2 An Introduction to

Data Analysis

CHAPTER OUTLINE LEARNING OBJECTIVES

Learning Objectives
• Discuss the goals and scope of
Overview the book.

ute
Motivating Data Analysis
• Explain the importance of
The Main Components of Data Analysis acquiring skills in data analysis.

trib
Developing Hypotheses by Describing Data
• List the components of data
Model Building and Estimation analysis and how they fit together.
Diagnostics
• Form hypotheses from

dis
Next Questions descriptions of data.
Summary
• Explain the connection between
Common Problems hypotheses, models, and
Review Questions
Practice on Analysis and Visualization or estimates.

• Define diagnostics and explain

st,
Annotated R Functions their role in data analysis.

Answers • Formulate new questions.

po
y,

Overview
op

Scholars, practitioners, and policymakers interested in explaining human behavior are drown-
tc

ing in data. This book is designed for those who want to swim safely to shore. The goal is to
introduce the method, logic, art, and practice of data analysis. Specifically, the book provides
no

the essential skills and tools necessary to examine data in the service of solving problems.
Toward that end, you will learn to marry the art and practice of data analysis with how experi-
enced practitioners approach problems, formulate hypotheses, estimate models, and present
Do

their results. The underlying philosophy is learning by doing. Learning statistics along with the
art and practice of data analysis is best achieved by doing data analysis.
This introductory chapter is organized as follows. First, I discuss some of the motivations
behind analyzing quantitative data. Second, I provide a short exploration of an important ques-
tion in the social sciences to illustrate the main components of data analysis. The example does
not fully explain each component. The goal here is to provide a road map for the rest of the book.
The book is organized according to the process of data analysis laid out in this chapter:
describing data and formulating hypotheses, building and estimating models, diagnostics,
and generating the next question. Presenting your results also forms an important part of this
book. The process of data analysis is iterative: it is an ongoing conversation that goes back and

Copyright ©2022 by SAGE Publications, Inc. 29

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
30 STATISTICS AND DATA VISUALIZATION USING R

forth between all of the different components. For example, the estimates we generate from our
statistical models rarely settle the matter. New and better questions are often the result.

Motivating Data Analysis

The motivations are many, but here are three of the most rehearsed. First, more things are
being measured and quantified. Not only are more data available on demographic and finan-
cial trends, our daily activity (driving, dating, shopping, and listening to music) is available for
analysis. Second, and partly related to the first, data analysis is in demand. Consequently, the
skills associated with analyzing quantitative data are marketable. Third, data analysis helps us

ute
separate facts from fiction, a defining characteristic of successful democratic societies.

Big Data Are Getting Bigger

trib
A significant part of our lives is spent on the Internet. The bread crumbs left behind are used
by government, private industry, and others. Politicians are driven (in part) by survey results.
CEOs make decisions based on data collected from their employees and their customers. Here

dis
are just a few examples of data analysis that reveal both its ubiquity and effectiveness.

1. Noticing an increasing amount of junk mail from Target advertising diapers, baby

or
formula, and onesies, the inundated recipient calls Target to ask why they’re sending
him so much baby-oriented advertising (it had been years since there was a baby in the
house). Target explained that recent purchasing data indicated there was a pregnant
st,
woman in his household. One week after calling Target, he discovers his daughter is
pregnant.
po

2. By constructing a meta-analysis of polls, Nate Silver (in his FiveThirtyEight column)

correctly predicted the outcome of the presidential race in 2012 in all 50 states.
y,

3. When diseases spread (e.g., the Ebola outbreaks in Africa), the U.S. Centers
for Disease Control and Prevention and other authorities previously relied on
op

­2-week-old data collected from hospitals and clinics. In a quickly developing

scenario like the H1N1 (bird flu), Ebola, Zika, or COVID-19 virus outbreaks, 2
tc

weeks is too late. A team from Google let the Internet figure out where outbreaks
occur. To develop their model, they traced the spread of H1N1 and correlated it
with search terms on the Internet: high fever, cough, and aches. They eventually
no

found 45 phrases that allowed them to track the outbreak, informing officials
when and where, exactly, the newest flu outbreak was occurring.1
Do

4. In an article entitled “China Invents the Digital Totalitarian State” (2016), The
Economist details the plans of a new Chinese program called the social-credit
project designed to increase trust in society by making social-credit scores available.
This example, more than the others, illustrates how big data can be abused. In what
is called the “judgment defaulter’s list” (a tally of citizens who have defied a court
order), Chinese authorities keep track of who they consider to be trustworthy in
society to help spur economic cooperation and trust among its citizens.

1
The examples above can be found in Big Data: A Revolution That Will Transform How We Live, Work, and Think by
Kenneth Cukier and Viktor Mayer-Schönberger (2013). Google’s attempts to track flu outbreaks have not come without
criticism. For a nice discussion of the shortcomings associated with big data, see the discussion in Lazer et al. (2014).

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 31

5. To help reduce crime in Chicago, a city plagued by high homicide rates, new data and
data analysis software help to reduce crime (“Violent Crime Is Down in Chicago,”
2018). Equipped with sensors that locate gunshots throughout the city, along with
maps of liquor stores and freeway on-ramps, analysts identify areas where crime will
most likely occur. That information, combined with data on televised sporting events,
increases the accuracy of locating potential problems. The new data allow police in
Chicago’s inner-city areas to patrol more effectively.

Whether in commerce, politics, government, health, or crime, data are the key to solving
important problems. These examples show that data are used increasingly to understand and

ute
influence human behavior or in the Chinese case exert social control. Those who know how to
collect, analyze, and explain data can influence important decisions.

Data Analysis Is a Marketable Skill

trib
Quantitative data analysis is a marketable skill. Confronted by an ocean of data, com-
panies, government offices, and nongovernmental organizations recognize the need for

dis
accurate and timely analysis. One need only to Google the phrase “jobs in R” to see the
possibilities.
In the private sector, an increasing number of companies rely on surveys to understand
their market, employees, and customers. The expanded use of Survey Monkey and Qualtrics

or
adds to that reliance. Conducting surveys is an art. Done poorly, surveys either mislead or say
nothing about the problem at hand. Understanding the basics of quantitative data analysis
st,
develops an appreciation and facility for survey research.
Government agencies need well-trained social scientists. Ten or 15 years ago, government
po

agencies would only pay lip service to their need for them. Over the last several years, they’ve
actually started to hire social scientists. Data skills will bolster a career in local, state, and fed-
eral government agencies; organizations know that a work force staffed with effective analysts
y,

produces better policy.

Postgraduate programs want students with a quantitative background. A firm grasp of
op

quantitative skills can help gain entrance into many top 20 departments, no matter the disci-
pline. Law school candidates with a specialty in engineering, computer science, or quantita-
tc

tive analysis have a unique skill that puts them head and shoulders above others, an important
­consideration given the competition.
In addition to law school, quantitative skills can open doors to policy schools and graduate
no

programs in the social sciences. The number of subjects we study in the social sciences that
involve examining quantitative evidence is growing. Over the last few decades, there’s just too
much quantitative data and evidence to be ignored. Applicants to graduate programs in these
Do

areas increase the likelihood of gaining entrance to top-ranked schools if they can demonstrate
a facility with data analysis.

Data Analysis Is a Public Good

Perhaps more important than these purely instrumental reasons, society depends on good
analysis. As wonderfully illustrated in the book A Mathematician Reads the Newspaper,
we deserve the kind of public policy we get (Paulos, 1995). In one example, Paulos recounts
a ­contested state senate race in Pennsylvania where fraud was detected by fitting a regres-
sion line to previous election results and determining that the race in question was simply
too anomalous to be credible. Given the current divisiveness that characterizes U.S. politics,
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
32 STATISTICS AND DATA VISUALIZATION USING R

understanding sampling and the inferences we can draw from it can either bolster our faith in
the results or suggest a new election is necessary.
Another characteristic of our political age is the growth of conspiracy theories. With so
much data floating around, it is easy to connect dots that are completely unrelated yet convince
many that dark forces are at work. In a prescient example, Paulos recounts an exercise under-
taken by John Leavy, a computer programmer from the University of Texas. Leavy fed data into
a computer to come up with similarities between different pairs of U.S. presidents. In that exer-
cise, he found some very interesting connections between Presidents William McKinley and
James Garfield who were both assassinated. Observe the similarities:

ute
It turns out that both of these presidents were Republicans who were born and
bred in Ohio. They were both Civil War veterans, and both served in the House
of ­Representatives. Both were ardent supporters of protective tariffs and the gold

trib
­standard, and both of their last names contained eight letters. After their assassina-
tions they were replaced by their vice presidents, Theodore Roosevelt and Chester Alan
Arthur, who were both from New York City, who both sported mustaches, and who

dis
both had names c­ ontaining seventeen letters. (Paulos, 1995, p. 91)

Whether citizens become data scientists or sophisticated consumers of quantitative data,

a more informed public is better able to determine fact from fiction. An important goal of this

or
book is to demonstrate how statistics are used and abused. Distinguishing between fact and fic-
tion depends on individuals with the skill and knowledge this book hopes to provide.
st,
KNOWLEDGE CHECK: Explain the importance of acquiring skills
po

in data analysis.
y,

1. Given big data’s growing influence on our lives, why is data analysis an important skill?
op

a. It helps in our careers.

b. We get better policy.
tc

c. It defines our roles as citizens and consumers.

d. With more skills we can realize higher salaries.
no

2. Data analysis as a skill (a private good) is helpful with which of the following?
a. We can realize a higher salary.
Do

b. Both government and the private sector need skilled individuals.

c. It helps with gaining entrance into postgraduate degree programs (e.g., law school,
business school, graduate school in general).
d. It helps society.
3. As a public good, which of the following describe the benefits of having data analytic skills.
a. It leads to better government policy.
b. It helps society determine fact from fiction.
c. It aids democracy.
d. It helps businesses understand their customers.
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 33

The Main Components of Data Analysis

There is no shortage of good textbooks on how to conduct social science research in which the
focus is on the process of hypothesis formation, theory testing, and drawing inferences. Here,
the goal is the same but the approach is different. In this book we learn those foundations of
inquiry through data visualization and by working on real problems with real data.
In that spirit, this section outlines four components of data analysis in the order they
should occur: (1) describing data and formulating hypotheses, (2) building and estimating
models, (3) diagnostics, and (4) generating the next question. There are concepts and tech-
niques (i.e., model building and estimation, transforming variables, diagnostics, etc.) that will

ute
be new. The purpose of this chapter is to introduce the broad outlines of good data analysis
with an example. When new techniques and concepts are introduced, I will indicate where
they are covered in detail later in the book rather than dwell on them here. Let me define the

trib
main components of data analysis.

