Sense and Nonsense About Surveys
Sense and Nonsense About Surveys
Sense and Nonsense About Surveys
http://ctx.sagepub.com/
Published by:
http://www.sagepublications.com
On behalf of:
time, George Gallup, using many fewer cases but a much better method, made the more accurate prediction that FDR
would win. Gallup used quotas in choosing respondents in
order to represent different economic strata, whereas the
Literary Digest had worked mainly from telephone and automobile ownership lists, which in 1936 were biased toward
wealthy people apt to be opposed to Roosevelt. (There were
other sources of bias as well.) As a result, the Literary Digest
poll disappeared from the scene, and Gallup was on his way
to becoming a household name.
Yet despite their intuitive grasp of the importance of representing the electorate accurately, Gallup and other commercial
pollsters did not use the probability sampling methods that
were being developed in the same decades and that are fundamental to social science surveys today. Probability sampling
in its simplest form calls for each person in the population to
have an equal chance of being selected. It can also be used in
more complex applications where the chances are deliberately made to be unequal, for example, when oversampling a
minority group in order to study it more closely; however, the
chances of being selected must still be known so that they can
later be equalized when considering the entire population.
40
Probability sampling theory reveals a crucial but counterintuitive point about sample size: the size of a sample needed
to accurately estimate a value for a population depends very
little on the size of the population. For example, almost the
same size sample is needed to estimate, with a given degree
of precision, the proportion of left-handed people in the
United States as is needed to make the same estimate for, say,
Peoria, Illinois. In both cases a reasonably accurate estimate
can be obtained with a sample size of around 1,000. (More
cases are needed when extraordinary precision is called for, for
example, in calculating unemployment rates, where even a
tenth of a percent change may be regarded as important.)
The link between population size and sample size cuts
both ways. Although huge samples are not needed for huge
populations like those of the United States or China, a handful of cases is not sufficient simply because ones interest is limited to Peoria. This implication is often missed by those trying
to save time and money when sampling a small community.
who is missing?
A good sample depends on more than probability sampling theory. Surveys vary greatly in their quality of implementation, and this variation is not captured by the margin of
error plus/minus percentage figures that accompany most
media reports of polls. Such percentages reflect the size of the
final sample, but they do not reveal the sampling method or
the extent to which the targeted individuals or households
were actually included in the final sample. These details are at
least as important as the sample size.
When targeted members of a population are not interviewed or do not respond to particular questions, the omissions
are a serious problem if they are numerous and if those missed
differ from those who are interviewed on the matters being
studied. The latter difference can seldom be known with great
confidence, so it is usually desirable to keep omissions to a minimum. For example, sampling from telephone directories is
undesirable because it leaves out those with unlisted telephones, as well as those with no telephones at all. Many survey
reports are based on such poor sampling procedures that they
may not deserve to be taken seriously. This is especially true of
reports based on focus groups, which offer lots of human
interest but are subject to vast amounts of error. Internet surveys
also cannot represent the general population adequately at
present, though this is an area where some serious attempts are
being made to compensate for the inherent difficulties.
The percentage of people who refuse to take part in a survey is particularly important. In some federal surveys, the percentage is small, within the range of 5 to 10 percent. For even
the best non-government surveys, the refusal rate can reach
25 percent or more, and it can be far larger in the case of poorly executed surveys. Refusals have risen substantially from earlier days, becoming a major cause for concern among serious
survey practitioners. Fortunately, in recent years research has
shown that moderate amounts of nonresponse in an otherwise careful survey seem in most cases not to have a major
effect on results. Indeed, even the Literary Digest, with its
abysmal sampling and massive nonresponse rate, did well predicting elections before the dramatic realignment of the electorate in 1936. The problem is that one can never be certain
as to the effects of refusals and other forms of nonresponse,
so obtaining a high response rate remains an important goal.
41
42
reference provided by the question. The form of a question provides the rules of the game for respondents, and this must
forbid them, a difference of 21 percentage points. This finding was replicated several times in later years, not only in the
United States but also (with appropriate translations) in
Germany and the Netherlands. Such survey-based experiments call for administering different versions of a question
to random subsamples of a larger sample. If the results
between the subsamples differ by more than can be easily
explained by chance, we infer that the difference is due to the
variation in wording.
