3. The cost-benefit imperative

К оглавлению1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 
17 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 

Our cost-benefit approach to epistemology takes elapsed time to be a surrogate

for epistemic costs and reliability to be a surrogate for epistemic benefits.

This approach has at least two important virtues. The first is that our

surrogates (time and reliability) are measurable. This means that the central

components of applied epistemology (or at least their rough approximations)

can be empirically determined in the following sense: (a) The costbenefit

curve is determined by observed outcomes in performance; and (b)

the curve can then be successfully used as a basis for predictions of performance.

So the central theoretical components of applied epistemology—

or at least rough approximations of them—can in principle be tested for

accuracy, rather than for their ability to stand up to imagined counterexamples.

The second virtue of our approach is that our surrogates roughly

track the properties of interest (the costs and benefits of reasoning). As a

result, the central tool of applied epistemology is reasonably accurate.

At this point, one might point out that successful SPRs have typically

been introduced without the explicit use of the formal machinery of costbenefit

analysis we have introduced here. So one might wonder whether

88 Epistemology and the Psychology of Human Judgment

this machinery is really required to address the efficient allocation of cognitive

resources. Are we shooting a mosquito with a bazooka?We don’t think

so, for two reasons. First, cost-benefit curves are good teaching tools. When

repairing individual reasoning strategies, it is helpful for the individual

to see, in stark and unapologetic terms, how poorly they are performing,

even by their own lights. And nothing does that like a curve—even if the

curve does not capture everything of value. Cost-benefit curves are at once

painfully accessible and mercifully impersonal. Consider the Goldberg

Rule we introduced in chapter 2. This rule predicts whether a psychiatric

patient is neurotic or psychotic on the basis of an MMPI profile. When

tested on a set of 861 patients, the Goldberg Rule had a 70% hit rate;

clinicians’ hit rates varied from a low of 55% to a high of 67%. We can set

out the choice between the Goldberg Rule and clinical prediction in terms

of what their cost-benefit curves might look like (see Figure 5.1).

The cost-benefit curve for the Goldberg Rule is very steep and hits its

near-maximum reliability after a rather modest expenditure of resources.

That’s because it doesn’t require many resources to achieve a high degree

of accuracy; spending more resources (by checking whether one has

plugged in the proper values and done the arithmetic correctly) is likely to

bring very small increments in reliability. Clinical prediction requires

greater resources than the Goldberg Rule but never achieves its reliability.

This point can be made vivid with a curve.

There is a second way that an explicit cost-benefit approach can be

useful. It can help to bring a certain kind of discipline to reasoners. Recall

the selective defection findings. Those who are given the Goldberg Rule

and allowed to selectively defect from it end up reasoning less reliably than

the rule itself; and many who know about the interview effect nonetheless

insist on doing unstructured interviews and drawing conclusions from

them. This hapless defection is typically undesirable. If we have two strategies

for solving a problem and one is more reliable, it is folly to use the

less reliable strategy to correct the more reliable one. There are some very

limited situations in which defection is warranted (see our discussion in

chapter 2). We can represent the costs and benefits of selective defection

with a cost-benefit curve, which might have something like the shape

shown in Figure 5.2. This curve suggests that with minimal cognitive resources

(i.e., those resources necessary to find the relevant scores, do the

simple arithmetic, and determine whether the sum is or is not greater than

45), the reasoner can attain a 70% accuracy rate. But by devoting more resources

to the problem (i.e., by using information from the MMPI profile

to try to improve on the Goldberg Rule’s prediction), the reasoner

degrades his reliability. This finding—that there is a point beyond which

the additional effort associated with considering more information degrades

performance—was also found in the interview effect (chapter 2). So when our cognitive limitations tempt us with a reasoning strategy that is both

subjectively seductive and systematically defective, it is time to lean on a

cognitive prosthetic. Cost-benefit analysis is that cognitive prosthetic.

We understand the temptations of defection. We know what it’s like

to use a reasoning strategy of proven reliability when it seems to give an

answer not warranted by the evidence. It feels like you’re about to make an

unnecessary error. And maybe you are. But in order to make fewer errors

overall, we have to accept that we will sometimes make errors we could

have corrected—errors that we recognized as errors before we made them

but made them nonetheless (Einhorn 1986). (Of course, the point is that

in these situations, more often than not, what we think will be an error

in fact won’t be.) People often lack the discipline to adhere to a superior

strategy that doesn’t ‘‘feel’’ right. Reasoning in a way that sometimes

‘‘feels’’ wrong takes discipline. And one way to impose that discipline is to

think about applied epistemology in terms of costs and benefits. All reasoning

strategies will lead to error (costs). When it comes to deciding

whether to defect from an SPR, the net benefits of defection (more errors

with greater effort) are typically outweighed by the net benefits of diffident

acquiescence to the SPR. When it comes to learning to reason better,

discipline is essential; and we can think of no more effective way to impose

that discipline than with a cost-benefit approach to applied epistemology.