1. Describing data and formulating hypotheses

dis
We describe data to better understand the problem and to ask better questions. At its base,
describing data focuses primarily on identifying the typical case (central tendency) and under-

or
standing how typical that typical case is (dispersion). Describing data, however, should go
much deeper than that. Observing where specific cases or entire classes of cases lie in relation
to others is an important part of the enterprise. The more we know about our data, the better
st,
questions we’ll generate and the better hypotheses we’ll formulate. The concepts and tools nec-
essary for describing data are found in Chapters 3–7.
po

Before we continue, here is a word about theories and hypotheses. Typically, hypotheses
are guesses and theories are interconnected hypotheses that form a greater whole and have
been tested with some success. The distinction is often fuzzy since in the vernacular theories
y,

and hypotheses are often used interchangeably. In this book, a hypothesis will refer to a specific
guess about how two things are related (e.g., religiosity and political ideology). A theory will
op

refer to a set of related hypotheses that explain successfully some empirical phenomenon.
tc

2. Building and estimating models

Once we have some familiarity with our data and some possible explanations have been
no

forwarded, we move on to building and estimating models. Models are simplified versions
of reality that help us understand our complex world. Models can be thought of as arguments
or explanations. They are arguments we make to explain an empirical problem or puzzle. For
Do

example, if we want to explain why some countries have high rates of homicide, we construct
a model or argument that might include income, age of the population, number of police, and
efficacy of the judicial system. There are a multitude of other possible causes we could include,
but it helps to keep things simple. We don’t want to recreate reality; we merely want to approx-
imate it. With a good model in hand (i.e., a model that contains the main causes but not every
single possible one), we can begin to understand how important each cause is and we can esti-
mate its impact.
Estimates can be very sensitive to the model we choose. How we look at the data ­influences
what we see. Consequently, model building and estimation is a process that should be
­performed as a back-and-forth between theory and evidence. The practice of model building is
treated in Chapters 10–14.
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
34 STATISTICS AND DATA VISUALIZATION USING R

3. Diagnostics
After we’ve constructed models and obtained some estimates, we turn to diagnostics.
Diagnostics are a set of tools we use to determine whether we’re using the right kind of model.
To ascertain whether our model is appropriate, we examine how well the predictions from our
model match reality. The difference between our prediction and reality is called the residual.
For example, if our model does a good job of predicting infant mortality rates in all countries
except for the oil-rich states of the Middle East, the resulting diagnostics will say so. That is,
the residuals for these cases will be relatively large. Perhaps our model estimates are overly
influenced by those Middle Eastern countries. Diagnostics help us determine whether our
estimates provide a good sense of how the world really works, are the product of some strange

ute
cases, or are the result of a poorly chosen model.
It is important to keep in mind that diagnostics can both detect problems and help uncover
interesting relationships, generating additional explanations or hypotheses. Diagnostics are

trib
explained in Chapters 15 and 16.
4. Generating the next question

dis
Finally, armed with our estimates and a sense of how well our predictions fit reality, addi-
tional questions and ideas inevitably surface. A useful way to construct or identify these
­additional questions and ideas is to use a simple if-then statement: if the estimates we obtained

or
are correct, then we would expect to see x. Following each set of estimates with that statement
helps to unearth possible explanations and additional hypotheses to test. Since it is impossible
st,
to prove anything with complete certainty, the exercise of generating additional hypotheses to
test is extremely important. Much like a prosecutor in a court of law, we must provide as much
­circumstantial evidence as possible to convince the jury. The more evidence we provide, the
po

more likely the jury will find the defendant guilty. The Appendix presents a way to approach
generating additional hypotheses with a few tips and tricks.
Now that we have a sense of what the main components of data analysis are, let’s dive into
y,

an example of what they look like in real life. The next section serves as an introduction to data
op

analysis by providing a very brief exploration of a problem, a puzzle that has occupied the
attention of the social sciences for decades.
tc

KNOWLEDGE CHECK: List the components of data analysis and

no

how they fit together.

Do

4. Which are components of data analysis?

a. Generating and forming hypotheses
b. Diagnostics
c. Describing the data
d. Collecting the data
5. What is the correct order of the components of data analysis?
a. Estimating models, forming hypotheses, diagnostics, forming new questions
b. Describing data, forming hypotheses, estimating models, diagnostics, forming
new questions
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 35

c. Forming hypotheses, describing data, estimating models, diagnostics, forming

new questions
d. Describing data, forming hypotheses, forming new questions, diagnostics
6. Which of the following are accurate statements about data analysis?
a. There is an agreed-upon method everyone follows.
b. Hypothesis formation always comes before describing data.
c. Describing data always comes before hypothesis formation.
d. There is an agreed-upon order of analysis.

ute
7. What is the difference between a theory and hypothesis?
a. There really is no difference.

trib
b. Hypotheses are guesses, theories are certain.
c. Theories are simply hypotheses that have stood the test of time.

dis
d. Theories define an integrated approach or set of hypotheses that have accumulated
corroborating evidence over time.
8. Which statements best describe models?
a. A model replicates reality.
b. A model is a simplification of reality. or
st,
c. A model is best judged by its ability to predict.
d. A model’s use is defined both by its ability to predict and its simplicity.
po

Developing Hypotheses by Describing Data

y,

The best analysis starts with a puzzle or a question. To illustrate the broad outlines of data analy-
op

sis, let’s begin with a tried-and-true example. Why are some countries richer than others? Work
drawn from development economics, political science, and sociology emphasizes education,
religion, ethnic fractionalization, gender equality, and political stability.2 These explanations rep-
tc

resent only some of the more common hypotheses scholars have forwarded to account for the
huge disparity we observe between countries. Let’s explore these hypotheses in a little more detail.
no

Art and Practice of Data Visualization

Do

D E S C R I B I N G DATA A N D F O R M U L ATI N G
H Y P OT H E S E S

Describing data and formulating hypotheses is an interactive process. Data

analysis never starts without some guess about what we’ll find. An honest
analysis, however, changes what we know. At its best, data analysis leads to
new questions we neglected to ask in the first place.

2
For an informative treatment of growth theory and the current scholarship on the empirics of growth, see Jones (2015).
For a recent review of the literature on long-run growth trends, see Lloyd and Lee (2018).
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
36 STATISTICS AND DATA VISUALIZATION USING R

Hypothesis I: Ethnolinguistic Fractionalization

What is the relationship between a country’s ethnic landscape and economic growth? Some have
argued there is a direct relationship between economic development and the number of ethnic
groups in a country. Endowed with a multitude of different languages, religions, and beliefs, indi-
viduals in heterogeneous societies may find it more difficult to organize production and markets.
Rivalries between different ethnic groups (ethnic conflict) can also inhibit economic development.
To explore the first hypothesis, we now move to more involved R code. Remember that the
best way to work with R is to save chunks of code so that they can be copied and pasted when
needed. Once you’ve established how you want a scatter plot or any other figure to look, it’s easy
to switch out variables, choose colors, or change a figure’s underlying features.

ute
So far, you’ve been introduced to just a few simple base R functions, hist() and plot(). To
get you up and running with publishable figures and tables, let’s dive in. Throughout the book,
I use a package in R called ‘ggplot2.’ The “gg” in ggplot stands for the grammar of graphics, a

trib
way of thinking about graphs based on layers. Let me explain.
The code in Code Chunk 2-1 starts with the ggplot() function. Within the parentheses,
we first indicate which data set to use: world. The next feature of the code is the aes() func-

dis
tion or “aesthetics.” The aesthetics indicate which variables to use and any shapes, colors, or
labels to add. Once that first layer is established, note that the rest of the code is simply a series
of additional layers. First, we add a layer of points with geom_point(). We then add a straight

or
line with geom_smooth(), a title with ggtitle(), x- and y-axis labels with xlab and ylab, a style
with theme_minimal() and theme(), and some text with geom_text_repel(). The code within
st,
the geom_text_repel() function is a little complex because we want R to only label the two data
points that represent Luxembourg (LUX) and Madagascar (MDG). Notice how each layer is
added by placing an addition symbol between each added layer. I encourage you to start with
po

the ggplot() function, execute it, and observe the result. Then, add each additional layer one by
one, observing how the plot evolves into the final product. That exercise will demonstrate the
logic of the grammar of graphics: the simple addition of layer upon layer.
y,

The code in Code Chunk 2-1 draws a nice scatter plot that illustrates the relationship
op

between ethnic heterogeneity and the logged value of gross domestic product (GDP) per capita.
tc

Code Chunk 2-1

ggplot(world, aes(ethfrac, log(gdppc), label = iso3c)) +
no

    geom_point(color=“#bf0000”) +
    geom_smooth(method=“lm”, se=FALSE, color=“#0000bf”) +
    ggtitle(“Figure 2-1: ELF Lowers Income”) +
Do

    ylab(“GDP per Capita (logged)”) +

    xlab(“Ethnolinguistic Fractionalization (ELF)”) +
    theme_minimal() +
    theme([Link] = element_text(size = 8, face = “bold”),
    [Link] = element_text(size = 8, face = “bold”)) +
    geom_text_repel(size = 2.8,
aes(label=ifelse(iso3c==“MDG” |
iso3c==“LUX”,
[Link](iso3c),’’),
   hjust = 0, vjust=-1), [Link]=FALSE)
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 37

FIGURE 2-1 ELF Lowers Income

Lux

11
GDP per Capita (logged)

10

ute
9

trib
8

MDG
7

dis
0.00 0.25 0.50 0.75
Ethnolinguistic Fractionalization (ELF)

or
st,
In Figure 2-1, I use GDP per capita to measure wealth (income). GDP is simply the total
amount of all goods and services in a country produced in a given year. Put simply, ethnolinguis-
po

tic fractionalization (ELF) indicates how likely your neighbor speaks a different language. The
higher the score, the more languages spoken over a country’s population. The scatter plot of GDP
per capita and the degree of ethnic heterogeneity3 suggests that as ethnic heterogeneity increases,
y,

GDP per capita decreases. Scatter plots show where a single observation (a country, in this case)
is situated on a two-dimensional grid delineated by two variables (GDP per capita [logged] and
op

ethnolinguistic fractionalization). Scatter plots will be discussed at greater length in Chapter 5.

