2.3. An alternative hypothesis: The world we care about consists of mostly monotone interactions
К оглавлению1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1617 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84
85 86 87 88 89 90 91 92 93 94
Reid Hastie and Robyn Dawes have offered a different account of the
success of linear models (2001, 58–62; see also Dawes 1988, 212–15). Their
explanation comes in three parts. Since we embrace and elaborate on the
third part of their explanation in section 3, we will focus only on the first
two parts of their explanation here. The first part of their explanation for
the success of SPRs is a principle about the relationship between proper
linear models and the world: Proper linear models can accurately represent
monotone (or ‘‘ordinal’’) interactions. We have already introduced linear
models—they are models in which the judgment made is a function of the
sum of a certain number of weighted variables. The best way to understand
what monotone interactions are is to consider a simple example.
Suppose a doctor has told you to reduce your body fat, and she recommends
a special diet D and an exercise regime E. Now, let’s suppose that D
alone, without the exercise regime, is effective at reducing body fat. This
would be the diet’s main effect. Suppose also that the exercise regime
alone, without the diet, is also effective at reducing body fat. Again, this
would be the main effect of exercise. Now let’s suppose Sam goes on the
diet D and the exercise regime E. If Sam gets the benefits of both—the
main effect of D and the main effect of E—then the interaction of D and E
is monotone. If, however, Sam gets the main effects of both plus an extra
benefit, then the interaction is not monotone. The extra benefit is often
called an interaction effect.
If we continue this absurdly simplistic example, it will be easy to see
why proper linear models can accurately represent monotone interactions.
Suppose that for a certain population of people, D will bring a loss of 1
2
pound per week while E will bring a loss of 3
4 pound per week. The
following linear model will predict how much weight loss one can expect:
W ј 1
2
d ю 3
4
e
where W is the number of pounds lost, d is the number of weeks on the
diet, and e is the number of weeks on the exercise regimen. It should be
clear that a proper linear model will do a reasonably good job of predicting
interactions that are not monotone, but for which the interaction
effects are not strong.
The second part of the Hastie-Dawes explanation is a speculation
about the world: In practical social settings (where linear models have proven
most successful), interactions are, near enough and in the main, monotone.
Those who study complex systems, nonlinear dynamics, and catastrophe
theory will note that not all of the world we’re interested in consists of
monotone interactions. The idea is that as long as we are not looking for
SPRs to predict the performance of nonlinear systems, linear models may
perform well—better than human experts. By restricting the explanation
of the success of linear models to practical, social settings, Hastie and
Dawes can take advantage of the flat maximum principle. From the reliability
of proper linear models, they can employ the flat maximum principle
to infer the reliability of improper linear models as well.
We have doubts about the Hastie-Dawes explanation for the success
of SPRs. Consider the linear model that represents the monotone weight
loss interaction. The reason this linear model is reliable is that it accurately
portrays the main causal agents and the relative influence of those agents
in subjects’ weight loss. But the robust reliability of SPRs can’t depend on
their reasonably accurate portrayal of causal reality. The reason is quite
simply that many SPRs are not even close to accurate portrayals of reality.
Consider a linear model that predicts academic performance on the basis
of grade point average and test scores. The student’s college GPA is not a
primary cause of her graduate school performance; same with her test
score. Rather, it is much more plausible to suppose that whatever complex
of factors goes into a student’s GPA and test scores is also heavily implicated
in a student’s success in graduate school. (Recall that the flat maximum
principle is operative when the cues employed by a linear model are
redundant.) So it seems unlikely that the success of SPRs depends on their
mirroring or reflecting monotone interactions. (Thanks to Michael Strevens
for this point.)
36 Epistemology and the Psychology of Human Judgment
We need to be a bit careful here. We’re not suggesting that we oppose
or doubt the possibility of successful SPRs that identify causes. (Just the
opposite.) Nor are we suggesting that successful SPRs do not depend for
their success on causal regularities. (Again, just the opposite.) Our point is
that even when we can’t ‘‘read off’’ anything like the causal structure of the
world from an SPR, it can still be highly reliable and worthy of being used.