Robyn Dawes (2001) notes that piloting an airplane by sight can give

you the powerful impression that you are flying right-side up when you

are actually upside down. So pilots learn to fly by instruments. This takes

some doing, but pilots are repaid with longer lives. To the extent that

individuals appreciate the benefits of averting costly error, they need to fly

by cost-benefit instruments. Adhering to a cost-benefit analysis may feel

wrong, but then again so does flying by instruments at first.

We fear this discussion may have made us appear like unduly strict

disciplinarians, so let us end on a gentler note. It would be irresponsible

for any applied epistemology that announces the importance of efficiency

to ignore the controllable inefficiency of reasoners. And we worry about

just how far discipline can really go in saving reasoners from ill-fated

temptation. For certain sorts of reasoning problems, applied epistemology

might well recommend strategies—reasoning or otherwise—that are designed

to develop and foster healthy reasoning dispositions. Why labor in

the construction of reasoning rules designed to correct our errant cognitive

impulses when we can cultivate actors who are not seduced by those

detrimental impulses in the first place? This point is reflected in good

advice for parents: It is probably less costly to cultivate your child’s many

The Costs and Benefits of Excellent Judgment 91

interests—keeping them very busy and focused on satisfying activities—

than to teach them how to control or correct their drug addictions later.

Similarly, when thinking about the various ways people reason badly, it

may be easier to cultivate new habits than to revise how we reason. If S’s

reasoning rashly discounts the future, rather than force her to adopt new

reasoning strategies, it might be more effective to set up an automatic

withdrawal from her bank account into a retirement fund that comes with

stiff penalties for early withdrawals. Suppose S is tempted to become an

active stock-market trader because he takes a short-term rise in his portfolio

to be evidence of his financial cleverness. Rather than fill his head

with a lot of theories and statistics, it is probably easier to force him to

focus on time horizons of at least 10 years in the stock market and thereby

avoid the temptation of active trading. And what about an academic

department tempted by the interview effect? Take the money the department

spends to send interviewers to conferences and transfer it into research

accounts for junior faculty. The result will likely be better hires and

happier, more productive junior faculty.

Our cost-benefit approach to epistemology takes elapsed time to be a surrogate

for epistemic costs and reliability to be a surrogate for epistemic benefits.

This approach has at least two important virtues. The first is that our

surrogates (time and reliability) are measurable. This means that the central

components of applied epistemology (or at least their rough approximations)

can be empirically determined in the following sense: (a) The costbenefit

curve is determined by observed outcomes in performance; and (b)

the curve can then be successfully used as a basis for predictions of performance.

So the central theoretical components of applied epistemology—

or at least rough approximations of them—can in principle be tested for

accuracy, rather than for their ability to stand up to imagined counterexamples.

The second virtue of our approach is that our surrogates roughly

track the properties of interest (the costs and benefits of reasoning). As a

result, the central tool of applied epistemology is reasonably accurate.

At this point, one might point out that successful SPRs have typically

been introduced without the explicit use of the formal machinery of costbenefit

analysis we have introduced here. So one might wonder whether

88 Epistemology and the Psychology of Human Judgment

this machinery is really required to address the efficient allocation of cognitive

resources. Are we shooting a mosquito with a bazooka?We don’t think

so, for two reasons. First, cost-benefit curves are good teaching tools. When

repairing individual reasoning strategies, it is helpful for the individual

to see, in stark and unapologetic terms, how poorly they are performing,

even by their own lights. And nothing does that like a curve—even if the

curve does not capture everything of value. Cost-benefit curves are at once

painfully accessible and mercifully impersonal. Consider the Goldberg

Rule we introduced in chapter 2. This rule predicts whether a psychiatric

patient is neurotic or psychotic on the basis of an MMPI profile. When

tested on a set of 861 patients, the Goldberg Rule had a 70% hit rate;

clinicians’ hit rates varied from a low of 55% to a high of 67%. We can set

out the choice between the Goldberg Rule and clinical prediction in terms

of what their cost-benefit curves might look like (see Figure 5.1).

The cost-benefit curve for the Goldberg Rule is very steep and hits its

near-maximum reliability after a rather modest expenditure of resources.

That’s because it doesn’t require many resources to achieve a high degree

of accuracy; spending more resources (by checking whether one has

plugged in the proper values and done the arithmetic correctly) is likely to

bring very small increments in reliability. Clinical prediction requires

greater resources than the Goldberg Rule but never achieves its reliability.