Note that while there is an overall negative relationship (the line slopes downward), know-
tc

ing the degree of ethnic heterogeneity in a country does little to help us predict the income
level with precision. For example, Madagascar (MDG) and Luxembourg (LUX) have roughly
the same ethnolinguistic heterogeneity score (ELF) but extremely different levels of per capita
no

income. Even though knowing the ELF score can give us a rough guess, there is still consider-
able variation in GDP per capita left to explain.
Although there does seem to be a negative relationship between the ELF score and GDP
Do

per capita, a downward sloping line may not be the best way to summarize the relationship.
A closer look at the plot indicates there is a cluster of countries that might be influencing the
line (the upper-left corner). If those cases were not there, it’s possible the line would have a flat-
ter slope. That cluster of countries might be pulling the left part of the line upward, resulting
in a negative slope. It turns out that those countries fall into a particular category labeled the

3
Per capita income is measured by GDP per capita. The GDP per capita measure is presented in its logged form for
reasons that will be explained in the chapter on transforming variables (Chapter 6). The measure of ethnicity used
is the ethnolinguistic fractionalization score constructed by Alesina et al. (2003). In earlier formulations, the score
was constructed to represent the probability that two randomly selected individuals in a country were from the same
ethnolinguistic group. Alesina and coauthors use the same underlying construct but combine racial and linguistic
characteristics. These data are from Pollock (2014).
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
38 STATISTICS AND DATA VISUALIZATION USING R

industrial countries, a collection of countries that includes Europe, North America, Australia,
New Zealand, and Japan.
In the following code, note how some changes to Code Chunk 2-1 allow me to color the
industrial countries blue, turn the original line into a dashed line, and fit a separate line (using
the smooth() function) to the nonindustrialized countries (those colored red). I can also place
some words in the figure with the annotate() function. We’re not going to get bogged down
with all of the code at this point, since we want to concentrate our efforts on the components
of data analysis. As the chapters unfold, you will become well acquainted with all of the code in
Code Chunk 2-2.

ute
Code Chunk 2-2
ggplot(world, aes(ethfrac, log(gdppc))) +

trib
   geom_point(color=ifelse(world$aclpregion==”Industrial Countries”,
“#0000bf”, “#bf0000”)) +
   geom_smooth(method=“lm”, col=“grey”, se=FALSE,

dis
linetype = “dashed”) +
   ggtitle(“Figure 2-2: Industrialized Countries Are Clustered”) +
   ylab(“GDP per Capita (logged)”) +
   xlab(“Ethnolinguistic Fractionalization”) +
   theme_minimal() +
or
st,
   theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”)) +
   geom_smooth(data=subset(world,
po

aclpregion != “Industrial Countries”),

color=“#bf0000”, se=FALSE, method = “lm”) +
   annotate(“text”, x = .15, y = 11.5,
y,

   label = “Industrial Countries”, col=“#0000bf”)

op
tc

Figure 2-2 demonstrates how sensitive the line is to the influence of the industrial coun-
tries. Note how the line fit for the rest of the world (the red line) is relatively flat compared to the
gray dashed line that accounts for the industrial countries.
no

To further emphasize the influence of the cluster, I asked R to fit a curved line to all of
the countries in the world that are not classified “industrial.” Smooths will be discussed and
defined in detail later (Chapter 5). For now, think of a smooth as simply a line or curve that R
Do

fits to the data. The curve demonstrates there may be another pattern lurking. In Figure 2-3, the
black line shows that when we describe the data with a nonlinear smooth (the curve), an inter-
esting pattern emerges. While the overall trend might be roughly negative (gray dashed line),
a majority of the data (the nonindustrialized countries) tell a different story. As we move from
left to right along the x axis, GDP per capita increases, reaches a maximum, then decreases.
Among nonindustrialized countries, there may be an optimal degree of ethnolinguistic hetero-
geneity that encourages economic dynamism.
In Code Chunk 2-3, I’ve only replaced the straight red line in Figure 2-2 with a curved
black line. The curved line is the default for geom_smooth(), so I simply removed the expression
method=“lm” that tells R to draw a straight line.
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 39

FIGURE 2-2 Industrialized Countries Are Clustered

Industrial Countries
11
GDP per Capita (logged)

10

ute
9

trib
8

dis
0.00 0.25 0.50 0.75
Ethnolinguistic Fractionalization (ELF)

or
st,
Code Chunk 2-3
po

ggplot(world, aes(ethfrac, log(gdppc))) +

  geom_point(color=ifelse(world$aclpregion==”Industrial Countries”,
“#0000bf”, “#bf0000”)) +
y,

  geom_smooth(method=“lm”, col=“grey”, se=FALSE,

  linetype = “dashed”) +
op

  ggtitle(“Figure 2-3: The Relationship Could Be Nonlinear”) +

  ylab(“GDP per Capita (logged)”) +
tc

  xlab(“Ethnolinguistic Fractionalization”) +
  theme_minimal() +
  theme([Link] = element_text(size = 8, face = “bold”),
no

[Link] = element_text(size = 8, face = “bold”)) +

  geom_smooth(data=subset(world,
aclpregion != “Industrial Countries”),
Do

color=“black”, se=FALSE) +
  annotate(“text”, x = .15, y = 11.5,
   label = “Industrial Countries”, col=“blue”)

The curved pattern depicted by the black line might lead us to conclude that if there is a
severe lack of ethnic heterogeneity or if there is too much, economic development suffers.
While the observation has some merit, an even closer look indicates the upside-down
U-shape pattern may be somewhat overblown. Note in Figure 2-4 that the four most impov-
erished and ethnolinguistically homogeneous countries are all poor, very small island
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
40 STATISTICS AND DATA VISUALIZATION USING R

FIGURE 2-3 The Relationship Could Be Nonlinear

Industrial Countries
11

GDP per Capita (logged)

10

ute
9

trib
8

dis
0.00 0.25 0.50 0.75
Ethnolinguistic Fractionalization (ELF)

or
st,
FIGURE 2-4 Small Countries Have Low GDP per Capita
po

11
y,
GDP per Capita (logged)

op

10
tc

9
no

8
KIR SLB
Do

HTI
7 COM

0.00 0.25 0.50 0.75

Ethnolinguistic Fractionalization (ELF)

nations: Comoros (COM), Haiti (HTI), Kiribata (KIR), and the Solomon Islands (SLB).
Given the small size of these countries and those in close proximity, perhaps size matters. 4

4
There is a literature in economics that examines whether there is an optimal state size for economic performance. For one
contribution in that debate, see Alesina (2003).

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 41

Among the developing countries, the smallest (regardless of heterogeneity) all have relatively
low per c­ apita incomes.
To highlight the four cases, I use the ifelse() function both to color them differently and to
include their labels.

Code Chunk 2-4

ggplot(world, aes(ethfrac, log(gdppc))) +
   geom_point(color=ifelse(world$iso3c==“COM” |

ute
world$iso3c==“HTI” |
world$iso3c==“KIR” |
world$iso3c==“SLB”, “#bf0000”, “grey”)) +

trib
   geom_smooth(method=“lm”, col=“grey”, se=FALSE,
linetype = “dashed”) +
   ggtitle(“Figure 2-4: Small Countries Have Low GDP per Capita”) +

dis
   ylab(“GDP per Capita (logged)”) +
   xlab(“Ethnolinguistic Fractionalization”) +
theme_minimal() +

or
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”)) +
geom_smooth(data=subset(world,
st,
aclpregion != “Industrial Countries”),
color=“black”, se=FALSE) +
po

geom_text_repel(size = 2.8, aes(label=ifelse(iso3c==“COM” |

iso3c==“HTI” |
iso3c==“KIR” |
y,

iso3c==“SLB”,
[Link](iso3c),’’)))
op
tc

Since we may want to avoid summarizing the relationship between GDP per capita and ELF
based on the small island nations of Comoros, Kiribata, Haiti, and the Solomon Islands, there
no

does seem to be a negative relationship between GDP per capita (logged) and the ELF score as
we travel from left to right across the x axis. There is, in other words, some preliminary evidence
that the hypothesis has some merit. Nevertheless, a great deal of variation remains and the rela-
Do

tionship is by no means perfectly linear. By linear, I mean a straight line with a single slope.
Before we move on to exploring the next hypothesis, let’s take stock to highlight the pro-
cess of formulating hypotheses and describing data. We started the section with a hunch about
the ethnic makeup of a country and its economy. A simple scatter plot showed there might
be a negative relationship between the two. We identified a cluster of countries that explains
the downward sloping line. We then fit a curve to the remaining data and found there was an
upside-down U-shaped pattern: countries with low and high levels of ethnolinguistic hetero-
geneity had lower levels of income compared to countries with medium levels. Further inves-
tigation revealed that the countries with low levels of income and low heterogeneity were small
island nations. That observation suggested the geographic size of a country might matter.
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
42 STATISTICS AND DATA VISUALIZATION USING R

Throughout this section, we saw how identifying different clusters of data led to new discover-
ies and suggested additional hypotheses to test.

