The Bactra Review: Occasional and eclectic book reviews by Cosma Shalizi   135

# Against Prediction

## by Bernard E. Harcourt

University of Chicago Press, 2007

#### Harcourt contra divinationem

Law professors — even law professors at the University of Chicago — are not known for mathematical modeling. Nonetheless, what Harcourt has done here is to step up a basic but compelling mathematical model of profiling, and solve it. Consider, he says, a population where some people are law-abiding and some people are not, and which is also divided into a majority and a minority by the presence of some observable trait. Further suppose that the members of the minority are more likely to be criminals. Question: should the police profile based on the trait, i.e., give more of their attention to members of the minority, on the grounds that they are more likely to catch criminals in this way?

A number of economists have set up similar models, and come to the conclusion that this sort of profiling is rational, because it maximizes the success of searches, i.e., the fraction of the time an investigation will uncover a criminal. This may, indeed, be good for the police officers, but Harcourt points out that it completely misses the point of having a police department; that is not to conduct successful investigations but to reduce crime. It also completely misses the fact that criminals, or potential criminals, are people with brains, and, as the conomists like to say in other contexts, "people respond to incentives", specifically the incentive to not get caught. Given fixed police resources, profiling members of the minority means anti-profiling members of the majority, so majority-group criminals become less likely to be caught. If potential criminals are at all risk-sensitive, this will mean a higher proportion of the majority group will commit whatever crime it is we're talking about. Since there are, after all, more people in the majority than in the minority, this can easily mean that accurate profiling leads to an increase in crime. The key has to do with the relative size of the groups, and with their relative elasticity of crime with respect to policing — if the minority group has a low elasticity (because, e.g., of limited non-criminal economic opportunities), a dramatic increase in the success rate of searches can be accompanied by an equally dramatic increase in over-all criminal activity.

This basic argument is related to the "carnival booth" algorithm for foiling security screening, which I like to assign as an exam problem. Harcourt doesn't mention carnival booth, but that's OK, because in some ways his proof is even nicer — it doesn't require any kind of conscious deliberation or coordination on the part of the criminals, just some sensitivity to the risks of getting caught. Within his model, it rises to the level of a proof.

In the real world, we have literally no idea what the elasticity of crime with respect to policing is, and needless to say advocates of profiling have made no attempt to gather such information. Of course, turning the logic around, it does show that there are some cases where profiling does make crime-control sense, e.g., focusing on men as rapists, because there we do know something about the elasticities. But these are so obvious that the only serve as sanity-checks on the calculation. In every controversial case, there is, because of Harcourt's demonstration, literally no reason to think that profiling actually helps.

The core of the argument is helpfully set out by Harcourt in this free working paper. The book adds a less-mathematical explanation of the argument, more elaboration on why profiling is a bad idea, both as a matter of policy and of justice, and some historical investigation into the rise of the use of statistical prediction ("actuarial methods") in policing and corrections. (Unfortunately, Harcourt puts this historical material first, before actually laying out his case against prediction, which I think is a mistake.) Many scholars, having made this case, would go on to say that the problem is that actuarial methods treat people as interchangable members of mass bodies, lacking the individualizing, humanizing touch, etc., etc.; refreshingly, Harcourt will have none of this, and is quite forthright that replacing actuarial methods with "clinical" ones, i.e., the intutive judgments of policemen, jailers, etc., would amount to replacing predictions which can be publicly scrutinized and critiqued with ones which can't. The solution, he says, is to stop trying to predict, and he describes many ways in which this can be profitably done.

264 pp., a few graphs

Currently in print as a hardback, US\$65, ISBN 0226316130 [buy from Powell's] and as a paperback, US\$25, ISBN 0226316149 [buy from Powell's]

30 April 2007