $13.13 [23,44].

13.3.1 State hypotheses

As before, we need to begin with stating the null hypothesis and alternative hy-
pothesis in terms of some parameter and that parameters relationship to some
value. In the one-sample test for proportions, we were interested in the popu-
lation proportion p. In this test, we are concerned with the population mean μ.
The null hypothesis still represents the status quo state of the world, assuming
that there is nothing interesting occurring. Usually, our null hypothesis takes the
form H0 : μ = μ0 . Our alternative hypothesis can take one of three forms: that
the population mean is greater than the null value, less than the null value, or
not equal to the null value. We notate these HA : μ > μ0 , HA : μ < μ0 , and
HA : μ = μ0 , respectively.
Say you wanted to test if the video slot machine true average payout was
equal to or less than the theoretical value of $13.13. Out hypotheses would be
144 Basic Statistics With R

H0 : μ = 13.13
HA : μ < 13.13.

13.3.2 Set significance level

The considerations for setting α are identical to what we previously discussed
in the one-sample test. With that in mind, we will not discuss this in detail, only
reminding that α needs to be set prior to collecting the data and conducting the
test.
For our example testing the average payout of the video lottery terminal, we
will use the standard significance level of α = 0.05.

13.3.3 Collect and summarize data

At this step, we are not so much concerned with actually collecting the data as
we are with the summaries of the data that we need. Similar to the one-sample
test for proportions, the first summary that we will need is the sample size n.
This is because, as we saw before, the sample size will affect the standard error
that we need for our hypothesis test.
In addition to the sample size, we will need the sample statistic that estimates
the population parameter of interest. In this case, that parameter is the population
mean μ, so the sample statistic that estimates this will be the sample average x̄.
Finally, we will need some assessment of the variability that exists in our data.
We will find this in either the sample standard variance s 2 or sample standard
deviation s. Either one will be sufficient, as knowing the variance is analogous
to knowing the standard deviation.
In order to test the slot machine hypotheses, a researcher played 345 games
on the machine, with a sample average payout of x̄ = 0.672 with a standard
deviation of s = 2.551.

13.3.4 Calculate test statistic

Our test statistic gives us an idea of how far our observed data is from our
null value, scaled be the standard error. In the tests that we are discussing, test
statistic t is defined to be
Sample Statistic – Null Hypothesis Value
t= .
Standard Error
For this test, the sample statistic is the sample mean x̄ as this is the sample
statistic that estimates our population mean. Our null hypothesis value is defined
to be μ0 , so our test statistic currently looks like

x̄ − μ0
t= .
s.e.(x̄)
 Hypothesis tests for a single parameter Chapter | 13 145

All that remains is the standard error of our sample average s.e.(x̄). Deriving
this is difficult, but it turns out that the standard error of the sample average is
σ
s.e.(x̄) = √
n

where σ is the population standard deviation and n is the sample size. Putting
this together, this would make our test statistic

x̄ − μ0
t= √ .
σ/ n

However, there is a problem with this test statistic, namely that it depends
on population parameter σ . As a population parameter, σ is assumed to be fixed
and unknown, so it would be impossible to calculate a test statistic based on this
unknown parameter. Unlike the one-sample test for proportions, assuming that
the null hypothesis is true does not help us. Knowing μ0 does not give us the
value of σ , which we need to calculate our test statistic. Despite this, there is
another option for our test statistic; we can find a sample statistic that estimates
σ . Such an estimate does exist, specifically the sample standard deviation s. We
can substitute s into our test statistic in place of σ for our calculations, with our
resulting test statistic for this hypothesis test being

x̄ − μ0
t= √ .
s/ n

In the video slots example, using our data and null hypothesis, we can cal-
culate our test statistic to be
0.672 − 13.13
t= √ = −90.7084.
2.551/ 343

13.3.5 Calculate p-values

When calculating p-values, we need to know what probability distribution our
test statistic follows. Once we know this, we can assess how rare our data was—
assuming the null hypothesis is true—based on this distribution. In general,
when our test statistics are of the form
Sample Statistic – Null Hypothesis Value
t=
Standard Error
the test statistic t will follow a standard Normal distribution. In the case of our
one-sample t-test for means, this implies that

x̄ − μ0
t= √ ∼ N (0, 1).
σ/ n
146 Basic Statistics With R

However, this is not our test statistic—specifically because we do not know

the value of σ . Our test statistic uses the sample standard deviation s in lieu of
the population standard deviation σ , resulting in

x̄ − μ0
t= √ .
s/ n

This means that our test statistic will not follow a standard Normal distribu-
tion. But what do we do in this case? To calculate the p-values, we need to know
what distribution our test statistic follows, which currently is a roadblock. With
this in mind, let us take a short detour to take a look at a new class of probability
distributions: t distributions.

13.3.6 A brief interlude: the t distribution

The t distribution is a class of probability distributions that we will use in our
hypothesis tests involving population means. The shape of a t-distribution is
unimodal and symmetric—just like the Normal distribution—but not fully bell-
shaped. It is very similar to the Normal in shape, but with higher probabilities
in the tails of the distribution. Like the Normal, it can take values from −∞ to
∞. Unlike the Normal distribution, which can move around and be centered at
any value, the t-distribution it is guaranteed to be centered at 0.
Like the Normal distribution, the probabilities of t-distributions are de-
scribed by a function, and will thus have parameters to define said function. The
t-distribution has only one parameter, called the degrees of freedom—notated
ν. If a random variable X follows a t-distribution with ν degrees of freedom, we
would write this as X ∼ tν . The degrees of freedom determines how (1) spread
out the t-distribution is and (2) how similar the t-distribution will look to the
Normal distribution. As the degrees of freedom gets larger, the t-distribution will
get closer and closer to a standard Normal. In fact, if ν = ∞, the t-distribution
will become the standard Normal. (See Fig. 13.1.)

Back to p-values
So, given our slight detour, it probably is no surprise to find that our test statistic
from Step 4 will follow a t-distribution. However, which one? In order to define
our t-distribution, we will need to define the degrees of freedom. It turns out that
our test statistic follows a t-distribution with n − 1 degrees of freedom, where n
is our sample size. That is,

x̄ − μ0
t= √ ∼ tn−1 .
s/ n

Now that we have a distribution, we can get p-values. Similar to previous

examples, our p-values are ultimately determined by our alternative hypothesis
 Hypothesis tests for a single parameter Chapter | 13 147

FIGURE 13.1 The t-distribution with various degrees of freedom and the standard Normal distri-
bution.

and our test statistic. Additionally, the rationale for each p-value will be similar
to that for our standard Normal p-values. To illustrate this, let us look at the p-
value for the alternative hypothesis HA : μ > μ0 . Our p-value is the probability
that we observe “more extreme” data, which in this case would mean that we
see a sample average that is greater than the value we actually observed. If the
sample average is greater than observed, this will result in a test statistic greater
than the one calculated from our observed data. This is equivalent to saying that
a tn−1 distribution is greater than our calculated test statistic, because our test
statistic formula follows a tn−1 distribution. The probability that this occurs is
our p-value, which is P (tn−1 > t) where t is our test statistic calculated in Step
4. The rationale for our other two alternative hypothesis is similar, and their
respective p-value is given below. (See Fig. 13.2.)

HA P-value
 
HA : μ < μ0 P tn−1 ≤ t
 
HA : μ > μ0 P tn−1 ≥ t
 
HA : μ = μ0 2 × P tn−1 ≥ |t|

So, in order to get our p-values, we will need to be able to get p-values from
out t-distributions. Like with the standard Normal, we can do this through a
simple function in R. Similar to how the pnorm function got us probabilities
from the Normal distribution, the pt function will get us probabilities from the
t-distribution. The pt function takes three arguments: the critical value of the
distribution desired t ∗ , the degrees of freedom ν, and whether
 or not we
 want
the
 lower or upper tail probability—that is, if we want P tn−1 < t or P tn−1 >
t , respectively. Our code for this function, assuming we want the lower tail
probability, will be
pt(q=t*, df=v, lower.tail=TRUE)
148 Basic Statistics With R

FIGURE 13.2 P-values based on the t-distribution for our three alternative hypotheses.

To get the upper tail probability, all we will have to change is our lower.tail
option to be FALSE.
One final important thing to note is the assumptions built into using our t-
distribution for p-values in this test. As with the one-sample test for proportions,
we are able to use the t-distributions only if the central limit theorem holds
for our sample statistic—in this case x̄. The central limit theorem will hold
only if one of two conditions are true: if our sample size is large enough or
our response variable follows a Normal distribution. This means that only for
large enough sample sizes—generally requiring that our sample size n be at
least 30—or Normal data can we use the t-distribution for p-values.
For the video slot, since our alternative hypothesis is HA : μ < 13.13, our p-
value will be P t344 < −90.7084). We use the following code in R to find that
our p-value is 1.834 × 10−247 Our sample size is sufficiently large with n = 345
for us to say that the central limit theorem will hold, and thus our p-value is valid.

13.3.7 Conclude
Just as in the one-sample test for proportions, we either will reject or fail to reject
our null hypothesis based on our p-value and the significance level α. In doing
so, we want to be sure to reference our decision—reject or fail to reject—the
reason for the decision—p-value less than or greater than α—and the context of
the decision in terms of the problem.
Finally, in our video slots example, since our p-value of 1.834 × 10−247
is decidedly less than the significance level α = 0.05, we will reject the null
hypothesis and conclude that the true average payout of the video slot machine
is less than $13.13.
Let us see all these steps in practice. Say there is a factory process for making
washers [45,46]. If the population average of washer size differs from 4.0 in any
 Hypothesis tests for a single parameter Chapter | 13 149

way, the process is reset. A quality control sample of 20 washers is taken and
the sample average was found to be x̄ = 3.9865 with sample standard deviation
s = 0.073. In this setting, to test if the process should be reset our hypotheses
would be

H0 : μ = 4
HA : μ = 4

with an α = 0.05. Our data summaries are n = 20, x̄ = 3.9865, s = 0.073, so

our test statistic is
3.9865 − 4
t= √ = −0.827.
0.073/ 20
Since our alternative hypothesis is HA : μ = 4, our sample
 size is n = 20, and

our test statistic is t = −0.827, our p-value will be 2 × P  t20−1 > | − 0.827|.
To get the p-value, we first need to get the probability P t20−1 > | − 0.827| .
We get this using the following code, and then double it to get our p-value of
0.4185027.
pt(0.827,19,lower.tail=FALSE)

To conclude, since our p-value will definitely be greater than α = 0.05, we

fail to reject the null hypothesis and conclude that it is plausible that the true
average of washer diameter is 4, so the process would not need to be reset.
However, it is important to note that our sample size of n = 20 is below the
sample size threshold, and a histogram of the data implies that our data is not
from a Normal distribution. Thus the central limit theorem seems unlikely to
hold, so the results of the hypothesis test may need to be viewed with a skeptical
eye. (See Fig. 13.3.)

FIGURE 13.3 Histogram of our washer data.

The t.test function allows us to do this hypothesis testing procedure in R. It

ultimately takes in our data, null hypothesis value, and alternative hypothesis to
return our test statistic and p-value.
150 Basic Statistics With R

13.3.8 Practice problems

The United States General Social Survey conducts a vocabulary assessment of
Americans on occasion by administering a 10-question test and counting the
number of correct responses. In 2000, the average score was 6. In 2004, the
group wanted to see if vocabulary had gone up. The survey was done again and
1438 respondents had an average score of 6.2 with a sample standard deviation
of 2.06 [47,48].
12. What are the null and alternative hypotheses for this situation.
13. What is your test statistic for this test?
14. What distribution will your test statistic follow in this test?
15. What will your p-value for this test be?
16. What conclusion do you draw from this test?
Normal concentrations of calcium in urine are typically 2.5 millimoles/liter.
Say you were interested in if the presence of kidney stones—formed by calcium
oxalate crystals—was marked by a higher average concentration of calcium in
the urine. Say that in a sample of n = 34 individuals, it is found that the sample
mean of calcium concentration x̄ = 6.143 with a sample variance of s 2 = 13.229
[102].
17. What are the null and alternative hypotheses for this situation.
18. What is your test statistic for this test?
19. What distribution will your test statistic follow in this test?
20. What will your p-value for this test be?
21. What conclusion do you draw from this test?

13.4 Conclusion
Our tests for single parameters are crucial to understanding our populations. As
the population mean provides the average value for our population, the test for
a single mean helps us to recognize where that central value for our popula-
tion lies. The test for proportions can help us understand the true probability
of success—however it is defined—in addition to allowing us to then estimate
the total number of successes in the population. Fortunately, hypothesis testing
has a well-defined set of procedures: defining hypotheses, set our significance
level, collect and summarize our data, calculate our test statistic, calculate our p-
value, and draw a conclusion. This process allows us to understand the process
of testing for a wide variety of tests. In addition to understanding this proce-
dure, we need to understand what makes these procedures work but recognizing
the assumptions that are essential to making our tests valid. These procedures
and assumptions will continue going forward as we learn more inferential tech-
niques, both tests and confidence intervals.
Chapter 14

Confidence intervals for a single

parameter
Contents
14.1 Introduction 151 14.4.2 Practice problems 159
14.2 Confidence interval for p 151 14.4.3 Confidence intervals for
14.2.1 Practice problems 153 μ and hypothesis testing 159
14.3 Confidence interval for μ 153 14.4.4 Practice problems 161
14.3.1 Practice problems 155 14.5 Conclusion 162
14.4 Other uses of confidence
intervals 156
14.4.1 Confidence intervals for
p and sample size
calculations 156

14.1 Introduction
In the previous chapter, we introduced the general ideas of interval estimation
and confidence intervals. In this chapter, we will expand further our confidence
intervals, calculating intervals for p and μ as well as providing additional uses
to these intervals beyond merely providing a range of plausible values.

14.2 Confidence interval for p

When creating the (1 − α)100% confidence interval for p, we previously came
to the formula
 
p̂ − z∗ s.e.(p̂), p̂ + z∗ s.e.(p̂)
where p̂ is the sample proportion, z∗ is the critical value chosen so that
P (N(0, 1) < z∗ ) = 1 − α/2, and s.e.(p̂) is the standard error of our sample
proportion. We are able to calculate our sample proportion from our data and
we can find our critical value using R, leaving just the standard error to fill in
the formula. Similar to hypothesis testing, this comes down to the results of the
central limit theorem. In deriving our steps in hypothesis testing, we found that
the central limit theorem stated that
 
p(1 − p)
p̂ ∼ N p,
n
Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00027-4 151
Copyright © 2022 Elsevier Inc. All rights reserved.
152 Basic Statistics With R

where p is the true value of the  population proportion. This implies that the
p(1 − p)
standard error of p̂ is s.e.(p̂) = . However, we do not know the
n
true value of p because the population proportion is a parameter and, therefore,
assumed to be unknown. We ultimately need to substitute something else into
our formula for the standard error that estimates our population proportion p.
We know that the sample proportion p̂ estimates our population proportion p,
so we can substitute p̂ for p in our standard error. This results in

p̂(1 − p̂)
s.e.(p̂) = .
n
With this change, our (1 − α)100% confidence interval for p will be
   
∗ p̂(1 − p̂) ∗ p̂(1 − p̂)
p̂ − z , p̂ + z
n n

where p̂ is the sample proportion, n is our sample size, and z∗ is our critical
value chosen so that P (N(0, 1) < z∗ ) = 1 − α/2. The final piece of the confi-
dence interval is the interpretation, which we discussed previously. We interpret
confidence intervals by saying we are (1 − α)100% confident that our calculated
interval covers the true value of p.
Similar to hypothesis testing, there is a key assumption in creating our con-
fidence interval. In order to use the Normal distribution to get our critical value
z∗ and our confidence interval, we need the central limit theorem to hold. For
the central limit theorem to hold, we stated that our sample size had to be suf-
ficiently large—generally n ≥ 30—and there needed to be a sufficient number
of successes and failures in our sample—at least 10 of each. So, for our confi-
dence intervals to be valid we need a sufficiently large sample size and enough
successes and failure in our sample.
As an example, Steph Curry made 325 of 362 free throws in 2016. Let us
create a 99% confidence interval (α = 0.01) for the true probability of Steph
Curry hitting a free throw p. Our first step is finding our critical value z∗ such
that P (N(0, 1) < z∗ ) = 1 − 0.01/2 = 0.995. Using the R code

qnorm(p=0.995, mean=0, sd=1)

we can get that our critical value is z∗ = 2.576. With our critical value, our
325
sample proportion p̂ = , and our sample size of n = 362 our 99% confidence
362
interval for p is
   
0.898(1 − 0.898) 0.898(1 − 0.898)
0.898 − 2.575 , 0.898 + 2.575
362 362
 
0.857, 0.939 .
 Confidence intervals for a single parameter Chapter | 14 153

We interpret this as we are 99% confident that (0.857, 0.939) will cover the
true probability that Steph Curry makes a free throw. Further, our sample size is
sufficiently large—n = 369—and we see sufficient successes and failure—325
successes and 362 − 325 = 37 failures—to trust that the results are valid.
The prop.test function mentioned earlier has the ability to calculate our
confidence intervals for p. By inputting our data as before and changing the
conf.level option to our desired confidence level, we can get the (1 − α)100%
confidence interval for p. Details and examples of this function will be discussed
in Chapter 17.

14.2.1 Practice problems

1. Mike Trout reached base 224 times in 507 plate appearances in 2017. Find
and interpret the 95% confidence interval for the true probability of Mike
Trout reaching base p.
2. In a game of craps, and individual won 8 of 20 games by betting the pass
line. Find and interpret the 99% confidence interval for the true probability
of winning by betting on the pass line p.
3. In a sample of 780 women, it is found that 349 of them participated in vol-
unteering [88]. Find and interpret the 90% confidence interval for the true
proportion of women who participate in volunteering.

14.3 Confidence interval for μ

In creating our confidence intervals for our population mean μ, we can go
through many of the same steps as we did when we derived our confidence
intervals for p. Our goal in calculating
 our (1 −α)100% confidence interval for
μ is to create an interval estimate lower, upper so that

 
P Lower < μ < Upper = 1 − α.

In order to do this, we are going to need to connect μ to a probability distri-

bution. From hypothesis testing, we found that

x̄ − μ
√ ∼ tn−1 .
s/ n

With this in mind, let us start with a probability statement related to the tn−1
distribution. Let us choose a value t ∗ —again called the critical value—such that

 
P − t ∗ < tn−1 < t ∗ = 1 − α.
154 Basic Statistics With R

x̄ − μ x̄ − μ
Now, since √ ∼ tn−1 , we can substitute in √ for tn−1 in the proba-
s/ n s/ n
bility statement, resulting in
 
∗ x̄ − μ ∗
P − t < √ < t = 1 − α.
s/ n
At this point, if we solve for μ using algebra inside our probability state-
ment, we will wind out with an interval estimate with our desired probabilistic
properties.
 
x̄ − μ
P − t∗ < √ < t∗ = 1 − α
s/ n
 
s s
P − t ∗ √ < x̄ − μ < t ∗ √ =1−α
n n
 
s s
P − x̄ − t ∗ √ < −μ < −x̄ + t ∗ √ =1−α
n n
 
s s
P x̄ − t ∗ √ < μ < x̄ + t ∗ √ = 1 − α.
n n
  
We now have an interval Lower, Upper such that P Lower < μ <
Upper = 1 − α. This means that in the end, our (1 − α)100% confidence inter-
val for μ is
 
s s
x̄ − t ∗ √ , x̄ + t ∗ √ .
n n
So, to calculate our confidence interval for μ, we need to get our sample
mean x̄, the sample standard deviation s, and sample size n, all of which is
gotten from our sample. This just leaves finding our critical value t ∗ . Similar to
our confidenceinterval for p, finding t ∗ as defined—choosing t ∗ such that P −
t ∗ < tn−1 < t ∗ = 1 − α—is problematic as it involves finding two probabilities.
However, since the t-distribution is symmetric, we can adjust the probability
used to choose t ∗ such that
 
P tn−1 < t ∗ = 1 − α/2.

To find our value of t ∗ , we again turn to R. Just as the pnorm function had a
t-distribution analogue in the pt function, the qnorm function—used to find our
critical values from the Normal distribution—has a t-distribution analogue in
the qt function. The qt function finds the critical value of a t-distribution based
on the given probability and degrees of freedom of the t-distribution. These two
values—the probability p associated with the critical value and the degrees of
freedom ν—are the inputs into the qt function, with the code given below.
qt(p=p, df=v)
 Confidence intervals for a single parameter Chapter | 14 155

Like all other confidence intervals, interpreting our confidence intervals is

important aspect of the inference of confidence intervals. As with our confi-
dence interval for p, our confidence interval for μ is interpreted in terms of our
confidence about if the interval is “right”—that is, covers the true value of our
parameter μ. In this case, to interpret our (1 − α)100% confidence interval for
μ, we say that we are (1 − α)100% confident that our calculated interval covers
the true value of our population mean μ.
Finally, just as in our confidence intervals for p and our hypothesis tests,
there is an important assumption implicit in our confidence intervals for μ.
Our confidence interval implicitly assumes that the central limit theorem holds,
which means that we have a sufficiently large sample size or we have Normal
data. For our confidence interval results to hold—in other words, for the central
limit theorem to hold for means—we need a sample size of n ≥ 30 or for the
data to be Normally distributed.
Say that we were interested in calculating the 96% confidence interval for the
amount of DDT [10,49] in parts per million found in kale μ. The dataset DDT
in the MASS library in R measures just that, with 15 different labs estimating
the amount of DDT in a sample of kale. The sample mean is x̄ = 3.328 with
a sample standard deviation of s = 0.4372. This just leaves our critical value
t ∗ . If we are calculating a 96% confidence interval our level of α is 0.04, as
(1− 0.04)100% = 96%. This means we need to select t ∗ such that P t15−1 <
t ∗ = 1 − 0.04/2 = 0.98. Using the R code
qt(p=0.98, df=14)

below we find that t ∗ = 2.263781. Putting this all together, our 96% confidence
interval for the population mean amount of DDT in the kale sample is
0.4372
3.328 ± 2.264 √ = (3.0724, 3.5836).
15
We are 96% confident that (3.0724, 3.5836) covers the true value of μ. How-
ever, it is important to note that our sample size is less than 30, implying that
there may be concerns for our results. However, checking our histogram of
the data shows reasonably symmetric data, implying reasonable Normality and
that the central limit theorem still holds. Thus, our results are trustworthy. (See
Fig. 14.1.)
Similar to how the prop.test function can be used to get confidence intervals
for p, the t.test function can be used to get confidence intervals for μ. As before,
we merely have to add information about our desired confidence level in order
to get this interval.

14.3.1 Practice problems

4. In New York City, data on the air quality of the city is collected daily [50].
In 116 observations, there was an average of x̄ = 42.13 ozone ppm with
156 Basic Statistics With R

FIGURE 14.1 Histogram of the DDT data.

a sample standard deviation of s = 32.99. Calculate and interpret the 99%

confidence interval for μ.
5. A study asked 182 people their height and then measured their actual height,
and calculated the difference between the two [89]. The sample average was
x̄ = 2.07 (actual height was 2.07 centimeters greater than reported) with a
sample standard deviation of s = 2.08. Calculate and interpret the 90% con-
fidence interval for μ.
6. In n = 1295, women who experienced a heart attack, it was found that the
sample average age was x̄ = 60.92 with a standard deviation of s = 7.042.
Calculate and interpret the 95% confidence interval for μ [23].

14.4 Other uses of confidence intervals

To this point, our confidence intervals have been used as a second form of statis-
tical inference, a complement to hypothesis tests that provides us with a range
of plausible values. However, it turns out that that the form and derivations of
our confidence intervals can be used for other aspects of the statistical process.
It turns out that our confidence intervals can help us determine required sample
sizes for future studies as well as providing us with a direct link to the results of
our hypothesis tests.

14.4.1 Confidence intervals for p and sample size calculations

One of the biggest challenges in designing studies and collecting data is decid-
ing on what sample size is sufficient to answer the questions of our study. As
we have seen, the sample size will affect our standard errors, our confidence
intervals, and hypothesis tests. In many instances, our sample size is determined
by the resources of a study, with more money meaning that we can survey more
individuals. However, our confidence intervals can also guide our decisions on
 Confidence intervals for a single parameter Chapter | 14 157

sample size. In initially defining interval estimates, we said that the most basic
interval estimate was defined by

Point Estimate ± Margin of Error

Let us define this in terms of our confidence interval for p. So, recall, our
confidence interval for p is

∗ p̂(1 − p̂)
p̂ ± z .
n
In the case of our population proportion p, the point estimate of that pop-
 parameter is the sample proportion p̂ and our margin of error is m =
ulation
p̂(1 − p̂)
z∗ . The margin of error is predicated on knowing the sample size n,
n
your confidence level determined through z∗ , and our sample proportion p̂. Say
that we wanted to know the sample size and we knew our margin of error m, our
sample proportion p̂, and the confidence level we want to have in our estimate—
which is conveyed through z∗ . We can use our equation for the margin of error
to solve for n through a little algebra, getting

(z∗ )2 p̂(1 − p̂)

n= .
m2
So how can we use this equation to design our studies? It relies on knowing
a lot of information that is not determined until after our data is collected. How-
ever, we can make decisions on a lot of this prior to our study. We can say that
we want to estimate our population proportion within a specific margin of error,
so we can know our value of m. Further, we can say exactly how confident we
want to be in our interval estimates—essentially setting how often our interval
estimates will be “right.” This is equivalent to knowing what value of z∗ we
want, and we can find that critical value using the qnorm function in R. Finally,
all we are missing to complete our calculations for sample sizes is the sample
proportion p̂.
This would seem to be the most difficult portion of the formula to account
for because we can’t set what sample proportion we want to see. However, in
some cases we may have information about what we think p̂ will be. In many
cases, studies ask questions repeated from past studies. As such, we may know
what the sample proportion was from a past study. We could use that as a proxy
for p̂, standing in for what we think we may see in our study.
However, there are many cases in which no previous study exists. What
can we do in those cases? If no previous survey exists, we have to assume the
“worst-case” scenario; we assume that the sample proportion will be the value
that makes our required sample size the largest. If the margin of error and confi-
dence level are constant values, the largest sample size will occur when p̂ = 0.5.
158 Basic Statistics With R

You can confirm this through either calculus or by calculating what the sample
size would be for various values of p̂. In the end, if we know that we want to es-
timate our population proportion within a margin of error m with (1 − α)100%
confidence, our sample size will have to be

(z∗ )2 p̃(1 − p̃)

n≥ .
m2
 
Where we set z∗ such that P N (0, 1) < z∗ = 1 − α/2 and our p̃ is set either
based on a previous study—if one exists—or the worst case scenario.

p̂ From previous study if it exists
p̃ =
0.5 If no previous study exists

When calculating this necessary sample size, often we will have n be a non-
integer. Of course, we cannot sample half a person, so we need to round up to
the next integer greater than or equal to your calculated n. If you round down,
the sample size will result in a larger margin of error than desired. By rounding
up, we ensure that our confidence intervals will have a slightly smaller margin
of error than initially desired.
For example, say that you wanted to estimate the true proportion p of Amer-
icans who believe that life is better for people like them today than it was 50
years ago. To design the survey, you determine that you want to estimate p
within a margin of error of 0.04 with 99.5% confidence. To begin, we know that
our margin of error is m = 0.04. Next, we need to find our critical value z∗ . For
99.5%
 confidence,
 our value of α will be α = 0.005, so we choose z∗ such that
P N (0, 1) > z∗ = 1 − 0.005/2 = 0.9975. Using the code qnorm(p=0.9975,
mean=0, sd=1) in R, we will find that z∗ = 2.807.
The final piece we are missing is p̃. Say that we do not have an idea of the
sample proportion from a previous study, so we will have to assume the worst
case scenario and set p̃ = 0.5. Putting all this together, our minimum sample
size that will estimate p within a margin of error of 0.04 with 99.5% confidence
is
(2.807)2 0.5(1 − 0.5)
n≥ = 1231.133.
0.042
So we will need a sample size of at least 1232 subjects. On the other hand,
say we did a little digging and found a Pew Research Survey from 2017 [51]
that found a p̂ = 0.37. In this case, our p̃ = 0.37 and our sample size calculation
would change to

(2.807)2 0.37(1 − 0.37)

n= = 1147.908.
0.042
 Confidence intervals for a single parameter Chapter | 14 159

So our sample size would need to be 1148 people. Notice that our necessary
sample size is lower than the worst case scenario, which is what we expect since
the sample size is maximized at p̃ = 0.5.