Hypothesis II: Women’s Suffrage

Another hypothesis concerns women and their role in both government and the economy. Are
women being educated? Do they have equal rights? We might hypothesize that women’s par-
ticipation in society is essential to economic development. As a very crude measure of wom-
en’s participation, we have data—a variable in the world data set—that record the year women
obtained the right to vote. What is the relationship between the year women gained suffrage
and GDP per capita? To answer the question, consider the following scatter plot that plots GDP

ute
per capita against the year women gained suffrage in the country (Figure 2-5).

trib
Code Chunk 2-5
ggplot(world, aes(womyear, log(gdppc))) +
geom_point(col=“#bf0000”) +

dis
geom_smooth(method=“lm”, se=FALSE, color=“#0000bf”) +
ggtitle(“Figure 2-5: The Later the Suffrage,
the Poorer the Country”) +

or
ylab(“GDP per Capita (logged)”) +
xlab(“Year Women Gained Suffrage”) +
theme_minimal() +
st,
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”))
po
y,

Although GDP per capita can take on very different values at each level of women’s
suffrage (there is considerable variation), there is a downward sloping trend. In other
op

words, as the year when women were granted suffrage increases (suffrage was granted
later), it appears that GDP per capita decreases. What is driving the result? Perhaps the
tc

downward sloping line merely describes the difference between the industrialized West
and Africa. The vote was not extended to anyone in Africa, after all, until independence
no

was achieved in the 1960s.

In Code Chunk 2-6, only a few refinements are added to the code in Code Chunk 2-5.
First, we use the ifelse() command to color the industrial countries blue and the sub-Saharan
Do

countries red. Second, we use the annotate() function to place the “Industrial Countries” and
“Sub-Saharan Africa” labels in the figure.
If the relationship is simply the difference between the industrial countries5 and sub-Saha-
ran Africa, there may be things other than women’s participation that distinguish the industrial

5
There is no good way to name this group of countries: the West, the industrialized West, developed countries, postindus-
trialized, high-income countries, and so forth. Rather than enter into a debate on nomenclature, I simply use the classifica-
tion given by Przeworski et al. (2000) whose data are used here. The countries considered industrial are Australia, Austria,
Belgium, Canada, Switzerland, Cyprus, Germany, Denmark, Spain, Finland, France, United Kingdom, Greece, Ireland,
Italy, Japan, Lichtenstein, Luxembourg, Malta, Netherlands, Norway, New Zealand, Portugal, San Marino, Sweden, and
the United States.

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 43

FIGURE 2-5 The Later the Suffrage, the Poorer the Country

11
GDP per Capita (logged)

10

ute
9

trib
8

dis
1900 1925 1950 1975
Year Women Gained Suffrage

or
st,
Code Chunk 2-6
ggplot(world, aes(womyear, log(gdppc))) +
po

geom_point(color=ifelse(world$aclpregion==
“Industrial Countries”, “#0000bf”,
ifelse(world$aclpregion==
y,

“Sub-Saharan Africa”, “#bf0000”, “grey”))) +

annotate(“text”, x = 1905, y = 11.75,
op

label = “Industrial Countries”,

col=“#0000bf”) +
tc

annotate(“text”, x = 1975, y = 6.2,

label = “Sub-Saharan Africa”, col=“#bf0000”) +
no

geom_smooth(method=“lm”, se=FALSE) +
ggtitle(“Figure 2-6: The Industrial Countries and
Sub-Saharan Africa Form Clusters”) +
Do

ylab(“GDP per Capita (logged)”) +

xlab(“Year Women Gained Suffrage”) +
theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”))

countries and sub-Saharan Africa from each other (e.g., a colonial past). The cluster of indus-
trial countries in the upper-left quadrant of the plot and the cluster of sub-Saharan African
countries in the lower-right quadrant certainly suggests that the date of independence or colo-
nialism is part of this story (see Figure 2-6).
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
44 STATISTICS AND DATA VISUALIZATION USING R

FIGURE 2-6 The Industrial Countries and Sub-Saharan Africa Form Clusters

12
Industrial Countries

GDP per Capita (logged)

10

ute
8

trib
dis
Sub-Saharan Africa
6
1900 1925 1950 1975
Year Women Gained Suffrage

or
st,
Although we’ve made some important discoveries so far, let’s not stop here. Part of the
challenge and fun of data analysis is to develop a keen eye that notices every nook and cranny.
po

Many times, that’s how the most interesting discoveries are made. Note, for example, there are
relatively few observations (dots) between 1935 and 1945 on the x axis. This implies that coun-
tries stopped granting suffrage during those years.
y,

In Code Chunk 2-7, I introduce the geom_vline() function, which creates a vertical line on
op

a specific part of the x axis. This is helpful when we want to highlight a very specific value.
tc

Code Chunk 2-7

no

ggplot(world, aes(womyear, log(gdppc))) +

geom_point(color=ifelse(world$womyear < 1935, “#bf0000”,
ifelse(world$womyear > 1944,
Do

“#0000bf”, “grey”))) +
geom_smooth(method=“lm”, se=FALSE) +
geom_vline(xintercept=1934, col=“grey”, linetype = “dashed”) +
geom_vline(xintercept=1944, col=“grey”, linetype = “dashed”) +
ggtitle(“Figure 2-7: Gap During World War II”) +
ylab(“GDP per Capita (logged)”) +
xlab(“Year Women Gained Suffrage”) +
theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”))

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 45

The pause between 1935 and 1945 might be a feature of the data worth exploring. The
gap during the period suggests that the fight for women’s rights took a back seat to fight-
ing World War II (see Figure 2-7). To illustrate, I colored the points before 1935 red and the
points after 1945 blue. After the war, a number of countries joined the community of nations

FIGURE 2-7 Gap During World War II

ute
11
GDP per Capita (logged)

trib
10

dis
8

7
or
st,
1900 1925 1950 1975
Year Women Gained Suffrage
po
y,

Code Chunk 2-8

ggplot(world, aes(womyear, log(gdppc))) +
op

geom_point(color=“#bf0000”) +
geom_smooth(method=“lm”, se=FALSE, color=“#0000bf”) +
tc

ggtitle(“Figure 2-8: Large Variation in GDP per Capita”) +

ylab(“GDP per Capita (logged)”) +
no

xlab(“Year Women Gained Suffrage”) +

theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
Do

[Link] = element_text(size = 8, face = “bold”)) +

geom_text_repel(size = 2.8,
aes(label=ifelse(iso3c==“ETH” |
iso3c==“BDI” |
iso3c==“SLE” |
iso3c==“HUN” |
iso3c==“CYP” |
iso3c==“LBY”,
[Link](iso3c),’’),
hjust = 0, vjust=-1), [Link]=FALSE)

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
46 STATISTICS AND DATA VISUALIZATION USING R

as independent entities. At their birth, many granted voting rights to women. Clearly, GDP per
capita could be driven by the political autonomy of a country rather than women’s rights. Both
the two clusters of countries and the gap separating pre- and postwar cases suggest that colo-
nialism and independence explains levels of GDP per capita.
Another feature of the relationship involves the huge variation in per capita income for
countries that granted suffrage at around the same time. By huge variation, I mean that at very
similar dates when suffrage was granted, levels of GDP per capita range from being fairly high
to fairly low. Note that Ethiopia (ETH), Burundi (BDI), and Sierra Leone (SLE) granted suf-
frage at around the same time as Hungary (HUN), Cypress (CYP), and Libya (LBY), yet the
two groups of countries have very different levels of per capita income (Figure 2-8). Clearly,

ute
women’s suffrage alone does not explain everything. If it did, all of the countries plotted in the
graph would be much closer to the line.
Refer back to Figure 2-6, which had the industrial countries in blue, the sub-­Saharan

trib
­African countries in red, and the rest of the world’s regions in gray. Given the pattern we
observed between the industrial countries and Africa in Figure 2-6, a very useful test might be
to fit a line to the sub-Saharan countries separately to see if a relationship exists between wom-

dis
en’s suffrage and income per capita for sub-Saharan Africa. By doing so, we’re asking whether
the negative relationship between GDP per capita and the year of suffrage is indeed negative
or if it is purely generated by the differences between the industrial countries and sub-Saharan

or
Africa. If there is a strong negative relationship between GDP per capita and women’s suffrage,
then we would expect to find it among the sub-Saharan African countries themselves.
To distinguish different groups in a figure, it’s often helpful to create a variable that indicates
st,
a 1 for a specific group and a 0 for all of the rest. The two first lines in Code Chunk 2-9 create
such a variable for sub-Saharan Africa. First, I use the ifelse() function to state that if the variable
po

world$region equals sub-Saharan Africa, then generate a 1; otherwise generate a 0. I then take that
new variable world$SSA and convert it to a factor. I’ll go into more detail on factors in Chapter 5.
y,

FIGURE 2-8 Large Variation in GDP per Capita

op
tc

11
no

CYP
GDP per Capita (logged)

