StatisticsA Review of Basic Statistical Concepts

A Review of Basic
Statistical Concepts
1
The record of a month’s roulette playing at Monte Carlo can
afford us material for discussing the foundations of knowledge.
—Karl Pearson

I know too well that these arguments from probabilities are
imposters, and unless great caution is observed in the use of them,
they are apt to be deceptive.
—Plato (in Phaedo)

Introduction

I t is hard to find two quotations from famous thinkers that reflect more
divergent views of probability and statistics. The eminent statistician
Karl Pearson (the guy who invented the correlation coefficient) was so
enthralled with probability and statistics that he seems to have believed that
understanding probability and statistics is a cornerstone of human under-
standing. Pearson argued that statistical methods can offer us deep insights
into the nature of reality. The famous Greek philosopher Plato also had
quite a bit to say about the nature of reality. In contrast to Pearson, though,
Plato was skeptical of the “fuzzy logic” of probabilities and central tenden-
cies. From Plato’s viewpoint, we should only trust what we can know with
absolute certainty. Plato probably preferred deduction (e.g., If B then C) to
induction (In my experience, bees seem to like flowers).
Even Plato seemed to agree, though, that if we observe “great caution,”
arguments from probabilities may be pretty useful. In contrast, some
modern nonstatisticians might agree with what the first author’s father,
Bill Pelham, used to say about statistics and probability theory: “Figures

1

2 INTERMEDIATE STATISTICS

can’t lie, but liars sure can figure.” His hunch, and his fear, was that “you
can prove anything with statistics.” To put this a little differently, a surpris-
ing number of thoughtful, intelligent students are thumbs-down on statis-
tics. In fact, some students only take statistics because they have to (e.g., to
graduate with a major in psychology, to earn a second or third PhD). If
you fall into this category, our dream for you is that you enjoy this book
so much that you will someday talk about the next time that you get to
take—or teach—a statistics class.
One purpose of this first chapter, then, is to convince you that Karl
Pearson’s rosy view of statistics is closer to the truth than is Bill Pelham’s
jaded view. It is possible, though, that you fully agree with Pearson, but
you just don’t like memorizing all those formulas Pearson and company
came up with. In that case, the purpose of this chapter is to serve as a
quick refresher course that will make the rest of this book more useful. In
either event, no part of this book requires you to memorize a lot of com-
plex statistical formulas. Instead, the approach emphasized here is heavily
conceptual rather than heavily computational. The approach emphasized
here is also hands-on. If you can count on your fingers, you can count
your blessings because you are fully capable of doing at least some of the
important calculations that lie at the very heart of statistics. The hands-
on approach of this book emphasizes logic over rote calculation, capital-
izes on your knowledge of everyday events, and attempts to pique your
innate curiosity with realistic research problems that can best be solved
by understanding statistics. If you know whether there is any connection
between rain and umbrellas, if you love or hate weather forecasters, and
if you find games of chance interesting, we hope that you enjoy at least
some of the demonstrations and data analysis activities that are contained
in this book.
Before we jump into a detailed discussion of statistics, however, we
would like to briefly remind you that (a) statistics is a branch of mathe-
matics and (b) statistics is its own very precise language. This is very fitting
because we can trace numbers and, ultimately, statistics back to the begin-
ning of human language and thus to the beginning of human written
history. To appreciate fully the power and elegance of statistics, we need to
go back to the ancient Middle East.

How Numbers and Language Revolutionized Human History

About 5,000 years ago, once human beings had began to master agricul-
ture, live in large city states, and make deals with one another, an
unknown Sumerian trader or traders invented the cuneiform writing
system to keep track of economic transactions. Because we live in a
world surrounded by numbers and written language, it is difficult for us

Chapter 1 A Review of Basic Statistical Concepts 3

to appreciate how ingenious it was for someone to realize that writing
things down solves a myriad of social and economic problems. When
Basam and Gabor got into their semimonthly fistfight about whether
Gabor owed Basam five more or six more geese to pay for a newly
weaned goat, our pet theory is that it was an exasperated neighbor who
finally got sick of all the fighting and thus proposed the cuneiform writ-
ing system. The cuneiform system involved making marks with a stylus
in wet clay that was then dried and fired as a permanent record of eco-
nomic transactions. This system initially focused almost exclusively on
who had traded what with whom—and, most important, in what quan-
tity. Thus, some Sumerian traders made the impressive leap of impress-
ing important things in clay. This early cuneiform writing system was
about as sophisticated as the scribbles of your 4-year-old niece, but it
quickly caught on because it was way better than spoken language
alone.
For example, it apparently wasn’t too long before the great-great-
great-grandchild of that original irate neighbor got a fantastically bril-
liant idea. Instead of drawing a stylized duck, duck, duck, duck to repre-
sent four ducks, this person realized that four-ness itself (like two-ness
and thirty-seven-ness) was a concept. He or she thus created abstract
characters for numbers that saved ancient Sumerians a lot of clay. We
won’t insult you by belaboring how much easier it is to write and verify
the cuneiform version of “17 goats” than to write “goat, goat, goat, goat,
goat, goat, goat, goat, goat, goat, goat, goat, goat, goat, goat, goat . . .” oh
yeah “. . . goat,” but we can summarize a few thousand years of human
technological and scientific development by reminding you that incredi-
bly useful concepts such as zero, fractions, p (pi), and logarithms, which
make possible great things such as penicillin, the Sistine Chapel, and
iPhones, would have never come about were it not for the development
of abstract numbers and language.
It is probably a bit more fascinating to textbook authors than to text-
book readers to recount in great detail what happened over the course of
the next 5,000 years, but suffice it to say that written language, numbers,
and mathematics revolutionized—and sometimes limited—human scien-
tific and technological development. For example, one of the biggest ruts
that brilliant human beings ever got stuck into has to do with numbers. If
you have ever given much thought to Roman numerals, it may have
dawned on you that they are an inefficient pain in the butt. Who thought
it was a great idea to represent 1,000 as M while representing 18 as XVIII?
And why the big emphasis on five (V, that is) in a base-10 number system?
The short answer to these questions is that whoever formalized Roman
numbers got a little too obsessed with counting on his or her fingers and
never fully got over it. For example, we hope it’s obvious that the Roman
numerals I and II are stand-ins for human fingers. It is probably less obvious


that the Roman V (“5”) is a stand-in for the “V” that is made by your
thumb and first finger when you hold up a single hand and tilt it outward
a bit (sort of the way you would to give someone a “high five”). If you do
this with both of your hands and move your thumbs together until they
cross in front of you, you’ll see that the X in Roman numerals is, essen-
tially, V + V. Once you’re done making shadow puppets, we’d like to tell
you that, as it turns out, there are some major drawbacks to Roman num-
bers because the Roman system does not perfectly preserve place (the way
we write numbers in the ones column, the tens column, the hundreds
column, etc.).
If you try to do subtraction, long division, or any other procedure
that requires “carrying” in Roman numerals, you quickly run into seri-
ous problems, problems that, according to at least some scholars,
sharply limited the development of mathematics and perhaps technol-
ogy in ancient Rome. We can certainly say with great confidence that,
labels for popes and Super Bowls notwithstanding, there is a good rea-
son why Roman numerals have fallen by the wayside in favor of the
nearly universal use of the familiar Arabic base-10 numbers. In our
familiar system of representing numbers, a 5-digit number can never be
smaller than a 1-digit number because a numeral’s position is even more
important than its shape. A bank in New Zealand (NZ) got a painful
reminder of this fact in May 2009 when it accidentally deposited
$10,000,000.00 (yes, ten million) NZ dollars rather than $10,000.00
(ten thousand) NZ dollars in the account of a couple who had applied
for an overdraft. The couple quickly fled the country with the money
(all three extra zeros of it).1 To everyone but the unscrupulous couple,
this mistake may seem tragic, but we can assure you that bank errors of
this kind would be more common, rather than less common, if we still
had to rely on Roman numerals.
If you are wondering how we got from ancient Sumer to modern New
Zealand—or why—the main point of this foray into numbers is that life
as we know and love it depends heavily on numbers, mathematics, and
even statistics. In fact, we would argue that to an ever increasing degree in
the modern world, sophisticated thinking requires us to be able under-
stand statistics. If you have ever read the influential book Freakonomics,
you know that the authors of this book created quite a stir by using statis-
tical analysis (often multiple regression) to make some very interesting
points about human behavior (Do real estate agents work as hard for you
as they claim? Do Sumo wrestlers always try to win? Does cracking down
on crime in conventional ways reduce it? The respective answers appear to
be no, no, and no, by the way.) So statistics are important. It is impossible
to be a sophisticated, knowledgeable modern person without having at
least a passing knowledge of modern statistical methods. Barack Obama
appears to have appreciated this fact prior to his election in 2008 when he


assembled a dream team of behavioral economists to help him get
elected—and then to tackle the economic meltdown. This dream team
relied not on classical economic models of what people ought to do but on
empirical studies of what people actually do under different conditions.
For example, based heavily on the work of psychologist Robert Cialdini,
the team knew that one of the best ways to get people to vote on election
day is to remind them that many, many other people plan to vote (Can you
say “baaa”?).2
So if you want a cushy job advising some future president, or a more
secure retirement, you would be wise to increase your knowledge of statis-
tics. As it turns out, however, there are two distinct branches of statistics,
and people usually learn about the first branch before they learn about the
second. The first branch is descriptive statistics, and the second branch is
inferential statistics.

Descriptive Statistics

Statistics are a set of mathematical procedures for summarizing and
interpreting observations. These observations are typically numerical or
categorical facts about specific people or things, and they are usually
referred to as data. The most fundamental branch of statistics is descrip-
tive statistics, that is, statistics used to summarize or describe a set of
observations.
The branch of statistics used to interpret or draw inferences about a set
of observations is fittingly referred to as inferential statistics. Inferential
statistics are discussed in the second part of this chapter. Another way of
distinguishing descriptive and inferential statistics is that descriptive sta-
tistics are the easy ones. Almost all the members of modern, industrialized
societies are familiar with at least some descriptive statistics. Descriptive
statistics include things such as means, medians, modes, and percentages,
and they are everywhere. You can scarcely pick up a newspaper or listen to
a newscast without being exposed to heavy doses of descriptive statistics.
You might hear that LeBron James made 78% of his free throws in 2008–
2009 or that the Atlanta Braves have won 95% of their games this season
when they were leading after the eighth inning (and 100% of their games
when they outscored their opponents). Alternately, you might hear the
results of a shocking new medical study showing that, as people age,
women’s brains shrink 67% less than men’s brains do. You might hear a
meteorologist report that the average high temperature for the past 7 days
has been over 100 °F. The reason that descriptive statistics are so widely
used is that they are so useful. They take what could be an extremely large
and cumbersome set of observations and boil them down to one or two
highly representative numbers.


In fact, we’re convinced that if we had to live in a world without
descriptive statistics, much of our existence would be reduced to a hellish
nightmare. Imagine a sportscaster trying to tell us exactly how well LeBron
James has been scoring this season without using any descriptive statistics.
Instead of simply telling us that James is averaging nearly 30 points per
game, the sportscaster might begin by saying, “Well, he made his first shot
of the season but missed his next two. He then made the next shot, the
next, and the next, while missing the one after that.” That’s about as effi-
cient as “goat, goat, goat, goat. . . .” By the time the announcer had docu-
mented all of the shots James took this season (without even mentioning
last season), the game we hoped to watch would be over, and we would
never have even heard the score. Worse yet, we probably wouldn’t have a
very good idea of how well James is scoring this season. A sea of specific
numbers just doesn’t tell people very much. A simple mean puts a sea of
numbers in a nutshell.