14.4.2 Practice problems

7. Say that you wanted to estimate the true proportion of American adults who
have experienced harassment online within a margin of error of m = 0.1 with
90% confidence. A previous study [52] found that p̂ = 0.41. What sample
size will you need for your study?
8. Say that you wanted to estimate the true proportion of people in Greece who
completed their college education at a foreign college within a margin of
error of m = 0.05 and 99% confidence. What sample size will you need for
your study?
9. Suppose that you wanted to estimate the true proportion of high school stu-
dents who took honors classes in high school within a margin of error of
m = 0.01 and 95% confidence. What sample size will you need for your
study?

14.4.3 Confidence intervals for μ and hypothesis testing

Hypothesis testing and confidence intervals are our two forms of statistical infer-
ence. When looking at the formulas involved, particularly tests and intervals for
means, it seems like there is some connection between the two. Both confidence
intervals and hypothesis testing are predicated on the central limit theorem,
they both involve the sample mean x̄ and they both involve the standard error
s
s.e.(x̄) = √ . There are too many similarities there for it to just be coincidence.
n
In fact, it does turn out that there is a direct connection between our (1−α)100%
confidence interval for μ and the results of one particular hypothesis test: when
our alternative hypothesis is HA : μ = μ0 .
In hypothesis testing, our goal is to determine whether we want to reject or
fail to reject the null hypothesis in favor of an alternative hypothesis. To reject
the null hypothesis, we need our p-value to be less than our significance level
α. For the alternative hypothesis that HA : μ = μ0 , this means that we reject the
null hypothesis if

2 × P (tn−1 ≥ |t|) < α or P (tn−1 ≥ |t|) < α/2

x̄ − μ0
where t is our test statistic and t = √ . If this is true and we reject the null
s/ n
hypothesis, this means that |t| > t , where we set t ∗ so that P (tn−1 ≥ t ∗ ) = α/2.
∗

We can see this in Figure 14.2, where the area under the curve (equal to α/2)
in the figure to the left is P (tn−1 ≥ t ∗ ) and the area under the curve (equal to
the p-value for a rejected null hypothesis) for the bottom is P (tn−1 ≥ |t|). (See
Fig. 14.2.)
160 Basic Statistics With R

FIGURE 14.2 If our test statistic t is greater than a critical value t ∗ , we will reject the null hypoth-
esis.

∗
 reject H0 :μ = μ0 ∗in favor of HA : μ = μ0 if |t|
 So, we ∗will  > t where
P tn−1 ≥ t = α/2. If P tn−1 ≥ t = α/2, this means that P tn−1 ≤ t ∗ =
1 − α/2. If this is our rule to reject the null hypothesis, we can define our other
possible decision by its opposite. That is, we will fail to reject H0 : μ = μ0 if
|t| < t ∗ , where we choose t ∗ so that P tn−1 ≤ t ∗ = 1 − α/2. We can expand
this a little more fully by writing out our test statistic that we calculate in our
hypothesis test. In doing so, we will fail to reject the null hypothesis if
 
 x̄ − μ0 
 √  < t ∗.
 s/ n 

Our values of x̄, s, and n will not change in a study regardless of the test that
we choose. Similarly, our t ∗ will not change once we set our significance level
α. In fact, the one thing that can change across these tests is the values of μ0 .
In fact, given a sample and a set significance level, we can determine if a value
of μ0 will result in rejection or failure to reject. We do this by solving for μ0 in
the equation above. Our first step is to recognize that if |t| < t ∗ , this implies that
−t ∗ < t < t ∗ . With this in mind, we can rewrite our above equation and solve
for μ0 using a little algebra. This will define under what conditions—in other
words, for what values of μ0 —we will fail to reject our null hypothesis.
x̄ − μ0
− t∗ < √ < t∗
s/ n
s s
− t ∗ √ < x̄ − μ0 < t ∗ √
n n
∗ s s
− x̄ − t √ < −μ0 < −x̄ + t ∗ √
n n
∗ s ∗ s
x̄ − x̄ − t √ < μ0 < x̄ + t √ .
n n
 Confidence intervals for a single parameter Chapter | 14 161

This means that we will fail to reject the null hypothesis if μ0 is in the
interval
 
s s
x̄ − t ∗ √ , x̄ + t ∗ √
n n
 
where we choose t ∗ such that P tn−1 ≤ t ∗ = 1 − α/2. If you look closely, you
will notice that this interval is our (1 − α)100% confidence interval for μ. This
is not a coincidence. If we create a (1 − α)100% confidence interval for μ and
then want to conduct a hypothesis test of H0 : μ = μ0 versus HA : μ = μ0 at
the same α significance level, we will fail to reject the null hypothesis if μ0 is
contained in our (1 − α)100% confidence interval. On the other hand, if μ0 is
not contained in the (1 − α)100% confidence interval, then we will reject the
null hypothesis at the α significance level.
Let us see this in practice, calculating our confidence interval, applying this
result, and confirming it through hypothesis testing. Recall a previous exam-
ple [45,46] in which a factory that manufactured washers wanted to test if the
population average of washer size differed from 4 in any way. The sample av-
erage of 20 washers was found to be 3.9865 with sample standard deviation of
s = 0.073. If we wanted to create a 95% confidence interval, we would find—
using qt(p=0.975, df=19)—that t ∗ is 2.093. Putting this and our sample data
together, our 95% confidence interval for μ would be

0.073
3.9865 ± 2.093 × √ = (3.952, 4.021).
20
With this in mind, if we were testing H0 : μ = 4 versus HA : μ = 4 at the α =
0.05, we could just look at our 95% confidence interval since the α level for a
95% confidence interval is α = 0.05. Since 4 is contained in the interval (3.952,
4.021)—our 95% confidence interval for μ—we would fail to reject the null
hypothesis. If we look back at our previous chapter, we found that our p-value
for this test was 0.4185027. This clearly implies that we should fail to reject the
null hypothesis, which matches our conclusion based on our confidence interval.

14.4.4 Practice problems

According to nutritional information, we expect to find 4.1 mg of copper in
wholemeal flour. To test this, a researcher collects data on 28 samples of whole-
meal flour. They find that the sample average of copper is x̄ = 4.28 with a sample
standard deviation or s = 5.297 [53].
10. Calculate the 90% confidence interval for μ. Would you reject the null
hypothesis in the test of H0 : μ = 4 versus HA : μ = 4 at the α = 0.1 level?
11. Say you created a 95% confidence interval for μ. Would you reject the null
hypothesis in the test of H0 : μ = 2 versus HA : μ = 2 at the α = 0.05
level?
162 Basic Statistics With R

12. Say you created a 99% confidence interval for μ. Would you reject the null
hypothesis in the test of H0 : μ = 4 versus HA : μ = 4 at the α = 0.005
level?

14.5 Conclusion
Our confidence intervals for a single parameter grow out of the same assump-
tions and foundational theorems as hypothesis testing. At the center is the central
limit theorem, which again allows us to connect our data to our parameters. This
connection brings additional uses to the confidence interval, allowing us to both
test with confidence interval as well as determine sample size for future studies.
This is all in addition to its main use: giving us the range of plausible values
for our parameter. Similar to hypothesis testing, the procedure for confidence
intervals remains consistent across intervals for both populations means and
proportions. We will see this continue going forward as we begin to compare
groups to see if they are similar.
Chapter 15

Hypothesis tests for two

parameters
Contents
15.1 Introduction 163 15.3.4 Calculate the test statistic 173
15.2 Two-sample test for proportions 164 15.3.5 Calculate p-values 175
15.2.1 State hypotheses 164 15.3.6 Conclusion 177
15.2.2 Set significance level 165 15.3.7 Practice problems 178
15.2.3 Collect and summarize 15.4 Paired t-test for means 179
data 165
15.4.1 State hypotheses 180
15.2.4 Calculate the test statistic 166
15.4.2 Set significance level 181
15.2.5 Calculate p-values 168
15.4.3 Collect and summarize
15.2.6 Conclude 169
data 181
15.2.7 Practice problems 170
15.3 Two-sample t-test for means 171 15.4.4 Calculate the test statistic 182
15.3.1 State hypotheses 171 15.4.5 Calculating p-values 182
15.3.2 Set significance level 172 15.4.6 Conclude 183
15.3.3 Collect and summarize 15.4.7 Practice problems 184
data 172 15.5 Conclusion 185

15.1 Introduction
To this point, we have talked about statistical inference for only a single param-
eter. Our questions have been limited to asking if a single population proportion
or population mean is plausibly equal to a value. However, the majority of
statistical questions are ones of comparisons: Does one group have a greater
population mean than another? Does one group have a population proportion
equal to another?
We see these sorts of questions all the time. In a 2016 pre-election survey
[54], respondents were asked if they had a candidate’s sign in their yard. Of
persons who had voted in the 2012 election, 26.6% had a sign in their yard,
while among those who had not voted in 2012 only 7.8% had a sign in their
yard. A reasonable question could be: “Does the population proportion of 2012
voters who have a candidate’s sign in their yard differ from those who didn’t
vote in 2012?”
This chapter looks into just these sorts of questions and the hypothesis tests
that can help answer them. There are three tests that we will talk about in
this chapter: the two-sample test for proportions, the two-sample t-test for

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00028-6 163

Copyright © 2022 Elsevier Inc. All rights reserved.
164 Basic Statistics With R

means, and the paired t-test for means. Each test is designed to work in a
specific scenario, so knowing which test to use is important. Ultimately, identi-
fying the correct test to use comes down to answering two questions about our
samples:
1. Is the response of interest categorical or quantitative?
2. Are the two samples independent? Put another way, is one sample affected
by or related to another sample in some way?
As we go through our test, we will go through the answers to these two
questions to choose the appropriate test, as well as the steps that go into each of
these hypothesis tests.

15.2 Two-sample test for proportions

In the two-sample test for proportions, we have are interested in if the propor-
tions of two separate populations are identical or different in some way. Because
we are dealing with proportions, we will have a categorical response, the only
one of the three tests to be for a categorical response. The steps for our two-
sample tests are similar to what we saw in the one-sample test, with the first
steps being...

15.2.1 State hypotheses

We begin by stating what we are ultimately interested in testing. As before, the
null hypothesis is our assumed status quo, that nothing interesting is happening.
In our two-sample test for proportions, we assume going in that the two popula-
tions have an identical population proportion. Mathematically, this means that

p 1 = p2

where p1 is the true population proportion of population 1 and p2 is the true

population proportion of population 2. However, our general form of our hy-
potheses is that our parameter—or some combination of parameters—equals
some numeric value. However, if we know that the two population proportions
are equal, or p1 = p2 , this means that

p1 − p2 = 0.

This is a combination of two parameters—p1 and p2 —where the combina-

tion equals a numeric value—0 in this case. This becomes our null hypothesis,
that the two populations have equal proportions, or

H0 : p1 − p2 = 0.

Our alternative hypothesis will again be one of three options. The first option
is that the population proportion for our first population is greater than popula-
tion proportion for our second population, or p1 > p2 . This is equivalent to
 Hypothesis tests for two parameters Chapter | 15 165

p1 − p2 > 0, leading to the alternative hypothesis

HA : p1 − p2 > 0.

Our second option is the opposite of the first, with p1 < p2 . Again, we can
do a little algebra to see that this is equivalent to p1 − p2 < 0, resulting in the
alternative hypothesis
HA : p1 − p2 < 0.
Last of all is the two-sided test, where we are interested in if the population
proportions of our two groups differs in any way, or p1 = p2 . In this case, our
alternative hypothesis will become

HA : p1 − p2 = 0.

Let us take an example to illustrate this and the following steps. In a 2018
survey [55], Pew research did a study to investigate America’s attitudes toward
space exploration. One of the key questions was: “Is it essential for the United
States to continue to be a world leader in space exploration?” In order to see
how opinions differed across generations, the researchers grouped responses
by generation. Say the researchers were interested in testing if the population
proportion of individuals who believed it was essential for the United States to
continue being a world leader in space exploration was the same or differed for
Millenials and Gen Xers. Our hypotheses for this test would be

H0 : pMillenial − pGenX = 0
HA : pMillenial − pGenX = 0.

15.2.2 Set significance level

As before, we need to consider your personal comfort level to make Type I Er-
rors when setting the significance level α in addition to the conventions of one’s
field. Again, unless otherwise stated, the common default significance level is
α = 0.05.
For the sake of seeing our hypothesis procedure with a different significance
level, during this particular example looking at opinions of space exploration,
we will work with a significance level of α = 0.1.

15.2.3 Collect and summarize data

In the one-sample test for proportions, we needed the sample proportion p̂ and
the sample size n of a single sample to conduct statistical inference. Here, we
have two samples to be concerned with, so we will need those two quantities
from both samples. We will need the sample proportions for our samples from
both populations one and two—p̂1 and p̂,2 respectively—and the sample size
166 Basic Statistics With R

for our samples from both populations one and two—n1 and n2 . Outside of these
summaries, we do not need to know anything else about the sample.
In the space exploration survey, 467 out of 667 Millenials believed it was
important for the United States to be a leader in space exploration while 407
our of 558 Gen Xers believed the same. Thus our sample proportions would be
467 407
p̂Millenial = = 0.7001 and p̂GenX = = 0.7293.
667 558

15.2.4 Calculate the test statistic

Calculating our test statistic in the two-sample test for proportions is a little
more involved than in the one-sample case. However, we can still get to it by
beginning with our general form of the test statistic, defined to be
Sample Statistic – Null Hypothesis Value
t= .
Standard Error
So, we first need to find the sample statistic that estimates the parameters—or
combination of parameters—in our null hypothesis. Remember, our null hypoth-
esis for the two-sample test for proportions is

H0 : p1 − p2 = 0.

So we will need a combination of sample statistics that estimates p1 − p2 .

We know that p̂1 estimates p1 and p̂2 estimates p2 , so it seems reasonable that
p̂1 − p̂2 would estimate p1 − p2 . Plugging this in, our test statistic to this point
would be
p̂1 − p̂2 – Null Hypothesis Value
t= .
Standard Error
We can easily fill in the null hypothesis value as well, as we assume in the
null hypothesis that p1 −p2 = 0, so the null hypothesis value for our test statistic
will be 0. This makes the test statistic
p̂1 − p̂2 − 0
t= .
Standard Error
The last component of our test statistic is to get the standard error of p̂1 −
p̂2 , or s.e.(p̂1 − p̂2 ). Like many of our other standard errors, we will get to
this through the central limit theorem. By the central limit theorem, we know
that for large enough sample sizes our sample proportion will follow a Normal
distribution, specifically
 
p(1 − p)
p̂ ∼ N p, .
n
This holds for any sample proportion that meets and sample size and success
threshold. So, we can apply this result to each of our two samples, assuming the
 Hypothesis tests for two parameters Chapter | 15 167

sample sizes are large enough for both groups. By the central limit theorem, this
means that
 
p1 (1 − p1 )
p̂1 ∼ N p1 ,
n1
 
p2 (1 − p2 )
p̂2 ∼ N p2 ,
n2

where p1 and p2 are the true population proportions for our two populations
and n1 and n2 are the sample sizes for our two samples. Now, we need to know
the distribution of p̂1 − p̂2 in order to get the standard error. Based on the fact
that p̂1 and p̂2 follow Normal distributions for sufficient sample sizes and what
we know about how we can combine Normal distributions, it turns out that the
distribution of p̂1 − p̂2 is a Normal distribution, specifically
 
p1 (1 − p1 ) p2 (1 − p2 )
p̂1 − p̂2 ∼ N p1 − p2 , + .
n1 n2

So, the standard error of p̂1 − p̂2 will be


p1 (1 − p1 ) p2 (1 − p2 )
s.e.(p̂1 − p̂2 ) = + .
n1 n2

Now, in hypothesis testing we always assume going in that the null hypoth-
esis is true, and thus we need to calculate this standard error assuming that the
null hypothesis is true. If that is the case, that means that p1 = p2 , so we can
drop the subscripts and use a common population proportion p. This means that
our standard error would become

p(1 − p) p(1 − p)
s.e.(p̂1 − p̂2 ) = +
n1 n2

where p is the true population proportion for both populations one and two.
However, this common population proportion p is a parameter and, therefore,
an unknown quantity. To calculate the test statistic, we would need to get an
estimate for p. Usually, we estimate the population proportion with the sample
proportion. However, we now have two samples that we are working with, so
which do we use? If p1 = p2 = p, that implies that the two populations are no
different from one another. If those two populations are no different from each
other, we can combine information from both samples to better estimate the
common population proportion p. This estimate for p using information from
both samples is called the pooled proportion because it pools together all the
information available to get a single estimate for p. The pooled proportion p̂
will be the combined successes y1 + y2 divided by the combined sample size
168 Basic Statistics With R

n1 + n2 , or
y1 + y2
p̂ = .
n1 + n2
Here, y1 and y2 are the number of successes in the sample from populations
one and two, while n1 and n2 are the sample sizes. If we define our pooled
proportion in this way, our standard error for p̂1 − p̂2 becomes

p̂(1 − p̂) p̂(1 − p̂)
s.e.(p̂1 − p̂2 ) = +
n1 n2
  
1 1
= p̂(1 − p̂) + .
n1 n2

This is the final component of our test statistic for this hypothesis test, re-
sulting in
p̂1 − p̂2
t=  .
1 1
p̂(1 − p̂) +
n1 n2

In our space exploration survey example, we first need to find our pooled
467 + 407
proportion p̂. Based on our data, this will be p̂ = = 0.7135. Putting
667 + 558
this together with our remaining sample data, we can calculate that our test
statistic is
0.7001 − 0.7293
t=   = −1.1273.
1 1
0.7135 × (1 − 0.7135) +
667 558

15.2.5 Calculate p-values

In our one-sample test for p, we found that our test statistic will follow a stan-
dard Normal distribution assuming that the null hypothesis is true. The same is
true here, as our test statistic for the two-sample test for proportions will follow
a standard Normal distribution when the null hypothesis is true. With this in
mind, we can use this distribution to get our p-values contingent on our alterna-
tive hypothesis.
Our rationale for each of our p-values—the probability that we see “more
extreme” data given the null hypothesis is true—is identical to what we saw in
the one sample case. Similarly, we will use the pnorm function to get do the
p-value calculation. With this mind, we will not reiterate it here and reference
you to that previous chapter.
 Hypothesis tests for two parameters Chapter | 15 169

HA P-value
 
HA : p1 − p2 < 0 P N (0, 1) ≤ t
 
HA : p1 − p2 > 0 P N (0, 1) ≥ t
 
HA : p1 − p2 = 0 2 × P N (0, 1) ≥ |t|

Again, we can only use our Normal distribution for p-values if the central
limit theorem holds, which is contingent on the sample size. We need for both
sample sizes n1 and n2 to be at least 30, and need to observe at least 10 successes
and failures in both our sample from population one and population two.
In the space exploration example, the alternative hypothesis is HA :
pMillenial − pGenX = 0 and
 our test  statistic is t = −1.1273, so our p-value
will be 2 × P N (0, 1) >  − 1.1273 . We can get this through the pnorm func-
tion, specifically with the code given below, to find that our p-value is 0.2596.

2*pnorm(1.1273, mean=0, sd=1, lower.tail=FALSE)

15.2.6 Conclude
Finally, once we have our p-value, we can draw our conclusions. As before, if
the p-value is less than the significance level α, we will reject the null hypothesis
and conclude in favor of our alternative hypothesis. Otherwise, we will fail to
reject the null hypothesis and conclude that it is plausible that the two population
proportions are equal.
To conclude our election example, since our p-value of 0.2596 is greater than
our significance level of α = 0.1, we would fail to reject the null hypothesis and
conclude it is plausible that the proportion of individuals who believe that it is
important for the United States to continue to be a leader in space exploration is
the same for Millenials and Gen Xers.
Let us see all these steps in practice all at once. A January 2018 poll done
by the Pew Research Group investigated the social media habits of Americans
[56]. One of the questions they were interested in was if the proportion of
Americans aged 18–29—population one with proportion p1 —were online “con-
stantly” more than ages 30–49—population two with proportion p2 . They found
that 137 out of 352 participants in the 18–29 sample were online “constantly”
as opposed to 190 out of 528 participants in the 30–49 sample. To answer this
question, our hypotheses would be

H0 : p1 − p2 = 0
HA : p1 − p2 > 0.

Here, we will assume that our significance level is α = 0.05. In collecting and
137
summarizing our data, we have that our sample proportions are p̂1 = =
352
170 Basic Statistics With R

190
0.3892 and p̂2 = = 0.3598. Additionally, we will need the pooled propor-
528
137 + 190
tion for our test, calculated using p̂ = = 0.3716. Now, to calculate
352 + 528
our test statistic we will have
p̂1 − p̂2
t=
1 1
p̂(1 − p̂) +
n1 n2
0.3892 − 0.3598
= = 0.8829.
 1 1 
0.3716(1 − 0.3716) +
352 528

So, based on our alternative hypothesis, our p-value will be P N (0, 1) >
0.889 . We can use the R code
pnorm(q=0.8829, mean=0, sd=1, lower.tail=FALSE)

to get that this probability and our p-value is equal to 0.1887. Since the p-value
is greater than α = 0.05, we fail to reject the null hypothesis and conclude that it
is plausible that the proportions 18–29-year-olds and 30–49-year-olds who are
online “constantly” are equal.
The previously mentioned prop.test function is also able to do our two-
sample test for proportions. Rather than just including information about a
single sample, we merely add in our information about our second sample. As
before, the function will give us our test statistic and p-values. More detail and
examples are given in Chapter 17.

15.2.7 Practice problems

In 2017, Pew Research conducted a survey to see what STEM job employees
valued when choosing a job [57]. Specifically, they wanted to see if men and
women valued different things when choosing a job in a STEM field. They found
that 794 men out of 1119 surveyed valued having a job that offered flexibility
to balance work and family, while 919 out of 1225 women valued that same
flexibility.
1. What will be the hypotheses if researchers want to test any difference be-
tween men and women’s preference about work/family balance?
2. What will be the pooled proportion p̂ if the null hypothesis is true?
3. What will be the test statistic for this hypothesis test?
4. What will be the p-value for this hypothesis test?
5. Assuming α = 0.05, what will our conclusion be?
A researcher is interested in the rate of criminal recidivism within a year of
release for those former inmates who received financial aid upon release was
equal to or lower than the rate of recidivism for those who did receive financial
 Hypothesis tests for two parameters Chapter | 15 171

aid [48,58]. Out of 216 former inmates who did not receive financial aid, 66
were arrested within a year of release. Of the 216 who did receive financial aid,
48 were arrested within a year of release.
6. Set up and test this hypothesis at the α = 0.01 level.
A researcher is interested in whether the proportion of individuals who re-
ceived corporal punishment as a child believe in corporal punishment for chil-
dren at a higher rate than those who did not receive corporal punishment [90,91].
Out of 307 individuals who remember receiving corporal punishment as a child,
263 believe in moderate corporal punishment of children. Of 1156 individuals
who do not remember receiving corporal punishment, 783 believe in moderate
corporal punishment of children.
7. Set up and test this hypothesis at the α = 0.1 level.

15.3 Two-sample t-test for means

Our second two-sample hypothesis test is our two sample t-test for means. The
two-sample t-test for means is used when our question of interest is if the means
of two populations are identical or differ in some way. Because we are dealing
with a question related to population means, this means that we will have a quan-
titative response. This trait is shared by both the two-sample t-test for means as
well as the paired t-test for means. With this in mind, we will need a factor to
distinguish between the two tests. The difference between the two-sample t-test
for means and the paired t-test is that the samples in our two-sample t-test for
means are independent from each other. That is, there is no overlap between
the two samples; they have no affect on each other. The paired test does have
overlap or some interaction between the two samples, a fact that we will discuss
more in our paired test section.
The steps for our two-sample tests are again similar to what we saw in the
one-sample tests. To illustrate these steps, we will look at the following example
[48,59]. A researcher set up a social-psychological experiment to see if individ-
uals conformed more to the opinions of individuals who were of higher or lower
status. To test this, the researcher set up 40 scenarios where the subject was
faced with a manipulated disagreement where their conversation partner was an
individual of high or low status. With this in mind, we will begin where we have
always begun.

15.3.1 State hypotheses

We begin by stating what we are ultimately interested in testing. As before, the
null hypothesis is our assumed status quo, that nothing interesting is happening.
In our two-sample t-test for means, we assume going in that the two popula-
tions have an identical population mean, or μ1 = μ2 . Mathematically, just like
we saw in the two-sample test for proportions, this is identical to saying that
172 Basic Statistics With R

μ1 − μ2 = 0. This will become our null hypothesis, that the two populations
have equal means, mathematically stated as

H0 : μ1 − μ2 = 0.

Our alternative hypothesis will be one of three options that we have seen
before. The first option is that population mean for population one is greater
than the population mean for population two, or μ1 > μ2 . With a little algebra,
this leads to the alternative hypothesis

HA : μ1 − μ2 > 0.

Option two has the population mean for population one being less than the
population mean for population two, or μ1 < μ2 . After subtracting μ2 from both
sides of the inequality, our alternative hypothesis becomes

HA : μ1 − μ2 < 0.

Last of all we have our two-sided test, used if we are interested in asking the
two population means differ in any way, or μ1 = μ2 . In this case, our alternative
hypothesis is
HA : μ1 − μ2 = 0.
For our conformity example, say we were interested in testing if partners of
high status resulted in equal versus higher conformity compared to partners of
low status. In this case, our hypotheses would be

H0 : μH igh − μLow = 0
HA : μH igh − μLow > 0.

15.3.2 Set significance level

As before, when setting the significance level α, we need to consider the conven-
tions of your field in addition to how comfortable you are in making Type I Er-
rors. Unless otherwise stated, the common default significance level is α = 0.05.
For our conformity example, for the sake of using various significance levels,
we will use the significance level of α = 0.01.

15.3.3 Collect and summarize data

In the one-sample t-test for means, we required the sample mean x̄, the sam-
ple standard deviation s, and the sample size n to calculate our test statistic
and p-values. Here, our question has expanded to two samples and we will
need summaries from both of these samples. We will be interested in the sam-
ple averages from both samples from population one and population two—x̄1
 Hypothesis tests for two parameters Chapter | 15 173

and x̄2 , respectively—the sample variances for both samples from population
one and population two—s12 and s22 , respectively—and the sample size for both
samples—n1 and n2 .
In the conformity example, 23 subjects experienced a partner of higher sta-
tus with the average number of conforming responses being x̄H igh = 14.217
igh = 19.087. For our second group, 22 subjects experienced a
with variance sH 2

partner of lower status with the average number of conforming responses being
x̄Low = 9.955 with a sample variance of sLow2 = 27.855.