In addition, answers to survey questions always depend on
the form in which a question is asked. If the interviewer presents a limited set of alternatives, most respondents will choose
one, rather than offering a different alternative of their own. In
one survey-based experiment, for example, we asked a national sample of Americans to name the most important problem
facing the country. Then we asked a comparable sample a parallel question that provided a list of four problems from which
to choose the most important; this list included none of the
four problems mentioned most often by the first sample but
instead provided four problems that had been mentioned by
fewer than 3 percent of the earlier respondents. The list question also invited respondents to substitute a different problem
if they wished (see Table 1). Despite the invitation, the majority of respondents (60 percent) chose one of the rare problems
offered, reflecting their reluctance to go outside the frame of
table 1
Experimental Variation Between Open and Closed Questions
A. Open Question
What do you think is the most important
problem facing this country today [1986]?
B. Closed Question
Which of the following do you think is the
most important problem facing this country
today [1986] the energy shortage, the quality of public schools, legalized abortion, or pollution or, if you prefer, you may name a different problem as most important.
1. Energy shortage.
2. Quality of public schools.
3. Legalized abortion.
4. Pollution.
Adapted from: H. Schuman and J. Scott, Problems in the Use of Survey Questions to Measure Public Opinion, Science v. 236,
pp. 957-959, May 22, 1987.
In a survey experiment, less than 3% of the 171 respondents asked the question on the left volunteered one of the four problems listed on the right. Yet, 60% of the 178 respondents asked the question on the right picked one of those four answers.
43
Interview on the digital divide. In a departure from conventional protocol, interviewer (on right) is sitting next to rather
than across from interviewee.
figure 1
Attitudes Toward Free Speech Against Democracy
100
90
80
75
70
60
54
50
46
40
% Forbid speeches
30
23
20
10
0
1940
1975
Year
44
Survey researchers should also ask several different questions about any important issue. In addition to combining
questions to increase reliability, the different answers can be
synthesized rather than depending on the angle of vision provided by any single question. A further safeguard is to carry
out frequent experiments like that on the forbid/allow wordings. By varying the form, wording, and context of questions,
researchers can gain insight into both the questions and the
relevant issues. Sometimes variations turn out to make no difference, and that is also useful to learn. For example, I once
expected support for legalized abortion to increase when a
question substituted end pregnancy for the word abortion in
the phrasing. Yet no difference was found. Today, more and
more researchers include survey-based experiments as part of
their investigations, and readers should look for these sorts of
safeguards when evaluating survey results.
Section of interview form used in the Surveys of Consumers conducted by the Survey
Research Center, University of Michigan.
45
Page from completed, self-administered questionnaire used to study high school students views of grading.
46
recommended resources
table 2
Percent of White Americans Approving or Disapproving
of Racial Intermarriage, 1958-1997
Year
Approve
Disapprove
Groves, Robert M. Survey Errors and Survey Costs. New York: Wiley,
1989. A sophisticated consideration of the sources of error in surveys.
Kalton, Graham. Introduction to Survey Sampling. Thousand Oaks,
Calif.: Sage Publications (Quantitative Applications in the Social
Sciences), 1983. A brief and lucid introduction to sampling.
1958
1978
1997
4
34
67
96
66
33
Page, Benjamin I., and Robert Y. Shapiro. The Rational Public: Fifty Years
of Trends in Americans Policy Preferences. Chicago: University of
Chicago Press, 1992. In part, a persuasive reply to Converses skepticism.
Schuman, Howard, and Stanley Presser. Questions and Answers in
Attitude Surveys: Experiments on Question Form, Wording, and
Context. San Diego, Calif.: Academic Press, 1981 (Reprint edition with
sampling and experiments in questioning. In the end, however, with surveys as with all research methods, there is no substitute for both care and intelligence in the way evidence is
gathered and interpreted. What we learn about society is
always mediated by the instruments we use, including our
own eyes and ears. As Isaac Newton wrote long ago, error is
not in the art but in the artificers. n
47