CENTRAL TENDENCY AND DISPERSION

Although descriptive statistics are everywhere, the descriptive statistics
used by laypeople are typically incomplete in a very important respect.
Laypeople make frequent use of descriptive statistics that summarize the
central tendency (loosely speaking, the average) of a set of observations
(“But my old pal Michael Jordan once averaged 32 points in a season”; “A
follow-up study revealed that women also happen to be exactly 67% less
likely than men to spend their weekends watching football and drinking
beer”). However, most laypeople are relatively unaware of an equally use-
ful and important category of descriptive statistics. This second category
of descriptive statistics consists of statistics that summarize the disper-
sion, or variability, of a set of scores. Measures of dispersion are not only
important in their own (descriptive) right, but as you will see later, they
are also important because they play a very important role in inferential
statistics.
One common and relatively familiar measure of dispersion is the range
of a set of scores. The range of a set of scores is simply the difference
between the highest and the lowest value in the entire set of scores. (“The
follow-up study also revealed that virtually all men showed the same
amount of shrinkage. The smallest amount of shrinkage observed in all
the male brains studied was 10.0 cc, and the largest amount observed was
11.3 cc. That’s a range of only 1.3 cc. In contrast, many of the women in
the study showed no shrinkage whatsoever, and the largest amount of
shrinkage observed was 7.2 cc. That’s a range of 7.2 cc.”) Another very
common, but less intuitive, descriptive measure of dispersion is the stan-
dard deviation. It’s a special kind of average itself—namely, an average


measure of how much each of the scores in the sample differs from the
sample mean. More specifically, it’s the square root of the average squared
deviation of each score from the sample mean, or

∑(x − m)2
S= .
n

S (sigma) is a summation sign, a symbol that tells us to perform the func-
tions that follow it for all the scores in a sample and then to add them all
together. That is, this symbol tells us to take each individual score in our
sample (represented by x), to subtract the mean (m) from it, and to square
this difference. Once we have done this for all our scores, sigma tells us to
add all these squared difference scores together. We then divide these
summed scores by the number of observations in our sample and take the
square root of this final value.
For example, suppose we had a small sample of only four scores: 2, 2, 4,
and 4. Using the formula above, the standard deviation turns out to be

(2 − 3)2 + (2 − 3)2 + (4 − 3)2 + (4 − 3)2
,
4
which is simply

1+1+1+1
,
4

which is exactly 1.
That’s it. The standard deviation in this sample of scores is exactly 1. If
you look back at the scores, you’ll see that this is pretty intuitive. The mean
of the set of scores is 3.0, and every single score deviates from this mean
by exactly 1 point. There is a computational form of this formula that is
much easier to deal with than the definitional formula shown here (espe-
cially if you have a lot of numbers in your sample). However, we included
the definitional formula so that you could get a sense of what the standard
deviation means. Loosely speaking, it’s the average (“standard”) amount
by which all the scores in a distribution differ (deviate) from the mean of
that same set of scores. Finally, we should add that the specific formula we
presented here requires an adjustment if you hope to use a sample of
scores to estimate the standard deviation in the population of scores from
which these sample scores were drawn. It is this adjusted standard devia-
tion that researchers are most likely to use in actual research (e.g., to make
inferences about the population standard deviation). Conceptually, how-
ever, the adjusted formula (which requires you to divide by n – 1 rather


than n) does exactly what the unadjusted formula does: It gives you an idea
of how much a set of scores varies around a mean.
Why are measures of dispersion so useful? Like measures of central
tendency, measures of dispersion summarize a very important property of
a set of scores. For example, consider the two groups of four men whose
heights are listed as follows:

Group 1 Group 2
Tallest guy 6′2″ 6′9″
Tall guy 6′1″ 6′5″
Short guy 5′11″ 5′10″
Shortest guy 5′10″ 5′0″

A couple of quick calculations will reveal that the mean height of the
men in both groups is exactly 6 feet. Now suppose you were a heterosexual
woman of average height and needed to choose a blind date by drawing
names from one of two hats. One hat contains the names of the four men
in Group 1, and the other hat contains the names of the four men in
Group 2. From which hat would you prefer to choose your date? If you
followed social conventions regarding dating and height, you would prob-
ably prefer to choose your date from Group 1. Now suppose you were
choosing four teammates for an intramural basketball team and had to
choose one of the two groups (in its entirety). In this case, we assume that
you would choose Group 2 (and try to get the ball to the big guy when he
posts up under the basket). Your preferences reveal that dispersion is a very
important statistical property because the only way in which the two
groups of men differ is in the dispersion (i.e., the variability) of their
heights. In Group 1, the standard deviation is 1.58 inches. In Group 2, it’s
7.97 inches.3
Another example of the utility of measures of dispersion comes from a
1997 study of parking meters in Berkeley, California. The study’s author,
Ellie Lamm, strongly suspected that some of the meters in her hometown
had been shortchanging people. To put her suspicions to the test, she con-
ducted an elegantly simple study in which she randomly sampled 50 park-
ing meters, inserted two nickels in each (enough to pay for 8 minutes), and
timed with a stopwatch the actual amount of time each meter delivered.
Lamm’s study showed that, on average, the amount of time delivered was
indeed very close to 8 minutes. The central tendency of the 50 meters was
to give people what they were paying for.
However, a shocking 94% of the meters (47 of 50) were off one way or
the other by at least 20 seconds. In fact, the range of delivered time was
about 12 minutes! The low value was just under 2 minutes, and the high


was about 14 minutes. Needless to say, a substantial percentage of the
meters were giving people way less time than they paid for. It didn’t matter
much that other meters were giving people too much time. There’s an obvi-
ous asymmetry in the way tickets work. When multiplied across the city’s
then 3,600 parking meters, this undoubtedly created a lot of undeserved
parking tickets.
Lamm’s study got so much attention that she appeared to discuss it on
the David Letterman Show. Furthermore, the city of Berkeley responded to
the study by replacing their old, inaccurate mechanical parking meters
with much more accurate electronic meters. Many thousands of people
who had once gotten undeserved tickets were presumably spared tickets
after the intervention, and vandalism against parking meters in Berkeley
was sharply reduced. So this goes to prove that dispersion is sometimes
more important than central tendency. Of course, it also goes to prove that
research doesn’t have to be expensive or complicated to yield important
societal benefits. Lamm’s study presumably cost her only $5 in nickels and
perhaps a little bit for travel. That’s good because Lamm conducted this
study as part of her science fair project—when she was 11 years old.4 We
certainly hope she won a blue ribbon.
A more formal way of thinking about dispersion is that measures of
dispersion complement measures of central tendency by telling you some-
thing about how well a measure of central tendency represents all the
scores in a distribution. When the dispersion or variability in a set of
scores is low, the mean of a set of scores does a great job of describing most
of the scores in the sample. When the dispersion or the variability in a set
of scores is high, however, the mean of a set of scores does not do such a
great job of describing most of the scores in the sample (the mean is still
the best available summary of the set of scores, but there will be a lot of
people in the sample whose scores lie far away from the mean). When you
are dealing with descriptions of people, measures of central tendency—
such as the mean—tell you what the typical person is like. Measures of
dispersion—such as the standard deviation—tell you how much you can
expect specific people to differ from this typical person.

THE SHAPE OF DISTRIBUTIONS

A third statistical property of a set of observations is a little more dif-
ficult to quantify than measures of central tendency or dispersion. This
third statistical property is the shape of a distribution of scores. One
useful way to get a feel for a set of scores is to arrange them in order from
the lowest to the highest and to graph them pictorially so that taller parts
of the graph represent more frequently occurring scores (or, in the case
of a theoretical or ideal distribution, more probable scores). Figure 1.1
depicts three different kinds of distributions: a rectangular distribution,


a bimodal distribution, and a normal distribution. The scores in a rect-
angular distribution are all about equally frequent or probable. An
example of a rectangular distribution is the theoretical distribution rep-
resenting the six possible scores that can be obtained by rolling a single
six-sided die. In the case of a bimodal distribution, two distinct ranges
of scores are more common than any other. A likely example of a
bimodal distribution would be the heights of the athletes attending the
annual sports banquet for a very large high school that has only two
sports teams: women’s gymnastics and men’s basketball. If this example
seems a little contrived, it should. Bimodal distributions are relatively
rare, and they usually reflect the fact that a sample is composed of two
meaningful subsamples. The third distribution depicted in Figure 1.1 is
the most important. This is a normal distribution: a symmetrical, bell-
shaped distribution in which most scores cluster near the mean and in
which scores become increasingly rare as they become increasingly
divergent from this mean. Many things that can be quantified are nor-
mally distributed. Distributions of height, weight, extroversion, self-
esteem, and the age at which infants begin to walk are all examples of
approximately normal distributions.
The nice thing about the normal distribution is that if you know that
a set of observations is normally distributed, this further improves your
ability to describe the entire set of scores in the sample. More specifically,
you can make some very good guesses about the exact proportion of
scores that fall within any given number of standard deviations (or frac-
tions of a standard deviation) from the mean. As illustrated in Figure 1.2,
about 68% of a set of normally distributed scores will fall within one
standard deviation of the mean. About 95% of a set of normally distrib-
uted scores will fall within two standard deviations of the mean, and well
over 99% of a set of normally distributed scores (99.8% to be exact) will
fall within three standard deviations of the mean. For example, scores on
modern intelligence tests (such as the Wechsler Adult Intelligence Scale)
are normally distributed, have a mean of 100, and have a standard devi-
ation of 15. This means that about 68% of all people have IQs that fall
between 85 and 115. Similarly, more than 99% of all people (again,
99.8% of all people, to be more exact) should have IQs that fall between
55 and 145.
This kind of analysis can also be used to put a particular score or
observation into perspective (which is a first step toward making infer-
ences from particular observations). For instance, if you know that a set
of 400 scores on an astronomy midterm (a) approximates a normal
distribution, (b) has a mean of 70, and (c) has a standard deviation of
exactly 6, you should have a very good picture of what this entire set of
scores is like. And you should know exactly how impressed to be when
you learn that your friend Amanda earned an 84 on the exam. She
scored 2.33 standard deviations above the mean, which means that she
probably scored in the top 1% of the class. How could you tell this? By


Figure 1.1 Rectangular Distribution, a Bimodal Distribution, and a Normal
A
Distribution

Mean = 3.50
Std. Dev. = 1.871
N=6
1.0
0.8
Frequency

0.6
0.4
0.2
0.0
1.00 2.00 3.00 4.00 5.00 6.00
die_rolls

Mean = 71.50
25
Std. Dev. = 11.508
N = 246

20
Frequency

15

10

5

0
40.00 50.00 60.00 70.00 80.00 90.00 100.00
new_height

12 Mean = 67.00
Std. Dev. = 4.671
N = 123
10
Frequency

8

6

4

2

0
55.00 60.00 65.00 70.00 75.00 80.00
height in inches


Figure 1.2
Percentage of Scores in a Perfectly Normal Distribution Falling Within
1, 2, and 3 Standard Deviations From the Mean

99.8%

95.4%

68.2%
0.4

0.3

0.2 34.1% 34.1%

0.1
2.1% 2.1%
0.1% 13.6% 13.6% 0.1%
0.0
−3α −2α −1α µ 1α 2α 3α

Source: Image courtesy of Wikipedia.

consulting a detailed table based on the normal distribution. Such a
table would tell you that only about 2% of a set of scores are 2.33 stan-
dard deviations or more from the mean. And because the normal dis-
tribution is symmetrical, half of the scores that are 2.33 standard
deviations or more from the mean will be 2.33 standard deviations or
more below the mean. Amanda’s score was in the half of that 2% that
was well above the mean. Translation: Amanda kicked butt on the
exam.
As you know if you have had any formal training in statistics, there is
much more to descriptive statistics than what we have covered here. For
instance, we skipped many of the specific measures of central tendency
and dispersion, and we didn’t describe all the possible kinds of distribu-
tions of scores. However, this overview should make it clear that descrip-
tive statistics provide researchers with an enormously powerful tool for
organizing and simplifying data. At the same time, descriptive statistics are
only half of the picture. In addition to simplifying and organizing the data
they collect, researchers also need to draw conclusions about populations
from their sample data. That is, they need to move beyond the data them-
selves in the hopes of drawing general inferences about people. To do this,
researchers rely on inferential statistics.