15.3.4 Calculate the test statistic

As with each of our hypothesis tests, we begin the process of calculating our
test statistic with our general formula for the test statistic.
Sample Statistic – Null Hypothesis Value
t= .
Standard Error
We will need to fill in each of these components from our data, our hypothe-
sis, and knowledge about our sample statistics. First, we need to find the sample
statistic that estimates the parameters—or combination of parameters—in our
null hypothesis. Remember, our null hypothesis for the two-sample t-test for
means is
H0 : μ1 − μ2 = 0.
So we need a combination of sample statistics that estimates μ1 − μ2 . We
know that x̄1 estimates μ1 and x̄2 estimates μ2 , so it should not be surprising
that x̄1 − x̄2 estimates μ1 − μ2 . Additionally, we know from our null hypothesis
that our null value is 0—as it will be for all our two-sample tests—so our test
statistic to the point is
x̄1 − x̄2 − 0
t= .
Standard Error
All that is missing is the standard error of our sample statistics. Similar to our
two-sample test for proportions, we will turn to the central limit theorem to find
our standard error. By the central limit theorem, we know that for either large
enough sample sizes—n ≥ 30—or for Normally distributed data our sample
mean will follow a Normal distribution, namely
 
σ2
x̄ ∼ N μ, .
n
With this result in mind, we can apply this result individually to both of our
samples, finding that each sample mean will follow its own Normal distribution.
 
σ2
x̄1 ∼ N μ1 , 1
n1
174 Basic Statistics With R

 
σ2
x̄2 ∼ N μ2 , 2 .
n2

However, we need the standard error of x̄1 − x̄2 . Based on what we know
about combining Normal distributions, we find that x̄1 − x̄2 also follows a Nor-
mal distribution, specifically
 
σ12 σ22
x̄1 − x̄2 ∼ N μ1 − μ2 , +
n1 n2

which would make the standard error, of x̄1 − x̄2 equal to


σ12 σ22
s.e.(x̄1 − x̄2 ) = + .
n1 n2

σ12 σ22
How do we estimate this quantity, specifically + ? There are a cou-
n1 n2
ple of ways to do this, but it comes down to one simple question: Does σ12 = σ22 ?
The answer to this question will affect not only our standard error, and thus our
test statistic, but also the distribution that the test statistic follows.
If the variances are equal, then we can combine their information into a sin-
gle estimate for a common pooled population variance, or σp2 . This is similar to
how we pooled our information to get our proportion estimate in our two-sample
test for proportions. If this is the case, we can estimate our pooled population
variance σp2 using the pooled sample variance sp2 , calculated using

(n1 − 1)s12 + (n2 − 1)s22

sp2 =
n1 + n2 − 2

where s12 and s22 are our sample variances for sample one and sample two, re-
spectively, and n1 and n2 are our sample sizes. This pooled sample variance
can be substituted for our pooled population variance in the standard error of
x̄1 − x̄2 , with a little factoring making the standard error now
  
1 1
s.e.(x̄1 − x̄2 ) = sp2 + .
n1 n2

We then substitute this standard error into the formula for our test statistic,
making our test statistic for equal variances

x̄1 − x̄2
t=  .
1 1
2
sp +
n1 n2
 Hypothesis tests for two parameters Chapter | 15 175

Our other option is if our variances are not equal, or σ12 = σ22 . If this is
the case, we need two statistics that estimate σ12 and σ22 in order to use these
statistics in our standard error. It seems reasonable that our two sample variances
should estimate these population variances so we can substitute these sample
variances for our population variances. This makes our standard error

s12 s2
s.e.(x̄1 − x̄2 ) = + 2.
n1 n2

So, in this case our test statistic for unequal variances

x̄1 − x̄2
t= .
s12 s22
+
n1 n2

This of course brings up a question: How do we know if the variances are

equal? They are parameters, so of course unknown. There are formal hypothesis
tests that look at if variances are equal, but in practice we will use a standard
rule of thumb. If one of the sample variances is more than four times larger than
s2 s2 1
the other—that is, 12 > 4 or 12 < —we will assume that the variances are not
s2 s2 4
equal. Otherwise, we will assume equivalent variances.
For our conformity example, we have all the various parts of our test statistic.
We first look at the ratio of the variances to see if we will conclude if variances
19.087 1
are equal or not. Since = 0.6852 is between and 4, we will say that
27.855 4
the variances are equal and use that formula. Thus, we need to calculate our
pooled variance sp2

(23 − 1) × 19.087 + (22 − 1) × 27.855

sp2 = = 23.37.
23 + 22 − 2
Now, we can plug this into our test statistic to get

14.217 − 9.955
t=   = 2.956.
1 1
23.37 +
23 22

15.3.5 Calculate p-values

In order to calculate our p-values we need to know what distribution our test
statistic follows in addition to what our alternative hypothesis is. In our one-
sample t-test for means, we found that our test statistic follows a t-distribution
with n − 1 degrees of freedom. In our two-sample case, the distribution that
176 Basic Statistics With R

our test statistic follows depends on the answer to the question that we posed
earlier: Are our variances equal or not? We saw how we can roughly determine
the answer to this question and how that answer can affect our test statistic. This
answer again comes into play as it will change our p-values.
In the case where we determine that our variances are equal, our test statistic
will follow a t-distribution, similar to what we saw before. However, our degrees
of freedom have changed. In this case, our degrees of freedom will be ν = n1 +
n2 − 2, where n1 and n2 are our sample sizes from our two samples.
When the variances are not equal, the distribution of our test statistic is a little
more difficult to define. It turns out that our test statistic follows approximately
a t-distribution. This is an approximation, not precise—the exact distribution
of this test statistic with unequal variances is one of the classic unsolved ques-
tions in statistics. However, knowing that the test statistic follows approximately
a t-distribution is not enough; we need our degrees of freedom as well. Exact
solutions for the degrees of freedom are hard to come by and difficult to imple-
ment, so more often approximations are used.
The most common of these approximations is referred to as the Satterthwaite
approximation, or the Welch approximation. In this case, our degrees of freedom
ν of our t-distribution used for p-values will be given by

 
s12 s2 2
+ 2
n1 n2
ν=  2 2  2 2 .
1 s1 1 s2
+
n1 − 1 n1 n2 − 1 n2

Once we know the distribution of our test statistic, we only need our alterna-
tive hypothesis. As before, calculating our p-values is dependent on our alterna-
tive hypothesis, with identical rationale to our one-sample t-test for means—and
therefore that rationale is omitted here. Under each of the three alternative hy-
potheses, our p-values will be

HA P-value
 
HA : μ1 − μ2 < 0 P tν ≤ t
 
HA : μ1 − μ2 > 0 P tν ≥ t
 
HA : μ1 − μ2 = 0 2 × P tν ≥ |t|

where ν is degrees of freedom determined by whether the variances are equal or

not. As before, these results only hold if the central limit theorem holds, which
is contingent on the sample size and distribution of our data. In this case, we
need for both sample sizes to be at least 30 or both samples of data to follow a
Normal distribution to be completely confident that our assumptions will hold.
 Hypothesis tests for two parameters Chapter | 15 177

That is,

n1 ≥ 30, n2 ≥ 30
OR
x1,i ∼ N (μ1 , σ12 ), x2,i ∼ N (μ2 , σ22 ).

In our conformity example, we determined that our variances are plausibly

equal. Thus, our test statistic will follow a t-distribution with ν = n1 + n2 −
2 = 23 + 22 − 2 = 43 degrees of freedom. This, combined with  our alternative

hypothesis of HA : μH igh − μLow > 0, our p-value will be P t43 > 2.956 . So,
we can use the R code pt(2.956, df=43, lower.tail=FALSE) to find that our p-
value is 0.00252.

15.3.6 Conclusion
Our conclusion step has not changed, as we again reject the null hypothesis if
the p-value is less than our significance level α, while failing to reject otherwise.
In our conformity example, since our p-value of 0.00266 is less than our signifi-
cance level of α = 0.01, we reject the null hypothesis and conclude that partners
of higher status result in higher levels of conformity than partners of lower sta-
tus. However, it is important to note that our sample size is smaller than 30 in
both samples and a look at histograms of our data brings Normality into doubt,
bringing the results of this test somewhat under scrutiny. (See Fig. 15.1.)

FIGURE 15.1 Histogram of conformity data.

Let us see all these steps in practice. A researcher is investigating the effect
of two different diets on the weight of newborn chickens [60]. Specifically, they
were interested in if there was any difference in weight gain for chicks on a diet
supplemented with sunflower seeds versus a diet supplemented with soybeans.
178 Basic Statistics With R

The hypotheses for this test would be

H0 : μsoybean − μsunf lower = 0

HA : μsoybean − μsunf lower = 0.

Say we tested this at an α = 0.05. Upon collecting the data, the a sam-
ple of 12 chicks on sunflowers had an average weight gain of x̄sunf lower =
lower = 2384.99. Meanwhile, amongst
2
328.92 with a sample variance of ssunf
14 chicks on soybeans, the sample average weight gain was x̄soybean =
2
ssunf lower
246.43 with a sample variance of ssoybean = 2929.96. Since 2
2 =
ssoybean
0.814, we can say that it is plausible that the variances are equal. Be-
cause of this, we can calculate our pooled sample variance to be sp2 =
(12 − 1) × 2384.99 + (14 − 1) × 2929.96
= 2680.182. Based on this data, our
12 + 14 − 2
test statistic would be
246.43 − 328.92
t=   = −4.050.
1 1
2680.812 +
14 12

Our test statistic will follow a t-distribution, with our degrees of freedom
being ν = 12 + 14 − 2 = 24. Based on our alternative hypothesis, our p-value
will be 2 × P (t24 > | − 4.050|). We turn to the R code pt(q=4.050, df=24,
lower.tail=FALSE) to get that this probability is 0.0002322, then double it to get
our p-value of 0.0004644. So, since our p-value is definitely less than α = 0.05,
we will reject the null hypothesis and conclude that there is a difference between
the two diets. The only concern with this result may be that since we have small
sample sizes for both of our two diets and a lack of Normality in our data makes
our results may not be fully reliable. (See Fig. 15.2.)

15.3.7 Practice problems

Researchers were interested in if population average clump thickness in biop-
sied tumors differ for benign and malignant tumors. In 458 benign tumors,
they found a sample average of x̄Benign = 2.96 with sBenign = 1.67, while in
241 malignant tumors they found a sample average of x̄Malignant = 7.2 with
sMalignant = 2.43 [10,61].
8. Set up the hypotheses for this test.
9. What is the test statistic for this test?
10. What are the degrees of freedom for this test?
11. What is the p-value for this test?
12. Assuming that α = 0.01, what is your conclusion for this test?
 Hypothesis tests for two parameters Chapter | 15 179

FIGURE 15.2 Histogram of the chicken diet dataset.

Researchers are interested in seeing if white-collar jobs are viewed as more

prestigious than blue-collar jobs. Based on a sample of 21 blue-collar jobs, the
sample average prestige score was found to be x̄BC = 22.762 with a variance
2 = 325.991. Of the six white-collar jobs, the sample average prestige
of sBC
score was found to be x̄W C = 36.667 with a variance of sW 2
C = 139.067 [48].
Histograms of the data show skewed histograms for both samples.
13. Test the hypothesis that white-collar jobs are viewed as the same versus
more prestigious than blue-collar jobs at the α = 0.1 level.
The cuckoo is a bird who often lays its eggs in the nests of other birds,
leaving their children to be raised by other species. A researcher is interested in
identifying if the cuckoos egg length was any different than the host egg length
when the host was specifically a meadow pipit. In the sample of 45 cuckoo eggs,
it was found that the average length was x̄C = 22.3 with a standard deviation of
sC = 0.89. Meanwhile, in 74 meadow pipit eggs the average length was x̄MP =
19.7 with a standard deviation of sMP = 1.25 [23,101].
14. Test the hypothesis that cuckoo egg length was the same as host meadow
pipit egg length versus different at the α = 0.05 level.

15.4 Paired t-test for means

Our final two-sample hypothesis test is the paired t-test for means. In conduct-
ing the paired t-test, we are again asking if the population means for two groups
are different. However, while this question is identical to the two-sample t-test
for means, there is an important distinction about our samples that needs to be
considered. We again have a quantitative response, as is the case with the two-
sample t-test. However, our samples in the paired t-test are not independent;
there is overlap between the two samples or some sort of direct connection be-
tween our two samples.
180 Basic Statistics With R

Why would we want overlap between our two samples? In most cases, we
want our samples or our variables to be independent, so why the exception now?
To answer this question, let us consider the following scenario. A shoe company
claims that their running shoe will help improve your mile run time. To test this
claim, you get a bunch of people together and randomly assign half the people
to use the new shoe in a mile run while the other half will run using their regular
old shoes. You then plan to do hypothesis testing using the two-sample t-test for
means to compare the two groups.
What is the problem with this setup? We wind up comparing the two groups,
so what is the problem? If we assign the groups in this way—half the partic-
ipants in each group—we do not account for each individual runner’s ability.
Under this scenario it is entirely possible that one group happens to have better
runners, resulting in that group’s times being lower, regardless of the effect of
the shoe.
So what setup would be better? It would be better to have each person run a
mile in both the new shoes and old shoes so that we can directly compare each
individual’s mile time under both conditions. In fact, we can take the difference
between their time with their old shoe and with the new shoe and conduct our
hypothesis test on this difference.
This will be the general idea of our paired t-test for means. We collect two
samples that are connected in some way—called a paired sample—and then
conduct what essentially amounts to a one-sample t-test for means on the dif-
ferences for each individual. The steps of this paired are identical to what we’ve
seen before, with an especial connection to our one-sample t-test.
As we work through these steps, we will illustrate them with the fol-
lowing example. A researcher is comparing the accuracy of two analysis
techniques determining the sugar concentration of a breakfast cereal: liquid
chromatography—a slow but accurate process—and an infra-analyzer 400 [23].

15.4.1 State hypotheses

In our paired t-test for means, we want to know if two population means differ.
However, the way that we do this involves collecting a paired sample and doing
inference on the differences. So we essentially want to look into the population
average of the differences between the populations, which we will notate as
μD . As usual, the null hypothesis represents our status quo, and for this test our
assumed state of the world is that there is no difference between the populations.
This means that the population average difference is zero, resulting in a null
hypothesis of

H0 : μD = 0.

Our alternative hypothesis comes in one of the three familiar forms. Either
the population average difference is less than 0, greater than 0, or just not equal
 Hypothesis tests for two parameters Chapter | 15 181

to 0. Leading to the three possible alternative hypotheses of

HA : μD < 0
HA : μD > 0
HA : μD = 0.

In the breakfast cereal example, say we wanted to test if there was any differ-
ence between the results of the two analysis techniques. Our hypotheses would
be

H0 : μD = 0
HA : μD = 0.

15.4.2 Set significance level

Our considerations when setting our significance level α—our personal comfort
with Type I errors as well as convention—remain the same, as does our default
significance level of α = 0.05.
We will use the baseline significance level of α = 0.05 for our breakfast
cereal example.

15.4.3 Collect and summarize data

Here, we need to consider what we are testing. We are seeing if the population
average difference is equal to zero or something else. With that in mind, it seems
reasonable that we’ll need to know the individual differences for each person.
So for each of the n subjects in the study, we will collect the observations for
each group and calculate the difference between the two.

Group 1 Group 2 Difference

x1,1 x2,1 xD,1 = x1,1 − x2,1
x1,2 x2,2 xD,2 = x1,2 − x2,2
. . .
.. .. ..
x1,n x2,n xD,n = x1,n − x2,n

After calculating the individual differences, we will need the sample aver-
age of the differences x̄D , the sample standard deviation (or variance) of the
differences sD , and the sample size n.
In the breakfast cereal analysis, the researcher collected the percentage of
sugar in 100 samples of breakfast cereal using both the liquid chromatography
method and infra-analyzer. The differences for each sample was calculated, with
the sample average being x̄D = −0.622 with a standard deviation of sD = 1.988.
182 Basic Statistics With R

15.4.4 Calculate the test statistic

When calculating our test statistic in our paired test, we will begin with our one-
sample t-test for means. This is because our paired t-test essentially boils down
to doing a one-sample t-test specifically on the differences. In the one-sample
t-test for means, we found that our test statistic was

x̄ − μ0
t= √ .
s/ n

As the paired t-test is essentially a one-sample t-test on the differences, we

can adjust this test statistic for the one-sample t-test in a few key places to work
for our paired t-test. Instead of using just any sample average x̄, we use the
sample average of the differences x̄D . Instead of any sample standard deviation
s, we use the sample standard deviation of differences sD . Finally, in the paired
t-test for means, our null hypothesis will always be H0 : μD = 0, so our μ0
value will always be equal to 0. With this in mind, if we make these changes to
our test statistic from the one-sample t-test, we get

x̄D
t= √ .
sD / n

This will be our test statistic for our paired t-test for means, with its direct
connections to the one-sample t-test for means. For breakfast cereal example,
we have our data, we just need to fill in the various parts of our test statistic.

−0.622
t= √ = −3.129.
1.988/ 100

15.4.5 Calculating p-values

In calculating our p-values, we need to know what distribution our test statistic
follows as well as the alternative hypothesis. However, a fair bit of the legwork
for getting this p-value has been done by the one-sample t-test for means. In
that one-sample t-test, we found that our test statistic followed a t-distribution
with n − 1 degrees of freedom. As the paired t-test for means is essentially
a one-sample t-test on the differences, this will again be the distribution that
our test statistic follows. Our p-value—the probability that we observe more
extreme data given that our null hypothesis is true—will still be dependent on
the alternative hypothesis with which we are working and will have identical
rationale to our one-sample t-test. As such, that rationale is omitted here and
our table for p-values is given below.
 Hypothesis tests for two parameters Chapter | 15 183

HA P-value
 
HA : μD < 0 P tn−1 ≤ t
 
HA : μD > 0 P tn−1 ≥ t
 
HA : μD = 0 2 × P tn−1 ≥ |t|

Just as with the one-sample t-test, this result will only hold if the central
limit theorem holds. This requires a large enough sample size—n ≥ 30 is the
standard—or that our differences are Normally distributed so that we are able to
trust the results of our test.
In the breakfast cereal example,
  our alternative
 hypothesis is HA : μD = 0,
so our p-value will be 2 × P t99 >  − 3.129 . Using the following R code, we
will find that our p-value is 0.0023.
2*pt(3.129, df=99, lower.tail=FALSE)

15.4.6 Conclude
As with all our other hypothesis tests, our conclusion step consists of the jus-
tification for our decision—our p-value will either be greater than or less than
our significance level α—the decision we reached—fail to reject or reject the
null hypothesis, respectively—and the resulting context about what this deci-
sion means for our parameter.
In the breakfast cereal analysis, since our p-value of 0.0023 is less than our
significance level α = 0.05, we will reject the null hypothesis and conclude there
is a difference between the two analysis techniques.
We can illustrate these steps using one of the early statistical datasets: Stu-
dent’s—a pseudonym of statistician William Gosset—sleep data [17]. Ten in-
dividuals were given one of two soporific drugs, also known as sleep aids, and
their hours of sleep were recorded. Then the process was repeated with the other
drug. They wanted to test if the two drugs—Dextro and Laevo—were any dif-
ferent in hours of sleep added. So, the hypotheses for this question are

H0 : μ D = 0
HA : μD = 0

which we will test at the α = 0.1 level. In the sample, it was found that the
average difference—calculated as Dextro minus Laevo—was x̄D = −1.58 with
a sample standard deviation of sD = 1.23. So the test statistic was
−1.58
t= √ = −4.06.
1.23/ 10
Because we are dealing with a two-tailed test, our p-value is 2 × P (t10−1 >
| − 4.06|). Using the R code pt(q=4.06, df=9, lower.tail=FALSE), we get that
this probability is 0.00142, which we double to get the p-value of 0.00284. Since
184 Basic Statistics With R

this p-value is very much below the α = 0.1 threshold, we will reject the null
hypothesis and conclude that there is some difference in the effectiveness of
the two drugs. While the sample sizes are small, a histogram of the differences
makes it appear as they are roughly Normal. (See Fig. 15.3.)

FIGURE 15.3 Histogram of the Sleep dataset differences.

The t.test function is adaptable to all our two-sample tests for means—
paired and unpaired, equal or unequal variances. By including both our samples
and changing a few of the functions options—namely the paired and var.equal
options—we are able to access all of these varied options.

15.4.7 Practice problems

Researchers were interested in whether children of parents who worked in lead-
related factories had higher levels of lead in their blood. To control for different
neighborhood environments, they paired each of the 33 children in the sample
with a corresponding child from their neighborhood whose parent did not work
in the lead-related factory. They found a sample average difference in blood lead
levels (Exposed minus Control) of x̄D = 15.97 and a sample standard deviation
of sD = 15.86 [62,63].
15. Set up the hypotheses for this test.
16. What is the test statistic for this test?
17. What is the p-value for this test?
18. Assuming that α = 0.01, what is your conclusion for this test?
A researcher is interested in the effect of shade versus full sun on the percent-
age of solids in grapefruit. Twenty-five grapefruits that were partially shaded—
one exposed side and one shaded side—were selected and their solids percent-
age measured for each half. The differences were calculated—shaded minus
exposed—and the sample average differences was found to be x̄D = 0.1936 with
standard deviation sD = 0.3066 [63,64]. Histograms of the differences show bi-
modal data.
 Hypothesis tests for two parameters Chapter | 15 185

19. Test if the population average percentage of solids is the same or different
for the shaded and exposed halves of the at the α = 0.05 level.
In 1876, Darwin tested corn plants to see if the height of cross-pollinated
corn plants was different than self-pollinated corn plants—grown in the same
pots under the same conditions. Of his n = 15 observations, he found a sample
average difference in heights—cross-pollinated minus self-pollinated—of x̄D =
2.617 with a sample difference variance of sD2 = 22.26 [83]. Histograms of the

differences show skewed left data.

20. Test if the population average height for cross-pollinated heights was taller
than the population average height of self-pollinated at the α = 0.1 level.

15.5 Conclusion
Comparing two populations is one of the most common and most important
problems in statistics. Whether it be proportions, means, or paired means, there
is a hypothesis test for each of these scenarios. As we saw previously, the pro-
cedures for these scenarios remain the same, with adjustments for the fact that
we now have two sets of data from two separate populations. We will be able to
apply these results to the other form of inference in confidence intervals.
Chapter 16

Confidence intervals for two

parameters
Contents
16.1 Introduction 187 16.3.4 Practice problems 196
16.2 Confidence interval for p1 − p2 187 16.4 Confidence intervals for μD 197
16.2.1 Practice problems 191 16.4.1 Practice problems 198
16.3 Confidence interval for μ1 − μ2 191 16.5 Confidence intervals for
16.3.1 Equal variances 192 μ1 − μ2 , μD , and hypothesis
16.3.2 Unequal variances 193 testing 199
16.3.3 Interpretation and 16.5.1 Practice problems 201
example 195 16.6 Conclusion 201

16.1 Introduction
When we first introduced one-sample hypothesis tests, we commented that these
tests did not represent the only form of statistical inference for one sample.
Confidence intervals, intervals giving us a range of plausible values for the pa-
rameter of interest, also were a valid form of inference that corresponded with
each hypothesis test. This is also the case for our two-sample tests. Just as we
were testing if the two population means differed, we are able to get a range of
plausible values for the difference between the two population proportions, two
population means, or a paired mean. In this chapter, we will derive and interpret
our interval as well as seeing some of the applications of the intervals beyond
providing ranges of plausible values.

16.2 Confidence interval for p1 − p2

In calculating confidence intervals for two proportions, we want a single interval
that gives us information about the relationship between our two population
proportions p1 and p2 . We know how to calculate a confidence interval for each
of these parameters individually, but this does not give us any information about
the relationship between these two intervals. However, we can accomplish our
goals similar to how we tested these two parameters. In our two-sample test for
proportions, rather than test the proportions individually we tested the difference
between proportions p1 − p2 . For our confidence intervals, we will do similarly,
creating our (1 − α)100% confidence interval for p1 − p2 . This will tell us about

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00029-8 187

Copyright © 2022 Elsevier Inc. All rights reserved.
188 Basic Statistics With R

the relationship between p1 and p2 , including the range of plausible values for
that difference.
As before, our goal is to come up with an interval for p1 − p2 of the form
(Lower, Upper) such that
 
P Lower < p1 − p2 < Upper = 1 − α.

To obtain this interval, were going to need a probability distribution—

courtesy of the central limit theorem—and a little bit of algebra. When creating
our confidence interval for p, we based our probability distribution on the stan-
dard Normal distribution—a good choice, since the central limit theorem was
able to connect our data to this distribution. As we are dealing with intervals for
a couple of proportions, the standard Normal again seems like a good place to
start. Let us choose some critical value z∗ so that
 
P − z∗ < N (0, 1) < z∗ = 1 − α.

Let us keep this probability statement in the back of our mind for now. Next,
we need to find some way to connect our data and our parameters p1 and p2 to
the standard Normal distribution. We will do this, as we have for previous inter-
vals and hypothesis tests, through the central limit theorem and the properties of
Normal distributions. According to the central limit theorem, if our sample size
is large enough and we have seen sufficient successes and failures, our two sam-
ple proportions p̂1 and p̂2 will each individually follow a Normal distribution
centered around the true population proportions p1 and p2 .
 
p1 (1 − p1 )
p̂1 ∼ N p1 ,
n1
 
p2 (1 − p2 )
p̂2 ∼ N p2 , .
n2
Now, we are interested in the interval for p1 − p2 . With that in mind, let us
look at what the distribution of p̂1 − p̂2 becomes. If we take the difference in
our sample proportions, the distribution will be a Normal distribution centered
at p1 − p2
 
p1 (1 − p1 ) p2 (1 − p2 )
p̂1 − p̂2 ∼ N p1 − p2 , + .
n1 n2

p1 (1 − p1 ) p2 (1 − p2 )
This implies that the standard error of p̂1 − p̂2 is + .
n1 n2
Knowing the value of this standard error involves knowing the values of p1 and
p2 , which is impossible since both p1 and p2 are parameters and, therefore,
unknown. We need to substitute in an estimate for each of the population pro-
portions in order for us to know the standard errors. These estimates will be the
sample proportions for each sample, substituting p̂1 for p1 and p̂2 for p2 .
 Confidence intervals for two parameters Chapter | 16 189

Now, since the distribution of p̂1 − p̂2 is a Normal distribution, we are able
to convert it to a standard Normal by subtracting off the mean p1 − p2 and
dividing by the standard error. That is,
p̂1 − p̂2 − (p1 − p2 )
 ∼ N (0, 1).
p̂1 (1 − p̂1 ) p̂2 (1 − p̂2 )
+
n1 n2

Now, because this follows a standard Normal distribution, we can substitute

it in for the standard Normal in our initial probability statement. When we do
this, that probability statement becomes

∗ p̂1 − p̂2 − (p1 − p2 ) ∗
P −z <  < z = 1 − α.
p̂1 (1 − p̂1 ) p̂2 (1 − p̂2 )
+
n1 n2

Now, to find our (1 − α)100% confidence interval, we need to take the above
probability and solve for p1 − p2 . Doing so requires a little algebra—omitted
here—and will result a confidence interval with the probability statement that
we desire.

