Robin argued in an earlier post that we should use stereotypes to the extent that they contain useful information. The impulse to use all relevant information is generally a good one, but there are solid grounds for certain rule-based refusals to do so. For example:
You say "1. The fact that jurors are supposed to assume that the guy in the dock committed the crime with probability 1/N, when everyone knows that it is really much higher than that, restrains cops who might otherwise exploit their power to get someone convicted on little evidence other than the fact that they were arrested."
It is because juries do not *convict* people based merely on the fact that they are in the dock that constraints the police, DA, etc. But this does not require them to assign probability 1/N, or in other words, to 'ignore plain facts', because a decision requires not just a probability distribution but also a loss function. To obtain the desired decision, the loss associated with convicting an innocent person must be greater than a threshold, the threshold being determined by the loss associated with freeing a guilty person and the probability of being guilty given only that you are in court.
Steve, I'm sure I'm not the only person to see the irony of posting a ridiculous post facto rationalisation laden with confirmation bias on a site dedicated to overcoming cognitive bias :-)
I'd be astonished if, having learned that the guy he checked in was Atta, the CSA didn't think he looked suspicious in retrospect.
We now know that the airline employee who checked in Mohammed Atta on the morning of 9/11/01 thought to himself that he looked more like an Arab terrorist than anyone he'd ever seen, but the employee then mentally criticized himself for political incorrectness and went on to let Atta proceed.
Would it have been more socially destructive to have avoided 3000 deaths plus the Iraq war by using ethnic profiling?
Regarding the Muslims-on-a-plane, there are good reasons to exclude some signals precisely in that case - there is evidence (http://www.firstmonday.org/... that, when the rate of real positives is very low as with terrorists, profiling actually decreases the likelihood of catching a terrorist. Also, when false positives have a significant cost (which can be a loss of information), it may be better to turn down the sensitivity.
What I'm driving at is that quite often, in the sort of intelligence/policing scenarios that this usually affects, it's beneficial to set the barrier for action high - like turning down the sensitivity of a radar in order to get rid of clutter or using a band-pass filter to get rid of radio noise.
"But if jurors start acting on the assumption that people against whom charges have been brought must be guilty, they destroy the incentive for cops and DAs to show restraint."
Well, to be accurate, jurors *do* act on that assumption. They're instructed not to, which (we hope) reduces the tendency. Also, the knowledge of the general population that cops and DAs *don't* show complete restraint (partly because of jurors' imperfect impartiality) reduces the tendency a bit more. So there's an equilibrium there, and a dynamic one, since jurors' level of impartiality is determined by their perception of cops' restraint, which has a substantial time lag from actual changing levels of police restraint. You might even expect there to be cycles around the equilibrium point.
Case 1 is potentially very interesting, because it might be that the frequency of guilt --prior to the evidence but conditioned on the fact of being before a jury -- is high only because we ignore the just mentioned frequency and instead aim to convict on evidence establishing guilt beyond a reasonable doubt. Suppose that we had a system in which the jury verdict took into account that frequency as a prior probability, or took into account prior convictions, etc. The police would perhaps get more convictions just on that basis. But would it still be true that most people tried before a jury are guilty? Wouldn't the police, as pointed out, go for more trials of innocent people when the evidence is weaker. But then the prior probability of guilt would be lower. Would a Bayesian approach to this (0) necessarily result in a prior different from what it is now; (1) be better at detecting guilt; (2) be more wasteful (more trials of innocent people; (3) be fairer?
#1 brings up the interesting issue that sometimes making use of knowledge changes it, so we need to think carefully about /why/ rules of thumb work before incorporating them into our decision structure.
There exists a rule that says: to get a conviction, the case must be proved in court beyond a shadow of a doubt. Because that rule exists, cops are reluctant to arrest and DAs unlikely to press charges unless the case can be proven on its own merits. Since cops and DAs act that way, it seems safe to assume most people who get arrested and accused of a crime are probably guilty.
But if jurors start acting on the assumption that people against whom charges have been brought must be guilty, they destroy the incentive for cops and DAs to show restraint. The more people make use of the "he's probably guilty" rule of thumb in /current/ cases, the less valid it will be in /future/ cases.
So we have some knowledge about what is probably true given our current set of social institutions, but it's not safe to act on that knowledge as a hard rule, because acting on it destroys the contingency that produced it.
Since we are trying to overcome bias, we must certainly exist that biases exist. Thus we might plausibly sometimes try to correct for one bias by introducing another one. But the question has to be what standards of evidence will we use here. With too high a standard, we might never make needed corrections, but with too low a standard, we might just reflexively justify all of our biases by making up biases they are supposed to be correcting.
The Muslims-on-a-plane analogy is often used because it would be such a severe example of socially destructive descrimination. But there are many more stereotypes that can be used as useful heuristics, perhaps if only in background cogitation rather than front-and-center decision-making processes.
#1 advances the policy of disgarding age or race descriminations in arrests, which is probably a good idea. People should be judged on what they are doing and where they are located at the scene of a crime, not external features like race.