Inferential Statistics

The basic idea behind inferential statistical testing is that decisions about
what to conclude from a set of research findings need to be made in a
logical, unbiased fashion. One of the most highly developed forms of logic
is mathematics, and statistical testing involves the use of objective, math-
ematical decision rules to determine whether an observed set of research
findings is “real.” The logic of statistical testing is largely a reflection of the
skepticism and empiricism that are crucial to the scientific method. When
conducting a statistical test to aid in the interpretation of a set of experi-
mental findings, researchers begin by assuming that the null hypothesis is
true. That is, they begin by assuming that their own predictions are wrong.
In a simple, two-groups experiment, this would mean assuming that the
experimental group and the control group are not really different after the
manipulation—and that any apparent difference between the two groups
is simply due to luck (i.e., to a failure of random assignment). After all,
random assignment is good, but it is rarely perfect. It is always possible that
any difference an experimenter observes between the behavior of partici-
pants in the experimental and control groups is simply due to chance. In
the context of an experiment, the main thing statistical hypothesis testing
tells us is exactly how possible it is (i.e., how likely it is) that someone
would get results as impressive as, or more impressive than, those actually
observed in an experiment if chance alone (and not an effective manipula-
tion) were at work in the experiment.
The same logic applies, by the way, to the findings of all kinds of
research (e.g., survey or interview research). If a researcher correlates a
person’s height with that person’s level of education and observes a modest
positive correlation (such that taller people tend to be better educated), it
is always possible—out of dumb luck—that the tall people in this specific
sample just happen to have been more educated than the short people.
Statistical testing tells researchers exactly how likely it is that a given
research finding would occur on the basis of luck alone (if nothing inter-
esting is really going on). Researchers conclude that there is a true associa-
tion between the variables they have manipulated or measured only if the
observed association would rarely have occurred on the basis of chance.
Because people are not in the habit of conducting tests of statistical
significance to decide whether they should believe what a salesperson is
telling them about a new line of athletic shoes, whether there is intelligent
life on other planets, or whether their friend’s taste in movies is “signifi-
cantly different” from their own, the concept of statistical testing is pretty
foreign to most laypeople. However, anyone who has ever given much
thought to how American courtrooms work should be extremely familiar
with the logic of statistical testing. This is because the logic of statistical


testing is almost identical to the logic of what happens in an ideal court-
room. With this in mind, our discussion of statistical testing will focus on
the simile of what happens in the courtroom. If you understand court-
rooms, you should have little difficulty understanding statistical testing.
As mentioned previously, researchers performing statistical tests begin
by assuming that the null hypothesis is correct—that is, that the research-
er’s findings reflect chance variation and are not real. The opposite of the
null hypothesis is the alternative hypothesis. This is the hypothesis that
any observed difference between the experimental and the control group
is real. The null hypothesis is very much like the presumption of innocence
in the courtroom. Jurors in a courtroom are instructed to assume that they
are in court because an innocent person had the bad luck of being falsely
accused of a crime. That is, they are instructed to be extremely skeptical of
the prosecuting attorney’s claim that the defendant is guilty. Just as defen-
dants are considered “innocent until proven guilty,” researchers’ claims
about the relation between the variables they have examined are consid-
ered incorrect unless the results of the study strongly suggest otherwise
(“null until proven alternative,” you might say). After beginning with the
presumption of innocence, jurors are instructed to examine all the evi-
dence presented in a completely rational, unbiased fashion. The statistical
equivalent of this is to examine all the evidence collected in a study on a
purely objective, mathematical basis. After examining the evidence against
the defendant in a careful, unbiased fashion, jurors are further instructed
to reject the presumption of innocence (to vote guilty) only if the evidence
suggests beyond a reasonable doubt that the defendant committed the
crime in question. The statistical equivalent of the principle of reasonable
doubt is the alpha level agreed upon by most statisticians as the reasonable
standard for rejecting the null hypothesis. In most cases, the accepted
probability value at which alpha is set is .05. That is, researchers may reject
the null hypothesis and conclude that their hypothesis is correct only
when findings as extreme as those observed in the study (or more extreme)
would have occurred by chance alone less than 5% of the time.
If prosecuting attorneys were statisticians, we could imagine them ask-
ing the statistical equivalent of the same kinds of questions they often ask
in the courtroom: “Now, I’ll ask you, the jury, to assume, as the defense
claims, that temperature has no effect on aggression. If this is so, doesn’t it
seem like an incredible coincidence that in a random sample of 40 college
students, the 20 students who just happened to be randomly assigned to
the experimental group—that is, the 20 people who just happened to be
placed in the uncomfortably hot room instead of the nice, comfortable,
cool room—would give the stooge almost three times the amount of shock
that was given by the people in the control group? Remember, Mr. Heat
would have you believe that in comparison with the 20 participants in the
control group, participants number 1, 4, 7, 9, 10, 11, 15, 17, 18, 21, 22, 24,


25, 26, 29, 33, 35, 36, 38, and 40, as a group, just happened to be the kind
of people who are inherently predisposed to deliver extremely high levels
of shock. Well, in case you’re tempted to believe this load of bullsh—.” “I
object, your Honor! The question is highly inflammatory,” the defense
attorney interrupts. “Objection overruled,” the judge retorts. “As I was say-
ing, in case any one of you on the jury is tempted to take this claim seri-
ously, I remind you that we asked Dr. R. A. Fisher, an eminent
mathematician and manurist, to calculate the exact probability that some-
thing this unusual could happen due to a simple failure of random assign-
ment. His careful calculations show that if we ran this experiment
thousands of times without varying the way the experimental and control
groups were treated, we would expect to observe results as unusual as these
less than one time in a thousand if the manipulation truly has no effect!
Don’t you think the defense is asking you to accept a pretty incredible
coincidence?”
A final parallel between the courtroom and the psychological labora-
tory is particularly appropriate in a theoretical field such as psychology. In
most court cases, especially serious cases such as murder trials, successful
prosecuting attorneys will usually need to do one more thing in addition
to presenting a body of logical arguments and evidence pointing to the
defendant. They will need to identify a plausible motive, a good reason why
the defendant might have wanted to commit the crime. It is difficult to
convict people solely on the basis of circumstantial evidence. A similar
state of affairs exists in psychology. No matter how “statistically signifi-
cant” a set of research findings is, most psychologists will place very little
stock in it unless the researcher can come up with a plausible reason why
one might expect to observe those findings. In psychology, these plausible
reasons are called theories. It is quite difficult to publish a set of significant
empirical findings unless you can generate a plausible theoretical explana-
tion for them.
Having made this “friendly pass” through a highly technical subject,
we will now try to enrich your understanding of inferential statistics by
using inferential statistics to solve a couple of problems. In an effort to
keep formulas and calculations as simple as possible, we have chosen
some very simple problems. Analyzing and interpreting the data from
most real empirical investigations require more extensive calculations
than those you will see here, but of course these labor-intensive calcula-
tions are usually carried out by computers. In fact, a great deal of your
training in this text will involve getting a computer to crunch numbers
for you using the statistical software package SPSS. Regardless of how
extensive the calculations are, however, the basic logic underlying infer-
ential statistical tests is almost always the same—no matter which spe-
cific inferential test is being conducted and no matter who, or what, is
doing the calculations.


PROBABILITY THEORY

As suggested in the thought experiment with American courtrooms, all
inferential statistics are grounded firmly in the logic of probability theory.
Probability theory deals with the mathematical rules and procedures used
to predict and understand chance events. For example, the important sta-
tistical principle of regression toward the mean (the idea that extreme
scores or performances are usually followed by less extreme scores or per-
formances from the same person or group of people) can easily be derived
from probability theory. Similarly, the odds in casinos and predictions
about the weather can be derived from straightforward considerations of
probabilities. What is a probability? From the classical perspective, the
probability of an event is a very simple thing: It is (a) the number of all
specific outcomes that qualify as the event in question divided by (b) the
total number of all possible outcomes. The probability of rolling a 3 on a
single roll with a standard six-sided die is 1/6, or .167, because there is
(a) one and only one roll that qualifies as a 3 and (b) exactly six (equally
likely) possible outcomes. For the same reason, the probability of rolling an
odd number on the same die is 1/2 or .50—because three of the six possible
outcomes qualify as odd numbers. It is important to remember that the
probability of any event (or complex set of events, such as the observed
results of an experiment) is the number of ways to observe that event
divided by the total number of possible events.
With this in mind, suppose the Great Pumpkini told you that he had
telekinetic powers that allow him to influence the outcome of otherwise
fair coin tosses. How could you test his claim? One way would be to ask
him to predict some coin tosses and to check up on the accuracy of his
predictions. Imagine that you pulled out a coin, tossed it in the air, and
asked Pumpkini to call it before it landed. He calls heads. Heads it is! Do
you believe in Pumpkini’s self-proclaimed telekinetic abilities? Of course
not. You realize that this event could easily have occurred by chance. How
easily? Fully half the time we performed the test. With this concern in
mind, suppose Pumpkini agreed to predict exactly 10 coin tosses. Let’s
stop and consider a number of possible outcomes of this hypothetical
coin-tossing test. To simplify things, let’s assume that Pumpkini always
predicts heads on every toss.
One pretty unremarkable outcome is that he’d make 5 of 10 correct pre-
dictions. Should you conclude that he does, indeed, have telekinetic abilities?
Or that he is half telekinetic (perhaps on his mother’s side)? Again, of course
not. Making 5 of 10 correct predictions is no better than chance. To phrase
this in terms of the results of the test, the number of heads we observed was
no different than the expected frequency (the average, over the long run) of a
random series of 10 coin tosses. In this case, the expected frequency is the
probability of a head on a single toss (.50) multiplied by the total number of
tosses (10). But what if Pumpkini made six or seven correct predictions


instead of only five? Our guess is that you still wouldn’t be very impressed
and would still conclude that Pumpkini does not have telekinetic abilities (in
statistical terms, you would fail to reject the null hypothesis). OK, so what if
he made a slightly more impressive eight correct predictions? What about
nine? You should bear in mind that Pumpkini never said his telekinetic pow-
ers were absolutely flawless. Pumpkini can’t always carry a glass of water
across a room without spilling it, but his friends usually allow him to carry
glasses of water unassisted. Despite your firmly entrenched (and justifiable)
skepticism concerning psychic phenomena, we hope you can see that as our
observations (i.e., the results of our coin-tossing test) depart further and
further from chance expectations, you would start to become more and
more convinced that something unusual is going on. At a certain point,
you’d practically be forced to agree that Pumpkini is doing something to
influence the outcome of the coin tosses.
The problem with casual analysis is that it’s hard to know exactly where
that certain point is. Some people might be easygoing enough to say they’d
accept eight or more heads as compelling evidence of Pumpkini’s teleki-
netic abilities. Other people might ask to see a perfect score of 10 (and still
insist that they’re not convinced). After all, extraordinary claims require
extraordinary evidence. That’s where inferential statistics come in. By
making use of (a) some basic concepts in probability theory, along with
(b) our knowledge of what a distribution of scores should look like when
nothing funny is going on (e.g., when we are merely flipping a fair coin
10 times at random, when we are simply randomly assigning 20 people to
either an experimental or a control condition), we can use inferential sta-
tistics to figure out exactly how likely it is that a given set of usual or not-
so-usual observations would have been observed by chance. Unless it is
pretty darn unlikely that a set of findings would have been observed by
chance, the logic of statistical hypothesis testing requires us to conclude
that the set of findings represents a chance outcome.
To return to our coin-tossing demonstration, just how likely is it that a
person would toss 9 or more heads by chance alone? One way to figure
this out is to use our definition of probability and to figure out (a) all the
specific ways there are to observe 9 or more heads in a string of 10 coin
tosses and (b) all the specific outcomes (of any kind) that are possible for
a string of 10 coin tosses. If we divide (a) by (b), we should have our
answer. Let’s begin with the number of ways there are to toss 9 or more
heads. At the risk of sounding like the announcer who was describing
LeBron James’s scoring history without using statistics, notice that one
way to do it would be to toss a tail on the first trial, followed by 9 straight
heads. A second way to do it would be to toss a head on the first trial and
a tail on the second trial, followed by 8 straight heads. If you follow this
approach to its logical conclusion, you should see that there are exactly
10 specific ways to observe exactly 9 heads in a string of 10 coin tosses.
And in case you actually want to see the 10 ways right in front of you, they


appear in Table 1.1—along with all of the unique ways there are to
observe exactly 10 heads. As you already knew, there is only one of them.
However, it’s important to include this one in our list because we were
interested in all of the specific ways to observe 9 or more heads in a series
of 10 coin tosses.5 So there are 11 ways.
But how many total unique outcomes are there for a series of 10 coin
tosses? To count all of these would be quite a headache. So we’ll resort to a
less painful headache and figure it out logically. How many possible ways
are there for 1 toss to come out? Two: heads or tails—which turns out to be
21 (2 to the first power). How about 2 tosses? Now we can observe 22 (2 × 2)
or four possible ways—namely,

HH, HT, TH, or TT.

