Thinking and Deciding

THE PSYCHOLOGY OF HYPOTHESIS TESTING 177
the necessity or the sufficiency of some condition, in order to get a “point” toward
their total score. For example, a rule about necessity was “I get a point only if the
light goes through the shaded square.” Here, if the rule were true, going through the
shaded square would be necessary in order to get a point. A rule about sufficiency
was, “If the light goes through the shaded square, I get a point.” Here, going through
the square is sufficient, but there may be some other way to get a point. Subjects
in the rule condition did well at testing both kinds of rules. Subjects in the reward
condition, however, did well at testing the rule about sufficiency but poorly at testing
the rule about necessity. Statements about necessity require tests with sequences that
do not move the light through the shaded square (to see whether there is any other
way of getting the reward). Subjects in the rule condition made far fewer such tests
than those in the reward condition, as though they were still interested in getting the
reward.
These experiments probably are analogous to real situations — for example, in
classrooms — in which people are induced to work for extrinsic rewards such as
grades. The reward may be effective in encouraging the work in question, but it may
reduce the commitment to other valuable goals — such as satisfying one’s curiosity
— that could otherwise motivate the same behavior. Once students find a way of
getting the reward of grades, they may be less inclined to try other ways of thinking
that might teach them something about the real world that would ultimately enable
them to obtain rewards much more important to them than good grades.
Information bias and the value of information
The value of relevant information can be calculated in advance, before we know what
the answer to a question will be. A normative model of hypothesis testing can then
be specified in these terms. The best test of the hypothesis is the one that yields the
most relevant information.
In essence, what we want to calculate is the extent to which specific information
that we have or can get will help us to decide correctly among actions, such as
which disease should be treated or which scientific hypothesis should be accepted as
a basis for further experimentation. Let me suggest how this might be done, in the
context of some examples from medical diagnosis. Medical diagnosis is useful for
this purpose because probability is obviously involved. Only rarely can a physician
diagnose a disease with absolute certainty. Scientific reasoning, I have argued, is
also probabilistic, but it is more difficult to analyze many scientific situations in this
way, because it is more difficult to list the possible hypotheses.
Consider the following diagnostic problems (Baron, Beattie, and Hershey, 1988)
for a group of fictional diseases:
1. A patient’s presenting symptoms and history suggest a diagnosis of
globoma, with about .8 probability. If it isn’t globoma, it’s either
popitis or flapemia. Each disease has its own treatment, which is
ineffective against the other two diseases. A test called the ET scan