Thinking and Deciding

THE MECHANISM OF JUDGMENT 377
pendent. Within this pool, GRE score and research experience are unrelated. Now
suppose that the GRE score is missing for one applicant, and we make a judgment
of acceptability on the basis of research experience alone. Since the applicant’s research
experience is near the top of our scale, we give the applicant a very high
rating, compared to others in the pool. Then the applicant takes the Graduate Record
Examination, and the scores turn out to be a little better than the average of those in
the pool. Should we raise our rating for this person or lower it?
Most people would lower it (Lichtenstein, Earle, and Slovic, 1975), but they
should raise it. The lowering of the rating is just another example of subjects behaving
roughly as though they were averaging the two cues; the second cue lowers the
subjective average of the two. (Any account that explains departures from the adding
model can explain this.)
Why should the rating be raised? Before we find out the applicant’s GRE, our
best guess is that it is at the average of the group. This is what we would guess if
we knew nothing at all, and since the applicant’s research experience is unrelated to
GRE, we should still guess that the GRE is average after we knew about research
experience. Therefore, when we find out that the student’s GRE is a little better
than average, we ought to think that this student is even better than we would have
predicted when we did not know her GRE. (This argument depends on the cues being
independent. If the cues are highly correlated, a moderate level of one should make
us suspect the validity of a very high value on another.)
Representativeness in numerical prediction
The representativeness heuristic (discussed in Chapter 6) seems to cause biases in
quantitative judgment just as it causes biases in probability judgment. In the example
I just gave, we expect the GRE to be above average because that kind of GRE score
is more similar to (more representative of) the applicant’s research experience than
an average GRE score would be. We seem to have made a prediction based on
similarity rather than on the normative model, which holds that we ought to ignore
research experience — which is useless information because it is unrelated to GRE
score — and to guess the average GRE score.
As another example, suppose you are told that a certain student ranked in the
ninetieth percentile in terms of grade point average (GPA) in his first year of college.
What is your best guess of the student’s GPA? Suppose you say 3.5 (between
B and A), because you think that this is the ninetieth percentile for GPA. (So far, you
have done nothing unreasonable.) Now what GPA would you predict for another
student, who scored in the ninetieth percentile on a test of mental concentration? or
for another student who scored in the ninetieth percentile on a test of sense of humor?
Many subjects give the same prediction of 3.5 (Kahneman and Tversky, 1973,
pp. 245–247). This may be because a GPA of 3.5 is most similar to the ninetieth
percentile on the other measures.
In fact — as I shall explain shortly — your prediction should regress toward
the mean (the average grade for the class, which was judged to be about 2.5 by