What about three tosses? Now we have 23 (2 × 2 × 2), or eight possible
ways:

HHH, HHT, HTH, THH, HTT, THT, TTH, or TTT.

Notice that our answers always turn out to be 2 (the number of unique
outcomes for an individual toss) raised to some power. The power to
which 2 is raised is the number of trials or specific observations we are
making. So the answer is 210 (2 × 2 × 2 × 2 × 2 × 2 × 2 × 2 × 2 × 2) or 1,024
possible unique outcomes for a series of 10 coin tosses. This value of 1,024
includes every possible number of heads (from 0 to 10) and every possible
order or position (1st through 10th) for all of these possible numbers of
heads. So now we have our probability. The probability of observing 9 or
more heads in a series of 10 truly random coin tosses is thus 11/1,024, or
.011. So for every hundred times we conducted our coin-tossing study,
you’d expect to see 9 or more heads only about once. That’s only 1% of the
time, and it’s pretty impressive. (In fact, it’s exactly as impressive as
Amanda’s score on the astronomy midterm, and we, for what it’s worth,
were very impressed with Amanda.) So if we had treated the study like a
real experiment, if we had set alpha at .05, and if we had observed 9 heads,
we would have had to conclude that Pumpkini does, in fact, possess the
ability to influence the outcomes of otherwise fair coin tosses.
Now perhaps you’re the literal type who is saying, “But wait a minute.
I still wouldn’t believe Pumpkini has telekinetic abilities, and I certainly
don’t think most scientists would, either.” You are correct, of course,
because the theory that you have been asked to accept flies in the face of
everything you know about psychology and physics. A much more reason-
able explanation for the observed findings is that Pumpkini has engaged
in some form of trickery, such as using a biased coin. However, this simply
means that, like any scientific practice, the practice of conducting statisti-
cal tests must be carried out using a little common sense. If someone is


Table 1.1 the Possible Ways to Toss Nine or
All
More Heads in 10 Tosses of a Fair Coin:
A Single Tail Can Come on Any of the
10 Trials, or It Can Never Come at All

1. THHHHHHHHH
2. HTHHHHHHHH
3. HHTHHHHHHH
4. HHHTHHHHHH
5. HHHHTHHHHH
6. HHHHHTHHHH
7. HHHHHHTHHH
8. HHHHHHHTHH
9. HHHHHHHHTH
10. HHHHHHHHHT
11. HHHHHHHHHH

making a truly extraordinary claim, we might want to set alpha at .001, or
even .0001, instead of .05. Of course, setting alpha at a very low value
might require us to design a test with a much greater number of coin
tosses (after all, 10 out of 10 tosses—the best you can possibly do—has a
probability higher than .0001; it’s 1/1,024, which is closer to .001), but the
point is that we could easily design the test to have plenty of power to see
what is going on. The exact design of our study is up to us (and, to some
extent, to our critics). If people are sufficiently skeptical of a claim, they
might also want to see a replication of a questionable or counterintuitive
finding. If Pumpkini replicated his demonstration several times by cor-
rectly predicting 9 or more heads, and if we enacted some careful control
procedures to prevent him from cheating (e.g., we let a group of skeptics
choose and handle the coins), even the most ardent anti-telekinetician
should eventually be persuaded. And if he or she weren’t, we would argue
that this person wasn’t being very scientific.
The logic of the coin-tossing experiment is the same as the logic under-
lying virtually all inferential statistical tests. First, a researcher makes a set
of observations. Second, these observations are compared with what we
would expect to observe if nothing unusual were happening in the exper-
iment (i.e., if the researcher’s hypothesis were incorrect). This comparison


is ultimately converted into a probability—namely, the probability that the
researcher would have observed a set of results at least this consistent with
his or her hypothesis if the hypothesis were incorrect. Finally, if this prob-
ability is sufficiently low, we conclude that the researcher’s hypothesis is
probably correct. Because inferential statistics are a very important part of
the research process, let’s look at another highly contrived but informative
question that could be answered only with the use of inferential statistics.

A STUDY OF CHEATING

Suppose we offered a group of exactly 50 students the chance to win a
very attractive prize (say, a large amount of cash, or an autographed copy
of this textbook) by randomly drawing a lucky orange ping-pong ball out
of a large paper bag. Assume that each student gets to draw only one ball
from the bag, that students return the drawn balls to the bag after each
drawing, and that the bag contains exactly 10 balls, only 1 of which is
orange. Because our university is trying to teach students the values of
honesty and integrity, university regulations require us to administer the
drawing on an honor system. Specifically, the bag of ping-pong balls is
kept behind a black curtain, and students walk behind the curtain—one at
a time, in complete privacy—to draw their balls at random from the bag.
After drawing a ball, each student holds it up above the curtain for every-
one else to see. Anyone who holds up an orange ball is a winner.
Suppose that we’re the curious types who want to find out if there was
a significant amount of cheating (peeking) during the drawing. At first
blush, it would seem like there’s nothing we could do. Unless we engage in
a little cheating ourselves (e.g., by secretly videotaping the drawings), how
can we figure out whether people were peeking as they selected their balls?
We’re at a complete loss to observe the unobservable—unless we rely on
inferential statistics. By using inferential statistics, we could simply calcu-
late the number of winners we’d expect to observe if no one was cheating.
By making a comparison between this expected frequency and the number
of winners we actually observed in our drawing, we could calculate the
exact probability (based on chance alone) of obtaining a result as extreme
as, or more extreme than, the result of our actual drawing. If the probabil-
ity of having so many winners were sufficiently low, we might reluctantly
reject the null hypothesis (our initial assumption that the students were all
innocent until proven guilty) and conclude that a significant amount of
cheating was happening during the drawing.
Let’s find out. To begin with, we need to assume that our suspicions
about cheating are completely unfounded and that no one peeked (as usual,
we begin by assuming the null hypothesis). Assuming that no one was
peeking, what’s your best guess about how many of the 50 students should
have selected a winning ball? If you are a little fuzzy on your probability


theory, remember that you can figure out the expected frequency of an
event by multiplying (a) the probability of the event on a single trial by
(b) the total number of trials in the series of events. This is how we knew
that 5 was the expected number of observed heads in a series of 10 coin
tosses. It was .50 × 10. The answer here is also 5 (it’s .10 × 50). Now imagine
that we had 6 winners. Or 9 winners, or 15—or 50. Hopefully, you can see,
as you did in the coin-tossing study, that as our observed frequencies depart
further and further from the frequency we’d expect by chance, we become
more and more strongly convinced that our observed frequencies are not
the product of chance.
For the purposes of actually seeing some inferential statistics in action,
let’s assume that we had exactly 10 winners in our drawing. Because our
outcome was a categorical outcome (“success” or “failure” at the draw),
and because we had a pretty large sample, we’d probably want to conduct
a χ2 (chi-square) test on these data. The formula for this test appears as
follows:
( f o − f e )2
χ2 = ∑ .
fe

Recall that Σ (sigma) is a summation sign that tells you to add together all
the appropriate examples of the basic calculation.

fo refers to the observed frequencies of each of the events you care about
(successes and failures when it comes to sampling a lucky orange ball).
fe refers to the expected frequencies for each of these same events.
You could think of a χ2 statistic as a “surprise index.” Notice that the
most important thing the formula does is to compare expected and
observed frequencies. Specifically, expected frequencies are compared
with (i.e., subtracted from) observed frequencies, and then a couple of
simple transformations are made on these difference scores. The more
our observed frequencies depart from what you’d expect if chance alone
were operating (i.e., the more surprising our results are), the bigger our
χ2 statistic becomes. And as our χ2 statistic grows, it tells us that it’s less
and less likely that we’re observing a chance process (and, in this case,
more and more likely that we’re observing cheating).
The χ2 value for 10 winners (out of 50) when only 5 were expected is
computed as follows:

(10 − 5)2 (40 − 45)2 .
χ2 = +
5 45

The 10 in the first half of the equation is the observed frequency of suc-
cesses, and the two 5s both refer to the expected frequency of successes.


The 40 in the second half of the equation is the observed frequency of
failures, and the two 45s both refer to the expected frequency of failures
(this has to be the sample size, which is 50, minus the expected number
of successes). When we do the math, we get 25/5 + 25/45, which works
out to 5.55. Notice that this isn’t a probability. The way most inferential
statistics work is that you generate both the statistic itself (e.g., a cor-
relation coefficient, a t value, an F ratio) and then use the exact value of
the observed statistic to determine a probability value (one that corre-
sponds to the value of your statistic). If you are doing your calculations
on a computer, the software program you are using will always do this
for you. That is, it will give you the exact p value (i.e., the exact proba-
bility) that corresponds to your results after they have been converted
to the unambiguous language of your statistic. However, if you are
doing your calculations by hand, as we have here, you will need to con-
sult some kind of statistical table to see what the critical values are for
your statistic. In the case of our study of cheating, the critical χ2 value
that corresponds to an alpha level of .05 is 3.841. Any χ2 value that
exceeds this score will have an associated p value that is lower than .05
and will thus be significant when alpha is set at .05. If we were a little
bit more stringent, we might set alpha at .02 or .01. Our χ2 table hap-
pens to include critical values for each of these levels of alpha (i.e., for
each of these probability values). In a study such as ours, the critical χ2
value for an alpha of .02 is 5.412, and the critical χ2 value for an alpha
of .01 is 6.635. By these criteria, our result is still significant even if
alpha is set at .02. However, if we move to the still more stringent alpha
level of .01, the number of winners we observed would no longer be
significant (because we’re effectively saying that it’d take more than
10 winners to convince us).
Suppose we followed standard practice and set alpha at .05. We’d have
to conclude that some people cheated. Notice, however, that we couldn’t
draw any safe inferences about exactly who cheated. Presumably about 5 of
our 10 winners just got lucky, and about 5 cheated. Realizing that only
about 5 people cheated provides a different sort of perspective on our
findings. Specifically, it highlights the fact that there is often more than
one way to look at a set of observations. Notice also that an alternate, and
equally correct, perspective on our observation is that people are signifi-
cantly honest! It appears to be the case that about 45 of our 50 students
were completely honest—even in a situation that allowed rampant cheat-
ing. Why did we say 45? Because we just decided that only about 5 people
are likely to have cheated. In light of how hard it is to win the game by
playing fairly, these 5 or so cheaters led to a significant amount of cheating.
However, if we had started out with the hypothesis that 49 of 50, or 98%,
of all people should be expected to cheat under these conditions, and if we
had taken 49 (nearly absolute dishonesty) as our standard of comparison


rather than 5 (absolute honesty), we would have obtained an extremely
large χ2 value:

(10 − 49)2 (40 − 1)2
χ2 = + ,
49 1

which is 1552.04, and which corresponds to an infinitesimally small
p value. Even if we set alpha at a very, very, very low level (say one in a
billion, or .000000001), this would still be significant. In other words, it’s
important to keep in mind that we appear to have observed a lot more
honesty than cheating.6
A final aspect of this exercise about drawing ping-pong balls from a bag
is that it provides a useful metaphor for thinking about what researchers
do when they draw inferences about people in their research. Notice that
in the lottery involving ping-pong balls, we could not directly observe the
phenomenon in which we were interested. The activities we cared about
were shrouded behind a black curtain—just as the activities that psycholo-
gists often care about (e.g., dissonance reduction, feelings of passionate
love, parallel distributed representations of language) are hidden inside
the black box of people’s minds. Inferential statistics work hand in hand
with things like operational definitions to allow us to make scientific infer-
ences. Operational definitions allow us to draw inferences about processes
that we cannot observe (those that occur inside the person), and inferen-
tial statistics allow us to draw inferences about people we can’t observe
(those we didn’t sample in our study). When we conclude that a research
finding is significant, we are concluding that it is real and thus that it
applies to people who did not take part in our study. This is one sense in
which the ping-pong ball demonstration is a little different from most
significance tests. Although it would probably be safe to generalize our
findings about cheating to other college students, what we really cared
about most in this particular test was finding out what was going on in our
particular sample.
Virtually every inferential statistic that you will ever come across will
be based on the logic that was explicated here. Of course, the particular
distributions of responses that researchers examine vary enormously
from one study to the next, and this, among other things, influences the
particular statistics that researchers use to summarize and draw infer-
ences about their data. Moreover, once a researcher has chosen a particu-
lar statistic, the specific calculations that she or he will have to carry out
(or get a computer to carry out) will typically be a good bit more involved
than those you have seen here. For example, in a two-way analysis of vari-
ance (ANOVA), there are separate calculations (and separate degrees of
freedom) for each of the two possible main effects as well as for the two-
way interaction. No matter what statistics they are computing, however,


researchers will always rely on the logic of probability theory to help them
make their case that something significant is at the root of their empirical
observations.

Things That Go Bump in the Light:
Factors That Influence the Results of Significance Tests

ALPHA LEVELS AND TYPE I AND II ERRORS

Now that you have a better feel for what it means for a research finding
to be statistically significant, we feel that it is our duty to warn you that
when we look at significance testing in the cold, hard light of day, it has a
couple of limitations. In other words, there are a few things that can go
wrong when people are conducting statistical significance testing. First of
all, it is important to remember that when a researcher conducts a statisti-
cal test and obtains a significant result, this does not always mean that his
or her hypothesis is correct. Even if a study is perfectly executed with no
systematic design flaws, it is always possible that the researcher’s results
were due to chance. In fact, the p value we observe in an experiment tells
us exactly how likely it is that we would have obtained results like ours
even if nothing but dumb luck were operating in our study. Statisticians
refer to this worrisome possibility—incorrectly rejecting the null hypoth-
esis when it is, in fact, correct—as a Type I error. The likelihood of making
a Type I error is a direct function of where we set our alpha level. As sug-
gested earlier, if we think it would be a practical or scientific disaster to
reject the null hypothesis in error, we might want to set alpha at a very
conservative level, such as .001. Then we would be taking only one chance
in a thousand of falsely rejecting the null hypothesis.
So why not set alpha at .001 (or even lower) all the time? Because we
have to strike a balance between being cautious and being so cautious that
we become downright foolish. In statistical terms, if we always set alpha
at an extraordinarily low level, we would decrease the likelihood of com-
mitting a Type I error at the expense of increasing the likelihood of com-
mitting a Type II error. A Type II error occurs when we fail to reject an
incorrect null hypothesis—that is, when we fail to realize that our study
has revealed something meaningful (usually that our hypothesis is cor-
rect). The reason it is useful to know about Type I and Type II errors is
that there are things we can do to minimize our chances of making both
of these troublesome mistakes. As suggested previously, one of the easiest
ways to minimize Type I errors is to set alpha at a pretty low level. Over
the years, most researchers have pretty well agreed that .05 is a reasonable
level for alpha (i.e., a reasonable risk for making a Type I error). And of
course, if we want to be a little more cautious, but we don’t want to ask


anyone to adjust any alpha levels, we can always insist on seeing a replica-
tion. In the grand scheme of things, replications are what tell us whether
an effect is real.

EFFECT SIZE AND SIGNIFICANCE TESTING

Although no one wants to make a Type I error, no one really wants to
make a Type II error either. Several things influence the likelihood that a
researcher will make a Type II error (and fail to detect a real effect). Some
of these are things over which researchers have little or no control, and
some of them are things over which researchers have almost complete
control. One thing that researchers can’t do too much about is their “effect
size,” the magnitude of the effect in which they happen to be interested. If
you collected a sample of 20 people and measured their heights and their
foot sizes, you could probably expect to observe a statistically significant
correlation between height and foot size, even though your sample was
pretty small. This is because there is a pretty robust tendency for big peo-
ple to have big feet. Of course there are exceptions, but they are relatively
rare. We doubt that you will ever meet a gymnast who squeezes into a size 14
(or an NBA center who slips comfortably into a size 9). On the other hand,
if you gave a sample of 20 people a measure of extraversion and a measure
of self-esteem, you might not necessarily observe a significant correlation.
Although self-esteem and extraversion do tend to go hand in hand, this
correlation is much more modest than the substantial correlation between
height and foot size. To return to our example about peeking and ping-
pong balls, it would have been much easier to detect an effect of cheating
if cheating had been rampant. In fact, notice that in this study, it was
quite easy to detect an effect of honesty—precisely because honesty was so
rampant.

MEASUREMENT ERROR AND SIGNIFICANCE TESTING

Although it’s obviously impossible to change the true size of an effect,
one thing that researchers can sometimes do to maximize their chances of
detecting a small effect is to conduct a within-subjects or repeated mea-
sures study. As we argue later in this text, within-subjects designs are usu-
ally more sensitive than between-subjects designs. One of the reasons this
is the case is that within-subjects designs cut down on extraneous sources
of variability that can mask an effect. A person in a cool room might
deliver high levels of shock to a confederate just because this person hap-
pens to be an unusually aggressive person. However, if we could observe
the behavior of the same person in both a hot and a cool room (and if we


could make sure the person didn’t know that she or he was being studied),
we would presumably see that the person would deliver even higher levels
of shock when the temperature was cranked up a bit. Of course, another
reason why within-subjects designs are more powerful than between-
subjects designs is that they simply increase the number of observations in
a study. If we measure the aggressive behavior of each of our 20 partici-
pants in both a hot and a cool room, it is almost as if we had 40 partici-
pants in our study rather than 20 (see Pelham, 1993, for a further
discussion of the advantages of within-subjects designs).

SAMPLE SIZE AND SIGNIFICANCE TESTING

When researchers are unable to make use of within-subjects designs, they
can still do a couple of things to maximize their chances of detecting a real
effect. One simple, albeit potentially expensive, thing that researchers can do
is to conduct studies with a lot of participants. Increasing your sample size
in a study (whether it be an experiment, a quasi-experiment, a survey, or an
archival study) can greatly increase the chances that you will detect a real
effect. For example, suppose that the true correlation between extroversion
and self-esteem among American adults is exactly .32. And suppose that you
conducted a survey of 27 randomly sampled American adults and observed
a correlation of exactly .32 in your study. Would this be statistically signifi-
cant? Unfortunately not. In a sample of only 27 people, a correlation of .32
would have a p value slightly greater than .10—at best a marginally signifi-
cant value. On the other hand, if you had sampled 102 people rather than
27, and if you happened to hit the nail on the head again by observing
another correlation of exactly .32, this result would be significant even if you
had set alpha at .001. That’s because when you have a sample as large as the
second, it’s quite unusual to observe a correlation as large as .32 when the
two variables in question are actually unrelated. If this doesn’t quite seem
right to you, consider your own intuitive conclusions when we asked you
earlier what you’d think if Pumpkini were able to correctly predict 6 heads
in 10 coin tosses. If he produced exactly the same proportion of heads (600)
in 1,000 tosses, you should be much more impressed.

RESTRICTION OF RANGE AND SIGNIFICANCE TESTING

Limits in the range or variability of the variables you are measuring or
manipulating (i.e., restriction of range) can also limit your ability to detect
a true effect. Wording your dependent measures carefully, choosing the
right population, and making sure that your independent variable is as
potent and meaningful as possible (which means not shooting yourself in
the foot by artificially diminishing your real effect size) are all potential


solutions to the problem of restriction of range, and thus, they are all
potential solutions to the problem of avoiding Type II errors. The particu-
lar statistical analysis that you conduct can also play an important role in
whether your research findings are significant. When you have a choice
between conducting a powerful test (one that can detect even relatively
small effects) and a less powerful test, you should always perform the more
powerful of the two. For example, performing a correlation between two
continuous variables (e.g., self-esteem and the number of minutes people
spent reading positive feedback about themselves) is usually more power-
ful than performing a median split (e.g., on self-esteem) and then con-
ducting an ANOVA or t test to see if the mean difference between the low
group and the high group is significant. Similarly, making use of continu-
ous (“How much did you like your partner?”) rather than dichotomous
(“Did you like your partner?”) dependent measures in an experiment usu-
ally allows for more powerful statistical tests. As a second example, when
you have a choice of conducting more than one separate between-subjects
analysis (e.g., three different between-subjects ANOVAs, one on each of
your three different dependent measures) versus a single within-subjects
or mixed-model analysis on the same set of research findings (e.g., because
you asked people to rate a target person on positive, neutral, and negative
traits), you will usually be better served by the analysis that incorporates
the within-subjects aspect of your design.
The issues discussed here can help you to conduct better research stud-
ies. Just as important, they can also help you to better interpret the find-
ings of other people’s studies. For example, if a team of experimenters
claims that they failed to replicate an important effect, you would do well
to ask a few questions about the nature of their manipulation, the nature
of their sample, the wording of their dependent measures, and the number
of participants they included in their between- or within-subjects study
before you abandon your own research on the same topic. If Dr. Snittle
noted that he failed to replicate Phillips’s archival research on suicide by
noting that none of the 23 people in his small Nebraska farming commu-
nity committed suicide after reading about a front-page suicide, this
wouldn’t be much cause for concern. However, if Dr. Snittle learned to speak
fluent Mandarin, traveled to China, gained access to media and suicide
records in several very large Chinese provinces, duplicated Phillips’s analyti-
cal strategies perfectly, and failed to replicate some aspect of Phillips’s find-
ings, we’d want to figure out why. Perhaps some aspect of Chinese culture
(or Chinese media coverage) is responsible for the difference. This way of
thinking about how to interpret statistics is consistent not only with com-
mon sense but also with the logic of the scientific method. It is important to
remember that statistics are simply a tool. When effectively applied to an
appropriate problem, statistics can be incredibly powerful and effective.
However, when misapplied or misinterpreted, statistics—like real tools—
can be useless or even dangerous.