∗ p̂1 − p̂2 − (p1 − p2 ) ∗
P −z <  <z
p̂1 (1 − p̂1 ) p̂2 (1 − p̂2 )
+
n1 n2
 
p̂1 (1 − p̂1 ) p̂2 (1 − p̂2 )
= P p̂1 − p̂2 − z∗ + < p1 − p 2 <
n1 n2

∗ p̂1 (1 − p̂1 ) p̂2 (1 − p̂2 )
p̂1 − p̂2 + z + .
n1 n2

This is our (1 − α)100% confidence interval for p1 − p2 , where we are

(1 − α)100% confident that this interval will cover the true value of p1 − p2 —in
other words, will be right. As before, getting z ∗ such that P − z∗ < N (0, 1) <

z∗ = 1 − α can be difficult,  as it involves two probabilities. We are able to
instead find z∗ such that P N (0, 1) < z∗ ) = 1 − α/2, where α is obtained from
the confidence level of our confidence interval.

p̂1 (1 − p̂1 ) p̂2 (1 − p̂2 )
p̂1 − p̂2 ± z∗ +
n1 n2

where p̂1 and p̂2 are our sample proportions, n1 and n2 are our sample sizes,
and z∗ is our critical value. In practice, we will find the critical value z∗ using
the qnorm function as previously described.
190 Basic Statistics With R

You might notice that this confidence interval formula does not use the
pooled proportion, something that was a key component of our two-sample test
for proportions. This discrepancy is entirely explainable by one of our funda-
mental assumptions of hypothesis testing. In hypothesis testing, we assume that
the null hypothesis is true, which in this case implies that p1 = p2 = p, leading
to the need for our pooled proportion estimate that we use in our hypothesis
test. In confidence intervals, we have no null hypothesis to assume to be true,
and thus there is no common proportion that is estimated by the pooled propor-
tion.
It’s important to note that as this confidence interval relies on the results
of the central limit theorem, we will need the assumptions of the central limit
theorem to hold for our confidence interval to be valid. As we are dealing
with two sample proportions p̂1 and p̂2 , the assumptions will have to hold for
both proportions. This means that we need both our sample sizes to be large
enough—generally n1 ≥ 30 and n2 ≥ 30—and we see enough successes and
failures in both samples—usually at least 10 each.
Let us look at this in practice. In a 2018 survey, the Pew Research group
asked 743 teens and 1058 adults the question, “Do you spend too much time
on your cellphone?” Four hundred and one of the teens and 381 of the adults
answered “Yes” [65]. Say that we wanted to create a 98% confidence interval
for the difference between the true proportions pT een and pAdult .
To fill in the various components of our confidence interval, we need our
sample proportions, our sample sizes, and our critical value. Our sample pro-
401 381
portions for the two groups are p̂T een = = 0.5397 and p̂Adult = =
743 1058
0.3601. Our sample sizes will be nT een = 743 and nAdult = 1058. All that
remains is to find the critical value z∗ . Our α level is α = 0.02 since (1 −
0.02)100%=98%, so we will choose so that P N (0, 1) < z∗ = 1 − 0.02/2 =
0.99. We use the code qnorm(0.99, mean=0, sd=1) to get our critical value
z∗ = 2.326348. So, our 98% CI for pT een − pAdult will be

p̂T een (1 − p̂T een ) p̂Adult (1 − p̂Adult )
p̂T een − p̂Adult ± z∗ +
nT een nAdult

0.5397(1 − 0.5397) 0.3601(1 − 0.3601)
0.5397 − 0.3601 ± 2.326 +
743 1058
= (0.1249, 0.2343).

So we are 98% confident that the interval (0.1249, 0.2343) covers the true
value of pT een − pAdult .
The prop.test function allows us to get our confidence interval for p1 −
p2 . By merely changing our confidence level—the conf.level option in the
function—R will get us the confidence interval we are interested in. Further
details and examples are given in Chapter 17.
 Confidence intervals for two parameters Chapter | 16 191

16.2.1 Practice problems

1. In a 2018 poll, 1750 of 2500 of Americans with a college degree said that
the products and services of major tech companies have had a “more good
than bad” impact on society while 1228 of 2082 of Americans without col-
lege degrees answered similarly. Calculate and interpret the 96% confidence
interval for pCollege − pN oCollege [66].
2. A researcher is interested in comparing the mortality of women suffering
myocardial infarctions based on whether or not they suffered from diabetes.
In a survey, it was found that 70 out of 248 women with diabetes died of their
myocardial infarction while 189 out of 978 women without diabetes died
of their myocardial infarction. Calculate and interpret the 80% confidence
interval for pDiabetes − pN oDiabetes [23].
3. In refugee cases where the claimant was rejected by the refugee board, the
claimant can appeal to a federal court for the overturning of that decision. In
those cases, an independent rater is called in to see if they believe the case
has merit. The case, depending on the Canadian province in which the case
takes place, is argued in either French or English. In 253 English-argued
cases, 86 appeals were granted, while in 131 French-argued cases, 44 appeals
were granted [96]. Calculate and interpret the 95% confidence interval for
pF rench − pEnglish , where p is the true proportion of successful appeals.

16.3 Confidence interval for μ1 − μ2

Similar to our confidence interval for p1 − p2 , our (1 − α)100% confidence
interval for μ1 − μ2 will provide us with a range of plausible values for the dif-
ference between our two population means μ1 and μ2 . In doing so, the interval
gives us an idea of the relationship between our two parameters. As before, for
our (1 − α)100% confidence interval we need to find an interval of the form
(Lower, Upper) so that
 
P Lower < μ1 − μ2 < Upper = 1 − α.

To find this probability, we will need a probability distribution to work with.

But which distribution? Let us start where we begin all our confidence intervals:
the central limit theorem. For means, the theorem states that for large enough
sample sizes or Normal data, our sample mean follows a Normal distribution
σ2
centered around the true population mean μ with variance . This holds for
n
both of our sample means x̄1 and x̄2 , resulting in
   
σ2 σ2
x̄1 ∼ N μ1 , 1 x̄2 ∼ N μ2 , 2 .
n1 n2
Similar to our confidence interval for p1 − p2 , we can combine these two
distributions to get the probability distribution for x̄1 − x̄2 , which will be another
192 Basic Statistics With R

Normal distribution.
 
σ2 σ2
x̄1 − x̄2 ∼ N μ1 − μ2 , 1 + 2 .
n1 n2

We can convert this Normal distribution to a standard Normal by subtracting

off its mean and dividing by its standard deviation, resulting in

x̄1 − x̄2 − (μ1 − μ2 )

 ∼ N (0, 1).
σ12 σ22
+
n1 n2

This result does come with some problems, though, as the denominator re-
quires knowing the true values of the population variances σ12 and σ22 . This is
impossible, as both values are
 population parameters and, therefore, unknown.
σ12 σ22
We need a substitution for + . In fact we know two substitutions, one
n1 n2
each depending on our answer to the question: “Are our population variances
equal?”

16.3.1 Equal variances

We saw this question in hypothesis testing previously, and depending on the
answer, we could use one of two estimates. If the variances were equal—
1 s2
determined by using the rule of thumb that if < 12 < 4 we would determine
4 s2
the variances were plausibly equal—we could pool our sample variances to-
gether to get a pooled sample variance of

(n1 − 1)s12 + (n2 − 1)s22

sp2 = .
n1 + n2 − 2

This pooled sample variance estimates the common population variance and
is then inserted into our standard error for x̄1 − x̄2 . We can do this exact same
procedure in this case, making our standard error for equal variances
  
1 1
s.e.(x̄1 − x̄2 ) = sp
2 + .
n1 n2

We can substitute this in for the standard error based on our unknown popu-
lation variances. However, when we do this, the probability distribution changes.
Instead of dealing with a standard Normal, our focus changes to a t-distribution
 Confidence intervals for two parameters Chapter | 16 193

with ν = n1 + n2 − 2 degrees of freedom.

x̄1 − x̄2 − (μ1 − μ2 )

   ∼ tν .
1 1
2
sp +
n1 n2

We will use this t-distribution as our probability distribution. So, for our
(1 − α)100% confidence interval we will want to choose a critical value t ∗ so
that
 
P − t ∗ < tν < t ∗ = 1 − α.
Now, as we saw above, we have something to connect our parameters μ1 −
μ2 and data to this distribution: our test statistic. So, we can sub this into the
probability statement, resulting in,

∗ x̄1 − x̄2 − (μ1 − μ2 ) ∗
P −t <    < t = 1 − α.
1 1
sp2 +
n1 n2

Now, all we have to do is solve for μ1 − μ2 inside the probability statement

and we will have our (1 − α)100% confidence interval for μ1 − μ2

∗ x̄1 − x̄2 − (μ1 − μ2 ) ∗
P −t <    <t =1−α
1 1
sp2 +
n1 n2
..
.
      
∗ 1 1 ∗ 1 1
P x̄1 − x̄2 − t sp2 + < μ1 − μ2 < x̄1 − x̄2 + t sp2 + .
n1 n2 n1 n2

This gives us our interval with lower and upper bounds so that P Lower <
μ1 − μ2 < Upper = 1 − α. Thus our (1 − α)100% confidence interval for μ1 −
μ2 when variances are equal will be
  
∗ 1 1
x̄1 − x̄2 ± t sp2 + .
n1 n2

16.3.2 Unequal variances

If our variances are not equal, we are unable to combine our information to-
gether into a single pooled estimate of our pooled population variance σ 2 .
194 Basic Statistics With R

However, we can substitute our sample variances s12 and s22 for σ12 and σ22 in
our standard error. Thus, our standard error of x̄1 − x̄2 becomes

s12 s2
s.e.(x̄1 − x̄2 ) = + 2.
n1 n2
We can then substitute this into our probability statement instead of us-
ing the standard error based on the unknown population variances. However,
when we use this standard error in our probability statement, the associated
probability distribution changes from a Normal distribution to approximately a
t-distribution. As with the hypothesis test, this is distribution an approximation,
as an exact answer does not exist. Like in our hypothesis test, we approximate
the degrees of freedom ν using the Satterthwaite approximation, resulting in a
degrees of freedom of
 2 
s1 s22 2
+
n1 n2
ν=  2 2  2 2 .
1 s1 1 s2
+
n1 − 1 n1 n2 − 1 n2
We will use this t-distribution as our probability distribution. So, for our
(1 − α)100% confidence interval we will want to choose a critical value t ∗ so
that
 
P − t ∗ < tν < t ∗ = 1 − α.
Now, as we saw above, we have something to connect our parameters μ1 −
μ2 and data to this distribution: our test statistic. So, we can sub this into the
probability statement, resulting in,

∗ x̄1 − x̄2 − (μ1 − μ2 ) ∗
P −t <  < t = 1 − α.
s12 s22
+
n1 n2
Now, all we have to do is solve for μ1 − μ2 inside the probability statement
and we will have our (1 − α)100% confidence interval for μ1 − μ2

∗ x̄1 − x̄2 − (μ1 − μ2 ) ∗
P −t <  <t =1−α
s12 s22
+
n1 n2
..
.
  
s 2 s 2 2 s22
∗ ∗ s1
P x̄1 − x̄2 − t 1
+ 2
< μ1 − μ2 < x̄1 − x̄2 + t + .
n1 n2 n1 n2
 Confidence intervals for two parameters Chapter | 16 195


This gives us our interval with lower and upper bounds so that P Lower <
μ1 − μ2 < Upper = 1 − α. Thus our (1 − α)100% confidence interval for μ1 −
μ2 will be

s2 s2
x̄1 − x̄2 ± t ∗ 1 + 2 .
n1 n2

16.3.3 Interpretation and example

Our interpretation for this confidence interval is similar to before, regardless of
whether the variances are equal or not. Namely, we are (1 − α)100% confident
that the true value of μ1 − μ2 is covered by the confidence interval calculated
from the formula above.
As usual, the reliance of our confidence intervals on the central limit theorem
means that we need sufficient sample sizes or Normally distributed data for the
theorem to hold and our intervals to be valid. I say samples because—like with
the two-sample test for proportions—we have two samples for which the central
limit theorem has to hold. Generally speaking, we need each of our sample sizes
to be at least 30—n1 ≥ 30 and n2 ≥ 30—or for our samples to follow a Normal
distribution for our intervals to be valid.
Let us see this in practice. Say apple tasters were interested in comparing the
taste of two variaties of apples—product 298 and product 493—by calculating
a 90% confidence interval for μ298 − μ493 [23]. Fifteen apples of each variety
were tasted, with the sample means being x̄298 = 68.333 and x̄493 = 94.067 with
2
sample variances of s2982 = 758.667 and s 2 = 1688.352. As s1 = 758.667 =
493
s22 1688.352
1
0.4494 is between and 4, we will assume that our variances are equal. Thus,
4
we can calculate a pooled sample variance

(15 − 1) × 758.667 + (15 − 1) × 1688.352

sp2 = = 1223.51.
15 + 15 − 2

 All that remains is to find our value of t ∗ , which is chosen so that P tν <
t ∗ = 1 − 0.1/2 = 0.95 where ν = 15 + 15 − 2 = 28, since our variances are
assumed equal. We can now use the R code qt(0.95, df=28) to find our critical
value t ∗ = 1.701. Now, we can plug this and our data into the confidence interval
formula to get
  
1 1
68.333 − 94.067 ± 1.701 1223.51 + = (−47.460, −4.008).
15 15

So, we are 90% confident that (−47.460, −4.008) will cover the true value
of μ298 − μ493 . However, our small sample and slight lack of Normality gives
196 Basic Statistics With R

us reason to be somewhat skeptical of the trustworthiness of the result. (See

Fig. 16.1.)

FIGURE 16.1 Histogram of the apple tasting data.

16.3.4 Practice problems

Say a researcher was interested in looking into if there was any difference in
the population average clump thickness of a tumor for benign and malignant tu-
mors. In 458 benign tumors, researchers found a sample average x̄Benign = 2.96
with a sample standard deviation of sBenign = 1.67. In 241 malignant tumors,
the sample average was x̄Malignant = 7.2 with a sample standard deviation of
sMalignant = 2.43 [10,61].
4. Calculate and interpret the 90% confidence interval for μMalignant −
μBenign .
A researcher is interested in the rate of absenteeism among male and female
students. Out of 80 female students, they had an sample average of x̄F = 15.225
days absent with a variance of sF2 = 253.7968. Of the 66 male students, we saw
a sample average of x̄M = 17.955 days absent with a variance of sM
2 = 276.6902

[10,28].
5. Calculate and interpret the 96% confidence interval for μF − μM .
Say a researcher is interested in the difference in atmospheric ozone levels
during the summer and winter in Leeds. In 578 summer days sampled, the av-
erage ozone level was found to be x̄S = 32.00 parts per billion with a variance
of sS2 = 106.81. Meanwhile, in 532 winter days, the average ozone level was
x̄W = 20.06 with a variance of sW2 = 118.72. [97,98]

6. Calculate and interpret the 92% confidence interval for μS − μW .

 Confidence intervals for two parameters Chapter | 16 197

16.4 Confidence intervals for μD

As we discussed in our hypothesis tests for means, there are times when using
our two-sample t-test for means—or our confidence interval for μ1 − μ2 —does
not sufficiently answer our question. This is when we have paired data, or data
in which our samples are not independent. It may be that our subjects appear in
both samples or each subject has a similar counterpart so as to define a baseline
to compare treatments. Regardless, in these scenarios our sample have some
direct connection that precludes using our two-sample t-test for means and its
confidence interval counterpart.
In these instances, rather than compare the group means, we do inference on
the differences between the paired samples. We find the difference between the
two groups for each subject and then conduct essentially our one-sample t-test
for means on the differenced data. The same will hold for the corresponding
confidence interval, as we are able to take our one-sample t-test for means and
convert it accordingly. In either instance, we will be doing inference on the
population mean of the differenced data μD .
Our (1 − α)100% confidence interval for a single mean μ relies only on the
sample mean x̄, the sample standard deviation s, our sample size n, and a critical
value t ∗ from a t-distribution with n − 1 degrees of freedom.
s
x̄ ± t ∗ √
n

Our goal again is to do inference on a single mean of differenced data μD .

Once we have converted our two samples of data to a single differenced sample,
we are doing the exact same inferential techniques on μD as tests and intervals
for a single mean μ. As such, we can find that our (1 − α)100% confidence
interval for μD will be
sD
x̄D ± t ∗ √ .
n
This is identical to our (1 − α)100% confidence interval for μ with added
subscripts on our parameters and sample data to denote that we are working with
the differenced data. We choose t ∗ such that P tn−1 < t ∗ = 1 − α/2, just as we
did before. Generally, the qt function in R is used to find this particular value.
Finally, our interpretation remains similar, stating that we are (1 − α)100% con-
fident that our calculated interval will cover the true value of μD —or, in other
words, will be right.
In our confidence interval for a single population mean μ, we commented
that we needed the sample size to be n ≥ 30 or for the data to follow a Normal
distribution for the central limit theorem to hold, and thus our confidence inter-
vals to be valid. As our paired confidence interval is essentially a one-sample
confidence interval on the differences, it stands to reason that these same as-
sumptions will have to hold in order for our confidence interval to be valid. That
198 Basic Statistics With R

is, we will again need a sample size of n ≥ 30 paired observations or the dif-
ferences to be Normally distributed in order for the central limit theorem—and
thus our confidence intervals—to hold.
To see this in practice, let us look at an example from the PairedData li-
brary in R. Say that we want to determine if two processes for determining the
percentage of iron in an ore differ in their results [63,67]. To do so, we take
10 samples of iron ore and measure the percentage of iron by method A and
method B. Since there is overlap between the two samples, this is a paired sam-
ple. In the sample, they find that the average difference (A-B) between the two
methods is x̄D = −0.13 with a sample standard deviation of sD = 0.177. Say
we wanted to create a 90% confidence interval for μD . They only thing we are
missing is the critical value t ∗ . To find this, we use the qt(0.95, df=9) function
to get t ∗ = 1.833. So, our 90% Confidence Interval for μD is
0.177
−0.13 ± 1.833 × √ = (−0.027, −0.233)
10
with the interpretation that we are 90% confidence that the true population dif-
ference is covered by the interval (−0.027, −0.233). However, the small sample
size and non-Normal differences may cast some doubts about the validity of the
results. (See Fig. 16.2.)

FIGURE 16.2 Histogram of the iron dataset differenced data.

As with our hypothesis tests, the t.test function is able to give us our con-
fidence intervals for μ across the board. Adjusting similar options that were
mentioned in our two-sample tests for μ and changing the conf.level option al-
lows us to get any (1 − α)100% confidence interval for μ1 − μ2 or μD that is
required.

16.4.1 Practice problems

7. A researcher wants to see if children with parents who work in a lead factory
have higher blood lead levels than children whose parents do not work in a
 Confidence intervals for two parameters Chapter | 16 199

lead factory. The researchers selected 33 children with parents working in

the lead factory, and then paired them with 33 children as controls from their
respective neighborhoods whose parents do not work in a lead factory. They
found that the average difference in blood lead levels (Exposed minus Con-
trol) of x̄D = 15.97 and a sample variance of sD 2 = 251.65. Histograms of the

differences appear skewed right. Calculate and interpret the 99% confidence
interval for μD [62,63].
8. A researcher is interested in the effectiveness of a drug at stabilizing CD4
counts in HIV-positive patients. A decrease in CD4 counts can mark the
onset of full-blown AIDS in a patient. A baseline CD4 counts was measured
for 20 patients and then the CD4 counts were remeasured a year later. They
found a sample average of x̄D = −0.805 with a standard deviation of sD =
0.8017. Histograms of the differences seem Normal. Calculate and interpret
the 90% confidence interval for μD [68–70].
9. Researchers were developing a device that generates electricity from wave
power at sea. In order to keep the device in one location, one of two mooring
methods was used—one of which was much cheaper than the other. Then a
series of simulations in a wave tank was done to test the bending stress on one
particular part of the device, with each of the two mooring methods receiving
the same simulated wave type. In 18 simulations, the average difference in
stress was x̄D = 0.062 with a standard deviation of sD = 0.29. Histograms
of the differences seem Normal. Calculate and interpret the 80% confidence
interval for μD [99,100].

16.5 Confidence intervals for μ1 − μ2 , μD , and hypothesis

testing
Because both confidence intervals and hypothesis testing is predicated on the
central limit theorem, it seems reasonable that there is some connection between
the two inferential techniques. This was seen through our one-sample t-test for
means and corresponding confidence interval, as the test statistic for said test
and the confidence interval are calculated from the same formula and distribu-
tion. We ultimately found that we could conduct a hypothesis test of H0 : μ = μ0
versus HA : μ = μ0 at the α significance level by creating a (1 − α)100% confi-
dence interval for μ. If μ0 was contained in the interval we would fail to reject
the null hypothesis, while if μ0 was not contained in the interval we would reject
the null hypothesis.
This logic extends to our two-sample hypotheses as well. We are able to
conduct a hypothesis test of H0 : μ1 − μ2 = 0 versus HA : μ1 − μ2 = 0 by
creating a (1 − α)100% confidence interval for μ1 − μ2 . If 0 is contained in
the interval, we would fail to reject the null hypothesis, while if 0 is outside the
interval we reject the null hypothesis.
Let us look back at a previous hypothesis test to confirm this, specifically
looking at the chicken weight dataset [10,60]. In this, we were looking at
200 Basic Statistics With R

whether or not two different dietary supplements—soybeans or sunflowers—

resulted in different weight gains for chickens, specifically testing it at the
α = 0.05 level. We ultimately found that we should reject the null hypothe-
sis and conclude that there is a difference between the two supplements. We
could also accomplish this by calculating our 95% confidence interval—or
α = 0.05—for μsoybean − μsunf lower . Our data saw that among the 14 chicks on
soybeans there was an average weight gain of x̄soybean = 246.43 with a variance
2
of ssoybean = 2929.96. Meanwhile, the 12 chicks on sunflowers had an average
weight gain of x̄sunf lower = 328.92 with a variance of ssunf
2
lower = 2384.99.
Due to the ratio of the variances, we assume that they are equal and we can
calculate the pooled variance of
(12 − 1) × 2384.99 + (14 − 1) × 2929.96
sp2 = = 2680.182.
12 + 14 − 2
 
All that remains to find our critical value t ∗ , chosen so that P tν < t ∗ =
1 − 0.05/2 = 0.975. As our variances are assumed equal, the degrees of freedom
will be ν = 14 + 12 − 2 = 24. We can then use the R code qt(0.975, df=24) to
find that t ∗ = 2.064. We can now plug all this into the formula to get our 95%
confidence interval
  
1 1
246.43 − 328.92 ± 2.064 ± 2680.182 + = (−124.526, −40.454).
14 12

Since 0 is not in this interval, we would conclude that we should reject the
null hypothesis at the α = 0.05 level. This matches up with the results from
our hypothesis test, as we would expect to see given the connection between
hypothesis testing and confidence intervals. However, we should note that our
sample size and Normality concerns remain from before.
The exact same procedure applies to the paired test as well. In order to do
a test of H0 : μD = 0 versus HA : μD = 0, we can calculate a (1 − α)100%
confidence interval with matching α levels. If 0 is contained in the interval we
fail to reject the null hypothesis, while if 0 is not contained in the interval we
will reject the null hypothesis.
Again let us confirm this with the results of a previous hypothesis test. Recall
our earlier example involving analyzing breakfast cereal [23]. We previously
tested our hypothesis of H0 : μD = 0 versus HA : μD = 0 at the α = 0.05 and
found that we would reject the null hypothesis. We could alternatively test this
by calculating a 95% confidence interval—again with α = 0.05—for μD . As
x̄ = −0.622, sD = 1.988, and n = 100, all we are missing is our critical value
of t ∗ . Our level of α will be α = 0.05, so we would use the R code qt(0.975,
df=99) to find that t ∗ = 1.9842. Putting all this together we get our 95% confi-
dence interval to be
1.988
−0.622 ± 1.9842 √ = (−1.016, −0.228).
100
 Confidence intervals for two parameters Chapter | 16 201

Since 0 is not in this interval, we would reject the null hypothesis at the
α = 0.05 level, the result that we got earlier when doing the entire hypothesis
testing procedure.

16.5.1 Practice problems

10. A researcher is seeing if two methods for determining the percentage of
fat in meats differ in their results. Twenty cuts of meat were selected,
and then each method was used to determine the percentage of fat. The
95% confidence interval for μD was found to be (−0.345, 0.265), with
the histograms of the differences appearing Normal. What would be the
conclusion of a hypothesis test of H0 : μD = 0 versus HA : μD = 0 at the
α = 0.05 level [63,71]?
11. Two judges were shown a series of 93 routines from an athletics com-
petition and each asked to score the routines on a 1–10 scale. The 90%
confidence interval for μD was (−0.557, −0.045). The histograms for the
differences appear Normal. What conclusion would be drawn of a hypoth-
esis test of H0 : μD = 0 versus HA : μD = 0 at the α = 0.1 level [63,72]?
12. A researcher is interested in comparing two plant growth methods effects
on the total yield (Measured in weight) of the crop. The 95% confidence
interval for μ1 − μ2 was calculated from a sample of n1 = 20 and n2 = 10
to be (0.215, 1.515). The histograms of the datasets appear Normal for
both samples. What would be the conclusion of a hypothesis test of H0 :
μ1 − μ2 = 0 versus HA : μ1 − μ2 = 0 at the α = 0.05 level [73]?

16.6 Conclusion
As single-parameter confidence intervals allow us to get plausible values for the
parameter, two-parameter confidence intervals allow us to understand the plau-
sible differences between two populations. These intervals still follow the same
procedure as single-parameter intervals, with adjustments for the two samples
that occur in this case. Now, these intervals are much more easily calculated
using R, which we will see in the following chapter.
Chapter 17

R tutorial: statistical inference

in R
Contents
17.1 Introduction 203 17.4 Inference for means 209
17.2 Choosing the right test 203 17.4.1 Inference for a single
17.3 Inference for proportions 204 mean 209
17.3.1 Inference for a single 17.4.2 Inference for two means 210
proportion 205 17.4.3 Paired inference for
17.3.2 Inference for two means 212
proportions 207 17.4.4 Practice problems 213
17.3.3 Practice problems 208 17.5 Conclusion 213

17.1 Introduction
As we have seen from the past few chapters, hypothesis tests and confidence
intervals are two statistical techniques that can be done mostly by hand, es-
pecially in small datasets. However, as is often the case when working with
modern datasets, real data often can be quite large and complex, often having
hundreds or thousands of observations. When this is the case, it is impractical,
if not impossible, to perform these inference techniques by hand. Fortunately,
in addition to providing the exploratory data analyses and plots that we previ-
ously discussed, R allows us to perform the whole of statistical inference—not
just probabilities and p-values—for all of the hypothesis tests and confidence
intervals that we discussed.

17.2 Choosing the right test

One important aspect of hypothesis testing is recognizing which test should be
used. As we said, knowing which test to use boils down to three questions, all
dealing with our response and samples. The three questions are:
1. Is our response categorical or quantitative?
2. Are there one or two samples in our data?
3. If there are two samples, are our samples independent—i.e., is there any
overlap or connection between our samples?
Let us take these questions one by one, beginning with whether our response
is categorical or quantitative. Our two tests for proportions—the one-sample and

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00030-4 203

Copyright © 2022 Elsevier Inc. All rights reserved.
204 Basic Statistics With R

two-sample tests for proportions—will deal with categorical responses, while

all three tests for means—the one-sample t-test, two-sample t-test, and paired
t-test—are a result of quantitative responses.
Second, let us look at the number of samples in our data. Clearly, our two-
sample tests—the two-sample test for proportions, two-sample t-test for means,
and paired t-test for means—all involve two samples. Our one-sample tests—
one-sample test for proportions and one-sample t-test for means—involve a
single data sample.
Finally, we have to consider if our samples are independent. Only one of our
two-sample tests are not independent: the paired t-test. In that one test, there
is either overlap between our two samples—such as subjects experiencing two
treatments—or there is some sort of connection between the samples. In general,
we look at the chart in Fig. 17.1 to find our way to the correct test.

