Sunday, February 20, 2011

Minority Report: Leon Neyfakh on Risk Prediction

Bernard E. Harcourt

Leon Neyfakh, over at the Ideas section of the Boston Globe, has a fascinating piece today on the increased use of risk prediction instruments, "Inside the new science of predicting violence." Neyfakh covers the ground extremely well and even-handedly, discussing the important contributions of John Monahan and his colleagues, Richard Berk at Penn, and others.

Neyfakh raises some criticisms of risk prediction, and the only thing I would add is that, even if you assume that these prediction tools work, in all likelihood their use undermines the goals that they are intended to promote. We usually adopt these risk prediction instruments with the goal of reducing future crime. The fact is, however, that profiling on risk likely is going to increase overall crime in society, either because (if you believe that people are “rational actors” and change their criminal behavior based on the amount of punishment) the profiled people are less likely to respond much to the increased punishment whereas everyone else will respond much more to the decreased punishment, or because (if you believe that people don’t change their behavior because of the amount of punishment) profiling is going to disproportionately affect the profiled group, which is going to have devastating effects on their employment, family, and social outcomes.

There arguments are a mouthful, I realize, and I spell them out in detail in Against Prediction, but just to be clear, let me give a paragraph to each. First, even if we assume that individuals respond to increased punishment by offending less (even if we adopt the economists’ assumptions of choice), the overall long-term effect is likely to be counter-productive if the profiled population is less responsive to punishment than the non-profiled population. Profiling may well deter crime among the profiled population, but if so, it will likely increase the offending rates of non-profiled groups. If those others are more responsive to changes in punishment, then the overall rate of crime in society will go up. And the truth is, we have no idea about the comparative responsiveness of different groups to punishment, but a reasonable and conservative guess is that, if a population has a higher offending rate, it probably is less elastic to punishment. Most explanations for why a group would have a higher offending rate than other groups would also suggest that the group is less responsive to punishment.

Second, even if people don’t change their behaviors in response to punishment, the problem is that profiling is going to have a disproportionate effect on members of the profiled population and this distortion is only going to get worse with every new annual crime statistic. It may surprise, but the only way to get a prison population that reflects the offending population is to police and punish randomly: to sample randomly. That’s why we use random sampling in the social sciences. Anytime you start picking and choosing, your sample is going to be skewed. And the point is, the skew has devastating consequences for the profiled group, not only on their criminal justice outcomes, but also on their job potential, on their educational outcomes, and on their family lives—which in turn fuels increased criminal activity.

The bottom line is that there is no good law enforcement argument for profiling on risk prediction: the practice is in all likelihood counter-productive to the very goals of law enforcement. The other big danger, of course, is that these risk instruments rely increasingly on prior criminal history and, because of disparate treatment in policing, prior criminal history is increasingly a proxy for race. Therefore, the use of these risk tools is likely going to worsen the already disproportionate effect of the criminal justice system on minorities.

Neyfakh ends his article with an interesting conundrum—the question of where all this will lead us. Neyfakh recounts: “Supposing I am able to tell a mother that her 8-year-old has a one in three chance of committing a homicide by age 18,” says Richard Berk at the University of Pennsylvania. “What the hell do I do with that information? What do the various social services do with that information? I don’t know.”

One answer, I take it, would be to divert more resources toward that youth. That’s exactly what the City of Chicago began to do over the past year, with school officials creating a “risk model” to predict who is more likely to be the victim of gun violence and then to target mentoring and other resources toward them. In other words, to affirmatively target higher-risk youngsters so as to reduce the odds.

As a critic of prediction, that’s the toughest case for me. Is it alright to use risk prediction to target treatment to youth in order to help them beat the odds? I’m afraid the answer, again, is no, because it’s likely that the diversion of resources away from the youths who are less at risk is likely to have a greater and more detrimental effect on them. The best bet is to invest equally in our youths. Equally, and heavily.

Older Posts
Newer Posts