HUH
10 LBY
Do

SLE
7
ETH BDI
1900 1925 1950 1975
Year Women Gained Suffrage

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 47

Code Chunk 2-9

world$SSA <- ifelse(world$region==“Sub-Saharan Africa”, 1, 0)
world$SSA <- [Link](world$SSA)

ggplot(world, aes(womyear, log(gdppc), col=SSA)) +

geom_point() +
theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”)) +

ute
geom_smooth(method=“lm”, se=FALSE) +
ggtitle(“Figure 2-9: No Relationship in Sub-Saharan Africa”) +
ylab(“GDP per Capita (logged)”) +

trib
xlab(“Year Women Gained Suffrage”) +
scale_color_manual(values=c(“lightblue”, “#0000bf”)) +
theme([Link] = “none”) +

dis
annotate(“text”, x = 1928, y = 6.8,
label = “Sub-Saharan Africa”, col=“#0000bf”)

or
In Figure 2-9, the sub-Saharan countries are colored dark blue. As the figure demonstrates,
there does not seem to be a relationship between women’s suffrage and economic development
st,
in sub-Saharan Africa (the line is flat). In the rest of the world, however, the relationship seems
strongly negative. Sub-Saharan Africa is different in this respect and accounting for it helps
po

clarify the strong relationship that exists between GDP per capita and women’s suffrage in the
other countries.
y,

FIGURE 2-9 No Relationship in Sub-Saharan Africa

op
tc

11
no GDP per Capita (logged)

10
Do

7
Sub-Saharan Africa

1900 1925 1950 1975

Year Women Gained Suffrage

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
48 STATISTICS AND DATA VISUALIZATION USING R

Think about what this implies for our investigation. Women’s suffrage occurred simul-
taneously with independence and democratization in many countries. To the degree that
other things are intertwined with women’s suffrage (independence, democratization, or
both), a measure that merely captures the year women were allowed to vote may only be the
result of a larger process. Consequently, these other important events—independence and
democratization—need to be accounted for in our explanation.
Before we examine the next hypothesis, let’s pause here again to take stock of how the pro-
cess of formulating hypotheses and describing data unfolds. Careful observation of simple
scatter plots provides considerable information. In this example, we again took advantage of
coloring the observations by region to make additional discoveries. We first explored whether

ute
the negative relationship between income and the year women were granted suffrage could
be explained by the differences between industrial countries and sub-Saharan Africa. We also
noticed a slight gap in the data during World War II. This suggested that other things might

trib
explain the relationship between women’s suffrage and income. Finally, we examined whether
the negative relationship among all the observations existed within the sub-Saharan cases.
Throughout the process, we can see how describing data and formulating hypotheses interact.

dis
Good descriptions of the data can help us evaluate hypotheses as well as formulate new
ones. With one scatter plot, we went from evaluating the hypothesis that women’s suffrage
influences economic growth to discovering how a country’s independence or move toward

or
democracy should be considered as well. Simple views of the data may provide what might
seem to be an endless number of patterns, discrepancies, or oddities that are worth investi-
gating. As the last few figures demonstrate, important insights can be gained by thoroughly
st,
examining each plot. I hope to have demonstrated here that a simple scatter plot can tell us a lot
about a relationship if we simply take the time to investigate, coloring groups of data differently
po

to learn what cases determine the slope of a line or the shape of a curve. Spotting interesting
aspects of the data through visualization in different ways is a useful skill worth developing.

Hypothesis III: Human Capital

y,

Continuing on, an important component of economic growth is human capital: education. As

op

a population accumulates knowledge, economists argue economic development follows. How

does a measure of human capital (an index of human capital constructed by Robert Barro and
tc

Jong-Wha Lee) relate to GDP per capita? According to Figure 2-10, it appears the relation-
ship is strong and positive: the more human capital per worker, the higher GDP per capita. Of
no

course, a developed economy may allow more people to attend school: the direction of causal-
ity may run the other way (from wealth to education). Nevertheless, we can say there’s a strong
association between the two variables. Note how the observations in this figure are much closer
Do

to the line, which indicates our predictions of GDP per capita at each level of the human capital
index. Also observe how the shape of this plot is better summarized by a line than the previous
two plots. Note my use of the word “summarize.” We use summaries to characterize the data.
When we use a line, we summarize the data with two numbers: the intercept of a line and its
slope.
The code in Code Chunk 2-10 should look familiar by now. Along with the ggplot()
­function that establishes a grid with two dimensions (GDP per capita and human capital),
it specifies a layer of dots with the function geom_point() and a line with the function geom_
smooth(). It also uses the ifelse() function to label the cases of Qatar (QAT), Kuwait (KWT), and
Zimbabwe (ZWE).

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 49

Code Chunk 2-10

ggplot(world, aes(pwthc, log(gdppc))) +
geom_point(col=“#bf0000”) +
geom_smooth(method=“lm”, se=FALSE, col=“#0000bf”) +
ggtitle(“Figure 2-10: Strong Relationship Between
GDP and Education”) +
ylab(“GDP per Capita (logged)”) +
xlab(“Barro and Lee’s Human Capital Index”) +
geom_text_repel(size = 2.8,

ute
aes(label=ifelse(iso3c==“QAT” |
iso3c==“KWT” |
iso3c==“ZWE”,

trib
[Link](iso3c),’’),
hjust = 0, vjust=-1),
[Link]=FALSE) +

dis
theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”))

or
st,
FIGURE 2-10 Strong Relationship Between GDP and Education
po

QAT
KWT
y,

11
GDP per Capita (logged)

op

10
tc

9
no

ZWE
Do

2 3
Barro and Lee’s Human Capital Index

Besides the distinct linear pattern, a noticeable feature of the plot is the outliers,
the cases far removed from the rest of the data. For example, Qatar (QAT) and Kuwait
(KWT) stand out as having been severely underpredicted—they lie considerably above

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
50 STATISTICS AND DATA VISUALIZATION USING R

the line (the line represents our prediction). Given the levels of the human capital index,
we would expect GDP per capita to be much lower in those countries. Of course, both
have significant proven oil reserves, perhaps accounting for our inaccurate predic-
tion. Note also one of the cases lying below the line, Zimbabwe (ZWE). At Zimbabwe’s
level of human capital, we would have predicted a higher level of per capita income. In
­Zimbabwe’s case (a primarily agricultural country), growing corruption in the 1990s and
2000s along with unstable property rights—disputes over land—might help explain its
underachievement. Even though the human capital index does fairly well in explaining
GDP per capita, the outliers help identify other possible explanations: natural resources,
property rights, and politics.

ute
In this section, we learned that identifying specific cases suggests possible hypotheses we
might want to test in the future. Outliers alert us to important causal factors that we hadn’t pre-
viously considered. In this example, while investigating the importance of human capital’s role

trib
in determining a country’s income, we found that oil (Qatar and Kuwait) and property rights
(Zimbabwe) might contribute to income levels as well. Again, we see how describing data gen-
erates additional hypotheses.

dis
Hypothesis IV: Political Stability
What role does politics play in the development of national economies? Countries with

or
unstable regimes are probably not the best venues for trade and investment. Is political insta-
bility associated with lower levels of wealth? To consider that possibility, GDP per capita is
st,
plotted (Figure 2-11) against the number of years since a significant change in the regime
(the variable durable). The durable variable is logged for the same reasons we logged GDP per
capita: transforming variables by taking logs will be explained in detail when I cover trans-
po

forming data in Chapter 6. There does seem to be a positive relationship between regime
stability and GDP per capita—countries with relatively stable regimes experience high levels
of income (Figure 2-11). As we move from left to right along the x axis, the level of income
y,

appears to increase.
op

Code Chunk 2-11 lists the set of commands used previously to draw a scatter plot. The only
wrinkle involves logging the durable variable. Since the durable variable contains some zeros,
we add a 1 to the variable before logging since the log of 0 is undefined. There’ll be more on
tc

­logging and transforming variables in Chapter 6.

no

Code Chunk 2-11

Do

ggplot(world, aes(log(durable + 1), log(gdppc))) +

geom_point(col=“#bf0000”) +
theme_minimal() +
geom_smooth(se=FALSE, col=“#0000bf”, method=“lm”) +
ggtitle(“Figure 2-11: Political Stability Increases Wealth”) +
ylab(“GDP per Capita (logged)”) +
xlab(“Durable (logged)”) +
theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”))

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 51

FIGURE 2-11 Political Stability Increases Wealth

11
GDP per Capita (logged)

10

ute
9

trib
8

dis
0 2 4
Durable (logged)

or
st,
While there does seem to be a positive relationship between stability and income, the blue
line may not be the best way to summarize the relationship. In particular, notice the points at 0
on the x axis. Save for one, they are all located above the line. Also notice that as we move from
po

left to right along the x axis between 0 and 2, most of the observations are located below the
line. Then, above 3, most of the points can be found above the line.
y,

To compare a straight line with a curved line, I added two lines to the figure with the
geom_smooth() function. One of the layers simply draws a curved line to the data (the
op

default), while the other specifies method=“lm”, which tells R to fit a straight line to the data
(Code Chunk 2-12).
tc

Code Chunk 2-12

no

ggplot(world, aes(log(durable + 1), log(gdppc))) +

geom_point(col=“grey”) +
Do

geom_smooth(se=FALSE, col=“#0000bf”) +
geom_smooth(method=“lm”, se=FALSE,
linetype = “dashed”, col = “grey”) +
ggtitle(“Figure 2-12: The Nonlinear Relationship Between
GDP per Capita and Political Stability”) +
ylab(“GDP per Capita (logged)”) +
xlab(“Durable (logged)”) +
theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
[Link] = element_text(size = 8, face = “bold”))

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
52 STATISTICS AND DATA VISUALIZATION USING R

FIGURE 2-12 The Nonlinear Relationship Between GDP per Capita and
Political Stability