Part of the problem is that race is the most readily distinguishable human characteristic aside from sex and general stage of life, so that it tends to be overweighted in many judgements, especially those of guilt. Us humans have a millenia-old history of xenophobic cultural tendencies, which can be tapped to turn anyone against anyone else if the situation calls for it, and causes major problems both a local and national leevel when allowed to flourish.
#2 is true, but also partially beside the point because there is always descrimination in employment, no matter what. So nothing can be changed.
#3 is the theory that racial profiling in the police force operates like a bigot magnet. I believe that the police force already has tendencies to behave that way, so allowing racial prolifing would only make it much worse. We all remember the recent incident in New York City, when a man was shot dead on his wedding day for no real reason, with over 10 bullets.
While I don't really understand what "policy" is being advocated in #1, #2 seems incomplete, and probably wrong. If some group is offered jobs based on statistical evidence of discrimination in some industry, absent any evidence of actual bias, then it is true that members of that group might invest in skills required for that industry. But if the discrimination is purely statistical, then why is that a good thing? The same preference in hiring that leads to more investment by the underrepresented group leads to less investment by overrepresented groups. And, if the over and under representation reflects underlying skill abilities differences rather than discrimination, then we simply have a less effective work force. I mean, there are no guys in in the NBA who are in their 40s and under 6 ft. - can't we do something about that?
The picture at the top of this blog shows Odysseus tied to the mast of his vessel so that he can hear the sirens' song but not respond to it, while his crew has wax in their ears to prevent them too from responding to the song. This is essentially what David is suggesting. In certain circumstances there may be benefits from ignoring signals. In a more economic context, this is the essence of Milton Friedman's rule for a fixed rate of growth of the money supply.
When someone else has preferences that differ from you, but is taking actions that effect you, then you can want to limit their actions, including limiting the kinds of signals that their actions depend on. For example, you might not want jurors to condition their decision on whether the accused lived in the same small town as they, as you expect people to favor people in the same town for reasons unrelated to whether they are guilty.
You're correct; there are, in some cases, benefits to ignoring reality. But that doesn't necessarily mean ignorance is a rational approach. Policymakers must weigh costs AND benefits when implementing policy. All three of the policies you listed come with large costs that must be compared to the benefits you gave, and it's not unlikely that all three policies are irrational when viewed this way.
Err, I'm not sure that I disagree with these positions, though I think I disagree with at least the first and third, but the arguments you gave are really so weak as to tend to move my opinion against them. I recommend that you try to strengthen them.
What Incentives Does Statistical Discrimination Give?
Suppose most economists believe that "Libertarian economists can't do math," and that, on average, they are correct. How does this...
Dear David,
You say "1. The fact that jurors are supposed to assume that the guy in the dock committed the crime with probability 1/N, when everyone knows that it is really much higher than that, restrains cops who might otherwise exploit their power to get someone convicted on little evidence other than the fact that they were arrested."
It is because juries do not *convict* people based merely on the fact that they are in the dock that constraints the police, DA, etc. But this does not require them to assign probability 1/N, or in other words, to 'ignore plain facts', because a decision requires not just a probability distribution but also a loss function. To obtain the desired decision, the loss associated with convicting an innocent person must be greater than a threshold, the threshold being determined by the loss associated with freeing a guilty person and the probability of being guilty given only that you are in court.
illywacker;
Steve, I'm sure I'm not the only person to see the irony of posting a ridiculous post facto rationalisation laden with confirmation bias on a site dedicated to overcoming cognitive bias :-)
I'd be astonished if, having learned that the guy he checked in was Atta, the CSA didn't think he looked suspicious in retrospect.
We now know that the airline employee who checked in Mohammed Atta on the morning of 9/11/01 thought to himself that he looked more like an Arab terrorist than anyone he'd ever seen, but the employee then mentally criticized himself for political incorrectness and went on to let Atta proceed.
Would it have been more socially destructive to have avoided 3000 deaths plus the Iraq war by using ethnic profiling?
For details, see:
http://isteve.blogspot.com/...
Regarding the Muslims-on-a-plane, there are good reasons to exclude some signals precisely in that case - there is evidence (http://www.firstmonday.org/... that, when the rate of real positives is very low as with terrorists, profiling actually decreases the likelihood of catching a terrorist. Also, when false positives have a significant cost (which can be a loss of information), it may be better to turn down the sensitivity.
What I'm driving at is that quite often, in the sort of intelligence/policing scenarios that this usually affects, it's beneficial to set the barrier for action high - like turning down the sensitivity of a radar in order to get rid of clutter or using a band-pass filter to get rid of radio noise.
"But if jurors start acting on the assumption that people against whom charges have been brought must be guilty, they destroy the incentive for cops and DAs to show restraint."
Well, to be accurate, jurors *do* act on that assumption. They're instructed not to, which (we hope) reduces the tendency. Also, the knowledge of the general population that cops and DAs *don't* show complete restraint (partly because of jurors' imperfect impartiality) reduces the tendency a bit more. So there's an equilibrium there, and a dynamic one, since jurors' level of impartiality is determined by their perception of cops' restraint, which has a substantial time lag from actual changing levels of police restraint. You might even expect there to be cycles around the equilibrium point.