The Changing State of the Art:
Alternate Perspectives on Statistical Hypothesis Testing

During the past three quarters of a century, statistical hypothesis testing
has become a methodological touchstone for evaluating specific research
findings. When a provocative research finding proves to be statistically
significant, it is considered scientifically meaningful. When an equally
provocative research finding proves to be nonsignificant, it usually is not
taken seriously in scientific circles. As we have just seen, however, an abso-
lute reliance on significance testing—when divorced from basic consider-
ations involving things such as effect size or sample size—can often lead
researchers to inappropriate conclusions. Another way of putting this is
that there is more to hypothesis testing than simple significance testing.
Critics of significance testing have pointed out, for example, that even
when a study is well designed, basing a decision about whether an effect is
real solely on the basis of statistical “significance” is not always advisable.
In actual practice, for example, when a researcher conducts a study whose
results are promising but not significant, the researcher will often run
additional participants—or modify the design and run the study again—
rather than concluding that the original hypothesis is incorrect. In fact,
some researchers have argued that the traditional use of significance test-
ing is an inherently misleading process that should be abandoned in favor
of other approaches (J. Cohen, 1994).
Although it seems unlikely that significance testing will be abandoned
any time in the near future, most researchers would probably agree that it
is often useful to complement significance testing with other indicators
of the validity, meaningfulness, or repeatability of an effect. A complete
review of the pros and cons of alternate approaches to significance testing
is beyond the scope of this book. However, it is probably worth noting that
researchers have recently begun to complement significance testing by
making use of special statistics to assess the practical or theoretical mean-
ingfulness of research findings. One way in which researchers have done
this is to compute estimates of effect sizes, that is, indicators of the
strength or magnitude of their effects. A second way is to compute esti-
mates of (a) the overall amount of existing support for an effect or (b) the
consistency or repeatability of the effect. The statistical approach most
suited to this second category of questions is referred to as meta-analysis.

ESTIMATES OF EFFECT SIZE

When researchers want to assess the practical or theoretical rather than
the statistical significance of a specific research finding—that is, when they
want to know how big or meaningful an effect is—they typically calculate


an effect size. Although there are many useful indicators of effect size, the
two most commonly reported indicators of effect size are probably r and
d. The statistic r is the familiar correlation coefficient, and thus, you prob-
ably have had some practice interpreting this frequently used indicator of
effect size. Psychological effects that are considered small, medium, and
large correspond respectively to correlations of about .10, .30, and .50. The
less familiar statistic d is more likely to be used to describe effect sizes from
experiments or quasi-experiments because d is based on the difference
between two treatment means. Specifically,

d = (mean 1 − mean 2)/D,

where D is simply the overall standard deviation of the dependent measure in
the sample being studied (see Rosenthal Rosnow, 1991, p. 302). Thus, d tells
us how different two means are in standard deviation units (or fractions
thereof). Because two means in a study can sometimes be more than one
standard deviation apart, this means that d, unlike r, can sometimes be larger
than 1. Otherwise, the interpretation of d is pretty similar to the interpretation
of r. The respective values of d that correspond to small, medium, and large
effects are about .20, .50, and .80 (see Rosenthal Rosnow, 1991, p. 444).
Notice that we used the word about when we listed the specific values of r
and d that correspond to different effect sizes. The reason we did so is that
what makes an effect big or small is partly a judgment call. Moreover, how
“big” an effect must be to qualify as meaningful varies quite a bit from one
research area to another. If a cheap and easy-to-administer treatment (e.g., a
daily vitamin C tablet) could reduce the risk of cancer and turned out to have
a “small” effect size (e.g., r = .10 or less), this could easily translate into mil-
lions of saved dollars in medical expenses (and thousands of saved lives).
Moreover, as we just noted, the size of an effect that researchers observe in a
particular study is as much a function of how carefully the study is crafted as
it is a function of the state of the world. Thus, considerations of effect size,
like considerations of statistical significance, should reflect the theoretical or
the practical significance of a given finding—regardless of its absolute mag-
nitude. If our easy-to-administer experimental treatment gets blood from
only 10% of the turnips that we treat, we will have to consider the relative
value of blood and turnips before deciding how meaningful the treatment is.
For many years, when researchers wanted to know how strongly two
variables were related, they would compute a coefficient of determination
by squaring the correlation associated with a particular effect. So if
researchers learned, for example, that people’s attitudes about a politician
correlated .40 with whether people voted for that politician, the research-
ers might note that attitudes about candidates account for only 16% of the
variance in voting behavior (.40 × .40 = .16, or 16%). Although this is a
technically accurate way of summarizing the association between two vari-
ables, some researchers have noted that it provides a misleading picture of


the true strength of the relation between two variables. In particular,
Rosenthal and Rubin (1982) developed the binomial effect size display as
a more intuitive way to illustrate the magnitude and practical importance
of a correlation. The binomial effect size display is referred to as binomial
because it makes use of variables that can take on only two values (success
or failure, survival or death, male or female) to illustrate effect sizes. As
matters of convenience and simplicity, Rosenthal and Rubin demonstrate
effect sizes using two dichotomous variables whose two values are equally
likely. To simplify matters further, they express binomial effect sizes using
samples in which exactly 100 people take on each of the two values of each
of the two dichotomous variables.
Consider a hypothetical example involving attendance at a review ses-
sion and performance on a difficult exam. Assume (a) that exactly 100 of
200 students attended the review session and (b) that exactly 100 of 200
students passed the exam. If we told you that the correlation between
attending the review session and passing the exam was .20 (meaning that
attendance at the review session accounts for only 4% of the variance in
exam performance), you might not bother to attend the review session.
However, if you examine the binomial effect size display that appears in
Table 1.2, you can see that a correlation of .20 corresponds to 20 more
people passing than failing the exam in the group of attendees (and
20 more people failing than passing the exam in the group of nonattend-
ees). More generally, when summarized using a binomial effect size display, a
correlation coefficient corresponds to the difference in success rates that exists
between two groups of interest on a dichotomous outcome. If the correlation
summarized in Table 1.2 had been .40, we would have seen that 70% of
those attending the review (and only 30% of those failing to attend) passed
the exam (70 – 30 = 40). Similarly, if we had observed a potential cookie
thief for 200 days, if the person had been present in the kitchen for exactly
100 of the 200 days, and if cookies had disappeared on exactly 100 of the
200 days, then a correlation of .66 would mean that when the potential
thief visited the kitchen, cookies disappeared on 83 out of 100 days (83 –
17 = 66). Even though the presence of this person accounts for only about
44% of the variance in cookie thefts (.662 = .436), notice that cookies are
almost five times more likely to disappear when the person is present than
when the person is absent (.17 × 5 = .85).
Regardless of what format researchers use to illustrate effect sizes,
reporting effect sizes provides a very useful complement to traditional
significance testing. For example, suppose we know that the effect size for
a specific research finding corresponds to a d of .43. If a researcher claims
that he failed to replicate this finding, it would be useful to consider the
effect size the researcher observed (rather than focusing solely on his
observed p value) before concluding that his finding is different from the
original (see Rosenthal Rosnow, 1991, for a much more extensive
discussion). In some cases, researchers have claimed that they failed to


Table 1.2
Performance on an Exam as a Function of Attendance at a Review
Session

Exam Performance
Attendance at review Passed Failed Total
Attended 60 40 100
Did not attend 40 60 100
Total 100 100 200

replicate findings when they observed effects that were just as large as, or
larger than, those observed by previous researchers (e.g., when the second
group of researchers had a much smaller sample than the first).

META-ANALYSIS

Estimates of effect size, such as r or d, provide a useful metric for describ-
ing and evaluating the magnitude of specific research findings. Regardless of
how big a specific finding is, however, researchers are often interested in ques-
tions that have to do with the consistency or repeatability of the finding.
Questions about the repeatability of a finding almost always have to do with
a group of studies (and perhaps even an entire literature) rather than a single
specific study. How many failed studies would have to exist to indicate that a
set of findings is a statistical fluke rather than a bona fide phenomenon? If a
phenomenon is bona fide, how consistently has it been observed from study
to study? Even more important, what are the limiting conditions of the effect?
That is, when is the effect most and least likely to be observed? Questions such
as these can rarely be answered by any single study. Instead, researchers need
systematic ways to summarize the findings of a large number of studies.
Fortunately, researchers have developed a special set of statistical tech-
niques to summarize and evaluate entire sets of research findings. Not
surprisingly, R. A. Fisher (1938), the person who popularized modern sta-
tistical testing, was one of the first researchers to address the question of
how to combine the results of multiple studies. In the days since Fisher
offered his preliminary suggestions, researchers have developed a wide
array of techniques for summarizing and evaluating the results of multiple
studies (see Rosenthal Rosnow, 1991, for an excellent conceptual and
computational review of such techniques). Statistical techniques that are
designed for this purpose are typically referred to as meta-analytic tech-
niques. The more commonly used term meta-analysis thus refers to the use
of such techniques to analyze the results of studies rather than the responses


of individual participants. From this perspective, meta-analyses are to
groups of studies what traditional statistical analyses are to groups of spe-
cific participants. Literally, meta-analysis refers to the analysis of analyses.
Prior to the development of meta-analysis, the only way researchers could
summarize the results of a large group of studies was to logically analyze and
verbally summarize all the studies. Meta-analyses complement such poten-
tially imprecise analyses by providing precise mathematical summaries of
different aspects of a set of research findings. For example, a meta-analysis of
effect sizes can provide a good estimate of the average effect size that has been
observed in all the published studies on a specific topic. Other meta-analytic
techniques can be used to indicate how much variability in effect sizes has
been observed from study to study on a specific topic (see Hedges, 1987).
Finally, meta-analysis can be used to determine the kinds of studies that tend
to yield especially large or small effect sizes (e.g., studies that did or did not
make use of a particular control technique, studies conducted during a par-
ticular historical era, studies conducted in a particular part of the country).
This final kind of meta-analysis can provide very useful theoretical and meth-
odological information about the nature of a specific research finding.
As an example of this third approach, consider a couple of meta-analyses
conducted by Alice Eagly. Eagly (1978) analyzed findings from a large num-
ber of studies of the effects of gender on conformity and social influence.
Many researchers had argued that women are more easily influenced than
men are. When Eagly looked at studies published prior to 1970 (i.e., prior to
the beginning of the women’s movement), this is exactly what she found.
However, when she focused on studies published during the heyday of the
women’s movement (during the early to mid-1970s), Eagly observed very
little evidence that women were more easily influenced than men. Furthermore,
in a second meta-analysis, Eagly and Carli (1981) found that (a) the gender
of the researcher conducting the study and (b) the specific topic of influence
under investigation were good predictors of whether women were more con-
forming than were men. When studies were conducted by men or when the
topic of influence was one with which women were likely to be unfamiliar
(e.g., football), most studies showed that women were more conforming than
were men. However, when studies were conducted by women or when the
topic of influence was one with which men were likely to be unfamiliar (e.g.,
fashion), men often proved to be more conforming than were women.
Although meta-analysis may be used for many different purposes, the
biggest contribution of meta-analysis to psychological research is probably an
indirect one. The growing popularity of meta-analytic techniques has encour-
aged researchers to think about research findings in more sophisticated
ways. Specifically, instead of treating alpha as an infallible litmus test for
whether an effect is real, contemporary researchers are beginning to pay
careful attention to the question of when a given effect is most (and least)
likely to be observed. Ideally, when a meta-analysis suggests that an effect
is magnified or diluted under certain conditions, researchers should


conduct a study in which they directly manipulate these conditions. Doing
so boils down to designing factorial studies in which at least two indepen-
dent variables are completely crossed. An example would be a single
experiment on persuasion that randomly assigned half of all men and half
of all women to read about a stereotypically masculine topic before seeing
how much they conform to others’ opinions on this topic. Of course, the
other half of men and the other half of women would be randomly
assigned to read about a stereotypically feminine topic before the research-
ers assessed conformity on this topic. If the results of the experiment con-
firmed the results of the meta-analysis, researchers could be even more
confident of the conclusion suggested by the meta-analysis.