11

GDP per Capita (logged)

10

ute
9

trib
8

dis
7

0 2 4
Durable (logged)

or
st,
In this example, the shape of the relationship may be best described by an s-shaped curve.
At the low end of durable, income is high (higher than our line predicts).6 As we move from
po

left to right on the x axis, income per capita increases sharply and we begin to systematically
underpredict the level of GDP per capita (the majority of cases lie above our prediction on the
right portion of the figure).
y,

Although the curvature of the s-shape is not severe, patterns like this suggest there are
thresholds. In this example, the s-shape suggests that economic development does not take
op

off until several years of political stability are achieved. At the high end of political stability,
the s-shape indicates that each additional year of political stability change does not improve
tc

economic development appreciably. In other words, once the same set of constitutional rules
have been in force for a number of years, each additional year does not have the same positive
no

influence.
While the relationship between regime stability and GDP per capita may not be strictly
linear (it may not be a straight line) and there remains plenty of variation, there is a positive
Do

­association between the two variables. In other words, as regime stability increases, so does
GDP per capita: the two variables increase together. Regime stability might be another variable
we should consider when trying to explain why some countries are rich and some are poor.
To summarize, in this section we saw another example of how describing data can inform
formulating hypotheses and vice versa. In earlier examples, we saw that certain techniques, such
as identifying clumps of data as well as individual outliers, help generate additional hypotheses
we might want to test. This example demonstrates that examining the functional form of the

6
Since the log of 0 is undefined, I added 1 to the durable variable before taking logs since some of the cases registered 0, a
common trick when transforming data. Notice the cases at 0 along the x axis; they represent the cases that were 0 in raw form
but then registered 1 when I added the constant 1 to the variable. The log of 1 is 0.

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 53

relationship—whether the observations form a straight line or s-shaped curve—also informs

our investigation. Now that we have a better sense of what the data look like and we’ve interro-
gated a few likely suspects, it’s time to move on to model building and estimation.

KNOWLEDGE CHECK: Form hypotheses from descriptions

of data.

9. Which of the following accurately describe the process of forming hypotheses?

ute
a. We always enter into the exercise with some preconceived notions or priors.
b. Never enter into the exercise with some preconceived notions or priors.

trib
c. Adding different colors and labeling cases can help us generate hypotheses.
d. Don’t start describing the data until you have generated a hypothesis to test.
10. When evaluating the hypothesis that ethnolinguistic heterogeneity explains income

dis
levels, what features of the data generated new hypotheses?
a. Incomes are high in Europe.

or
b. The industrialized countries of the world are clustered together.
c. The relationship may not be linear.
st,
d. There are some poor island nations that have low heterogeneity.
11. What features of the data generated the hypothesis that independence has an
po

important role to play in explaining GDP per capita?

a. There is a gap in the data during World War II.
y,

b. There is a clear linear relationship between when women were allowed to vote and
GDP per capita.
op

c. Small island nations were early in allowing women to vote.

d. There is a linear relationship between income and human capital.
tc

12. What visualization techniques were used to make discoveries in this section?
no

a. Representing groups of data with different shapes

b. Representing groups of data with different colors
c. Singling out specific cases with labels
Do

d. Singling out specific cases with color

13. What did you notice in the preceding section in terms of describing data and
generating hypotheses?
a. Viewing the data can prove hypotheses.
b. Describing data usually settles the matter.
c. Describing the data often amends the hypothesis or leads to new questions.
d. Hypotheses are rarely proven true or completely dismissed by useful descriptions
of the data.
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
54 STATISTICS AND DATA VISUALIZATION USING R

Model Building and Estimation

Once the data are described and hypotheses have been generated, it’s time to develop our
argument or to construct a model. We’re now moving into material that will be covered in
­Chapters 10–14. Regression tables will appear in this section without an in-depth discussion
of their mechanics. They are presented here to help identify some important tools you will
develop that figure prominently in the process of data analysis.
The scatter plots we’ve examined so far suggest that there are multiple causes of economic
development. Ethnolinguistic heterogeneity, women’s suffrage, human capital, and regime sta-
bility all seem to matter. To obtain the best estimate of their relationship with GDP per c­ apita,

ute
we need a good model. For our purposes, let’s define a model as simply a list of factors we want
to include in an explanation. Well-reasoned hypotheses and a familiarity with the data should
be combined toward that end. A good model identifies only the essential elements that help

trib
accurately predict outcomes—the outcome in this example is a country’s GDP per capita.
Before we continue, it is important to acknowledge that both scholars and practitioners
disagree on how to build models. Important philosophical debates lurk beneath the surface.

dis
Put simply, the debate concerns whether theory or the data should be our guide. One side
doesn’t trust the data, arguing that knowing all of the facts in the world does us no good unless
we know how things fit together—the theory. The other side doesn’t trust the theory, argu-

or
ing that the best theory comes from knowing as much about the world as possible—the data.
Devotion to one approach to the exclusion of the other, in my opinion, often misleads. Relying
on a conversation between the two is more fruitful. So far, we’ve seen how looking at the data
st,
can inform theory. Next we find how theory can influence the conclusions drawn from the
data. Recognizing how both approaches work in practice is a primary concern of this book.
po

Start with a simple model: a model that uses one factor—the year women were
granted suffrage—to explain GDP per capita (Table 2-1). Models with two variables,
a dependent variable (GDP per capita) and an independent variable (women’s suf-
y,

frage; womyear), are called bivariate regression models. Models with more than two
­v ariables—a dependent variable (GDP per capita) and more than one independent
op

­variable—are called multiple regression models. In the bivariate case, the model uses a
line to summarize the relationship between GDP per capita and women’s suffrage without
tc

accounting for anything else.

The code in Code Chunk 2-13 defines an object called mod1 and then presents the
no

results from mod1 in a regression table using the stargazer() function from the ‘stargazer’
package. The mod1 object is defined as a simple bivariate regression model using the lm()
command. The lm in the lm() function stands for linear model. In the regression, the log of
Do

GDP per capita (gdppc) is regressed on the year women were granted suffrage in a coun-
try (womyear). An important convention in statistics is to say that the dependent variable is
regressed on the independent variable.

Code Chunk 2-13

mod1 <- lm(log(gdppc)~womyear, data=world)

stargazer(mod1, header=FALSE,
title = “Table 2-1: Estimates for a Bivariate Model”,
type = “html”, out = “[Link]”)
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 55

TABLE 2-1 Estimates for a Bivariate Model

DEPENDENT VARIABLE:
log(gdppc)
womyear −0.032***
(0.005)
Constant 70.750***
(8.831)

ute
Observations 143
R 2
0.257
Adjusted R 2
0.252

trib
Residual SE 1.046 (df = 141)
F statistic 48.874*** (df = 1; 141)

dis
Note: ***p < 0.01.

The stargazer function requires the right kind of an object as an argument. In this case, the

or
object mod1—defined by a linear model—is appropriate. We can also specify the title of the
table, the type of output (HTML in this case), and a file where the output will appear. Note that
st,
when you knit to an HTML file or PDF, a nice table will appear. Unfortunately, as of this writ-
ing, stargazer does not work when knitting a Microsoft Word document.
With the bivariate model, we find there is a negative association between GDP per cap-
po

ita and women’s suffrage—the number (−.032). In other words, as the women’s suffrage vari-
able increases (the later the date suffrage was granted), GDP per capita decreases. The number
(−.032) represents the slope of the line that describes the relationship between women’s suf-
y,

frage and GDP per capita.

op

Before moving to the next section, let’s recap. After describing the data and exploring
a few hypotheses, we constructed a model to explain GDP per capita. We first constructed a
very simple model that fit a line summarizing the relationship between GDP per capita and
tc

the year women were granted suffrage in a country. We found that the slope of that line (our
­estimate) was −.032, implying that women’s participation in politics does play an important
no

role in e­ conomic development.

Do

KNOWLEDGE CHECK: Explain the connection between

hypotheses, models, and estimates.

14. Which of the following statements about model building are true?
a. Theory should always come before the facts.
b. The facts should always come before the theory.
c. There should be a conversation, a back-and-forth between theory and facts.
d. In the author’s opinion, there should be a conversation, a back-and-forth between
theory and facts. Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
56 STATISTICS AND DATA VISUALIZATION USING R

15. Indicate which statements are true.

a. Models with two variables are called bivariate regression models.
b. Models with more than two variables are called bivariate models.
c. To obtain the best estimate of a relationship between two variables, we need a good
model.
d. Models with more than two variables are called multiple regression models.

Diagnostics

ute
Now that we’ve obtained estimates, we need to ascertain how much trust to place in our results.
Tables with model estimates can evince an air of confidence and authority. Those estimates

trib
aligned so neatly in nice-looking columns and rows could be based on some questionable
assumptions, outlying cases, or very poor measures. Much like when being confronted by
a pushy salesman who leads you over to a bright, shiny, Corvette on a used-car lot, it might

dis
be good to check under the hood. Diagnostics provide the tools to determine whether our
estimates are reasonable or highly misleading. Will the Corvette go from 0 to 60 in less than
4 ­seconds or will it die once we’ve driven it off the lot?

Stability of the Results

or
Checking the stability of the results and whether we’ve met certain assumptions are only two
st,
possible ways to perform diagnostics. Chapters 15 and 16 are devoted to the enterprise. Here,
I simply want to illustrate how our choice of model (theory) can influence our estimates of the
po

relationship (the data).