If that’s so, then the success of SPRs can’t depend on their representing
(even approximately) the interactions that produce the item of interest.
Reid Hastie and Robyn Dawes have offered a different account of the
success of linear models (2001, 58–62; see also Dawes 1988, 212–15). Their
explanation comes in three parts. Since we embrace and elaborate on the
third part of their explanation in section 3, we will focus only on the first
two parts of their explanation here. The first part of their explanation for
the success of SPRs is a principle about the relationship between proper
linear models and the world: Proper linear models can accurately represent
monotone (or ‘‘ordinal’’) interactions. We have already introduced linear
models—they are models in which the judgment made is a function of the
sum of a certain number of weighted variables. The best way to understand
what monotone interactions are is to consider a simple example.
Suppose a doctor has told you to reduce your body fat, and she recommends
a special diet D and an exercise regime E. Now, let’s suppose that D
alone, without the exercise regime, is effective at reducing body fat. This
would be the diet’s main effect. Suppose also that the exercise regime
alone, without the diet, is also effective at reducing body fat. Again, this
would be the main effect of exercise. Now let’s suppose Sam goes on the
diet D and the exercise regime E. If Sam gets the benefits of both—the
main effect of D and the main effect of E—then the interaction of D and E
is monotone. If, however, Sam gets the main effects of both plus an extra
benefit, then the interaction is not monotone. The extra benefit is often
called an interaction effect.
If we continue this absurdly simplistic example, it will be easy to see
why proper linear models can accurately represent monotone interactions.
Suppose that for a certain population of people, D will bring a loss of 1
2
pound per week while E will bring a loss of 3
4 pound per week. The
following linear model will predict how much weight loss one can expect:
W ј 1
2
d ю 3
4
e
where W is the number of pounds lost, d is the number of weeks on the
diet, and e is the number of weeks on the exercise regimen. It should be
clear that a proper linear model will do a reasonably good job of predicting
interactions that are not monotone, but for which the interaction
effects are not strong.
The second part of the Hastie-Dawes explanation is a speculation
about the world: In practical social settings (where linear models have proven
most successful), interactions are, near enough and in the main, monotone.
Those who study complex systems, nonlinear dynamics, and catastrophe
theory will note that not all of the world we’re interested in consists of
monotone interactions. The idea is that as long as we are not looking for
SPRs to predict the performance of nonlinear systems, linear models may
perform well—better than human experts. By restricting the explanation
of the success of linear models to practical, social settings, Hastie and
Dawes can take advantage of the flat maximum principle. From the reliability
of proper linear models, they can employ the flat maximum principle
to infer the reliability of improper linear models as well.
We have doubts about the Hastie-Dawes explanation for the success
of SPRs. Consider the linear model that represents the monotone weight
loss interaction. The reason this linear model is reliable is that it accurately
portrays the main causal agents and the relative influence of those agents
in subjects’ weight loss. But the robust reliability of SPRs can’t depend on
their reasonably accurate portrayal of causal reality. The reason is quite
simply that many SPRs are not even close to accurate portrayals of reality.
Consider a linear model that predicts academic performance on the basis
of grade point average and test scores. The student’s college GPA is not a
primary cause of her graduate school performance; same with her test
score. Rather, it is much more plausible to suppose that whatever complex
of factors goes into a student’s GPA and test scores is also heavily implicated
in a student’s success in graduate school. (Recall that the flat maximum
principle is operative when the cues employed by a linear model are
redundant.) So it seems unlikely that the success of SPRs depends on their
mirroring or reflecting monotone interactions. (Thanks to Michael Strevens
for this point.)
36 Epistemology and the Psychology of Human Judgment
We need to be a bit careful here. We’re not suggesting that we oppose
or doubt the possibility of successful SPRs that identify causes. (Just the
opposite.) Nor are we suggesting that successful SPRs do not depend for
their success on causal regularities. (Again, just the opposite.) Our point is
that even when we can’t ‘‘read off’’ anything like the causal structure of the
world from an SPR, it can still be highly reliable and worthy of being used.
If that’s so, then the success of SPRs can’t depend on their representing
(even approximately) the interactions that produce the item of interest.