Summary

This chapter provided a brief review of statistics. We noted that statistical
procedures can be broken down into descriptive statistics and inferential sta-
tistics. As the name suggests, descriptive statistics simply describe (i.e., illus-
trate, summarize) the basic properties of a set of data. Along these lines,
measures of central tendency describe the typical or expected score in a
given data set. In contrast, measures of dispersion reveal how much the
entire set of scores varies around the typical score. The most common mea-
sures of central tendency are the mean, the median, and the mode, and the
most common measures of dispersion are the range and standard deviation.
Of course, psychologists interested in testing psychological theories are
typically interested in inferential statistics as well as descriptive statistics.
This second branch of statistics applies probability theory to determine
whether and to what degree an observed data pattern truly differs from a
chance pattern. Inferential statistics thus provide a basis for determining
whether an observed research finding reveals a systematic association that is
likely to be true in a population of interest or whether it merely reflects noise
or error. For instance, if a treatment group differs from a control group to a
degree that would not be expected by chance alone, then researchers will
view this as evidence that the treatment is actually causing changes in the
outcome. As another example, if people who tend to score high in self-
esteem also tend to score high on a measure of aggression, inferential statis-
tics can tell us whether this tendency for the scores to go hand in hand could
have happened easily by chance or whether the tendency is strong and con-
sistent enough that it probably reflects a true association between these two
variables in the general population (or at least the population that most
closely resembles the researcher’s sample). Both of these examples reveal the
logic of significance testing. More recently, statisticians have begun to com-
plement traditional statistical tests with indicators of effect size. Whereas
statistical significance tells you whether an effect is likely to exist, estimates
of effect size tell you how large an effect is likely to be.


Appendix 1.1: Some Common Statistical Tests and Their Uses

In Cervantes’s classic novel Don Quijote, there is a point at which Don
Quijote expresses tremendous self-satisfaction when he learns that he has
been speaking prose his entire life. Unlike Don Quijote, most statisticians
are more impressed with proofs than with prose. Thus, many statistics
texts offer readers flowcharts, formulae, or decision trees to help them
decide what kind of statistical analysis to perform on different kinds of
data. Because of our abiding love of prose as well as our pathological fear
of decision trees, we offer an alternative in this appendix. That is, instead
of a decision tree, we offer a series of definitions and concrete examples
that are much richer than a decision tree. If you spend a little while reading
over the list of analytic techniques covered in this appendix, we hope that
you’ll have a good sense of how to analyze most basic data sets while also
gaining a good sense of all of the major topics we cover in this textbook.
The one way in which this list does vaguely resemble a decision tree is that
it is loosely organized in increasing order of the complexity or sophistica-
tion of the research question. It thus begins with a couple of simple
descriptive statistics and progresses through a series of increasingly com-
plex inferential statistical tests. Readers who are interested in a true deci-
sion tree can find an excellent one in Chapter 2 of Tabachnick and Fidell
(2007, pp. 28–31).

Mean, median, and mode: Very often, researchers who have collected
data on a continuous (i.e., interval or ratio) scale simply want to sum-
marize what the typical score is like. When the scores are normally dis-
tributed with very little skew (and modest to high kurtosis), the mean is
an excellent indicator of the typical score. Height, SAT scores, and the
highway mileage for one’s car are all normally distributed without too
much skew or kurtosis (the Hummer and Prius notwithstanding). In
contrast, variables such as personal income, number of criminal convic-
tions, and number of depressive symptoms only approximate a normal
distribution because they are typically highly skewed. In such cases, the
median and/or mode are often better indicators of central tendency
because the median and the mode are influenced very little by extreme
outliers. When researches wish to make inferences about populations
rather than merely summarize a set of scores, they will want to report the
standard error of the mean and/or a 95% or 99% confidence interval for
the mean—to give others some idea of how far from the observed sam-
ple mean the true population mean is likely to fall. The standard error of
the mean is a function of the observed sample standard deviation (see
below) and the sample size, and it grows smaller as the sample size gets
larger. All else being equal, this means, for example, that a researcher
who randomly samples 1,000 people will be able to make a more precise
statement about the likely range of the population mean than will a


researcher who randomly samples only 100 people. We discuss measures
of central tendency in great detail in Chapter 2. In that same chapter, we
also discuss the important topic of variability (especially the standard
deviation) while also covering important topics such as skewness and
kurtosis. These last two topics set the stage for subsequent chapters on
inferential statistics and data cleaning.

Standard deviation: The standard deviation is an indicator of the vari-
ability of a set of continuous scores around the mean. Whereas central
tendency tells us what the typical score is like, the standard deviation tells
us how well that typical score describes all of the scores in the distribution—
because the standard deviation is an indicator of the average amount by
which all of the scores in a distribution vary around the mean. We dis-
cuss both traditional and nontraditional (creative) uses of the standard
deviation in Chapter 2.

Variance: The variance of a set of scores is simply the standard deviation
of that set of scores squared.

Pearson’s r: describes the strength and direction of the linear association
between two continuous (interval or ratio) variables. An r of zero means
that there is no linear association whatsoever between two variables.
Absolute values of r closer to 1.0 indicate a stronger association. If r is
negative, it means that as scores on one variable (X) increase, scores on the
other variable (Y) decrease. If r is positive, it means that as scores on one
variable (X) increase, scores on the other variable (Y) also increase. The
concept of correlation is closely linked to prediction. Thus, for example, if
height and weight are correlated r = .70, one can minimize errors of pre-
diction by predicting that a person who is exactly one standard deviation
above the mean in height is 0.70 standard deviation units above the mean
in weight. We discuss the correlation coefficient in Chapter 3, including a
brief discussion of how to assess curvilinear as well as linear associations
between continuous variables.

Phi coefficient (j): The phi coefficient is very similar to r except it has no
sign because it is used to describe the strength of association between two
nominal or categorical variables (variables that do not indicate quantity or
amount). It thus ranges between zero and 1.0. We discuss both the phi
coefficient and the chi-square test of association in Chapter 4.

Chi-square test of association: A chi-square (χ2) test of association is
conceptually identical to a phi coefficient (j) because, like phi, this test of
association indicates whether two categorical variables are related. In fact,
it is very easy to convert a chi-square test of association to a phi coefficient
using the simple formula ϕ = (χ2 / N ).
obt


Single-sample t test: This very simple test is designed to test see whether
the mean score for a continuous, normally distributed variable (e.g., IQ
score, height) in a specific sample differs from some known or hypothe-
sized (e.g., theoretically meaningful) population value. For example, the
SAT scores for the students at a particular high school might be compared
with the known U.S. population mean for SAT scores to see if the students
at that high school tend to be more academically prepared than the aver-
age American high school student. We discuss this test in the beginning of
Chapter 6.

Independent samples t test: This common test is used to assess the reli-
ability (statistical significance) of mean differences on a continuous vari-
able between two independent groups or categories of people. Some
examples of the use of this test are (a) drawing inferences about the results
of a lab experiment that has one experimental and one control group,
(b) assessing gender differences on a continuously measured emotional
performance test, and (c) comparing people with versus without a disease
on a suspected health consequence of the disease. The main assumption of
the test is that the dependent measure is normally distributed (although
the test is pretty robust to many, but not all, violations of this assumption).
We discuss this test in some detail in the latter portion of Chapter 6.

One-way analysis of variance (ANOVA): This test is used to assess the
reliability (statistical significance) of mean differences between three or
more groups. This test is very similar to an independent samples t test (and
shares the same assumption of a normally distributed dependent mea-
sure). However, the difference is that a one-way ANOVA controls for the
experiment-wise error rate that occurs as researchers consider all of the
many possible comparisons that can be made between specific groups
when there are multiple experimental or naturally existing groups (three
or more levels of the independent variable). Some examples of the use of
this statistic are (a) outcomes in a lab experiment that has three different
conditions (e.g., three dosage levels of a drug), (b) comparing kids in four
different grades on an intellectual outcome, and (c) comparing the atti-
tudes of soldiers from five different military ranks. When the researcher
has a clear a priori reason to expect the various conditions to yield results
that follow a specific pattern, the researcher can greatly increase statistical
power by conducting a planned comparison based on this specific expec-
tation. We discuss one-way ANOVA, including planned comparisons, in
the first half of Chapter 7.

Factorial ANOVA: This technique is used to assess joint effects of two or
more fully crossed categorical independent variables on a continuously
scored outcome. It allows for the statistical separation of main effects of all
independent variables and, if desired, interactions between two or more


independent variables. This technique is used frequently to draw conclu-
sions about the results of laboratory experiments. It also assumes a nor-
mally distributed dependent measure. We discuss factorial ANOVA,
including follow-up comparisons such as simple effects tests, in the second
half of Chapter 7.

Analysis of covariance (ANCOVA): Sometimes researchers may wish to
control for a confound or nuisance variable in an ANOVA and—rather
than reporting the raw, between-groups means—hold the different, natu-
rally occurring or experimental groups constant on that nuisance variable.
This is both conceptually and mathematically identical to a simultaneous
multiple regression analysis with at least one categorical variable (the
independent variable or variables) and at least one continuous variable
(the covariates in an ANCOVA). The main advantage of ANCOVA over a
regression analysis is the fact that ANCOVA readily yields covariate-
adjusted means, which look very much like regular means and thus are
very easy to interpret.
One of the crucial assumptions of ANCOVA is homogeneity of covari-
ance, meaning that the covariate for which the analysis makes a statistical
adjustment should have roughly the same association with the dependent
measure in all of the various experimental conditions. For example, a
researcher studying gender differences in aggressive behavior in the lab
might wish to control statistically for the fact that people who more
strongly believe in the concept of defending one’s honor (reported, say, on
a 9-point scale) behave more aggressively than people who do not believe
in the concept of defending one’s honor (e.g., see D. Cohen Nisbett,
1994). So long as the association between beliefs about honor and labora-
tory aggressive behavior was about equally strong for women and men, it
would be appropriate to reduce the noise associated with this belief to see
if a significant gender difference remained after controlling for the belief.
The test for a gender difference could thus be more powerful than it would
have been otherwise after controlling for any effects of this nuisance vari-
able that is more or less independent of gender. In Chapter 11, we compare
and contrast ANCOVA with ANOVA and with multiple regression analy-
sis.We emphasize that whereas ANCOVA is mathematically identical to a
simultaneous multiple regression analysis, the two techniques yield very
different kinds of outputs. For example, it is often much easier for people
to understand covariate-adjusted means (because they look just like tradi-
tional means) than to understand standardized regression coefficients.

Reliability analysis: A reliability analysis is used to determine the degree
to which the multiple items in a scale all behave in the same fashion (i.e.,
are positively correlated with one another). Cronbach’s alpha (a) is a very
useful and easy-to-calculate statistic. However, a very high alpha statistic
for a scale does not always guarantee that the items in the scale form a


single factor. On the other hand, low corrected item-total rs for specific
items in a scale are a useful indicator that the specific items are not cor-
related with the other scale items (and thus should probably be excluded
from the scale). Treating individual raters of a specific behavior or judg-
ment as if they were specific items in a scale can allow an assessment of the
reliability of individual raters. A rater whose item-total correlation is low
is in disagreement with the average of all of the other raters and can either
be retrained, if possible, or dropped. We discuss reliability analysis (along
with principal components analysis and factor analysis) in Chapter 5.