FIGURE 17.1 The process of identifying the correct hypothesis test to use.

17.3 Inference for proportions

In discussing hypothesis tests, we learned about two hypothesis tests related
to proportions. The one-sample test for proportions dealt with testing if a sin-
gle population proportions was equal to or differs from a single value. The
two-sample test for proportions dealt with testing if the population proportions
for two groups is the same or is different from some value. Similarly, when
discussing confidence intervals we learned about two confidence intervals in-
 R tutorial: statistical inference in R Chapter | 17 205

volving population proportions. Each of these intervals corresponded to one of

the previously mentioned hypothesis tests. Our (1 − α)100% confidence interval
for p gave us an interval estimate for a single population proportion while the
(1 − α)100% confidence interval for p1 − p2 looked into the plausible values
for p1 − p2 .
These intervals and tests—four in total—represent our options to do statis-
tical inference on population proportions. Fortunately, R makes implementing
any of these inferential methods simple by having them all done by the same
function. The prop.test function allows you to do hypothesis testing and confi-
dence intervals for proportions. The code to execute it depends on whether you
are doing a one or two-sample test and interval. To begin, we will look at our
one-sample test for proportions.

17.3.1 Inference for a single proportion

For a one-sample test for proportions, the prop.test function takes on five ar-
guments: the number of successes in our sample x, the sample size n, the null
hypothesis value p, our alternative hypothesis alternative, and whether or not
to apply a continuity correction correct.
prop.test(x, n, p, alternative, correct)

Two notes on the function. The correct argument is a TRUE/FALSE option

stating whether or not we want to apply a continuity correction to our sample.
In small samples, this is sometimes done to allow the central limit theorem to
hold. However, this adjustment sometimes overcorrects, making the correction
generally undesirable. Thus, we will set this argument to FALSE throughout
our applications. Second, the alternative argument is defined what our alter-
native hypothesis is, with three possible options based on our three alternative
hypotheses.

HA R-code
HA : p < p0 alternative=“less”
HA : p > p0 alternative=“greater”
HA : p = p0 alternative=“two.sided”

So, let us use an example to see how the function works. Let us use the
example we used to walk through the one-sample test for proportions. In this
example, we were asking if the proportion of Sweet Briar College legacy stu-
dents differed from the national average of 14%. To test this, we collected a
sample of 79 students and found that 12 of them were legacy students. In this
case, our hypotheses were

H0 : p = 0.14
HA : p = 0.14.
206 Basic Statistics With R

In order to conduct this test, we need our number of successes—x =

12—our sample size—n = 79—our null value—p = 12—and our alternative
hypothesis—alternative=“two.sided.” Note that if the dataset had been loaded
into R via a csv file we could find the number of successes using the table func-
tion. The code to conduct this test in R would then be
prop.test(x=12, n=79, p=0.14,
alternative="two.sided", correct=FALSE)

R will then output several results, all relevant to our hypothesis test proce-
dure.
1-sample proportions test without continuity correction

data: 12 out of 79, null probability 0.14

X-squared = 0.092897, df = 1, p-value = 0.7605
alternative hypothesis: true p is not equal to 0.14
95 percent confidence interval:
0.08908268 0.24699854
sample estimates:
p
0.1518987

There is a lot of output here to sift through, some of which are intuitive. The
first important output is the X-squared value, which gives us our test statistic
squared. If we had calculated our test statistic using the formula, we are familiar
with by hand, we would either get the positive square root of X-squared—if
p̂ is greater than p0 —or the negative square root of X-squared—if p̂ is less
than p0 . So, for our example, since our p̂ = 0.1519 is √ greater than the null
hypothesis value of 0.14, our test statistic would be + 0.0929 = 0.3048, the
value we got when we did the work by hand. The reason why prop.test reports
the test statistic squared instead of the test statistic is due to the function’s direct
connection to a more general form of our test for proportions—the chi-squared
test for independence—that can be learned in a second course in statistics.
Next, we have the p-value, which is just what it says it is; the p-value for
our hypothesis test based on our test statistic and alternative hypothesis. Here,
our p-value will be 0.7605, which would likely be greater than any reasonable
significance level that we would choose. Again, this is the result we got when
going through our test step-by-step.
The function returns a few other results—the sample proportion p̂, the al-
ternative hypothesis—but one is of particular interest. We see that the function
outputs a 95% confidence interval for p based on our results. This is helpful
when we want a 95% confidence interval, but what about when we want a 90%
or 99% confidence interval? It turns out that there is one additional option in the
prop.test function that will give us any confidence interval we want: the confi-
dence level conf.level.
 R tutorial: statistical inference in R Chapter | 17 207

prop.test(x, n, p, alternative, correct, conf.level)

In the function, conf.level is always set by default to 0.95, thus returning a

95% confidence interval unless we change it. If we instead want a 90% Confi-
dence Interval, we set conf.level=0.9. For 99%, we use conf.level=0.99. Say in
our previous example we wanted a 99% confidence interval for p. We use the ex-
act same code and add conf.level=0.99 to find that the 99% confidence interval
for p. The result—found in the function output—is (0.07538655, 0.28235191).
One important rule to note with these confidence intervals is that for prop.test
to return the correct confidence interval in the one-sample case, we need to
set the alternative option to “two.sided” If not, the function output will return
an incorrect interval based on the upper or lower (1 − α)100% of plausible
values—depending on what the alternative argument is set to.

17.3.2 Inference for two proportions

So that takes care of our one-sample test and confidence interval. What about
the two-sample cases? We earlier said that the prop.test function could be used
for two-sample tests and intervals, and ultimately it turns out that the code
needed for our two-sample test is nearly identical. The arguments in this case
are the successes in both samples x, the sample sizes n, the alternative hypoth-
esis alternative, the continuity correction correct, and our confidence level
conf.level for our confidence interval—if so desired.

prop.test(x, n, alternative, correct, conf.level)

The important change here comes in our successes x and sample sizes n.
Previously, they were single values as we only were concerned with a single
sample. Here—as we have two groups to compare—both x and n will be vectors
of values. It is important to ensure that the ordering of successes in x matches
up with the ordering of sample sizes in n. Additionally, the alternative hypoth-
esis and alternative option will be affected by this ordering. For example, the
“greater” option implies HA : p1 > p2 , where group one—with population pro-
portion p1 —will be defined by the successes and sample size entered first into
x and n. So we must make sure that the ordering of x and n matches up with
what our alternative argument implies.
For example, let us look back at the example we walked through while learn-
ing the steps of our two-sample test for proportions [55]. In this example, the
Pew Research Group asked 667 Millenials and 558 Generation Xers if they
thought it was essential for the United States to be a world leader in space explo-
ration. 467 Millenials and 407 Generation Xers answered “Yes” to the question.
We wanted ultimately to test if there was any difference between the population
proportions for those two groups. The hypotheses in this case would be

H0 : pMillenial − pGenX = 0
208 Basic Statistics With R

HA : pMillenial − pGenX = 0

leading to .” Here, our successes would be x = c(467, 407), and our sample
sizes would be n = c(667, 558). All put together, the code would be
prop.test(x=c(467, 407), n=c(667, 558),
alternative="two.sided", correct=FALSE)

The resulting output for this code is given below. Again, our test statistic will
be the positive or negative square root of X-squared: positive if p̂1 > p̂2 and
negative if p̂1 < p̂2 . Since p̂1 < √
p̂2 in this case, our test statistic is the negative
square root of X-squared, so − 1.2707 = −1.127. The p-value, determined
by our test statistic and alternative hypothesis, is 0.2596, likely higher than our
significance level.
2-sample test for equality of proportions without continuity
correction

data: c(467, 407) out of c(667, 558)

X-squared = 1.2707, df = 1, p-value = 0.2596
alternative hypothesis: two.sided
95 percent confidence interval:
-0.07991559 0.02143408
sample estimates:
prop 1 prop 2
0.7001499 0.7293907

Just as in our one-sample case, we can find our (1 − α)100% confidence

interval for p1 − p2 by changing our conf.level options. Also as before, for
prop.test to return the correct confidence interval, we need to set our alternative
option to “two.sided” So, for our example, to find the 90% confidence interval
for pMillenial − pX , all we need to add is conf.level=0.9 to our code to get the
90% confidence interval of (−0.07176842, 0.01328690).

17.3.3 Practice problems

1. A survey wanted to look at if men and women are left-handed at different
rates. They found 7 of 117 women and 10 of 118 men were left-handed. Run
a hypothesis test with HA : pF − pM = 0 at the α = 0.1 level and an 87%
confidence interval for pF − pM [10].
2. In a 2018 Pew Research Group survey of 1500 Americans, 705 held an un-
favorable view of China. Run a hypothesis test with HA : p < 0.5 at the
α = 0.05 level and a 93% confidence interval for p [74].
3. A 1987 survey found that out of 212 wives who attended college, 171 of
their paired husbands also attended college. Meanwhile, of 541 wives that
did not attend college, 124 of their paired husbands also attended college.
 R tutorial: statistical inference in R Chapter | 17 209

Run a hypothesis test for with HA : pW if eCollege − pW if eN one > 0, where p

is the probability that the husband attended college at the α = 0.05 level. Ad-
ditionally, calculate a 98% confidence interval for pW if eCollege − pW if eN one
[48,104].
4. The same survey found that our of 753 married women, 428 were partic-
ipants in the labor force. Run a hypothesis test with HA : p > 0.5 at the
α = 0.01 level and a 84% confidence interval for p [48,104].

17.4 Inference for means

In addition to our hypothesis tests for proportions, we discussed the various in-
ferential techniques for means as well. These consisted of the one-sample t-test
for means, the two-sample t-test for means, and the paired t-test for means—in
addition to their corresponding confidence intervals. Similar to our inference for
proportions, all of these various options are able to be accomplished with one
single function: t.test. To discuss the variations of code in order to do all these
various tests and intervals, let us go through how the function works for each of
our tests.

17.4.1 Inference for a single mean

For our one-sample t-test for means, the t.test function requires three arguments:
the data sample x, the null value mu, and the alternative hypothesis alternative.
t.test(x, mu, alternative)

Our data set x is a vector—either taken from a data frame or manually

entered—containing our data values. The null value mu is the numeric value
found in the null hypothesis—the equivalent of p in the prop.test function.
Finally, our alternative argument is defined by the alternative hypothesis iden-
tically as in our prop.test function.

HA R-code
HA : μ < μ0 alternative=“less”
HA : μ > μ0 alternative=“greater”
HA : μ = μ0 alternative=“two.sided”

To illustrate this, let us look at the example we used to illustrate the one-
sample test for means. We tested if a video slot machine’s average payout was
equal to or less than the theoretical payout of $13.13. This data can be found
in the vlt dataset in the DAAG library [23,44]. The payout amount for the 345
games played are found in the prize variable in the dataset. So, to do our hy-
pothesis test we will use the code
t.test(vlt$prize, mu=13.13,
alternative="less")
210 Basic Statistics With R

And R will return the following output. Just like the prop.test function, the
t.test output covers all the bases for our hypothesis test. The value of t in the
output is the value of our test statistic, using the formula we saw in previous
chapters. The df value is the degrees of freedom for this test, calculated as
before using df = n − 1 where n is our sample size. The p-value is of course our
p-value, calculated with respect to our test statistic and alternative hypothesis.
As we can see, all of these match up with our results that we saw before.

One Sample t-test

data: vlt$prize
t = -90.718, df = 344, p-value < 2.2e-16
alternative hypothesis: true mean is less than 13.13
95 percent confidence interval:
-Inf 0.8989483
sample estimates:
mean of x
0.6724638

Just like in the prop.test function, we are able to create our (1 − α)100%
confidence interval—in this case for μ—using the same function with which we
do hypothesis testing. By default, the t.test function returns a 95% confidence
interval. However, we can change the desired confidence level for our interval
using the same conf.level argument that we saw in the prop.test function.

t.test(x, mu=0, alternative, conf.level)

Just as in the prop.test function, we need to set our alternative hypothesis

to alternative = “two.sided” in order to get the confidence interval using the
formula we are familiar with. So, say that we wanted the 99% confidence inter-
val for the true average payout of the machine. Our 99% confidence interval for
μ of (0.3167728, 1.0281548) can be generated using the following code:

t.test(vlt$prize, mu=13.13, alternative="two.sided",

conf.level=0.99)

17.4.2 Inference for two means

In our two-sample t-test for means, our goal is see if the population means for
two groups are the same or differ in some way. We are able to use the t.test
function to accomplish this with slight changes to our existing code. The argu-
ments for the t.test function to accomplish this two-sample test are the samples
from our two populations x and y and our alternative hypothesis alternative.
All of these arguments are familiar to us, having seen them in our one-sample
test for means. The only differences are the addition of our second sample y, the
removal of our null value—as it is 0 in our two-sample test for means—and the
 R tutorial: statistical inference in R Chapter | 17 211

option var.equal which is set to either TRUE or FALSE depending on whether

our rule of thumb says our variances are equal.
t.test(x, y, alternative, var.equal)

For example, let us look at our conformity example from the previous chap-
ter. The data is stored in the Moore dataset in the carData library [48,59]. We
were interested in testing if partners of a higher status resulted in equal ver-
sus higher conformity than partners of lower status, or H0 : μH igh − μLow = 0
s2
versus HA : μH igh − μLow > 0. Based on the ratio of our variances— 12 =
s2
27.855
= 1.459—we can say that our variances are plausibly equal. Thus, we
19.087
can use the t.test function to test this using the following code:
t.test(Moore$conformity[Moore$partner.status=="high"],
Moore$conformity[Moore$partner.status=="low"],
alternative="greater", var.equal=TRUE)

The function gives us the same set of output as our one-sample t-test for
means with our test statistic and p-value among other items. If we look at the
output of the function, we see that they all match up with the results we previ-
ously calculated by hand.
Two Sample t-test

data: Moore$conformity[Moore$partner.status == "high"] and

Moore$conformity[Moore$partner.status == "low"]
t = 2.957, df = 43, p-value = 0.002515
alternative hypothesis: true difference in means is greater than 0
95 percent confidence interval:
1.839379 Inf
sample estimates:
mean of x mean of y
14.217391 9.954545

Additionally, the t.test function returns a 95% confidence interval—in this

case for μ1 − μ2 —by default. We can change this by adding and adjust-
ing the conf.level argument in the function—keeping in mind that we need
alternative = “two.sided” to get the interval that we are familiar with. So—
from of conformity example—say we wanted to calculate our 98% confidence
interval for μH igh − μLow we would use the following code to get that the con-
fidence interval is (0.7795, 7.7462):
t.test(Moore$conformity[Moore$partner.status=="high"],
Moore$conformity[Moore$partner.status=="low"],
alternative="two.sided", var.equal=TRUE, conf.level=0.98)
212 Basic Statistics With R

17.4.3 Paired inference for means

Our final category of inference is paired inference for means. This inference is
used when we have paired data, or data where our samples are not independent.
That is, our samples have overlap or affect each other in some way. We still want
to compare the population means for two groups, but we need control for some
factor specifically through paired data.
Our t.test function handles this scenario as well. We again include our two
samples x and y as well as our alternative hypothesis option alternative. How-
ever, to ensure that the function does not think that we are conducting a two-
sample t-test for means, we need to tell the function that we are using paired
data. We do this through the paired argument in the function, which takes on
the value TRUE for paired data and FALSE for nonpaired data. So, for the paired
test for means, we set paired = TRUE.

t.test(x, y, alternative, paired=TRUE)

It is important to note that when entering our x and y samples for the paired
t-test for means we need our samples to be the same length and match up ac-
cordingly. Otherwise, R will return an error if our samples are different lengths
or incorrect results if our samples don’t match up correctly. Let us illustrate this
through the example we saw when learning about this test.
Let us see this applied to Student’s sleep data [17]. This is stored in the sleep
dataset in the base package of R. We were interested in testing if there was any
difference between the two sleep aids, or H0 : μD = 0 versus HA : μD = 0. The
code to test this will be

t.test(x=sleep$extra[sleep$group==1],y=sleep$extra[sleep$group==2],
alternative="two.sided", paired=TRUE)

The function returns the same results that we expect to see, among them
the test statistic t, our degrees of freedom df , and the p-value of the test. As
we can see, these results match up with the results that we calculated by hand
previously.

Paired t-test

data: sleep$extra[sleep$group == 1] and

sleep$extra[sleep$group == 2]
t = -4.0621, df = 9, p-value = 0.002833
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
-2.4598858 -0.7001142
sample estimates:
mean of the differences
-1.58
 R tutorial: statistical inference in R Chapter | 17 213

In addition to our hypothesis test, the t.test function provides us the confi-
dence interval for μD as well. By default a 95% confidence interval, we can
change to our specified confidence level by adding the conf.level option to our
function. So, if we wanted the 90% confidence interval for μD , we would use
the following code to find that the interval is (−2.2930053, −0.8669947):

t.test(x=sleep$extra[sleep$group==1],y=sleep$extra[sleep$group==2],
alternative="two.sided", paired=TRUE, conf.level=0.9)

17.4.4 Practice problems

5. Using the DDT dataset in the MASS library, test if the level of DDT is greater
than 3 ppm at the α = 0.01 level and create a 99% confidence interval [10,
49].
6. The monica dataset in the DAAG library looks at the mortality outcomes of
individuals experiencing heart attacks, along with certain variables about the
subjects. Test to see if the age at onset of cardiovascular disease (age) is
lower for men than women (sex) at the α = 0.05 level. Also, create a 90%
confidence interval [23].
7. The cd4 dataset in the boot library looks at the CD4 counts of HIV posi-
tive patients initially on admission to a clinical trial (baseline) and after a
year (oneyear) on an experimental drug. Test if the population mean CD4
level decreases after a year for patients taking the drug in the clinical trial
compared to the initial measurement. Also, create a 60% confidence interval
[68,69].
8. The Guyer dataset in the carData library looks at the cooperation of a group
(cooperation) based on whether or not the group’s decision was made by
public or anonymous vote (condition). Test whether or not the population
mean of cooperative decisions is the same versus different for the public and
anonymous voting systems. Additionally, create a 97% confidence interval
for μAnon − μP ub [48,105].
9. The pair65 dataset in the DAAG library looks at the stretchiness of elastic
bands—measured in length of stretch when a weight was hung from the
band—depending on whether they were warmed. The bands were paired
based on similar initial elasticity. The variable heated represents the nine
bands that were placed in hot water, while ambient is the nonwarmed band.
Test whether or not the population mean stretchiness was the same or differ-
ent for the warmed group versus nonwarmed group [23].

17.5 Conclusion
While all forms of statistical inference can be done by hand with simple formu-
las and a little code, it can be a tedious process. Instead of going through this
process R provides two functions that allow us to calculate confidence intervals
214 Basic Statistics With R

and do hypothesis tests of all varieties. However, it is important to recognize

and know how we can apply each of the functions appropriately. Once that is
recognized, the work of statistical inference becomes much easier.
Chapter 18

Inference for two quantitative

variables
Contents
18.1 Introduction 215 18.2.7 Practice problems 221
18.2 Test for correlations 216 18.3 Confidence intervals for
18.2.1 State hypotheses 216 correlations 222
18.2.2 Set significance level 217 18.4 Test for correlations in R 223
18.2.3 Collect and summarize 18.5 Confidence intervals for
data 217 correlations 224
18.2.4 Calculate the test statistic 217 18.6 Practice problems 224
18.2.5 Calculate p-values 218 18.7 Conclusion 225
18.2.6 Conclusion 219

18.1 Introduction
When we first introduced the definition of variables, we said that we would focus
on two main types of variables: categorical and quantitative. We then worked
our way through the process of statistics, doing exploratory analyses and then
conducting statistical inference. Our previous inferential techniques focused on
one of two situations: a quantitative response with a categorical predictor and a
categorical response with a categorical predictor. In both instances, we have a
categorical predictor which defined our one or two groups that we intended to
compare.
However, not all problems in statistics are limited to this idea of a categorical
predictor. Consider the following situation: you are interested in whether or not
the speed a vehicle was traveling at the time of braking affects the stopping
distance of the vehicle. In this case, your response is the stopping distance of
the vehicle which is quantitative. More importantly, your predictor variable is
the speed of the vehicle at the time of braking, again a quantitative variable.
To begin investigating this question, we start with collecting data, which
we can find in the cars dataset in R [31]. We then could do some exploratory
analyses, such as creating a scatterplot. Through this, we could see that there
appears to be a positive linear relationship between the speed of the vehicle and
its stopping distance that is moderately strong with no outliers. (See Fig. 18.1.)
Even with this scatterplot, our most powerful tool to describe the associa-
tion between two quantitative variables is the correlation. As a reminder, the

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00031-6 215

Copyright © 2022 Elsevier Inc. All rights reserved.
216 Basic Statistics With R

FIGURE 18.1 A scatterplot of driving speed and stopping distance of a vehicle.

correlation is a numeric summary of a linear relationship between two variables

that describes the direction and strength of the relationship. We can find that the
correlation between the speed and stopping distance is r = 0.8069.
What does this result mean? Yes, we can note that this implies a strong, pos-
itive relationship between our variables, but what can we conclude definitively
about our variables? Ultimately, not much. As we have discussed, any statistic
calculated from our random sample will thus be a random variable and will have
a probability distribution associated with it, referred to as the sampling distribu-
tion. The sample correlation is no different, as it does indeed have a probability
distribution. Because of that, we need to assess our sample statistic in light of
its variability.
This of course is the basis of statistical inference, both for hypothesis test-
ing and confidence intervals. In order to make definitive conclusions about the
relationship between our two variables, we need to run hypothesis tests and
confidence intervals for our sample correlation. This is our focus for the chap-
ter, and we will illustrate these techniques using the question speed and stopping
distance.

18.2 Test for correlations

Our test for correlations focuses on the question if two variables are correlated.
While the correlation is a more complicated statistic to calculate, many of the
procedures and mindsets that we use for our test for correlations are identical to
the previously learned procedures seen in our tests for proportions and means.
We begin our test for correlations where we begin all our hypothesis tests...

18.2.1 State hypotheses

In our test for correlations, we want to know if our two variables are correlated
with each other. The status quo in this case is assuming that these two variables
 Inference for two quantitative variables Chapter | 18 217

are uncorrelated, which implies that the true value of our population correlation
is ρ = 0. Thus, our null hypothesis will be

H0 : ρ = 0.

Our alternative hypothesis is one of the three forms that we have seen be-
fore: the population correlation is greater than 0, less than 0, or just not equal
to 0. This leads to the three possible alternative hypotheses—with the not equal
option being the most common option

HA : ρ < 0 HA : ρ > 0 HA : ρ = 0.

In our example dealing with vehicle speed and stopping distance say we are
interested in testing if the variables are correlated versus not correlated. With
this in mind, our hypotheses would be

H0 : ρ = 0
HA : ρ = 0.

18.2.2 Set significance level

Our considerations when setting our significance level α have not changed.
These remain our personal comfort with Type I errors as well as convention.
As before, our default significance level of α = 0.05 remains the same.
We will use the baseline significance level of α = 0.05 for our speed-
stopping distance example.

18.2.3 Collect and summarize data

For the test of correlations, we ultimately need two data summaries. We will of
course need our sample correlation r to create our test statistic. Additionally,
we will need the sample size n as we know the sample size affects the standard
error of our sample statistics—including our sample correlation.
For our example, we previously stated that the correlation between speed
and stopping distance is r = 0.8069. This comes with a sample size of n = 50.

18.2.4 Calculate the test statistic

We begin the process of calculating our test statistic for the test for correlations
with the general formula for the test statistic.
Sample Statistic – Null Hypothesis Value
t= .
Standard Error
Now in our previous tests we were able to fill in the various portions of
this formula through our knowledge of our sample or the central limit theorem.
218 Basic Statistics With R

However, for the sample correlation the central limit theorem does not hold.
Does this mean that we have no test statistic that we can use? Not at all. It turns
out that there exists a test statistic that directly relates our sample correlation and
sample size to a specific—and familiar—probability distribution. Specifically,
the test statistic for our test for correlations is
r
t= .
1 − r2
n−2

In our speed and stopping distance example, our test statistic would be

0.8069
t= = 9.464.
1 − 0.80692
50 − 2

18.2.5 Calculate p-values

Just as before, we need to know the distribution that our test statistic follows
in order to calculate our p-values. It turns out that our test statistic follows the
previously-seen t-distribution, specifically with n − 2 degrees of freedom where
n is our sample size. However, just like with every other instance from our pre-
vious hypothesis tests, there are specific conditions that must hold for the test
statistic to follow a distribution. This is the case for this test as well.
For our test statistic in the test for correlations, there assumption are two as-
sumptions that must hold. One we have discussed before, as it is the assumption
that must hold for our correlation to be interpretable. Namely, the relationship
between our two variables must be linear. The new additional assumption that
must hold is that our predictor and response must be normally distributed. If this
is true and the null hypothesis is true—which we always assume is true—our
test statistic will follow a t-distribution with n − 2 degrees of freedom. There
are formal tests for a variable being normally distributed, but we will not dig
into them in this book. For now, we will check to see if the histograms of our
predictor and response are roughly normally distributed. It is important to note
that even if the data are not normally distributed the test statistic given above
will still follow a t-distribution, albeit with lower degrees of freedom—making
it more difficult to reject the null hypothesis. (See Fig. 18.2.)
We know from the earlier scatterplot that the relationship between speed and
stopping appears to be linear. However, looking at the histograms of the driv-
ing speed and stopping distance they are not completely normal distributions.
Again, we can continue with the analysis, but we need to acknowledge that the
results could have potential flaws.
The other thing we need to know in order to calculate our p-values is our
alternative hypothesis. Depending on the alternative hypothesis, we will have
 Inference for two quantitative variables Chapter | 18 219

FIGURE 18.2 Histograms of driving speed and stopping distance of a vehicle.

one of three p-values, with the rationale for each of the p-values being identical
to previous arguments. As such, we will omit the explanation here. Under each
of the three alternative hypotheses, our p-values will be

HA P-value
 
HA : ρ < 0 P tn−2 ≤ t
 
HA : ρ > 0 P tn−2 ≥ t
 
HA : ρ = 0 2 × P tn−2 ≥ |t|

We of course can calculate this in R using the pt function. For our vehicle
speed and stopping distance example, we had the alternative hypothesis that
HA : ρ = 0 with a test statistic of t = 9.464 and n = 50. Thus our p-value will be
2 × P (t50−2 > |9.464|). We use the R code pt(9.464, df=48, lower.tail=FALSE)
to get our p-value of 7.45 × 10−13 .