To illustrate, I estimate a multiple regression model with the lm() function and then con-
struct a table displaying the results using the stargazer() command (Code Chunk 2-14).
y,
op

Code Chunk 2-14

mod2 <- lm(log(gdppc)~womyear + ethfrac + durable + pwthc,
tc

data=world)
no

stargazer(mod2, header=FALSE,
title = “Table 2-2: Estimates Are Unstable”,
type = “html”, out = “[Link]”)
Do

How stable are the results from the bivariate model? By stability, I mean how much the
results change when we alter the model slightly. If small changes in the model (adding a few
variables) produce noticeable changes, we say the results are unstable. Remember that in the
bivariate regression, the slope was −.032. Once other variables are added (see Table 2-2), the
slope on women’s suffrage changed to −.001, decreasing in magnitude by a factor of 32! Clearly,
our estimate of women’s suffrage depends on the particular model we use.
To see how this translates visually, I generate a plot that shows how women’s suffrage is
related to GDP per capita when we account for human capital, political instability, and ethno-
linguistic fractionalization. This is called an added variable plot, which shows the relationship
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 57

TABLE 2-2 Estimates Are Unstable

DEPENDENT VARIABLE:
log(gdppc)

womyear −0.001

(0.004)

ethfrac −0.698**

(0.291)

ute
durable 0.007***

(0.002)

trib
pwthc 1.416***

(0.163)

dis
Constant 6.682

(8.201)

or
Observations 105

R2 0.725
st,
Adjusted R2 0.714

Residual SE 0.657 (df = 100)

po

F statistic 65.813*** (df = 4; 100)

y,

Note: **p < .05; ***p < 0.01.

op

between two variables when other variables are accounted for. Again, you’ll learn how to
generate this very useful view of the data in Chapters 15 and 16. Until then, consider the rela-
tc

tionship between GDP per capita (logged) and women’s suffrage when all of the variables are
included in the model (Figure 2-13).
no

An easy way to visualize results from regression analysis involves using an R package called
‘visreg.’ In Code Chunk 2-15, I use the visreg() function to plot the predicted values from the
multiple regression model I defined in Code Chunk 2-14. In this example, I want to visualize
Do

the relationship between women’s suffrage and GDP per capita, accounting for the other vari-
ables in the model. The ‘visreg’ package makes this easy. I simply indicate which model I want
to use (mod2, in this case) and specify the independent variable I’m interested in (womyear).

Code Chunk 2-15

visreg(mod2, “womyear”, ylab=“Predicted Values GDP (logged)”,
xlab=“Year Women Gained Suffrage”,
main=“Figure 2-13: Women’s Suffrage Has No Effect”,
band=FALSE)

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
58 STATISTICS AND DATA VISUALIZATION USING R

FIGURE 2-13 Women’s Suffrage Has No Effect

11.0

Predicted Values GDP (logged)

10.5

10.0

9.5

9.0

ute
8.5

8.0

trib
7.5
1900 1920 1940 1960 1980
Year Women Gained Suffrage

dis
The blue line in the figure represents what level of GDP per capita model 2 predicts based

or
on women’s suffrage. The dots represent the actual values as they exist for each country.
While this doesn’t necessarily settle the issue, the added variable plot suggests there is not
st,
much of a relationship between model 2’s prediction of income (GDP per capita) and wom-
en’s suffrage when controlling for the other variables. Once human capital, political stability,
po

and ethnolinguistic fractionalization have been accounted for, knowing when women were
granted suffrage doesn’t really tell us much about the level of GDP per capita. This exercise
suggests that the relationship between GDP per capita and women’s suffrage is unstable with
y,

respect to the inclusion of political stability, ethnolinguistic fractionalization, and human

­capital in our model. It illustrates an important lesson: our estimates depend on the model we
op

choose.

Residual Plots
tc

Finally, we can check whether our predictions get better or worse as we range over our pre-
dictions, from low to high levels of GDP per capita. If our predictions are systematically better
no

or worse for rich or poor countries, we know our model could be improved. A residual plot
(Figure 2-14) tells us very quickly if the difference between each case and our prediction (what
we call the residual) follows any pattern. Figure 2-14 plots the residuals against the predictions.
Do

The horizontal dashed line represents perfect predictions. If a case rests on that line, our model
has correctly predicted the actual level of income for that country. Cases further away from the
dashed line, such as Singapore (SGP), are those that the model predicted poorly. This r­ esidual
plot indicates that the model severely underpredicted the level of income in Singapore and
Luxembourg.
To produce residuals, we first need to estimate the model and collect the residuals. In Code
Chunk 2-16, you can see I use the lm() function again to estimate a linear model. This time it
includes more than one independent variable; it includes four: womyear, ethfrac, durable, and
pwtch. Once the model is estimated, I create two variables (world$res and world$pred) using
the resid() function and the predict() function.
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 59

Code Chunk 2-16

mod2 <- lm(log(gdppc)~womyear + ethfrac +
durable + pwthc, [Link] = [Link],
data=world)

world$res <- resid(mod2)

world$pred <- predict(mod2)

ute
Now that I have a variable with the predicted values from the model and the residuals, I’m
ready to plot the two against each other. I do this using the code in Code Chunk 2-17. Code
Chunk 2-17 defines a scatter plot that colors the points red, labels the points of five countries

trib
using the ifelse() function, and draws a horizontal line that represents the predicted values
(where the residuals equal 0).

dis
Code Chunk 2-17
ggplot(world, aes(pred, res)) +
geom_point(col=“#bf0000”) +
geom_text_repel(size = 2.8,
aes(label=ifelse(iso3c==“SGP” | or
st,
iso3c==“TJK” |
iso3c==“ZWE” |
po

iso3c==“USA” |
iso3c==“LUX”,
[Link](iso3c),’’),
y,

hjust = 0, vjust=-1), [Link]=FALSE) +

geom_hline(yintercept = 0, linetype = 2) +
op

theme_minimal() +
theme([Link] = element_text(size = 8, face = “bold”),
tc

[Link] = element_text(size = 8, face = “bold”)) +

xlab(“Predicted or ‘Fitted’ Values”) +
ylab(“Residuals”) +
no

ggtitle(“Figure 2-14: Some Possible Outliers”)

Do

Here we look for patterns. As we’ll learn in subsequent chapters, an important assumption
in model estimation is that our predictions are just as good or bad for low or high values. In this
example we’re making the assumption that the model’s accuracy does not change when we’re
making predictions for relatively poor countries or relatively rich ones. The plot in Figure 2-14
can be characterized as a formless cloud: our predictions are just as good or bad as we make
predictions for rich and poor countries. There are no obvious observable patterns. From this
we would conclude that an important assumption of our analysis is met.7

7
An explanation of the assumptions associated with ordinary least squares (OLS) regression analysis, which is used in this
example, will be presented in the chapter on diagnostics (Chapter 15).

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
60 STATISTICS AND DATA VISUALIZATION USING R

FIGURE 2-14 Some Possible Outliers

LUX
SGP

Residuals

ute
0

trib
USA
–1

dis
ZWE TJK

7 8 9 10 11
Predicted or ‘Fitted’ Values

or
st,
The residual plot also indicates if any outliers exist (cases far removed from the rest of the
data). With this regression model, Singapore (SGP), Luxembourg (LUX), Tajikistan (TJK),
po

the United States (USA) and Zimbabwe (ZWB) are potential outliers. These outlying cases
are ones our model did not predict very well. Why? As we learned with the scatter plots in the
previous section, investigating those cases may provide clues as to what other factors might be
y,

involved.
op

KNOWLEDGE CHECK: Define diagnostics and explain their role

tc

in data analysis.
no

16. Which are good analogies for diagnostics?

a. Checking underneath the hood of a car
Do

b. Examining the X-ray of a patient

c. Cross-examining a witness in a trial
d. Questioning a politician at a press conference
17. What are we looking for when we perform diagnostics?
a. Do the results show a strong relationship?
b. Is the hypothesis confirmed?
c. Do the results change with small changes to the model?
d. Are there odd cases that are unduly influencing the results?
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 61

18. What are we looking for in a residual plot?

a. Odd cases.
b. Odd patterns
c. Any patterns
d. The typical case

Next Questions
One of the most important aspects of data analysis involves formulating new questions. What

ute
do the estimates imply and how can that be tested? In the regression analysis described ear-
lier, including human capital, ethnolinguistic fractionalization, and political stability in the
regression changed the estimate of women’s suffrage. Why is that? Perhaps women’s political

trib
rights and the human capital index are closely related (we can test that). Women’s influence on
GDP might also depend on human capital. The right to vote may have little impact in countries
with low levels of education. When women are educated, the right to vote could be extremely

dis
powerful. In addition to these possibilities, the model should probably contain other variables
(e.g., date of independence, colonial past, and democracy). As we include those variables in our
model, the estimates will change, leading to new discoveries.

or
Asking the next question performs two functions. First, it can help confirm our hypotheses.
Generating the next question, forming a hypothesis around it, and testing it can bring new evi-
st,
dence to bear on our original hypothesis. Second, it can lead to new discoveries. Often, we find
that the original question we asked was not the most interesting, or that it sent us barking up the
wrong tree. In either case, asking the next question is a fundamental aspect of data analysis.
po

How do we formulate new questions? I find that a helpful way to spark new questions starts
with an if-then statement: “if ” the results are true, “then” what else would we expect to see in
the data? This exercise tests our knowledge of the problem, our ability to think creatively, and
y,

our ability to think logically. This is surely part of what we mean by “critical thinking.” As with
op

anything, it requires practice. It is a muscle that, if used often, will grow big and strong. If sel-
dom used, it will atrophy.
A final feature of new questions that I’ll mention is that they often help us “think outside
tc

the box.” While this is an overused and ill-defined directive often muttered when the boss is
completely out of ideas, thinking outside the box comes directly from asking new questions.
no

In other words, asking new questions is thinking outside the box. New questions force us to
consider “what if ”? When we think outside the box, we’re simply exploring beyond our current
model—the proverbial box.
Do

KNOWLEDGE CHECK: Formulate new questions.