Multiple regression analysis: Multiple regression analysis is used to assess
the strength and direction of the unique linear association between mul-
tiple, continuous predictors of a continuous outcome (a criterion) and
that outcome. Because each predictor is controlled statistically for the
association between that predictor and all other predictors, an assessment
of the relative strength of each predictor is possible. Each predictor is thus
assessed controlling for the natural confounds between that predictor and
all other predictors. The primary statistical indicator is a B or b (an
unstandardized regression coefficient) or beta (b), a standardized regres-
sion coefficient that is conceptually very similar to r. Each coefficient has
its own associated p value that indicates the reliability (statistical signifi-
cance) of that unique association. Although both univariate (single-vari-
able) normality and multivariate normality of the continuously measured
predictors are assumed, this analysis is usually quite robust to the inclu-
sion of one or more categorical variables, especially when these variables
are not highly skewed. Gender, for example, is often dummy-coded with-
out any problems in a multiple regression analysis that also includes sev-
eral continuous, normally distributed predictors. We discuss the basics of
multiple regression analysis, with a conceptual emphasis on how multiple
regression identifies the unique association between different predictor
variables and a criterion variable, in the first part of Chapter 9.
In most cases of multiple regression analysis, researchers expect the
zero-order association between a predictor variable and the criterion of
interest to grow smaller (and sometimes even disappear altogether) when
all the other predictor variables are statistically held constant. However, a
multiple regression analysis can also reveal that a predictor variable that
appeared to be unrelated (at the simple or zero-order level) to a criterion
variable is actually associated with the criterion variable once statistical
adjustments are made for the effects of one or more additional predictors.
This unusual situation is referred to as suppression, and it is discussed in
great detail in Chapter 12.

Moderator analysis (multiple regression analysis of statistical interac-
tions): If a researcher also wishes to know whether the association between
a predictor and the criterion differs at different levels of some other predictor,


it is possible to conduct a moderator analysis by examining the cross-
product(s) of the two or more predictors of interest and following up a
significant interaction with simple slopes tests (analogous to simple main
effect tests in ANOVA). A moderator analysis in multiple regression could
thus be conducted to see if the association between the number of times
people moved as children and their physical health as adults is stronger for
introverts than for extraverts. (Introverts who moved a lot as children
often have poorer than average adult health, but extraverts seem to show
no such association.) A moderator variable can also be categorical while
the other predictor and the criterion variable are both continuous. For
example, the association between implicit self-esteem and explicit self-
esteem (both continuous variables) seems to be stronger (more positive)
for women than for men (Pelham, Koole, et al., 2005). We discuss mod-
erator analyses in multiple regression, including simple slopes tests to
elucidate the exact nature of a significant interaction, in Chapter 10.

Logistic regression: Logistic regression analysis is conceptually identical
to a standard multiple regression analysis except that the criterion variable
(and sometimes one or more of the predictors) is categorical rather than
continuous. The primary output statistic is a predictor-adjusted odds ratio
that is the rough conceptual equivalent of a B or a b. Unlike a simple odds
ratio, however, the odds ratio from a logistic regression analysis refers to
the association between one categorical variable and another while hold-
ing all other predictor variables in the model constant. We discuss the
basics of logistic regression in the last section of Chapter 9.

Principal components analysis and factor analysis: These two closely
related techniques are designed to uncover underlying dimensions along
which a set of many separate responses vary. These numerous individual
responses might be answers to individual personality questions, specific
political or social attitudes, or self-reported liking for many different kinds
of foods or many different specific types of music. For example, contem-
porary research in human personality suggests that hundreds of individual
personality questions all boil down to five core personality dimensions:
openness to experience, conscientiousness, extraversion, agreeableness,
and neuroticism (see Goldberg, 1990). A factor analysis of dozens of spe-
cific personality traits might thus reveal, for example, that trait terms such
as energetic, friendly, outgoing, outspoken, and loud would all load heavily
on the basic dimension of extraversion. In contrast, specific trait terms
such as reliable, punctual, obedient, organized, and honest might all load
heavily on the core dimension of conscientiousness. One key difference
between principal components analysis and factor analysis is that princi-
pal components analysis is usually a bottom-up, purely empirical way to
distill a large set of observations into a smaller number of dimensions. In
contrast, factor analysis is more likely to be used when the researcher has


a clear a priori theory about how the different items ought to load together
and how many factors there ought to be in a data set. We discuss principal
components analysis and factor analysis in Chapter 5, including a dis
cussion of how these methods are related to the concept of reliability
(specifically, the internal consistency of a multiple-item scale).

Paired-samples t test: When two measures come from the same organism
(or similar organisms), the two different measures are likely to be highly
correlated with one another. For example, a specific child who excels in
reading is also likely to excel in spelling. A similar lack of independence in
specific observations often occurs when genetically related or experimen-
tally yoked members of a pair are tested on the same outcome. When
exactly two such repeated measures are obtained, a paired-samples t test
can be used to assess the statistical significance of any observed behavioral
or performance difference between the two measures. This test is also very
useful when the same person fills out the same measure or task under dif-
ferent (manipulated) experimental conditions. For example, a child might
be given the same intellectual measure by two different experimenters, one
of whom expresses a positive expectancy about her performance and one
of whom expresses no expectations. If appropriate experimental controls
are used (e.g., counterbalancing the order of the two expectancies across
participants), a paired-samples t test can reveal whether performance varies
reliably with experimental condition, with a very high level of statistical
power. Like the independent samples t test, this test assumes that the
dependent measure (which in this case is the difference between two scores)
is normally distributed. Importantly, the increased power that usually
comes with this test comes precisely to the degree that the two measures of
interest are strongly correlated with one another. It is this strong correlation
between two related measures that effectively reduces the variance that
serves as the error term for this analysis. We discuss the paired-samples
t test in Chapter 8, with a special emphasis on how the correlation between
two repeated measurements plays a crucial role in the power of a repeated
measures t test to reveal differences between paired means.

Repeated measures ANOVA: If three or more within-subjects conditions
(or measurement periods) rather than two are collected from the same (or
related) participants, a repeated measures ANOVA is the appropriate statisti-
cal test for differences between the means. Just as a one-way ANOVA replaces
an independent samples t test once you graduate from two to three or more
groups, the one-way repeated measures ANOVA replaces the paired-samples
t test once you graduate from two to three or more within-subjects condi-
tions. For example, if children are exposed to three different expectancies
rather than two, a repeated measures ANOVA could be conducted on the
mean performance scores in the three within-subjects conditions. Repeated
measures studies can also involve complete factorial designs. For example, a


completely within-subjects experiment might separately study reactions to
sexist and nonsexist jokes that also vary independently in how funny the
jokes are pretested to be. In its simplest form, this study would be a 2 (Level
of Sexism: High vs. Low) × 2 (Level of Humor: High vs. Low) completely
within-subjects study, analyzed using a within-subjects ANOVA. We discuss
repeated measures ANOVAs in Chapter 8.

Mixed model ANOVA: If a study includes at least one within-subjects
variable and at least one between-subjects variable (whether measured or
manipulated), a mixed model ANOVA can test simultaneously for both
between-subjects and within-subjects effects. Mixed model ANOVAs can
also test for statistical interactions between one or more between-subjects
variables and one or more within-subjects variables. For example, a cogni-
tive psychologist might manipulate cognitive load on a between-subjects
basis while assessing both implicit and explicit memory for studied mate-
rial. She might predict, for example, that the cognitive load manipulation
(e.g., rehearsing a 7-digit number) will have a large effect on explicit
memory (recall memory) while having little or no effect on implicit
memory (e.g., based on performance on a word fragment completion
task). Thus, the researcher would expect to observe a Load × Memory–
type interaction in this mixed model design. We discuss mixed model
ANOVAs in Chapter 8.

Mediation analysis: In both laboratory experiments and passive observa-
tional studies, researchers often wish to know why one variable is related to
another. For example, research on frustration and aggression suggests that
one of the main reasons why frustration often leads to aggression is because
frustration leads to anger, which then leads to aggression. In the language of
mediation, this is to say that anger mediates the simple association between
frustration and aggression. Mediation analyses are merely variations on a
multiple regression analysis with an emphasis on assessing the degree to
which the association between the original independent variable and the
dependent variable disappears or gets weaker once you statistically control
for the significant effects of the mediator on the dependent measure.
Prototypically, if anger fully mediates the association between frustration
and aggression, (a) frustration should predict aggression, (b) frustration
should predict anger, (c) anger should predict aggression, and finally (d) the
association between anger and aggression should completely disappear once
you statistically control for the effects of anger on aggression (because frus-
tration affects aggression indirectly through the route of increased anger).
We discuss both mediation analysis and path analysis in Chapter 13.

Path analysis: Path analysis is a special version of multiple regression
analysis that is designed to assess the plausibility of a proposed causal chain
leading from one or more source variables to an ultimate (“downstream”)


outcome variable. In fact, the simplest possible kind of path analysis is a
three-variable mediation model with the mediator representing the causal
step between just one source variable and just one outcome variable. In
most cases, however, path analyses involve four or more variables, ideally
measured at different carefully selected time points (e.g., in a longitudinal
or prospective design). Moreover, researchers do not always expect every
variable in the middle of a causal chain to mediate the associations between
the source variables and the ultimate outcome variable. Instead, some of
the source variables might be expected to have a direct (nonmediated) as
well as an indirect (mediated) connection to the downstream variable. Path
analysis is the historical and conceptual precursor of modern structural
equation modeling, which can be thought of as a hybrid combination of
path analysis and factor analysis. In fact, some researchers refer to structural
equation modeling as confirmatory (aka “theory-driven”) factor analysis.
A detailed discussion of structural equation modeling is beyond the scope
of this intermediate text.

Notes

1. For more details, see http://www.news.com.au/business/story/0,27753,255
15799-462,00.html.
2. See the Time magazine story at http://www.time.com/time/magazine/arti
cle/0,9171,1889153,00.html.
3. We adapted this example of men of varying heights from an illuminating
statistics lecture by Daniel Gilbert, who probably adapted it from a lecture by
Plato.
4. http://imgs.sfgate.com/cgi-bin/article.cgi?f=/c/a/1998/12/28/MN9307
.DTLtype=printable
5. Computing the probability of an event as extreme as or more extreme than
an observed event (or set of events) is standard practice for most statistical tests.
At first blush, paying attention to events even more extreme than an observed
event may seem a little odd. However, if we care about events as unusual as or more
unusual than our observed event—which we almost always do—it makes a lot of
sense. If you think of the unusualness of a set of observations (e.g., a lot of heads
tossed, a pair of means that are noticeably different) as a standard of experimental
performance that a researcher hopes to meet or exceed, this may help make sense
of this practice. If we set a high-jump bar at exactly 6 feet and Amanda clears it,
the set of outcomes that Amanda, the judges, and the fans all care about is jumps
of exactly 6 feet or higher. Furthermore, if we tried to calculate the probability of
a specific observation or event, probabilities would almost always be pretty low—
because the probability of any specific event is always quite low. For example, the
probability of tossing a fair coin 20 times and observing exactly 10 heads is .176,
even though this is the most likely of all the possible outcomes. Once we move to
continuous rather than discrete events, this is even truer. The probability that a
particular high jump would be exactly 6 feet—even for a very good jumper who
was trying to jump exactly 6 feet—is extremely low.


6. Speaking of cheating, we cheated. Unless we increased our sample size to
about 250 people, we couldn’t actually conduct this second χ2 analysis. That’s
because we’re allowed to use the χ2 statistic only in situations in which all our
expected frequencies have a value of at least 5.0. With values lower than 5, the χ2
values that are generated can be pretty unstable and pretty inaccurate. In an
extreme case such as this one, however, it’s safe to say that people were signifi-
cantly honest. If nothing else, we could always choose to make a very conservative
comparison and set 90% (instead of 98%) dishonesty as our standard of com-
parison. This would yield 5 rather than 1 as the expected number of nonwinners.
In case you want to practice your calculations, the value you should get if you do
the analysis this more conservative (but legal) way is χ2 (1, N = 50) = 272.22. The
1 in the parentheses indicates the degrees of freedom you’d report in an actual
research report in which you conducted this analysis. We come back to this in the
section on reporting commonly used statistics.

StatisticsA Review of Basic Statistical Concepts

StatisticsA Review of Basic Statistical Concepts

More Related Content

Similar to StatisticsA Review of Basic Statistical Concepts (20)

Recently uploaded (20)

StatisticsA Review of Basic Statistical Concepts