CLIFFORD KONOLD AND ALEXANDER POLLATSEK
most basic questions in analyzing data involve looking at group differences to
determine whether some factor has produced a difference in the two groups.
Typically, the most straightforward and compelling way to answer these questions is
to compare averages. We believe that much of statistical reasoning will elude
students until they understand when a comparison of two averages makes sense and,
as a corollary, when such a comparison is misleading. If they do not understand this,
students’ explorations of data (i.e., “data snooping”) will almost certainly lack
direction and meaning.
SIGNALS IN NOISY PROCESSES
A statistician sees group features such as the mean and median as indicators of
stable properties of a variable system—properties that become evident only in the
aggregate. This stability can be thought of as the certainty in situations involving
uncertainty, the signal in noisy processes, or, the descriptor we prefer, central
tendency. Claiming that modern-day statisticians seldom use the term central
tendency, Moore (1990, p. 107) suggests that we abandon the phrase and speak
instead of measures of “center” or “location.” But we use the phrase here to
emphasize conceptual aspects of averages that we fear are often lost, especially to
students, when we talk about averages as if they were simply locations in
distributions.
By central tendency we refer to a stable value that (a) represents the signal in a
variable process and (b) is better approximated as the number of observations
grows.3 The obvious examples of statistics used as indicators of central tendency are
averages such as the mean and median. Processes with central tendencies have two
components: (a) a stable component, which is summarized by the mean, for
example; and (b) a variable component, such as the deviations of individual scores
around an average, which is often summarized by the standard deviation.
It is important to emphasize that measures of center are not the only way to
characterize stable components of noisy processes. Both the shape of a frequency
distribution and global measures of variability, for example, also stabilize as we
collect more data; they, too, give us information about the process. We might refer
to this more general class of characteristics as signatures of a process. We should
point out, however, that all the characteristics that we might look at, including the
shape and variability of a distribution, are close kin to averages. That is, when we
look at the shape of a particular distribution, we do not ordinarily want to know
precisely how the frequency of values changes over the range of the variable.
Rather, we tame the distribution’s “bumpiness.” We might do this informally by
visualizing a smoother underlying curve or formally by computing a best-fit curve.
In either case, we attempt to see what remains when we smooth out the variability.
In a similar manner, when we employ measures such as the standard deviation or
interquartile range, we strive to characterize the average spread of the data in the
sample.
CONCEPTUALIZING AN AVERAGE
Implicit in our description of central tendency is the idea that even as one speaks
of some stable component, one acknowledges the fundamental variability inherent in
that process and thus its probabilistic nature. Because of this, we claim that the
notion of an average understood as a central tendency is inseparable from the notion
of spread. That average and variability are inseparable concepts is clear from the
fact that most people would consider talking about the average of a set of identical
values to be odd. In addition, it is hard to think about why a particular measure of
center makes sense without thinking about its relation to the values in the
distribution (e.g., the mean as the balance point around which the sum of the
deviation scores is zero, or the median as the point where the number of values
above equals the number of values below).
Not all averages are central tendencies as we have defined them above. We
could compute the mean weight of an adult lion, a Mazda car, and a peanut, but no
clear process would be measured here that we could regard as having a central
tendency. One might think that the mean weight of all the lions in a particular zoo
would be a central tendency. But without knowing more about how the lions got
there or their ages, it is questionable whether this mean would necessarily tell us
anything about a process with a central tendency. Quetelet described this distinction
in terms of true means of distributions that follow the law of errors versus arithmetic
means that can be calculated for any assortment of values, such as our hodgepodge
above (see Porter, 1986, p. 107).
Populations versus Processes
In the preceding description, we spoke of processes rather than populations. We
contrast these two ways of thinking about samples or batches of data, as shown in
Figure 1. When we think of a sample as a subset of a population (see the left
graphic), we see the sample as a piece allowing us to guess at the whole: The
average and shape of the sample allow us perhaps to estimate the average and shape
of the population. If we wanted to estimate the percentage of the U.S. population
favoring gun control, we would imagine there being a population percentage of
some unknown value, and our goal would be to estimate that percentage from a
well-chosen sample. Thinking in these terms, we tend to view the population as
static and to push to the background questions about why the population might be
the way it is or how it might be changing.
From the process perspective (as depicted in the right graphic of Figure 1), we
think of a population or a sample as resulting from an ongoing, dynamic process, a
process in which the value of each observation is determined by a large number of
causes, some of which we may know and others of which we may not. This view
moves to the foreground questions about why a process operates as it does and what
factors may affect it. In our gun control example, we might imagine people’s
opinions on the issue as being in a state of flux, subject to numerous and complex
influences. We sample from that process to gauge the net effect of those influences
at a point in time, or perhaps to determine whether that process may have changed
over some time period.
CLIFFORD KONOLD AND ALEXANDER POLLATSEK
For many of the reasons discussed by Frick (1998), we have come to prefer
thinking of samples (and populations, when they exist) as outputs of processes.4 One
reason for this preference is that a process view better covers the range of statistical
situations in which we are interested, many of which have no real population (e.g.,
weighing an object repeatedly). Another reason for preferring the process view is
that when we begin thinking, for example, about how to draw samples, or why two
samples might differ, we typically focus on factors that play a role in producing the
data. That is, we think about the causal processes underlying the phenomena we are
studying. Biehler (1994) offered a similar analysis of the advantages of viewing data
as being produced by a probabilistic mechanism—a mechanism that could be altered
to produce predictable changes in the resultant distribution. Finally, viewing data as
output from a process highlights the reason that we are willing to view a collection
of individual values as in some sense “the same” and thus to reason about them as a
unity: We consider them as having been generated by the same process.
Figure 1. Data viewed as a sample of a population (left) versus data viewed as output of a
noisy process (right).
This notion of process is, of course, inherent in the statistician’s conception of a
population, and we expect that most experts move between the process and
population perspectives with little difficulty or awareness.5 However, for students
new to the study of statistics, the choice of perspective could be critical. To illustrate
more fully what we mean by reasoning about processes and their central tendencies,
we discuss recent results of the National Assessment of Educational Progress
(NAEP).