18.2.6 Conclusion
Our conclusion step has not changed, as we again reject the null hypothesis if
the p-value is less than our significance level α, while failing to reject otherwise.
In our speed and stopping distance example, this means that since our p-value
is 7.45 × 10−13 which is less than α = 0.05, we reject the null hypothesis and
conclude that the correlation between the vehicle’s speed and its stopping dis-
tance is not zero. However, as noted earlier, the assumptions required for this
test are not met. Therefore, it is possible that the conclusions of this test may be
suspect.
It is important to note what our conclusion is. We do not make any claim
about our variables being associated, independent, or anything along those lines.
Our hypothesis test—and, therefore, our conclusion—is about our parameter ρ,
the population correlation. This is the only claim we can rule on based on our
test; the claim that was stated in our hypotheses.
220 Basic Statistics With R

In practice, our test for correlations are done using the cor.test function in R.
This function takes in three key arguments, the two variables used to calculate
our correlation x and y, as well as the alternative hypothesis stated in alternative.
The general form of this function is

cor.test(x,y,alternative)

We will discuss the details of this function in a section later in this chapter.
For now, let us see all these steps in a single run. Say we were interested in if
the heights of parents are correlated with children’s height. In order to test this,
we set up the following hypotheses to be tested at the α = 0.05 level:

H0 : ρ = 0
HA : ρ = 0.

Data exists on this in the Galton dataset in the HistData library [75,76]. The
two variables in the dataset are the average height of the two parents parent and
the child’s height child. Looking at a scatterplot of this data, we can see that
there appears to be a linear relationship between our two variables. Also, we
should note that the data values seem to be constrained to particular values. (See
Fig. 18.3.)

FIGURE 18.3 Scatterplot of parent average height and child height.

We then find that in the sample of n = 928 parent-child trios, the sample
correlation is r = 0.4588. This results in a test statistic of
0.4588
t= = 15.713.
1 − 0.45882
928 − 2

Now, based on our data and alternative hypothesis, our p-value will be
2 × P (t928−2 > |15.713|). We calculate this through the R code pt(15.713,926,
 Inference for two quantitative variables Chapter | 18 221

lower.tail=FALSE), finding a value of 8.46 × 10−50 . In checking the assump-

tions, we find that the data are sufficiently Normal to meet our assumptions. (See
Fig. 18.4.)

FIGURE 18.4 Histogram of parent average height and child height.

Since our p-value is decidedly less than our significance level of α = 0.05,
we would reject the null hypothesis and conclude that the correlation between
the average height of two parents and their child’s height will be nonzero.

18.2.7 Practice problems

1. Say that a researcher is interested in the relationship between body mass
index and white blood cell count in athletes [77]. They find in a sample of
n = 202 athletes that there is a sample correlation of r = 0.177, where there
appears to be a linear relationship between BMI and white blood cell count
and the variables appear to be normally distributed. What would be the result
of the test for correlations with the alternative hypothesis of HA : ρ = 0 at
the α = 0.05 level?
2. Say that a researcher is interested in the relationship between a newborn’s
weight and the weight of its mother [33]. In a sample of n = 189, the re-
searcher finds a sample correlation of r = 0.1857, with a linear relationship
between the variables and both appear normally distributed. What would
be the result of a test for correlations with the alternative hypothesis of
HA : ρ = 0 at the α = 0.01 level?
3. Statistician William Gosset looked into the relationship between height
(height) and finger length (finger), specifically looking at the distribution
of the sample correlation. In a sample size of 3000, they found a sample cor-
relation between height and finger length of r = 0.6557, with both variables
appearing normally distributed. Conduct a hypothesis test of correlations
with HA : ρ = 0 at the α = 0.05 level [106].
222 Basic Statistics With R

18.3 Confidence intervals for correlations

While hypothesis testing gives us a powerful tool to test claims about our pop-
ulation correlation, in several instances we want to know the range of plausible
values for this parameter. In these instances, finding the confidence interval for
the population correlation ρ will answer this questions.
Let us recall our goal for confidence intervals for the population correlation.
A (1 − α)100% confidence interval for our population correlation ρ should be
constructed such that the interval is “right” (1 − α)100% of the time. We define
an interval as being “right” when the interval covers the true value of ρ. That
is, we want to create an interval for ρ of the form (Lower, Upper) such that
P (Lower < ρ < Upper) = 1 − α.
In constructing our confidence interval for ρ, we take a similar route to an
answer. However, the math that we use is a little more difficult. With this in
mind, I will briefly go through the concepts seen in constructing this interval
without requiring you to compute this interval by hand. We will then mention
the R code used to calculate these intervals, code that uses the same format
discussed below.
In order to calculate our confidence interval so that it has an attached prob-
ability statement, we will need to start with a probability distribution. For a
(1 − α)100% confidence interval for ρ, let us start with the standard Normal
distribution and choose a critical value z∗ so that
 
P − z∗ < N (0, 1) < z∗ = 1 − α.

With this in mind, we need a way to relate our data—particularly our sample
correlation—and our population correlation ρ to the standard Normal distribu-
tion. It turns out that R.A. Fisher, the same statistician at the lunch with the lady
tasting tea, found that a transformation of the sample correlation approximately
will follow the Normal distribution. Specifically,
     
1 1+r 1 1+ρ 1
ln ∼N ln ,
2 1−r 2 1−ρ n−3

where r is our sample correlation, ρ is the population correlation, n is our sam-

ple size, and ln is the natural log—a mathematical function. This is the case
regardless of sample size, sample correlation, or population correlation. How-
ever, I will note that the approximation improves with larger sample sizes and
correlations closer to 0. With this relationship, we can standardize this function
of r—by subtracting off the mean and dividing by the standard deviation—we
will get a standard Normal distribution. Substituting this into our probability
statement above, we get
  1+r  1  1+r  
2 ln 1−r − 2 ln 1−r
1
∗
P −z < √ < z∗ = 1 − α.
1/ n − 3
 Inference for two quantitative variables Chapter | 18 223

We would then solve this equation for ρ, where we would get an inter-
val of the form P (Lower < ρ < Upper) = 1 − α. Our interpretation of our
(1 − α)100% confidence interval for ρ remains similar to our previous intervals.
Namely, we are (1 − α)100% confident that our calculated interval will cover the
true value of ρ. We should note that for this interval to be valid, the relationship
between our two variables needs to be linear. There is no Normality assumption
necessary.
As can be seen from our equation above, doing this math would be tedious
to do. As such, we will let R do the work for us. The cor.test function can
create a confidence interval of this form, with only one added argument to the
function: the desired confidence level defined in conf.level. Let us take a look at
this function in some detail.

18.4 Test for correlations in R

As we mentioned, all of our inference needs for the population correlation in
this book are met through the cor.test function in R. Through this function, we
can get both our hypothesis tests and confidence intervals described above. We
will first focus on the hypothesis test for correlations and then our confidence
interval. Both of these will use the example of the relationship between a child’s
height and the average height of their parents.
The cor.test function takes in three key arguments to do our hypothesis test:
our two variables x and y, as well as our alternative hypothesis alternative.

cor.test(x, y, alternative)

Our two variables x and y are entered as vectors that we define manually or
take from a data frame. Our alternative argument is defined by the alternative
hypothesis we are interested in, exactly like in our prop.test and t.test functions.

HA R-code
hA : ρ < 0 alternative=“less”
HA : ρ > 0 alternative=“greater”
HA : ρ = 0 alternative=“two.sided”

So let us see this function at work with the Galton dataset—found in the
HistData library. As discussed, this looks at the relationship between parent
height and child height. Assuming, as before, that we are testing H0 : ρ = 0
versus HA : ρ = 0 at the significance level α = 0.05, our code would be

cor.test(x=Galton$parent, y=Galton$child,
alternative="two.sided")

R returns the following output below. cor.test includes all the output we have
discussed above. We see our test statistic given in t matching the value above.
224 Basic Statistics With R

The degrees of freedom for our t-distribution are given in df and our p-value
is also provided. The only thing missing to definitively draw a conclusion is
checking our assumptions for the test, which we can do by creating a scatterplot
of our variable—plot(Galton$parent,Galton$child)—and histograms of the two
variables—hist(Galton$parent) and hist(Galton$child). These graphs are seen
in examples above, so I will not repeat them here, but they show that our as-
sumptions are reasonably satisfied.
Pearson’s product-moment correlation

data: Galton$parent and Galton$child

t = 15.711, df = 926, p-value < 2.2e-16
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
0.4064067 0.5081153
sample estimates:
cor
0.4587624

Thus, since our p-value is decidedly less than α = 0.05, we reject the null
hypothesis and conclude that the correlation between average parent height and
child height is not equal to 0.

18.5 Confidence intervals for correlations

As discussed earlier, the confidence interval for the population correlation ρ
is a little more difficult to compute by hand. As such, we will use the cor.test
function to calculate these intervals. Similar to other instances, the only change
required to our cor.test function to create our desired confidence levels is setting
the confidence level, which we can do using the conf.level argument.
cor.test(x, y, alternative)

As with our other instances of the phenomenon, we need to set our

alternative = “two.sided” to get the desired confidence interval. So, if we
were to calculate the 90% confidence interval for the population correlation ρ
between parent average height and child height, we would use the following
code to get the interval (0.4150196, 0.5003891). Thus we are 90% confident
that (0.4150196, 0.5003891) covers the true value of ρ.
cor.test(x=Galton$parent, y=Galton$child,
alternative="two.sided", conf.level=0.9)

18.6 Practice problems

Say you are interested in the relationship between stock market trading volume
(Volume) and percentage return (Today). This data can be found in the Weekly
dataset in the ISLR package in R [78,79].
 Inference for two quantitative variables Chapter | 18 225

4. Run the test for correlations with the alternative hypothesis HA : ρ = 0 at the
α = 0.1 level.
5. Find and interpret the 90% confidence interval for ρ.
Say a researcher wants to investigate the relationship between academic staff
pay at universities (acpay) and the number of academic grants the university
receives (resgr). Data to investigate this question can be found in the University
dataset in the Ecdat package in R [12,80].
6. Run the test for correlations with hypotheses H0 : ρ = 0 versus HA : ρ > 0
at the α = 0.05 level.
7. Find and interpret the 99% confidence interval for ρ.
The Caschool dataset in the Ecdat library looks at school district perfor-
mance on a standardized test best on various factors [12,107]. Say a researcher
wants to test the correlation between test score (testscr) and expenditure per
student (expnstu).
8. Run the test for correlations with the hypotheses H0 : ρ = 0 versus HA : ρ >
0 at the α = 0.01 level.
9. Find and interpret the 91% confidence interval for ρ.

18.7 Conclusion
Our test for correlations and confidence interval for correlations give us an
important tool to understanding the relationship between two quantitative vari-
ables. However, it does not tell the whole story. While correlations can tell us the
direction and strength of the relationship between the variables, it does not tell
us how one variable affects the other. This can only be done through methods
that describe the linear relationship between our predictor and response, esti-
mating the slope of the line that describes the data. With this in mind, we have
to take one further step to have a fuller understanding of our variables.
Chapter 19

Simple linear regression

Contents
19.1 Introduction 227 19.9.4 Assumption 4:
19.2 Basic of lines 228 homoskedasticity 239
19.3 The simple linear regression 19.10 Inference for regression 240
model 229 19.10.1 State hypotheses 240
19.4 Estimating the regression model 230 19.10.2 Set significance level 241
19.5 Regression in R 232 19.10.3 Collect and
19.6 Practice problems 234 summarize data 241
19.7 Using regression to create 19.10.4 Calculate our test
predictions 234 statistic 241
19.8 Practice problems 235 19.10.5 Calculate p-values 242
19.9 The assumptions of regression 236 19.10.6 Conclusion 243
19.9.1 Assumption 1: linearity 236 19.10.7 Inference for
19.9.2 Assumption 2: regression in R 243
independence 239 19.11 How good is our regression? 244
19.9.3 Assumption 3: zero 19.12 Practice problems 247
mean 239 19.13 Conclusion 247

19.1 Introduction
In conducting inference for correlations, we begin to understand the linear re-
lationship between two quantitative variables, both in terms of direction and
strength. However, inference for correlations tells us nothing about how the vari-
ables affect each other. Consider the scatterplots in Fig. 19.1; in both cases, the
correlation between our x and y variables is approximately r ≈ 0.91. However,
in the scatterplot on the left, every time the predictor x increases by 1, the re-
sponse y increases by ≈ 0.4. On the right, every time the x increases by 1, the
y increases by ≈ 1.8. These are two very different effects, but with identical
correlations. We need a way to distinguish these two graphs and more fully un-
derstand the relationship between our variables.
Simple linear regression is a method to do this. It takes the two variables that
are linearly related—one predictor and one response—and finds the line that
best describes that relationship. This line will be related to the correlation, but
give us more information than just the correlation alone. From this technique,
we can look at a variety of aspects and consequences of this line, including pre-
dicting our response. This opens up a variety of possibilities in terms of analysis
and questions that we can answer.

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00032-8 227

Copyright © 2022 Elsevier Inc. All rights reserved.
228 Basic Statistics With R

FIGURE 19.1 Two scatterplots with identical correlations but different slopes.

While much of the calculations for simple linear regression can be derived
by hand, it is highly impractical to complete them given the sizes of datasets. As
such, we will discuss how the estimates are derived in general terms but not go
through the full derivation. As an example, let us look at the following dataset. It
is generally understood that there is a connection between the length of eruption
time for the Old Faithful geyser and the waiting time until the next eruption. We
will use the faithful dataset in R to look at this phenomenon in order to develop
a simple linear regression to predict waiting time [19].

FIGURE 19.2 Eruption time and wait time for the Old Faithful geyser.

19.2 Basic of lines

In math, the slope-intercept form of a line is

y = β0 + β1 x

where y is our dependent variable—or the variable that we plan to predict—and

x is our independent variable—the variable we want to use for prediction. There
 Simple linear regression Chapter | 19 229

are two key parts to this form of the line: the slope β1 and the intercept β0 . The
slope describes how much our independent variable increases when the depen-
dent variable increases by a single unit. If the slope is positive, the relationship
between independent and dependent variable is positive. If the slope is negative,
the relationship is negative. Finally, if the slope is zero, the independent variable
has no effect on the dependent variable.
The intercept tells us the value of our independent variable when our depen-
dent variable is equal to zero. We can see this clearly by substituting x = 0 into
our line. In this case, we would get

y = β0 + β1 × 0 = β0 + 0 = β0 .

19.3 The simple linear regression model

Many of the basic concepts of lines extends to simple linear regression. In sim-
ple linear regression, we say that our response yi —where this is the ith value
of our response variable—has a linear relationship with our predictor variable
xi —the ith value of our predictor variable. Putting this in the form of a line, this
would be
yi = β0 + β1 xi .
Again, our intercept β0 is the value that we expect our response to be when
our predictor is equal to zero, though this value may not make logical sense.
Our slope β1 is how much we expect our response to increase if our predictor
increases by a single unit.
Now, the way the relationship between our predictor and response as cur-
rently described says that our response values fall perfectly on a line, meaning
that we should be able to perfectly predict our response. Looking at the plot of
the Old Faithful geyser data in Fig. 19.2, this is clearly not the case and generally
speaking will not be. Because of this fact, we need to acknowledge that there
is error in our response. That is, while our line can do a good job of predicting
our response, there is always a certain amount of unexplainable, random error.
With this in mind, we need to make an addition to the equation of our line to
represent this error.
yi = β0 + β1 xi + i .
This is our standard simple linear regression model. In this new model, our
random error is represented by i —the ith random error. It is a random variable,
meaning that it follows a probability distribution of some form. In some cases,
we assume that the error follows a Normal distribution, but we will not do so
here. However, we do place several assumptions on our errors i , which we will
now discuss.
• The errors i are independent.
• The expected value of the errors is zero.
230 Basic Statistics With R

• The variance of the errors is constant, notated σ 2 .

Each of these assumptions can affect our regression line if they are not true.
If our errors are not independent our inferences for regression will be biased.
If the average value of the errors is not zero, our estimate of the intercept
will change to accommodate this. Finally, if the variance of our errors is not
constant—a case that is more common than might be expected—the inference
for our regression model will again be incorrect. As these assumptions are es-
sential to our regression model working properly, we will need to check these
assumptions, which we will go into greater detail later.

19.4 Estimating the regression model

In our regression model, we have three parameters: the intercept β0 , the slope
β1 , and the variance of our errors σ 2 . Each of these are fixed and unknown and
thus must be estimated. Determining the slope and intercept will give us our
estimated regression line, so this is where we will begin.
In estimating our regression line, our goal is to try and make our points as
a whole as close to our regression line as possible. In other words, we want to
make the total distance between our points and line as small as possible. This
distance is represented through our squared errors, and the total is called the
Sum of Squared Errors, or SSE. So in order to find the SSE, we first need to find
our errors. Based on our regression line, if we solved for our error i we would
get
i = yi − β0 − β1 xi .
Once we have solved for the error, we can easily find the SSE. We will square
each of the errors and then add them together. (See Fig. 19.3.)


n 
n
SSE = i2 = (yi − β0 − β1 xi )2 .
i=1 i=1

Now that we have defined the SSE, we want to choose the values of β0 and
β1 that make the SSE as small as possible. We will not go through how this is
done, but for those with a Calculus background it is accomplished in a similar
way to minimizing any function. If we went through this process of deriving our
estimates for our slope and intercept—β̂1 and β̂0 , respectively—we will find
they are
sy
β̂1 = rx,y ×
sx
β̂0 = ȳ − β̂1 x̄

where rx,y is the correlation between our predictor and response, sx and sy are
our sample standard deviations for predictor and response, respectively, and x̄
 Simple linear regression Chapter | 19 231

FIGURE 19.3 Visualizing errors in regression. The vertical lines represent the error in our regres-
sion model, which is squared and summed to make our SSE.

and ȳ are the sample mean of the predictor and response. In finding the estimate
for our slope particularly, we can see the connection between our correlation
and regression. If the correlation is positive, the slope will be positive. If the
correlation is negative, the slope will be negative. As the correlation grows, so
too does the slope. Finally, when the correlation is close to zero—which implies
that the association between the two variables is weak—the estimated slope will
also be close to zero—meaning the predictor has very little influence on the
response.
Let us apply this to the Old Faithful dataset. In this dataset, we can find
in R that our correlation between eruption time and wait time is r = 0.9008,
the standard deviation for our predictor eruption time is sx = 1.1414, and the
standard deviation for our response wait time is sy = 13.5950. Thus, the estimate
13.595
for our slope in the regression line will be β̂1 = 0.9008 × = 10.7296.
1.1414
This implies that for every minute longer that the eruption of old faithful lasts
we would expect to wait 10.7296 extra minutes for the next eruption.
Our next step is to estimate our intercept β̂0 . In order to do this, we need our
estimate for the slope—ββ1 = 10.7296, as we as the means for eruption time
x̄ = 3.4878 and wait time ȳ = 70.8971. This means that our intercept will be
β̂0 = 70.8971 − 10.7296 × 3.4878 = 33.4744. Thus, our estimated regression
line is

yi = 33.4744 + 10.7296xi .

The last parameter that we have to estimate for our model is the variance
of our errors σ 2 . Given our errors, we could estimate the true variance of the
errors by using the sample variance of those value. The difficulty with this task
is, of course, we do not know our errors. Since the formula of the errors is
i = yi − β0 − β1 xi , in order to know the true value of the errors we would need
to know the true value of β0 and β1 —which is impossible.
232 Basic Statistics With R

While we do not know β0 and β1 , we do have estimates β̂0 and β̂1 . We

can use these coefficient estimates to get an approximation of our errors. These
estimates of our errors are called residuals, notated ei .

ei = yi − β̂0 − β̂1 xi .

We can use these residuals to estimate the variance of our errors σ 2 . This
estimate is called the mean squared error (MSE), is notated σ̂ 2 and is calculated

1  2
n
σ̂ 2 = ei .
n−2
i=1

Calculating this by hand can be quite tedious, so we will turn to R to do this

calculation for us.

19.5 Regression in R
While it is possible to calculate the components of our regression model—
sample correlation, standard deviations, and means—in R and then use them
to calculate the slope and intercept, R provides a considerably easier method.
The lm function in R is used to estimate simple linear regression models.
It takes in our response and predictor variables and our dataset and outputs a
variety of important statistics, including our slope, intercept, and MSE. The code
for the lm function is

lm(Response ~ Predictor, data=dataset)

where Response is our response variable, Predictor is our predictor variable, and
dataset is the dataset where our variables are stored. If we were to apply this to
our Old Faithful example—where our dataset is named faithful, our response
variable is the wait time waiting, and the predictor variable is the eruption time
eruptions—our code would be

lm(waiting~eruption, data=faithful)

R then will return the output below. The (Intercept) column gives us the
value of our intercept estimate β̂0 , while the eruptions column gives the slope
estimate β̂1 . This column name will change depending on the name of the pre-
dictor variable.

Call:
lm(formula = waiting ~ eruptions, data = faithful)

Coefficients:
(Intercept) eruptions
33.47 10.73
 Simple linear regression Chapter | 19 233

In several cases, we will need to call parts of the R output for the lm function.
In order to do so, we will need to save the results of the lm function. We can do
this like we would save any variable, vector, or data frame in R.
save name=lm(...)

For example, say we save the Old Faithful regression as slr. The code would
be
slr=lm(waiting~eruption, data=faithful)

One instance where we need to call on the output of the lm is to get the
MSE, our estimate of the error variance. The MSE is stored in the summary
of our lm object, accessible through the summary function in R. We have seen
this function previously, where it was used to get our five-number summary of
a variable. When the input into the summary function is an lm object, it returns
a summary of the regression.
summary(lm object)

This summary will meet several of our needs when evaluating our regression.
If we wanted to access the summary of our Old Faithful regression—saved as
slr—the code would be
summary(slr)

And R will return the summary below. The MSE, or a statistic derived from
it, is stored in the Residual standard error portion of the summary. The resid-
ual standard error is the square root of the MSE, making it our estimate of the
standard deviation of our errors, or σ̂ . In our Old Faithful regression, our resid-
ual standard error is 5.914, making our MSE and estimate of our error variance
σ̂ 2 = 5.9142 = 34.9755.
Call:
lm(formula = waiting ~ eruptions, data = faithful)

Residuals:
Min 1Q Median 3Q Max
-12.0796 -4.4831 0.2122 3.9246 15.9719

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 33.4744 1.1549 28.98 <2e-16 ***
eruptions 10.7296 0.3148 34.09 <2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 5.914 on 270 degrees of freedom

Multiple R-squared: 0.8115, Adjusted R-squared: 0.8108
F-statistic: 1162 on 1 and 270 DF, p-value: < 2.2e-16
234 Basic Statistics With R

19.6 Practice problems

1. Load the cars dataset, a part of the base package of R. What is the regression
model that best predicts the stopping distance of the vehicle (dist) using the
speed of the vehicle (speed) [31]?
2. Load the Icecream dataset from the Ecdat package into R. What is the best
regression model that predicts the monthly amount of ice cream consumption
(cons) using the average monthly temperature (temp)? What is the estimate
of the error variance? [12,92]
3. Load the humanpower1 dataset from the DAAG library [23,108]. This dataset
looks at the relationship between oxygen uptake and watts per kilogram of
power output. What is the best regression model that predicts the watts per
kilogram of power output (wattsPerKg) using oxygen uptake (o2) as a pre-
dictor. What is the estimate of the error variance?

19.7 Using regression to create predictions

One of the major goals of our regression is to generate predictions. We want to
predict stopping distance, we want to predict the waiting time for old faithful.
As such, we need a way to create these predictions using our regression model.
Say you had just observed an eruption of Old Faithful that lasted 2 min-
utes. How would you predict the waiting time? We generally assume that
our estimated regression line is the best guess at the relationship between our
predictor—eruption time—and response—wait time. That line runs through all
the possible values of our eruption time, so it seems reasonable that our best
guess at wait time—our prediction—should be the value of the regression line
at our predictor of interest. So, if Old Faithful had an eruption of 2 minutes, we
would expect that the wait time would be 33.47 + 10.73 × 2 = 54.93 minutes.
In general, we notate our predictions as ŷ, and set these predictions given a
specific value of xi to be ŷi = β̂0 + β̂1 xi . When our value of xi is one that is
contained in our dataset, ŷi is referred to as a fitted value. If xi is a new value of
our predictor, ŷi is referred to as a prediction.
We can also use our fitted values to generate our residuals too. Recall that our
residuals were defined as ei = yi − β̂0 − β̂1 xi and our fitted value ŷi = β̂0 + β̂1 xi .
Thus our residuals can be calculated using

ei = yi − ŷi .

In this way, our residuals are often referred to being calculated as “observed
minus predicted.”
While our predictions are relatively easy to create by hand for an individual,
when fitted values or predictions are needed en masse R provides several options
to get predictions. If you need the fitted values for our observations, they are
stored in the saved lm object under the name fitted.values. In order to access
them for plotting or analyzing, we would use the code
 Simple linear regression Chapter | 19 235

lm object$fitted.values

In the Old Faithful regression saved as slr, we access our fitted values using
the code slr$fitted.values. The residuals can be accessed in a similar way as
they are stored in the lm object under the name residuals. For the Old Faithful
regression, we would access the residuals using
slr$residuals

If our desired prediction represents a new value of our predictor, we generate

these values using the predict function. The predict function takes in our saved
lm object—representing our linear regression—and a data frame containing the
new value or values of our predictor.
predict(lm object, newdata)

In our data frame for our new predictor values, there is a key point to recog-
nize. As our predictor variable is specifically named in our lm formula, the name
of our variable in the newdata data frame must match. For example, in the Old
Faithful regression—saved as slr—say that we wanted to generate a prediction
for wait time assuming that the previous eruption was 2.75 minutes long. Our R
code for this would be
predict(slr, newdata=data.frame(eruptions=2.75))

And R will return the prediction of 62.9809. We can confirm with a little
arithmetic that this is correct, up to rounding of course.
One important thing to note about our predicted values specifically is that
there are values for which we should not generate predictions. In general, we
should not generate predictions for values of outside of our predictor range—
what is called extrapolating. Why is extrapolation such a problem? We generally
assume that the relationship between predictor and response is linear in the
range of our predictor. Outside of that range we do not know how the predictor
and response are related. It could remain linear, or it could change drastically.
We can see an example of this in Fig. 19.4, where a small range of the pre-
dictor appears linear but the overall relationship between predictor and response
remains highly nonlinear. If we were to assume linearity based on the smaller
range and extrapolate, we would have drastically different predictions than real-
ity.

19.8 Practice problems

4. Based on the cars dataset in R—using speed to predict dist—what would be
your predicted stopping distance for a speed of 30 MPH [31]?
5. Load the Galton dataset from the HistData library into R. Run the regression
predicting the child height (child) using the average parent height (parent)
and predict the child height if the average parent height was 68.5 inches tall
[75,76].
236 Basic Statistics With R

FIGURE 19.4 An example of a relationship where a small range of the predictor is linear on the
left, but the overall relationship on the right is not.

6. Load the water dataset from the HSAUR library [99,102]. This dataset looks
at the relationship between water hardness and mortality. If the calcium con-
centration of the water (hardness) is equal to 60 parts per million, what
would be the expected mortality per 100,000 male inhabitants (mortality)
based on the simple linear regression.

19.9 The assumptions of regression

Like every statistical technique, regression has assumptions that must be true
in order for the results of the technique to be valid. We’ve discussed a few of
these already, but we have to add one more assumption: the assumed model is
appropriate, or the relationship between predictor and response is indeed linear.
This brings our assumptions up to four in total:
• The linear model is appropriate.
• The errors are independent.
• The expected value of the errors is zero.
• The variance of the errors is a constant, σ 2 .
In order to trust the results of our regression, these assumptions must hold.
Therefore we need to check each of these, which will do in order.