This point can be made vivid with a curve.

There is a second way that an explicit cost-benefit approach can be

useful. It can help to bring a certain kind of discipline to reasoners. Recall

the selective defection findings. Those who are given the Goldberg Rule

and allowed to selectively defect from it end up reasoning less reliably than

the rule itself; and many who know about the interview effect nonetheless

insist on doing unstructured interviews and drawing conclusions from

them. This hapless defection is typically undesirable. If we have two strategies

for solving a problem and one is more reliable, it is folly to use the

less reliable strategy to correct the more reliable one. There are some very

limited situations in which defection is warranted (see our discussion in

chapter 2). We can represent the costs and benefits of selective defection

with a cost-benefit curve, which might have something like the shape

shown in Figure 5.2. This curve suggests that with minimal cognitive resources

(i.e., those resources necessary to find the relevant scores, do the

simple arithmetic, and determine whether the sum is or is not greater than

45), the reasoner can attain a 70% accuracy rate. But by devoting more resources

to the problem (i.e., by using information from the MMPI profile

to try to improve on the Goldberg Rule’s prediction), the reasoner

degrades his reliability. This finding—that there is a point beyond which

the additional effort associated with considering more information degrades

performance—was also found in the interview effect (chapter 2). So when our cognitive limitations tempt us with a reasoning strategy that is both

subjectively seductive and systematically defective, it is time to lean on a

cognitive prosthetic. Cost-benefit analysis is that cognitive prosthetic.

We understand the temptations of defection. We know what it’s like

to use a reasoning strategy of proven reliability when it seems to give an

answer not warranted by the evidence. It feels like you’re about to make an

unnecessary error. And maybe you are. But in order to make fewer errors

overall, we have to accept that we will sometimes make errors we could

have corrected—errors that we recognized as errors before we made them

but made them nonetheless (Einhorn 1986). (Of course, the point is that

in these situations, more often than not, what we think will be an error

in fact won’t be.) People often lack the discipline to adhere to a superior

strategy that doesn’t ‘‘feel’’ right. Reasoning in a way that sometimes

‘‘feels’’ wrong takes discipline. And one way to impose that discipline is to

think about applied epistemology in terms of costs and benefits. All reasoning

strategies will lead to error (costs). When it comes to deciding

whether to defect from an SPR, the net benefits of defection (more errors

with greater effort) are typically outweighed by the net benefits of diffident

acquiescence to the SPR. When it comes to learning to reason better,

discipline is essential; and we can think of no more effective way to impose

that discipline than with a cost-benefit approach to applied epistemology.

Robyn Dawes (2001) notes that piloting an airplane by sight can give

you the powerful impression that you are flying right-side up when you

are actually upside down. So pilots learn to fly by instruments. This takes

some doing, but pilots are repaid with longer lives. To the extent that

individuals appreciate the benefits of averting costly error, they need to fly

by cost-benefit instruments. Adhering to a cost-benefit analysis may feel

wrong, but then again so does flying by instruments at first.

We fear this discussion may have made us appear like unduly strict

disciplinarians, so let us end on a gentler note. It would be irresponsible

for any applied epistemology that announces the importance of efficiency

to ignore the controllable inefficiency of reasoners. And we worry about

just how far discipline can really go in saving reasoners from ill-fated

temptation. For certain sorts of reasoning problems, applied epistemology

might well recommend strategies—reasoning or otherwise—that are designed

to develop and foster healthy reasoning dispositions. Why labor in

the construction of reasoning rules designed to correct our errant cognitive

impulses when we can cultivate actors who are not seduced by those

detrimental impulses in the first place? This point is reflected in good

advice for parents: It is probably less costly to cultivate your child’s many

The Costs and Benefits of Excellent Judgment 91

interests—keeping them very busy and focused on satisfying activities—

than to teach them how to control or correct their drug addictions later.

Similarly, when thinking about the various ways people reason badly, it

may be easier to cultivate new habits than to revise how we reason. If S’s

reasoning rashly discounts the future, rather than force her to adopt new

reasoning strategies, it might be more effective to set up an automatic

withdrawal from her bank account into a retirement fund that comes with

stiff penalties for early withdrawals. Suppose S is tempted to become an

active stock-market trader because he takes a short-term rise in his portfolio

to be evidence of his financial cleverness. Rather than fill his head

with a lot of theories and statistics, it is probably easier to force him to

focus on time horizons of at least 10 years in the stock market and thereby

avoid the temptation of active trading. And what about an academic

department tempted by the interview effect? Take the money the department

spends to send interviewers to conferences and transfer it into research

accounts for junior faculty. The result will likely be better hires and

happier, more productive junior faculty.