Case 1 is potentially very interesting, because it might be that the frequency of guilt --prior to the evidence but conditioned on the fact of being before a jury -- is high only because we ignore the just mentioned frequency and instead aim to convict on evidence establishing guilt beyond a reasonable doubt. Suppose that we had a system in which the jury verdict took into account that frequency as a prior probability, or took into account prior convictions, etc. The police would perhaps get more convictions just on that basis. But would it still be true that most people tried before a jury are guilty? Wouldn't the police, as pointed out, go for more trials of innocent people when the evidence is weaker. But then the prior probability of guilt would be lower. Would a Bayesian approach to this (0) necessarily result in a prior different from what it is now; (1) be better at detecting guilt; (2) be more wasteful (more trials of innocent people; (3) be fairer?
#1 brings up the interesting issue that sometimes making use of knowledge changes it, so we need to think carefully about /why/ rules of thumb work before incorporating them into our decision structure.
There exists a rule that says: to get a conviction, the case must be proved in court beyond a shadow of a doubt. Because that rule exists, cops are reluctant to arrest and DAs unlikely to press charges unless the case can be proven on its own merits. Since cops and DAs act that way, it seems safe to assume most people who get arrested and accused of a crime are probably guilty.
But if jurors start acting on the assumption that people against whom charges have been brought must be guilty, they destroy the incentive for cops and DAs to show restraint. The more people make use of the "he's probably guilty" rule of thumb in /current/ cases, the less valid it will be in /future/ cases.
So we have some knowledge about what is probably true given our current set of social institutions, but it's not safe to act on that knowledge as a hard rule, because acting on it destroys the contingency that produced it.
Since we are trying to overcome bias, we must certainly exist that biases exist. Thus we might plausibly sometimes try to correct for one bias by introducing another one. But the question has to be what standards of evidence will we use here. With too high a standard, we might never make needed corrections, but with too low a standard, we might just reflexively justify all of our biases by making up biases they are supposed to be correcting.
The Muslims-on-a-plane analogy is often used because it would be such a severe example of socially destructive descrimination. But there are many more stereotypes that can be used as useful heuristics, perhaps if only in background cogitation rather than front-and-center decision-making processes.
#1 advances the policy of disgarding age or race descriminations in arrests, which is probably a good idea. People should be judged on what they are doing and where they are located at the scene of a crime, not external features like race.
Part of the problem is that race is the most readily distinguishable human characteristic aside from sex and general stage of life, so that it tends to be overweighted in many judgements, especially those of guilt. Us humans have a millenia-old history of xenophobic cultural tendencies, which can be tapped to turn anyone against anyone else if the situation calls for it, and causes major problems both a local and national leevel when allowed to flourish.
#2 is true, but also partially beside the point because there is always descrimination in employment, no matter what. So nothing can be changed.
#3 is the theory that racial profiling in the police force operates like a bigot magnet. I believe that the police force already has tendencies to behave that way, so allowing racial prolifing would only make it much worse. We all remember the recent incident in New York City, when a man was shot dead on his wedding day for no real reason, with over 10 bullets.
While I don't really understand what "policy" is being advocated in #1, #2 seems incomplete, and probably wrong. If some group is offered jobs based on statistical evidence of discrimination in some industry, absent any evidence of actual bias, then it is true that members of that group might invest in skills required for that industry. But if the discrimination is purely statistical, then why is that a good thing? The same preference in hiring that leads to more investment by the underrepresented group leads to less investment by overrepresented groups. And, if the over and under representation reflects underlying skill abilities differences rather than discrimination, then we simply have a less effective work force. I mean, there are no guys in in the NBA who are in their 40s and under 6 ft. - can't we do something about that?
The picture at the top of this blog shows Odysseus tied to the mast of his vessel so that he can hear the sirens' song but not respond to it, while his crew has wax in their ears to prevent them too from responding to the song. This is essentially what David is suggesting. In certain circumstances there may be benefits from ignoring signals. In a more economic context, this is the essence of Milton Friedman's rule for a fixed rate of growth of the money supply.
When someone else has preferences that differ from you, but is taking actions that effect you, then you can want to limit their actions, including limiting the kinds of signals that their actions depend on. For example, you might not want jurors to condition their decision on whether the accused lived in the same small town as they, as you expect people to favor people in the same town for reasons unrelated to whether they are guilty.
David,
You're correct; there are, in some cases, benefits to ignoring reality. But that doesn't necessarily mean ignorance is a rational approach. Policymakers must weigh costs AND benefits when implementing policy. All three of the policies you listed come with large costs that must be compared to the benefits you gave, and it's not unlikely that all three policies are irrational when viewed this way.
Err, I'm not sure that I disagree with these positions, though I think I disagree with at least the first and third, but the arguments you gave are really so weak as to tend to move my opinion against them. I recommend that you try to strengthen them.