19.9.1 Assumption 1: linearity

This assumption is born out of how we defined our regression model. Recall that
we stated that our regression model was

yi = β0 + β1 xi + i .

This only allows for a linear component in our model. If the true relationship
between our predictor and response was quadratic—that is, it has an xi2 compo-
nent in it—our regression model would be inadequate. To check this assumption,
 Simple linear regression Chapter | 19 237

we plot the regression residual for each observation versus the observation’s fit-
ted value. If linearity holds, the residuals will show no pattern, looking like
random scatter. If there is any pattern in the residuals, this implies that the lin-
earity assumption does not hold. (See Fig. 19.5.)

FIGURE 19.5 An example of random scatter in the residuals—so the assumption holds—and pat-
tern in the residuals—so the assumption is violated.

In order to check this in R, we need to find our fitted values and residuals.
Recall that our fitted values ŷi = β̂0 + β̂1 xi are our predicted values for a given
observation in our dataset and our residuals ei are our estimates of our random
errors i , calculated as our observed value minus our predicted value, or ei =
yi − ŷi .
Our fitted values are stored in our lm object under the name fitted.values,
while our residuals are stored in the same place under the name residuals. In
order to access them we can use the $ notation, similar to how we call a variable
from a data frame. For example, recall that we saved our simple linear regres-
sion for the Old Faithful dataset under the name slr. If we wanted to see these
residuals and fitted values, we would use

slr$fitted.values

slr.$residuals

To check the assumption, we would use the base code

plot(slr$fitted.values,slr$residuals)

To produce the plot in Fig. 19.6. Note that this plot has several additional
options included in the code to change—among other things—the plotting char-
acter and axis labels. Looking at this plot, it appears that the linearity assumption
does hold.
In plotting the residuals versus fitted values to check the assumption, it
seems like there would be an easier method. Namely, we could merely plot
238 Basic Statistics With R

FIGURE 19.6 Checking the linearity assumption for the Old Faithful regression.

our predictor versus our response. If we see linearity in this relationship, the as-
sumption would be satisfied. However, plotting the predictor versus response
can be misleading. Consider the following data: James Forbes was a Scot-
tish researcher looking at the effect of altitude—represented through baromet-
ric pressure—on the boiling point of water. The data is stored in the forbes
dataset in the MASS library [10,93]. Plotting the barometric pressure—our
predictor—versus the boiling point—our response—it appears that the rela-
tionship is linear, albeit with one odd outlier point. However, if we run a
linear regression and plot the residuals versus fitted values we clearly see a
curved pattern in the residuals. The residuals are much better at picking up
these patterns, making them more suitable for checking these assumptions. (See
Fig. 19.7.)

FIGURE 19.7 Forbes’ boiling point data plotted on left and residuals versus fitted for Forbes’ data
on the right. A clear pattern exists in the residuals, implying the linearity assumption is violated.

If the linearity assumption is violated, you can often transform the response
to create a model where linearity exists. Forbes’ data is an example of this,
 Simple linear regression Chapter | 19 239

where modeling the log of the boiling point with the atmospheric pressure is an
appropriate model.

19.9.2 Assumption 2: independence

When errors are dependent on each other, there are often one of two causes: the
errors are related to each other because the predictor is a time-related variable—
what is referred to as longitudinal or time-series data—or the predictors are
related due to a physical proximity to other observations—called spatial data.
These cases are outside the scope of this book, and so their checks and remedies
are not discussed here. For these simple cases—nonlongitudinal and nonspa-
tial data—that we are discussing here, the assumption of independence is often
accepted as true and, therefore, not formally checked.

19.9.3 Assumption 3: zero mean

In checking that the expected value of our errors is zero, we have to recall that
our errors i are unknown random variables. We therefore need a proxy for these
errors, which we do have in the form of our residuals ei . With this in mind, if
we can find an estimate for the expected value of our residuals we can use it
to check our assumption. Usually, we can estimate our expected value using a
simple sample average. So, what is the sample average of our residuals ē? With
a little math, we can actually show that ē = 0, and this will be the case for any
dataset and any regression. Because of this, we will generally accept that this
assumption holds as well.

19.9.4 Assumption 4: homoskedasticity

The assumption that the variance of the errors is constant—referred to as
homoskedasticity—is violated more often than one might think. In many real
datasets, as the values of the response get larger the variability in the response
also increases. Checking this error variance, unsurprisingly, involves our resid-
uals as they are our proxies for the regression errors. As we did for checking
linearity, to check homoskedasticity we plot our residuals versus the fitted val-
ues. As with linearity, if the errors are homoskedastic the residuals plotted
against the fitted values will show a random scatter pattern. If we see a pat-
tern, specifically a “fan pattern” similar to what is seen in Fig. 19.8, the errors
are heteroskedastic and the assumption is violated.
As with linearity, the residuals more clearly show the violation of the as-
sumption than a plot of the raw data. In the Old Faithful regression, looking
at Fig. 19.6 shows no fan pattern and thus the errors are homoskedastic. If the
errors are heteroskedastic, in several cases a transformation of the response—
similar to the linearity instance—will make the errors homoskedastic, thus sat-
isfying the assumption.
240 Basic Statistics With R

FIGURE 19.8 A simulated dataset with the plot of predictor versus response and the residual plot
showing heteroskedastic errors.

19.10 Inference for regression

Now that we have fit our regression, generated predictions, and checked assump-
tions, we have come to a key question about our regression: is our predictor
useful in our regression model? Put another way, does it help predict our re-
sponse?
This is a question that is of the utmost importance. We need a way to deter-
mine the answer to this question, to test the claim that the predictor is useless
versus useful. Fortunately, statistics provides such a mechanism through hypoth-
esis testing. Our steps remain the same for this hypothesis test, and we will go
through them from the point of view of testing whether or not the eruption time
of the Old Faithful geyser predicts the wait time for the next eruption.

19.10.1 State hypotheses

In hypothesis testing, our goal is to test a claim about our parameters that an-
swers our question of interest. In this case, our question is: “Does our predictor
variable help predict our response?” Our null hypothesis—which represents our
status quo or “Nothing interesting going on” result—would be that our predic-
tor variable isn’t useful in predicting our response. Our alternative hypothesis
would then be that our predictor is indeed useful in predicting our response.
The challenge is that we need to put these claims in terms of our parameters.
Let us think about what each of these hypotheses implies. If the null hypothesis
is true and our predictor does not affect our response, this means that no matter
what our predictor value is our response stays the same. The only way this is
possible is if our slope β1 is equal to 0. This will become our null hypothesis. If
our predictor is indeed useful, then the value of the predictor affects our response
and thus our slope is nonzero. Thus our two hypotheses will be

H0 : β1 = 0
 Simple linear regression Chapter | 19 241

HA : β1 = 0.

This set of hypotheses will hold for our Old Faithful regression as well,
testing if slope is zero or not.

19.10.2 Set significance level

As with our other hypothesis tests, our significance level α is set using your
comfort level of making a Type I error and field convention, with a default of
α = 0.05. In our example, we will work with the default α = 0.05.

19.10.3 Collect and summarize data

For this particular hypothesis test, we will need the results of our regression—
our estimated slope β̂1 and mean squared error σ̂ 2 particularly—as well as the
sample size n and variance of our predictor sx2 . For our example, we know our
estimated slope is β̂1 = 10.7296 and our MSE is σ̂ 2 = 34.9755. Our sample size
is n = 272 and the variance of our predictor eruption time is sx2 = 1.3027.

19.10.4 Calculate our test statistic

As with each of our previous hypothesis tests, we begin the process of calculat-
ing our test statistic with our general formula for the test statistic.

Sample Statistic – Null Hypothesis Value

t= .
Standard Error
We will need to fill in each of these components from our data, our hypothe-
sis, and knowledge about our sample statistics. First, we need to find the sample
statistic that estimates the parameter in our null hypothesis. In the regression
case, our null hypothesis was

H0 : β1 = 0.

Thus we need a sample statistic calculated from our data that estimates β1 . It
seems fairly clear that the slope from our simple linear regression β̂1 is a reason-
able choice for our sample statistic. Additionally, looking at our null hypothesis
we can see that our null hypothesis value is equal to 0. Thus, our test statistic at
this stage is

β̂1 − 0
t= .
Standard Error
All that remains is the standard error of our sample statistic, or s.e.(β̂1 ).
It turns out that under certain conditions—we will discuss them shortly—the
242 Basic Statistics With R

regression slope estimate β̂1 follows a Normal distribution, specifically

 
σ̂ 2
β̂1 ∼ N β1 , .
(n − 1)sx2

σ̂ 2
This implies that our standard error for β̂1 will be s.e.(β̂1 ) = ,
(n − 1)sx2
where σ̂ 2 is our mean squared error from our regression, n is our sample size,
and sx2 is the sample variance of our predictor. Thus our test statistic for this
hypothesis test will be
β̂1
t= .
σ̂ 2
(n − 1)sx2
For our Old Faithful regression, plugging in the various components of our
test statistic gives us

10.7296
t= = 34.09.
34.9755
271 × 1.3027

19.10.5 Calculate p-values

As with our other hypothesis tests, in order to calculate p-values we need to
know what distribution our test statistic follows in addition to our alternative
hypothesis. In the inference for regression case, if our regression slope estimate
β̂1 follows a Normal distribution, our test statistic will follow a t-distribution
with n − 2 degrees of freedom.
As always, calculating our p-values is dependent on our alternative hypoth-
esis, with similar rationale to our previous tests. However, since our focus in
this test is always HA : β = 0, there will only be a single p-value that we are
interested in.

HA P-value
 
HA : β1 = 0 2 × P tn−2 ≥ |t|

As with any hypothesis test, we have assumptions to worry about. We al-

ready know some of our assumptions, namely our regression assumptions that
we discussed earlier. These need to hold in order for our inference to be valid.
However, we are now focused on the assumptions needed for our test statistic to
follow a t-distribution. We mentioned that the key assumption is that our regres-
sion slope estimate β̂1 must follow a Normal distribution. A reasonable question
 Simple linear regression Chapter | 19 243

then is what is required for this to be true? In this case, either one of two con-
ditions must be true. Either our errors must come from a Normal distribution
or sample size must be sufficiently large—usually n ≥ 30—for the central limit
theorem to kick in.
We will know whether or not our sample size is large enough, but what
about the Normally distributed errors? It seems reasonable that we can use our
residuals in lieu of our errors to determine this. But how do we go about this?
Formal tests exist for seeing if a sample of data follows a Normal distribution.
However, in this book we will focus on informal techniques, namely looking at
a histogram of our residuals. If the histogram seems roughly Normal, we will
considered this assumption met.
In order to calculate our p-value, we again turn to the pt function in R. In
our Old Faithful example, our sample size is decidedly greater than 30, so the
central limit theorem will hold and our test statistic will follow a t-distribution
with 272 − 2 = 270 degrees of freedom. Thus we can use the following R code
to find that our p-value is 8.08 × 10−100 .

2*pt(34.09, 270, lower.tail=FALSE)

19.10.6 Conclusion
Our conclusion step has not changed, as we reject the null hypothesis if the p-
value is less than the significance level α. Otherwise, we would fail to reject the
null hypothesis. In our Old Faithful example, since our p-value is decidedly less
than our α = 0.05, we will reject the null hypothesis and conclude that β1 = 0,
and thus eruption time does help predict the wait time until the next eruption.

19.10.7 Inference for regression in R

While we can do much of this hypothesis testing by hand, R provides the full
results of our hypothesis testing in a function that we have seen before. The
summary function, when given an lm object as an input, will give us the full
results of our hypothesis test. The results of the hypothesis test are seen in the
Coefficients section of the summary. Consider the Old Faithful example, saved
as slr. In order to access the summary of the regression, we will use the code
summary(slr) to get the following results:

Call:
lm(formula = waiting ~ eruptions, data = faithful)

Residuals:
Min 1Q Median 3Q Max
-12.0796 -4.4831 0.2122 3.9246 15.9719

Coefficients:
244 Basic Statistics With R

Estimate Std. Error t value Pr(>|t|)

(Intercept) 33.4744 1.1549 28.98 <2e-16 ***
eruptions 10.7296 0.3148 34.09 <2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 5.914 on 270 degrees of freedom

Multiple R-squared: 0.8115, Adjusted R-squared: 0.8108
F-statistic: 1162 on 1 and 270 DF, p-value: < 2.2e-16

In the Coefficients section of this summary, we have a large amount of in-

formation, including our regression slope and intercept estimates stored under
the Estimate column, the standard errors of our slope and intercept in the Std.
Error column, the test statistic under t value, and our p-value under Pr(>|t|).
The last column provides a quick reference to be able to identify a statistically
significant variable—one that rejects the null hypothesis H0 : βj = 0—based on
the code given below the Coefficients section.
Let us look at this in the Old Faithful example. We were particularly inter-
ested if the slope for the eruptions variable was zero versus nonzero. We see in
the eruptions row of the Coefficients section that our regression slope estimate
was β̂1 = 10.7296, which we knew before. From working out this hypothesis
test by hand we knew that our standard error was s.e.(β̂1 ) = 0.3148 and the
test statistic was t = 34.09. Finally, while we do not have the exact p-value that
we calculated before, we do know—according to our table—that it is less than
2.2 × 10−16 . It would be exceedingly rare that our significance level would be
this small, thus we will still be able to draw conclusions about our parameters.

19.11 How good is our regression?

Now that we have determined whether or predictor is useful in modeling our
response, we come to the final question about our regression: how good is it?
This is not trivial, though it may seem to be. Even if a predictor is significant
and we reject the null hypothesis that H0 : β1 = 0, this does not mean that our
predictor creates a good regression model.
We need a way to evaluate our model, a way that helps us understand how
well our predictor does. A good starting place would seem to be our residu-
als, as they represent how far off our regression model’s predictions were from
observed reality. In evaluating our residuals we want to summarize them in a
single number. We know that the average of the residuals is always zero, mak-
ing it not an appropriate summary. Rather, we will summarize the residuals by
adding up all our squared residuals. By squaring the residuals, we make them
all positive, with values close to zero indicating that the regression model was
highly accurate. Adding these squared residuals together gives us the Sum of
 Simple linear regression Chapter | 19 245

Squared Residuals, or SSR.


n
SSR = ei2 .
i=1

SSE
Our SSR is directly related to our mean squared error σ̂ 2 , as σ̂ 2 = for
n−2
our simple linear regression. In general, if the SSR is close to 0, this implies that
our model is doing a good job fitting the data. If the SSR is large, the model is
doing a poor job fitting the data.
Unfortunately, there is a problem with this interpretation: it is dependent on
how much variability exists in our response. Consider the following two simu-
lated datasets presented in Fig. 19.9. In looking at the two graphs, it looks like
the regression on the right would do a better job predicting its response as the
observed values are more tightly clustered about the line. However, the SSR for
the regression on the left is 2421.9, while the SSR for the regression on the right
is 12050.5.

FIGURE 19.9 Two simulated regressions. The regression on the right appears to predict its re-
sponse better, despite a larger SSR.

Why is this the case? The variance of the response in the dataset on the
right—13516.02—is much larger than the variance of the response in the dataset
on the left—36.29. We need a statistic about our regression that both takes into
account our total variability and our SSR.
The most common statistic used that accounts for both of these is the coef-
ficient of determination, or R 2 . Our R 2 value represents the percentage of the
total variability in our response that is explained by the regression model. Since
R 2 is a percentage, this means it is bounded between 0 and 1, with larger val-
ues indicating that the regression explains more variability in the response—and
thus is doing a better job.
The formula for R 2 is given by
SSR
R2 = 1 −
SST
246 Basic Statistics With R

where SSR is our previously mentioned sum of squared errors and SST is the
total sum of square. We calculate this total sum of squares through the formula
n
SST = (yi − ȳ)2 , where ȳ is the average of our response. The SST is directly
i=1
connected to the variance of our response, just as the SSR is connected to our
mean squared error. Thus we can restate our R 2 formula using this information.

(n − 2)σ̂ 2
R2 = 1 − .
(n − 1)sy2

In this formula, σ̂ 2 is our mean squared error and sy2 is the variance of our
response. In our Old Faithful regression, we know that our mean squared error
is σ̂ 2 = 34.9755, the variance of the wait time response is sy2 = 184.8233, and
our sample size is n = 272. Using this information, our R 2 value is
270 × 34.9755
R2 = 1 − = 0.8115.
271 × 184.8233
Thus we can say that our regression using eruption time to predict wait time
until the next eruption accounts for 81.15% of the total variability in wait time.
The coefficient of determination has a specific connection to our correlation
for simple linear regressions. It turns out that for a simple linear regression in-
volving a response y and a predictor x, the R 2 value of that regression will be
R 2 = rx,y
2 , where r
x,y is equal to the correlation. We can see this in our Old
Faithful regression, as the correlation between eruption time and wait is equal
to rx,y = 0.9008, which squared matches up to our R 2 = 0.90082 = 0.8115.
Similar to our hypothesis test for regressions, we can find our value of R 2 us-
ing the summary function with an lm input in R. The R 2 is stored in the Multiple
R-squared section of our summary, beneath the Coefficients section. In looking
at the summary of our Old Faithful regression saved as slr—using the code sum-
mary(slr)—we can easily see that the value in the Multiple R-squared section
matches up with our calculated value of R 2 = 0.8115:
Call:
lm(formula = waiting ~ eruptions, data = faithful)

Residuals:
Min 1Q Median 3Q Max
-12.0796 -4.4831 0.2122 3.9246 15.9719

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 33.4744 1.1549 28.98 <2e-16 ***
eruptions 10.7296 0.3148 34.09 <2e-16 ***
---
 Simple linear regression Chapter | 19 247

Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 5.914 on 270 degrees of freedom

Multiple R-squared: 0.8115, Adjusted R-squared: 0.8108
F-statistic: 1162 on 1 and 270 DF, p-value: < 2.2e-16

19.12 Practice problems

Load the cabbages dataset from the MASS library into R. It looks at the rela-
tionship between cabbage head size and vitamin C content [10,94].
7. Run the regression using head size (HeadWt) to predict vitamin C content
(VitC). Do the assumptions for this regression hold?
8. Based on the inference for this regression, does the head weight predict the
vitamin C content of the cabbage head?
9. What percentage of the total variability in vitamin C content is explained by
the regression?
Load the Cigar dataset from the Ecdat library into R. This dataset is a survey
of cigarette prices and sales in the United States [12,109].
10. Run the regression using cigarette price (price) to predict total sales (sales).
Based on the inference for this regression, does the cigarette price predict
total sales?
11. What is the R 2 for this regression?
12. Do the assumptions for this regression hold?
Load the fruitohms dataset from the DAAG library. This dataset looks at the
electrical resistance of a slab of kiwi fruit based on its juice content [23,110].
13. Run the regression using juice percentage (juice) to predict electrical re-
sistivity (ohms). Based on the inference for this regression, does the juice
predict resistivity?
14. What is the R 2 for this regression?
15. Do the assumptions for this regression hold?

19.13 Conclusion
Regression compliments our inference for correlations by giving us a more com-
plete understanding of the relationship between predictors and responses. By
making a few assumptions about the random errors that exist in our data, we are
able to estimate the line that best describes our data. From this line, we can make
predictions, do inference on our line components, and understand how much of
our variability is described by the regression line. In doing all this, we have laid
down the foundations of linear models; one of the most useful techniques in
statistics.
Chapter 20

Statistics: the world beyond this

book
Contents
20.1 Questions beyond the 20.2 The answers statistics gives 252
techniques of this book 249 20.3 Where does this leave us? 254

There is a thread that connects through every idea and technique that exists in the
field of statistics. This book is the beginning of that thread. Through the course
of this book, we have been developing the mindset of statistics as a field of study
that is all about translating data into decisions, theories, and knowledge. We have
talked briefly about data collection and initially exploring our datasets. Then we
saw why these exploratory analyses are not enough to draw conclusions, leading
us finally to statistical inference.
We have been able to answer a variety of questions using these techniques,
but even so, these questions represent only a beginning in the grander scheme
of research. The complexity of questions has only increased as we have moved
forward, as has the size and scope data. These questions require answers, and
as such statistics and its techniques must be there to meet these questions with
ideas. And while our techniques learned in this book put us on the right track
toward answering many of these complex questions, they often leave us a little
short.

20.1 Questions beyond the techniques of this book

Through correlation, we are able to describe the linear relationship between
two quantitative variables. We can observe whether or not two variables are
positively or negatively associated. We can even predict a response based on a
single variable through simple linear regression, but it still leaves us short of
some very important and useful questions.
Say that, instead of dealing with only a single predictors, we wanted to pre-
dict a response using multiple predictors. Imagine trying to predict the price of
a laptop based only on its screen size or the value of a house based only on the
acreage of the lot. Those predictors may help you understand their respective re-
sponses, but they leave a hole in the total variability that can only be described
using other variables.

Basic Statistics With R. https://doi.org/10.1016/B978-0-12-820788-8.00033-X 249

Copyright © 2022 Elsevier Inc. All rights reserved.
250 Basic Statistics With R

Even when a technique seems well designed to answer a question, there are
times when generalizations are needed. We talked extensively about the two-
sample t-test for means, as well we should. It is the test to work with when
we are comparing a quantitative response across two groups. When we wanted
to compare the weight gain in chickens to look for the results of two different
dietary supplements—soybeans and sunflowers—this test supplied us with an
answer that we could trust.
However, this test limits us two comparing two groups at a time. What if we
wanted to compare six groups? What is to be done then? This is not a trivial
question, as it is directly related to our chicken weights question. This dataset
[60] that we presented—looking at soybeans and sunflowers—is just a subset
of the complete dataset which contains the weight gains of chickens on one of
six feed supplements—casein, horsebean, linseed, meatmeal, soybean, and sun-
flower. How would we know if there is any difference between the six different
supplements? Looking at the side-by-side boxplots of the weight gain for the six
supplements it seems like there is a difference, but how can we run statistical
inference to more definitively answer the question? We could compare each pair
of supplements in a series of two-sample t-tests for means, but this results in 15
comparisons and thus would seem more likely that we would make a mistake
along the way. It would seem like there is a more efficient and downright better
way to answer this question. (See Fig. 20.1.)

FIGURE 20.1 Side-by-side boxplots of chicken weight gain for six feed supplements.

This sort of scenario can extend to categorical responses as well. Right now,
our two-sample test for proportions allows us to see if the sample proportion of
a binary response differs across two groups. This can answer a wide array of
questions and give us a variety of useful information. For example, we could
look at if the proportion of liberals differs for individuals with at least a Bach-
elor’s degree than those without a Bachelor’s degree [54]. Data and inference
like this can help us understand and confirm or disprove conventional wisdom.
 Statistics: the world beyond this book Chapter | 20 251

Not Liberal Liberal

No Bachelor’s 538 280
Bachelor’s 185 132

However, this can be limiting. Similar to what we saw in our two-sample

t-test for means, we may want to compare our proportions across more than two
groups. For example, instead of looking only at No Bachelor’s versus Bachelor’s
we could expand our educational categories to include the variety of educational
levels seen in the population.

Not Liberal Liberal

No HS 72 24
HS Grad 270 107
Some College 138 103
2-Year Grad 58 46
4-Year Grad 118 80
Post-Grad 67 52

Now, this example is still a binary response. Say we were to expand this
to a multiary response, with as many options in our response as we want. In
our education and political leanings question, this is equivalent to going from
our Not Liberal versus Liberal—hardly a representation of the variety of the
political spectrum—to more complete view of the spectrum ranging from Very
Conservative (Represented as 1 in the table) to Moderate (4) to Very Liberal (7).

1 2 3 4 5 6 7
No HS 11 11 8 42 5 9 10
HS Grad 57 57 36 120 25 33 49
Some College 24 35 21 58 23 38 42
2-Year Grad 10 7 12 29 6 20 20
4-Year Grad 21 32 28 37 17 34 29
Post-Grad 16 16 14 21 13 23 16

This data is more granular than what we started with and can offer us a
wider variety of insights. Our original question—does the population propor-
tion of liberals differ among those who have a bachelor’s degree and those who
do not—can now be generalized to the question: “How does education affect po-
litical leaning?” However, the two-sample test for proportions is ill-equipped to
answer this question meaning that we need more complex inferential techniques
to develop an answer.
This need for more complex techniques is not confined to inference. There
are instances when the exploratory analyses discussed in the book are not suf-
ficient to understand a question. Say we wanted to try and classify a series of
252 Basic Statistics With R

pitches in a baseball game based on data about each individual pitch: the speed
of the pitch, spin rate, vertical movement, and horizontal movement. If we were
only dealing with two variables, we could plot them in a scatterplot and try to
identify clusters in the positions of the plotted points. However, with four vari-
ables this is impossible; at most we could visualize three variables at a time. We
need a way to visualize our data so that we can identify clusters within the data.
We can even extend this need for new ideas to the entire philosophy of statis-
tics. The methods of statistics discussed in this book fall into the paradigm
of statistics called frequentist statistics, in which we assume—among other
things—the frequentist or long-run view of probability and that our parameters
are fixed values. This mindset led us to hypothesis testing, in which we eval-
uate claims about our parameters through p-values. Say we want to know the
probability that one of our hypotheses is true. This seemingly simple question
is impossible in the framework of frequentist statistics and hypothesis testing,
as our p-values do not work in this way. Because frequentist statistics does not
have the tools to attack these problems, a different mindset is necessary.

20.2 The answers statistics gives

Each of these scenarios and the questions they ask just cannot be answered with
the techniques that we have learned in this book. The techniques we learned
may put us on the right path or even answer a portion of the question, but they
are unable to provide a complete and satisfactory answer. Of course, that hardly
means that Statistics lacks answers to these questions. In fact, each of these
questions can still be answered using statistical techniques, some of which are
commonly found in a second statistics course.
Consider our example of predicting the value of a house using multiple
variables—the acreage, house size, bedrooms, baths, etc. In predicting a re-
sponse with multiple predictors, it seems like the technique of simple linear
regression should be able to be extended to account for the new variables. Noth-
ing we talked about in terms of regression seems to be explicitly limited to a
single variable. This is indeed the case, as the technique of multiple regression
allows for the inclusion of several predictors in modeling a single response.
Many of the techniques discussed in simple linear regression extend directly
to multiple regression, including how our coefficients are estimated—if not the
exact values—inference for the regression slopes, and how we evaluate the re-
gression through R 2 .
Looking again at our chicken weight dataset, it seems like there should
be some technique that is able to allow for comparing multiple groups—
specifically more than two groups—at a time. Analysis of Variance—usually
called ANOVA—is just such a technique. ANOVA takes a look at the variabil-
ity that is inherent in our response and tries to break it down into its various
components—the individual treatments, their interactions, and error. In doing
so, ANOVA is able to identify if there is any difference among all the treat-
ments, not merely comparing a single pair of treatments.
 Statistics: the world beyond this book Chapter | 20 253

The same sort of ability to compare multiple groups is found the in the chi-
squared test of independence. In this test, we take our data—two categorical
variables with two or more categories apiece—and compare it to what we would
expect to see if the two variables had no affect on each other. In doing so, we
can find if any association exists between the two variables, such as if education
level and political leaning are associated in any way. This is a question we have
been asking—if two categorical variables are associated—just with restricting
our comparisons to merely two categories. This test allows us to expand our
possibilities, answering questions that are much more complex and allowing us
to gain further insight.
In the instance when we want to visualize more than two variables, a tech-
nique called Principal Component Analysis (PCA) exists. Similar to ANOVA,
PCA tries to partition the variability that exists in our data. PCA does this by
finding the independent combinations of variables that best explain the variabil-
ity. If we plot the two best combinations, this will help us identify clusters in the
data—should they exist—while maintaining the structure of the data. Consider
the baseball pitch data we mentioned earlier. In plotting the two best principal
components—the name for the linear combinations of variables—we see that
there seems to be three clusters within our data. With a little more informa-
tion about the data, this makes intuitive sense. Our data is the pitches thrown in
2019 by Mychal Givens, a reliever on the Baltimore Orioles. He throws three
pitches: a fastball, curveball, and change-up. We’d expect to see three clusters—
corresponding to those three pitch types—in our data, exactly what we do see in
the plot of our PCA. (See Fig. 20.2.)