19. Why do we formulate new questions?

a. They challenge our old questions.
b. They force us to reconsider our model.
c. They provide additional evidence.
d. They help us think “outside the box.”
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
62 STATISTICS AND DATA VISUALIZATION USING R

20. How do we formulate new questions?

a. Get to know the subject.
b. Use an if-then statement.
c. Practice.
d. Exercise.

SUMMARY

ute
In this chapter, the goal was to motivate the enterprise are many things we want to explain (wealth, inequality,
and to introduce the process of data analysis. The grow- ­violence, voting, etc.), and there are a lot of data to explore.

trib
ing amount of information and its use in our everyday
lives requires we know how to use and understand quan- We have many tools at our disposal to understand the
titative data. In particular, I hope to have shown the back- many characteristics of our data. For example, if our

dis
and-forth between theory and evidence that undergirds problem is to reduce homicide rates in the United States,
the approach taken in the following pages. Although it would be helpful to know where the lowest and high-
all scholars and practitioners take slightly different est rates are found. Do most states have similar rates or
approaches, the process outlined in this chapter is a good
way to start: (1) describing data and formulating hypoth-
or
do they vary dramatically? Are similar rates clustered in
the South, Northeast, and West? Answering these ques-
st,
eses, (2) building and estimating models, (3) diagnostics, tions involves describing the data. The more we know
and (4) generating the next question.
about our data, the better questions we’ll ask and the
po

This book is less about probability theory and statistics better models we’ll construct. In Chapter 3, I discuss the
and more about ways to discover patterns that exist in our importance of describing data and provide examples that
data and the inferences we can draw from them. There ­illustrate why it is so important.
y,
op

COMMON PROBLEMS
tc

•• Exploration versus presentation. Understanding the underpredicted and dots below the horizontal line
difference between exploration and presentation is represent instances where the model overpredicted.
no

usually not articulated in most books on statistics.

•• Theory versus hypothesis. Students and practi-
Much of what was presented in this chapter shows
tioners often fall into using these two terms inter-
what goes on behind the scenes. Following paths that
changeably. Theories are frameworks or systems of
Do

ultimately lead nowhere is not usually presented.

relationships used to explain a variety of different
Exploring data for the purpose of discovery and pre- phenomenon. Hypotheses are more specific and
senting data for the purpose of persuasion are two tentative. Hypotheses represent our best guesses
very different things. about how the world works. Those guesses are tested
to confirm or challenge existing theory. As evidence
•• Understanding the residual plot. Spend time under-
accumulates indicating that the theory is incorrect,
standing the analytics of the residual plot. The most
it’s time to change our theory.
difficult aspect is to understand that the flat horizon-
tal line represents residuals with a value of 0, indi- •• Patience. Patience, or rather the lack thereof, is
cating that the cases were perfectly predicted. Dots a common theme in data analysis. Students and
above the line represent instances where the model practitioners get impatient. Their impatience
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 63

manifests itself in three primary ways during without spending adequate time with diagnos-
the process of data analysis. First, they often tics. Finally, relatively little time is spent thinking
jump past providing a good description of the about how to present data effectively. As you’ll
data right to estimating models. Second, stu- see in Chapter 7, considerations concerning data
dents and practitioners report their results ­presentation are crucial.

REVIEW QUESTIONS
1. Describe the struggle between theory and evidence. 6. What is meant by stability of the results?

ute
2. What are the main components of data analysis? 7. How would you describe the grammar of graphics
used in ggplot?
3. Explain how colors, shapes, and labels aid in data

trib
analysis. 8. What is a model?

4. Why is generating the next question such a useful 9. What are diagnostics?

dis
exercise?
10. What purpose do diagnostics serve?
5. What are we looking for when we examine a residual
plot?

PRACTICE ON ANALYSIS AND VISUALIZATION

or
st,
po

1. What does the ifelse() function below say in English? 3. Which of the following statements accurately
describe the purpose of diagnostics?
ifelse(world, region ==”Europe”, 1, 0)
a. They help identify important cases.
y,

a. If the world variable equals “region,” then

b. They indicate whether we are using the appropri-
assign 1.
op

ate model.
b. If the world variable equals “region,” then
assign 0. c. They generate additional questions.
tc

c. If world$region equals Europe, assign 1 other- d. They can help generate additional hypotheses.
no

wise 0. 4. When we examined the relationship between wom-

d. If world$region equals Europe, assign 0 other- en’s suffrage and income levels, what if-then state-
wise 1. ment did we generate?
Do

2. Which of the following statements accurately a. If the relationship between women’s suffrage
describe the process of data analysis? and GDP per capita is just the result of the
important differences between Europe and
a. Model estimation benefits from describing data.
sub-Saharan Africa, then we would expect to
b. Diagnostics should always represent the end of see the same relationship exist among the coun-
our analysis. tries of sub-Saharan Africa.
c. Discovery comes once model estimates have b. If the relationship between women’s suffrage
been generated. and GDP per capita is not just the result of the
d. If a model is properly formulated, diagnostics are important differences between Europe and
not necessary. sub-Saharan Africa, then we would expect to see
Copyright ©2022 by SAGE Publications, Inc.
This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
64 STATISTICS AND DATA VISUALIZATION USING R

the same relationship exist among the countries b. A plot that shows the predicted values against the
of sub-Saharan Africa. residuals
c. If the relationship between suffrage and income c. A plot showing the relationship between one
is strong, the relationship should exist between variable, accounting for all of the others
Europe and Africa. d. A plot showing the relationship between the
d. If the relationship between women’s suffrage dependent and independent variables
and GDP per capita is not just the result of the
8. What observation led us to ask whether the rela-
important differences between Europe and
tionship between ethnic heterogeneity and GDP per
sub-Saharan Africa, then we would not expect to
­capita was a linear pattern?

ute
see the same relationship exist among the coun-
tries of sub-Saharan Africa. a. There were many poor island nations.
5. What did the outlying cases of Kuwait, Qatar, and b. The clump of countries we identified as being

trib
Zimbabwe illustrate when looking at the relationship “industrialized” may be generating the linear
between human capital accumulation and GDP per pattern.
capita? c. The slope of the line declined when ignoring the

dis
a. They indicated the possible importance European cases.
of natural resources, property rights, and d. There seems to be a nonlinear relationship
politics. between income and ethnolinguistic heterogene-
b. They should be removed from the data since they
are not representative cases. or ity when all of the cases are considered.
9. If there is large variation in y at different levels of x in a
st,
c. They all point to the importance of oil as an scatter plot, what does that indicate?
explanation of GDP per capita.
po

a. The relationship is not linear.

d. Corruption is clearly an important variable that b. There is no relationship between the x and y
needs to be considered. variables.
y,

6. What did exploring the stability of our results c. The x variable does a good job of explaining the
reveal? variation in y.
op

a. There are important thresholds we need to d. There could be other variables that might help
account for in our model estimation. explain the variation.
tc

b. The wrong estimation technique is being used. 10. Which of the following are true in the context of
no

c. Regardless of the variables we include in this residual plots?

model, the results always indicate the same a. Points above the horizontal line represent cases
thing. that are underpredicted.
Do

d. Our estimates for a variable in the model can b. Points above the horizontal line represent cases
depend heavily on how we specify the model— that are overpredicted.
what variables we include in the model.
c. Patterns in the data suggest there is a problem
7. What is an added variable plot? with the model.
a. A plot that shows the residuals against the pre- d. Patterns in the data suggest the model is
dicted values appropriate.

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
 Chapter 2 An Introduction to Data Analysis 65

ANNOTATED R FUNCTIONS
The following functions appear in this chapter. They are xlab(): labels the x axis in ggplot. (2-1)
listed in order of their first appearance (with the code theme_minimal(): specifies a minimalist style for
chunk number in parentheses) and annotated here to give ggplot. (2-1)
a very brief description of their use. Some are not stand-
alone functions and only work in combination with other theme(): specifies font, size, and so forth in a ggplot.
commands. As a reminder, the code in every ­chapter (2-1)
will work properly if executed in the order it appears. annotate(): allows the placement of text in the figure.

ute
Proper execution also depends on typing the author-­ (2-2)
defined libraries() command, which loads the required R geom_vline(): places a horizontal line in the figure.
packages. (2-7)

trib
ggplot(): defines the basic structure of a plot (usually scale_colour_manual: a function that allows the
the x and y variables). (2-1) user to specify exactly what colors to use in the figure.
(2-9)

dis
aes(): the aes (called “aesthetics”) function is used in
ggplot to define the basic structure of the plot, which stargazer(): a function from the ‘stargazer’ package
often includes the variables you want to use and any that helps create professional-looking tables. (2-13)
shapes or colors. (2-1)

or
lm(): a function that specifies a linear regression. The
ylab(): labels the y axis in ggplot. (2-1) “lm” stands for linear model. (2-13)
st,
ANSWERS
po

KNOW L E D G E CH ECK
y,

1. a 11. a
op

2. a, b, c 12. b, c, d
tc

3. a, b, c 13. c, d

4. a, b, c 14. d
no

5. b 15. a, c, d

6. c 16. a, b, c, d
Do

7. d 17. c, d

8. b, d 18. a, c

9. a, c 19. a, b, c, d

10. b 20. a, b, c, d

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.
66 STATISTICS AND DATA VISUALIZATION USING R

PR ACTICE ON ANA LYSI S AN D VI SUALIZ ATION

1. c 6. d

2. a 7. c

3. a, b, c, d 8. b

4. b 9. d

5. a 10. a, c

ute
trib
Access digital resources, including datasets, at
[Link]

dis
or
st,
po
y,
op
tc
no
Do

Copyright ©2022 by SAGE Publications, Inc.

This work may not be reproduced or distributed in any form or by any means without express written permission of the publisher.