FIGURE 20.2 Principal component analysis of Mychal Givens’s pitches from 2019.

Finally, even the question about the probability of a hypothesis being true
is able to be answered with new philosophies. Bayesian statistics is a philoso-
phy of statistics that is based around the ideas of subjective probability—that
probability is the quantification of a degree of belief—and the idea that we can
determine the probability of events given that some other event has occurred.
254 Basic Statistics With R

In this paradigm, our parameters are random variables as opposed to fixed. As

random variables, our parameters have probability distributions and much of
Bayesian inference is aimed at identifying the probability distribution of our pa-
rameters after we have observed our data. Another consequence of this change
is that we can make claims in inference that would couldn’t previously; claims
like the probability that a hypothesis is true or the probability that our parameter
lives in an interval. Most techniques that we have discussed have a Bayesian ana-
logue and many advanced techniques such as AI have foundations in Bayesian
statistics.

20.3 Where does this leave us?

As we can see, as our questions continue to get more complex, Statistics rises
to meet these questions with new techniques and sufficient answers. Even the
techniques and philosophies discussed above—multiple regression, ANOVA,
chi-squared test for independence, PCA, and Bayesian statistics—can be fur-
ther combined and built upon to develop even more useful methodologies. This
pattern of rising complexity and statistics developing new methods continues
onward, even to the highest levels of research.
However, this can leave us with a question: “If the techniques learned in this
book are unable to answer so many questions, what is their purpose?” With the
complexity that exists in both our questions and answers, it seems like many of
the things learned in the previous 250 pages are trivial and lacking in use. What
is their point if they cannot be used?
Despite the techniques in this book being limited in some respects, they pro-
vide a foundation on which many of our techniques can be built. Every one of
those second-level techniques used to answer the questions all have some basis
in the techniques of this book. In order to understand multiple regression, we
need to know simple linear regression. For us to implement either ANOVA or
the chi-squared test for independence, we need to fully grasp the mechanisms
of hypothesis testing laid down in this book. PCA is all about decomposing the
variability in our data which of course requires understanding variability, but it
also involves the correlation between our variables. Even Bayesian statistics is
based around probability and probability distributions, which are first discussed
in the previous chapters.
Throughout all of statistics runs a golden thread. No matter the complexity of
the question we want to answer or how advanced the technique, it all comes back
to wanting to answer a question with data—either already available or about
to be collected. Implementing the techniques needed to answer these question
requires an understanding of the basics. Essentially, all of statistics is linked
together, with all of those links tracing back to the topics of this book.
Appendix A

Solutions to practice problems

Chapter 2
1. Movie=c(“Citizen Kane,” “The Godfather,” “Casablanca,” “Raging Bull,”
“Singing in the Rain”)
2. Year=c(1941, 1972, 1942, 1980, 1952)
3. RunTime=c(119, 177, 102, 129, 103)
4. RunTimeHours=RunTime/60
5. MovieInfo=data.frame(Movie, Year, RunTime, RunTimeHours)
6. Title=c(“The Secret of Monkey Island,” “Indiana Jones and the Fate of
Atlantis,” “Day of the Tentacle,” “Grim Fandango”)
7. Release=c(1990,1992,1993,1998)
8. Release-1982
9. Rank=c(14,11,6,1)
10. AdventureGames=data.frame(Title, Release, Rank)

Chapter 3
1. Due to social desirability, individuals might not be willing to admit their
depression in person.
2. This could introduce sampling bias because individuals may not be available
during the survey times selected.
3. The response is the change in blood pressure. The explanatory variables are
the medication and exercise level. The treatments are Drug A-1 hour, Drug
A-5 hours, Drug B-1 hours, and Drug B-5 hours.
4. The response is the improvement of the student measured by the difference
in pre- and post-exam scores. The explanatory variable is the teaching style.
The treatments are Team-based method or Traditional method.
5. This study is an observational study, and observational studies are not able
to establish causality.
6. The subjects are the 200 participants, the explanatory variable is their amount
of exercise, and the response is their weight.
7. Possible confounding variables could be dietary choices and family history,
among many others.
8. Design an experiment where subjects are assigned to different exercise regi-
mens in order to control for the various confounding factors.

255
256 Solutions to practice problems

Chapter 4
1. Colleges$TopSalary[c(1, 3, 10, 12)]
2. Colleges$MedianSalary[Colleges$TopSalary>400 000]
3. Colleges[Colleges$Employees<=1000,]
4. Colleges[sample.int(n=14, size=5),]
5. Countries[Countries$GDPcapita<10000&Countries$Region!=“Asia.”]
6. Countries[sample.int(n=10, size=3),]
7. Countries$Nation[Countries$PctIncrease>1.5]
8. Olympics[Olympics$Host==Olympics$Leader,]
9. Olympics[Olympics$Competitors/Olympics$Events>35,]
10. Olympics[Olympics$Type==“Winter”&Olympics$Nations>=80,]

Chapter 5
1. library(Ecdat)
2. data(Diamond)

Chapter 6
1. The population of interest is American adults.
2. n = 2458 + 1796 + 378 + 94 = 4726
2458
3. p̂ = = 0.5202
4726
4. n = 48 + 334 + 66 + 990 + 62 + 848 + 11 + 166 = 2525
334 + 990 + 848 + 166
5. p̂ = = 0.9259
2525
None of it Not much Some of it Most of it
6.
382 1056 910 177
7. x1 = 0.7, y8 = 1.6, x (2) = −1.2, y (6) = 1.9
8. ȳ = 2.33
9. x̃ = 0.35
10. DDT (3) = 3.06, DDT (14) = 3.78
11. DDT ¯ = 3.328
12. DDT ˜ = 3.22
13. x̄ = 101.425
14. x̃ = 100.9
15. s 2 = 1.898
16. s = 1.378
17. I QR = 102.9 − 100.4 = 2.5
18. x̄ = 94.14
19. x̄ = 92
20. s 2 = 197.476
21. s = 14.053
22. I QR = 106 − 78 = 28
 Solutions to practice problems 257

23. Our z-scores are 0.74, −2.33, −0.34, −0.13, −0.4, 0.05, 0.19, 0.53, −0.18,
0.21, 0.93, 0.08 1.43, 1.04, and −1.83, so according to z-scores, none of
our data points are potential outliers.
24. As any observation less than 1 − 1.5 × (6.125 − 1) = −6.6875 or greater
than 6.125 + 1.5 × (6.125 − 1) = 13.8125 are potential outliers, we would
say −8.375 is a potential outlier.
25. Our z-scores are −0.03, −0.03, 2.05, −0.59, 1.49, −0.73, −0.31, −0.31,
0.38, −1.84, 0.38, −0.17, 0.52, −1.15, 0.8, −1.84, 0.52, −0.59, 0.24, and
1.22, so according to z-scores, none of our data points are potential outliers.
26. As any observation less than 810 − 1.5 × (890 − 810) = 690 or greater than
890 + 1.5 × (890 − 810) = 1010 is a potential outlier, so according to the
IQR, there are no potential outliers.
27. Top left: Bimodal, symmetric, centered around 0.5. Top right: Bimodel,
skewed to the right, centered near 0. Bottom left: Unimodal, skewed
left, centered around −2. Bottom right: Unimodal, skewed right, centered
around 0.75, potential outlier near −2.
28. See figure.

29. See figure.

258 Solutions to practice problems

30. Based on the boxplots, there does seem to be an association between

species and sepal length.
31. Top left: Linear, positive trend, strong association. Top right: Linear, nega-
tive trend, moderately strong association. Bottom left: Quadratic, negative
tend, strong association. Bottom right: No trend, weak association.
32. r = −0.7708
33. r = 0.9882

Chapter 7
1. table(leuk$ag)
2. median(leuk$time), mean(leuk$time)
3. range(leuk$wbc), IQR(leuk$wbc), var(leuk$wbc), sd(leuk$wbc)
4. cor(leuk$wbc, leuk$time)
5. table(survey$Smoke, survey$Exer)
6. mean(survey$Pulse), median(survey$Pulse)
7. range(survey$Age), IQR(survey$Age), var(survey$Age), sd(survey$Age)
8. cor(survey$Wr.Hnd, survey$NW.Hnd)
9. table(Housing$recroom, Housing$fullbase)
10. mean(Housing$lotsize), median(Housing$lotsize)
11. range(Housing$price), IQR(Housing$price), var(Housing$price), sd(Hous-
ing$price)
12. cor(Housing$price, Housing$bedrooms)
13. plot(Star$tmathssk, Star$treadssk)
14. hist(Star$totexpk)
15. Star$totalscore=Star$tmathssk+Star$treadssk
16. boxplot(totalscore∼classk, data=Star)
17. plot(survey$Height, survey$Wr.Hnd)
18. hist(survey$Age)
19. boxplot(Pulse∼Exer, data=survey)

Chapter 8
1. S = {Ace of Spades, 2 of Spades...Queen of Hearts, King of Hearts}
12
2. P (A) =
52
1 3
3. P (AC ) = 1 − P (A) = 1 − =
4 4
4. S = {1, 2, ..., 100}
10 1
5. P (A) = =
100 10
6. X can be equal to 0, 1, 2, 3.
3
7. P (X = 2) =
8
8. Binomial(19, 0.4)
9. X ∼ Binomial(8, 0.492)
 Solutions to practice problems 259

10. 8 × 0.492 = 3.936

X − (−2)
11. Z = √
7
X−3
12. Z = √
1
13. X + Y ∼ N (2, 5), X − Y ∼ N (−8, 5)
14. X + Y ∼ N (0, 2), X − Y ∼ N (0, 2)

Chapter 9
1. The sampling distribution would be centered at 0.5.
2. Your colleague will have a smaller standard error.
3. As the sample size increases, the sampling distribution of the sample average
will approach a Normal distribution.
4. The standard error will get smaller and eventually go to 0. The sampling
distribution will approach a Normal distribution.

Chapter 10
1. H0 : p = 1/6, HA : p > 1/6.
2. One-sided test
3. P (Type I) = 0.05
4. Type II Error
5. We would reject the null hypothesis because our p-value is less than our
significance level.
6. H0 : p = 0.5, HA : p = 0.5.
7. Two-sided test
8. The significance level α does not affect the probability of a Type II error.
9. We would fail to reject the null hypothesis because our p-value is greater
than our significance level.

Chapter 11
1. 0.9772499
2. 0.01222447
3. 0.9875338
4. 0.9583677
5. 0.001349898
6. H0 : p = 0.47, HA : p < 0.47
20
7. p̂ = = 0.4
50
0.4 − 0.47
8. t= = −1
0.07
9. p̂ ∼ N (0.47, 0.072 )
10. 0.1586553
260 Solutions to practice problems

11. H0 : p = 0.4546, HA : p < 0.4546

42
12. p̂ = = 0.4
100
0.42 − 0.4546
13. t = = −0.692
0.05
14. p̂ ∼ N (0.4546, 0.052 )
15. 0.2444687

Chapter 12
1. We are 90% confident that (0.438486, 0.461514) covers the true value of p
2. We are 95% confident that (0.4362803, 0.4637197) covers the true value
of p
3. We are 99% confident that (0.4319692, 0.4680308) covers the true value
of p
4. We are 99.5% confident that (0.4303508, 0.4696492) covers the true value
of p
5. We are 80% confident that (0.7610291, 0.7789709) covers the true value
of p
6. We are 90% confident that (0.758486, 0.781514) covers the true value of p
7. We are 95% confident that (0.7562803, 0.7837197) covers the true value
of p

Chapter 13
1. H0 : p = 0.68, HA : p > 0.68
1270
2. p̂ = = 0.8199
1549
0.8199 − 0.68
3. t =  = 11.8036
0.68(1 − 0.68)
1549
4. 1.87 × 10−32
5. Because our p-value is less than α, we reject the null hypothesis and con-
clude that the proportion of Americans who say children should be required
to be vacinnated is greater than 0.68.
6. H0 : p = 0.5, HA : p = 0.5
122
7. p̂ = = 0.442
276
0.442 − 0.5
8. t =  = −1.927
0.5(1 − 0.5)
276
9. 0.0540
10. Because our p-value is less than α, we reject the null hypothesis and con-
clude that the proportion of e-mails that are spam is not equal to 0.5.
11. H0 : p = 0.79, HA : p < 0.79, α = 0.05
 Solutions to practice problems 261

1082
p̂ = = 0.7204
1502
0.7204 − 0.79
t= = −6.622
0.79(1 − 0.79)
 1502 
P N (0, 1) < −6.622 = 1.77 × 10−11
Because our p-value is less than α, we reject our null hypothesis and con-
clude that the proportion of American who have read a book in the last year
is less than 0.79.
12. H0 : μ = 6, HA : μ > 6
6.2 − 6
13. t= √ = 3.683
2.06/ 1439
14. t distribution with 1438 degrees of freedom.
15. 0.000120
16. Because our p-value is less than α, we will reject the null hypothesis and
conclude μ > 6.
17. H0 : μ = 2.5, HA : μ > 2.5
6.143 − 2.5
18. t=√ √ = 5.84
13.229/ 34
19. t distribution with 33 degrees of freedom.
20. 7.7 × 10−7
21. Because our p-value is less than α, we will reject the null hypothesis and
conclude μ > 2.5.

Chapter 14
1. We are 95% confident that (0.3991859, 0.4853184) covers the true propor-
tion of times Mike Trout gets on base.
2. We are 99% confident that (0.1791346, 0.6706865) covers the true proba-
bility of winning at craps betting the pass line. However, the small sample
size may cast doubt on the results.
3. We are 90% confident that (0.4181515, 0.4767203) covers the true propor-
tion of women who volunteer.
4. We are 99% confident that (34.1071, 50.1529) covers the true mean ozone
ppm.
5. We are 95% confident that (1.815092, 2.324908) cover the true mean dif-
ference in actual minus reported height.
6. We are 95% confident that (60.5361, 61.3039) covers the true age of
women who have a heart attack.
1.6452 × 0.41 × (1 − 0.41)
7. n ≥ = 65.4471
0.12
2.5762 × 0.5 × (1 − 0.5)
8. n ≥ = 663.4897
0.052
1.962 × 0.5 × (1 − 0.5)
9. n ≥ = 9604
0.012
262 Solutions to practice problems

10. (2.427, 6.134). Since 4 is in the interval, we would fail to reject the null
hypothesis.
11. (2.044, 6.517). Since 2 is outside the interval, we would reject the null
hypothesis.
12. (1.506, 7.054). Because our confidence level does not match up with our
significance level, we cannot determine the result of our hypothesis test
based on this confidence interval.

Chapter 15
1. H0 : pM − pW = 0, HA : pM − pW = 0
794 + 919
2. p̂ = = 0.7308
1119 + 1225
0.7096 − 0.7502
3. t =    = −2.219
1 1
0.7308(1 − 0.7308) +
1119 1225
4. 0.02648
5. Since our p-value is less than α, we reject the null hypothesis and conclude
that men and women value work-life balance at different rates.
6. H0 : pAid − pN oAid = 0, HA : pAid − pN oAid < 0
α = 0.01
48 66
p̂Aid = = 0.2222, p̂N oAid = = 0.3056
216 216
48 + 66
p̂ = = 0.2639
216 + 216
0.2222 = 0.3056
t=   = −1.676934
1 1
0.2639(1 − 0.2639) +
216 216

P N (0, 1) < −1.676934) = 0.04677769
Becase our p-value is greater than α, we fail to reject the null hypothe-
sis and conclude it’s plausible that the rate of recidivism is equivalent for
former inmates with and without financial aid.
7. H0 : pY es − pN o = 0, HA : pY es − pN o > 0
α = 0.1
263
p̂Y es = = 0.8567, p̂N o = 0.6773
307
263 + 783
p̂ = = 0.7150
307 + 1156
0.8567 − 0.6773
t=   = 6.189747
1 1
0.7150(1 − 0.7150) +
307 1156
 
P N (0, 1) > 6.189747 = 3.01 × 10−10
 Solutions to practice problems 263

Because our p-value is less than α, we reject the null hypothesis and con-
clude that people who remember receiving corporal punishment are in
favor of moderate corporal punishment for children at a higher proportion
than those who do not remember receiving corporal punishment.
8. H0 : μBenign − μMalignant = 0, HA : μBenign − μMalignant = 0
9. Ratio of variances is between 0.25 and 4, so
1.672 × (458 − 1) + 2.432 × (241 − 1)
sp2 = = 3.862
458 + 241 − 2
2.96 − 7.2
t=   = −27.112
1 1
3.862 +
458 241
10. df = 458
 + 241  − 2 = 697 
11. 2 × P t697 >  − 27.112 = 4.35 × 10−111
12. Since our p-value is less than α, we reject the null hypothesis and conclude
that clump thickness differs for benign and malignant tumors.
13. H0 : μW C − μBC = 0, HA : μW C − μBC > 0
α = 0.1
sW2
C 139.067
2
= = 0.4266
sBC 325.991
325.991 × (21 − 1) + 139.067 × (6 − 1)
sp2 = = 288.6062
21 + 6 − 2
36.667 − 22.762
t=   = 1.76816
1 1
288.6062 +
6 21
 
P t25 > 1.76816 = 0.04462
Because our p-value is less than α, we reject the null hypothesis and con-
clude that white collar jobs are viewed as more prestigious than blue collar
jobs. However, the small sample sizes and lack of Normality cast some
doubt on the results.
14. H0 : μC − μMP = 0, HA : μC − μMP = 0
α = 0.05
sC2 0.892
2
= = 0.5069
sMP 1.252
0.892 × (45 − 1) + 1.252 × (74 − 1)
sp2 = = 1.2728
45 + 74 − 2
22.3 − 19.7
t=   = 12.19107
1 1
1.2728 +
45 74
 
P t117 > 12.19107 = 7.22 × 10−23
Because our p-value is less than α, we reject the null hypothesis and con-
clude that the length of cuckoo and host meadow pipit eggs differ.
15. H0 : μD = 0, HA : μD > 0
264 Solutions to practice problems

15.97
16. t = √ = 5.7844
 15.86/ 33
17. P t32 > 5.7844 = 1.01 × 10−6
18. Because our p-value is less than α, we reject the null hypothesis and con-
clude that the blood lead levels are higher for children of parents exposed
to lead in their factory work.
19. H0 : μD = 0, HA : μD = 0
α = 0.05
0.1936
t= √ = 3.157
0.3066/
 25 
2 × P t24 > 3.15 = 0.00426
Because our p-value is less than α, we reject the null hypothesis and con-
clude there is a difference in the percentage of solids for shaded and ex-
posed grapefruits. However, the sample size and lack of Normality might
cause one to doubt the results.
20. H0 : μD = 0, HA : μD > 0
α = 0.1
2.617
t=√ √ = 2.14826
22.26/ 15
P (t14 > 2.14826) = 0.02484
Because our p-value is less than α, we reject the null hypothesis and con-
clude that cross-pollinated corn plants are taller than self-pollinated corn
plants. However, the small sample size and lack of Normality casts some
doubt on the analysis.

Chapter 16
1. We are 96% confident that (0.08112345, 0.1392416) covers the true value
of pCollege − pN oCollege .
2. We are 80% confident that (0.07300417, 0.1050089) covers the true value
of pDiabetes − pN oDiabetes .
3. We are 95% confident that (−0.1037826, 0.09569645) covers the true value
of pF rench − pEnglish .
4. We are 90% confident that (3.982428, 4.497572) covers the true value of
μMalignant − μBenign .
5. We are 96% confident that (−8.331023, 2.871023) covers the true value of
μF − μM .
6. We are 92% confident that (10.82324, 13.05676) covers the true value of
μS − μW .
7. We are 99% confident that (8.407745, 23.53225) covers the true value
of μD .
8. We are 99% confident that (−1.113974, −0.494026) covers the true value
of μD .
9. We are 80% confident that (−0.02914, 0.15314) covers the true values
of μD .
 Solutions to practice problems 265

10. Since 0 is in the 95% confidence interval, we would fail to reject the null
hypothesis.
11. Since 0 is not contained in the 90% confidence interval, we would reject
the null hypothesis.
12. Since 0 is not contained in the 95% confidence interval, we would reject
the null hypothesis.

Chapter 17
1. prop.test(x=c(7,10), n=c(117,118), alternative=“two.sided,” conf.level=0.87,
correct=FALSE)
2. prop.test(x=705, n=1500, p=0.5, alternative=“less,” correct=FALSE),
prop.test(x=705, n=1500, p=0.5, alternative=“two.sided,” conf.level=0.93,
correct=FALSE)
3. prop.test(x=c(171,124), n=c(212,541), alternative=“greater,” correct=
FALSE) prop.test(x=c(171,124), n=c(212,541), alternative=‘two.sided,” cor-
rect=FALSE)
4. prop.test(x=428, n=753, p=0.5, alternative=“greater,” conf.level=0.98, cor-
rect=FALSE), prop.test(x=428, n=753, p=0.5, alternative=“two.sided,”
conf.level=0.84, correct=FALSE)
5. t.test(DDT, mu=3, alternative=“greater”), t.test(DDT, mu=3, alternative=
“two.sided,” conf.level=0.99)
6. t.test(x=monica$age[monica$sex==“men”], y=monica$age[monica$sex==
“women”], alternative=“less”, var.equal=TRUE), t.test(x=monica$age[mon-
ica$sex==“men”], y=monica$age[monica$sex==“women”], alternative=
“two.sided,” conf.level=0.9, var.equal=TRUE)
7. t.test(x=cd4$baseline, y=cd4$oneyear, alternative=“less,” paired=TRUE,
var.equal=TRUE), t.test(x=cd4$baseline, y=cd4$oneyear, alternative=
‘two.sided,” paired=TRUE, var.equal=TRUE, conf.level=0.6)
8. t.test(x=Guyer$cooperation[Guyer$condition==“anonymous”], y=Guyer$co-
operation[Guyer$condition==“public”], alternative=“two.sided,”
var.equal=TRUE, conf.level=0.97)
9. t.test(x=pair65$heated, y=pair65$ambient, alternative=“two.sided,”
var.equal=TRUE)

Chapter 18
1. H0 : ρ = 0, HA : ρ = 0
α = 0.05
0.177
t= = 2.5433
1 − 0.1772
202 − 2
2 × P (t200 > |2.5433|) = 0.01173731
266 Solutions to practice problems

Because our p-value is less than α, we reject the null hypothesis and conclude
that the correlation between body mass index and white blood cell count in
athletes is not 0.
2. H0 : ρ = 0, HA : ρ = 0
α = 0.01
0.1857
t= = 2.58436
1 − 0.18572
189 − 2
2 × P (t187 > |2.58436|) = 0.0105183
Because our p-value is greater than α, we fail to reject the null hypothesis and
conclude that the correlation between a newborn’s weight and its mother’s
weight is plausibly 0.
3. H0 : ρ = 0, HA : ρ = 0
α = 0.05
0.6557
t= = 47.55122
1 − 0.65572
3000 − 2
2 × P (t2998 > |47.55122|) = 0
Because our p-value is less than α, we reject the null hypothesis and conclude
that the correlation between height and index finger length is not equal to 0.
4. cor.test(x=Weekly$Volume, y=Weekly$Today, alternative=“two.sided,”
conf.level=0.9)
5. We are 90% confident that (−0.08281261, 0.01682141) covers the true value
of ρ.
6. cor.test(x=University$acpay, y=University$resgr, alternative=“two.sided,”
conf.level=0.99)
7. We are 99% confident that (0.7564774, 0.9300306) covers the true value of
ρ. 0.7564774 0.9300306
8. cor.test(x=Caschool$testscr, y=Caschool$expnstu, alternative=“two.sided,”
conf.level=0.91)
9. We are 91% confident that (0.11018480.2698314) covers the true value of ρ.

Chapter 19
1. y = −17.579 + 3.932x
2. ˆ 2 = 0.042262 = 0.001786
y = 0.2068 + 0.0031x, sigma
3. ˆ 2 = 0.18612 = 0.03463
y = 0.154010 + 0.065066, sigma
4. ŷ = −17.579 + 3.932 × 30 = 100.381
5. ŷ = 23.94153 + 0.64629 × 68.5 = 68.2124
6. ŷ = 1676.3556 − 3.2261 × 60 = 1482.79
7. lm=lm(VitC data=HeadWt,cabbages)
plot(lmf itted.values, lmresiduals)
 Solutions to practice problems 267

Based on the plot generated by this code, it appears that the regression
assumptions hold.
8. As the p-value from our lm summary is 9.75 × 10−9 , we would reject the
null hypothesis and conclude that the slope coefficient in our regression is
not zero.
9. R 2 = 0.4355
10. As the p-value from our lm summary is less than 2.2 × 10−16 , we would
reject the null hypothesis and conclude that the slope coefficient in our
regression is not zero.
11. R 2 = 0.09688
12. lm=lm(sales price,data=Cigar)
plot(lmf itted.values, lmresiduals)
Based on the plot generated by this code, it appears that the regression
assumptions do not hold, specifically our assumption about the constant
variance of the errors.
13. As the p-value from our lm summary is less than 2.2 × 10−16 , we would
reject the null hypothesis and conclude that the slope coefficient in our
regression is not zero.
14. R 2 = 0.6387
15. lm=lm(ohms juice,data=fruitohms)
plot(lmf itted.values, lmresiduals)
Based on the plot generated by this code, it appears that the regression
assumptions do not hold, specifically our assumption about the relationship
between our predictor and response being linear.
Appendix B

List of R datasets
Note: Datasets may appear in multiple locations in the book. The listed chapter
is their first occurrence.

Chapter Dataset Library Numeric Categorical

variables variables
5 nlschools MASS Yes Yes
6 sleep Base R Yes Yes
6 faithful Base R Yes No
6 iris Base R Yes Yes
6 cats MASS Yes Yes
6 seedrates DAAG Yes No
7 quine MASS Yes Yes
7 Housing Ecdat Yes Yes
7 Pima.te MASS Yes Yes
7 cars Base R Yes No
7 leuk MASS Yes Yes
7 geyser MASS Yes No
7 birthwt MASS Yes Yes
7 Star Ecdat Yes Yes
12 JobSat vcdExtra No Yes
13 vlt DAAG Yes Yes
13 Vocab carData Yes Yes
14 DDT MASS Yes No
14 airquality Base R Yes No
14 Davis carData Yes Yes
14 chem MASS Yes No
15 Rossi carData Yes Yes
15 Moore carData Yes Yes
15 biopsy MASS Yes Yes
15 frostedflakes DAAG Yes No
continued on next page

269
270 List of R datasets

Chapter Dataset Library Numeric Categorical

variables variables
15 BloodLead PairedData Yes No
15 Grapefruit PairedData Yes No
16 appletaste DAAG Yes Yes
16 cd4 boot Yes No
16 chickwts MASS Yes Yes
16 Meat PairedData Yes No
16 Rugby PairedData Yes No
17 monica DAAG Yes No