Thursday, 25 July 2013

Predictive Policing and Race

On facing pages in The Economist's print edition this week: a piece on predictive policing and a piece on Trayvon Martin. The juxtaposition raises questions that are worth exploring. Will predictive policing help keep future Trayvons, Oscar Grants, and Amadou Diallos alive? Or will it make matters worse?

Predictive policing purports to identify high-risk locations for crime within neighbourhoods, based, in part, on recent crimes there and nearby. It also promises to help identify likely perpetrators (see also this unfortunately gated New Yorker piece on predicting domestic violence in Massachusetts). I want to start with the direct effects on police interactions with the public, before turning to vigilantes like Zimmerman and how predictive policing might change regulation.

The first Economist piece outlines some problems with prediction:

Misuse and overuse of data can amplify biases. It matters, for example, whether software crunches reports of crimes or arrests; if the latter, police activity risks creating a vicious circle. And report-based systems may favour rich neighbourhoods which turn to the police more readily rather than poor ones where crime is rife. Crimes such as burglary and car theft are more consistently reported than drug dealing or gang-related violence.

However, the Economist is still overall optimistic about the effects of predictive policing on race:

But mathematical models might make policing more equitable by curbing prejudice. A suspicious individual’s presence in a “high-crime area” is among the criteria American police may use to determine whether a search is acceptable: a more rigorous definition of those locations will stop that justification being abused. Detailed analysis of a convict’s personal history may be a fairer reason to refuse parole than similarity to a stereotype.

In sum, prediction will allow the cops to do profiling better, and more fairly. I think this is overly optimistic. When a model identifies a high-risk location, the police show up and look around. But what are they looking for? Many appear to be looking for "suspicious-looking" black male teenagers with "furtive" movements. A mathematically-derived predictive model won't stop them doing so, no more than a hunch or prejudice about a violent neighbourhood.

It might, indeed, make matters worse. We're not just bad at predicting. We are bad at understanding prediction. We don't think enough about margins for error, we distort extreme probabilities badly, and we have very weird patterns of risk acceptance and risk aversion, especially at extremes of probability. What's more, the ex ante likelihood of crime on any given night in any given location is pretty small. So even if Block A is more likely than Block B to experience a robbery, we can't say with much confidence that a robbery will occur at Block A. You could make money betting that robbery is more likely to occur on Block A than Block B, but not by betting, at even odds, that a robbery was going to happen on Block A on a given night. But I don't know that we understand all that well: we may become inclined to believe that because Block A is "risky" in the sense of "more risky than Block B," it is then risky in the sense of "having an absolutely, not relatively, high risk of crime." That would be a non-sequitur, but it's one I think people will commit. Under the illusion of computers and their cachet, police may well become more inclined to shoot first. And what goes for a police goes for a jury.

Ultimately, if a cop thinks a location is high-risk and sees a black kid looking "furtive," whatever the hell that is supposed to mean, I don't think it matters much whether a computer told them the location was high-risk or whether their hunch told them. Whether we have a dead kid or not comes down to the cop's ethics, training, professionalism, and prejudice. How a jury responds depends on their prejudices and those of the legal system. There are massive problems with each.

As for vigilantes: it is probably not encouraging to contemplate Zimmerman having access to predictive crime models. Even absent that scenario, I think I'd ideally want the predictive models that are being used to be able to predict when a vigilante is about to be a Zimmerman. I have no idea how feasible this is. It may not even matter: Zimmerman was told by a police dispatcher that he shouldn't chase Travyon, and he did anyway. A cop on the scene may have helped, or, going by all of the above, may have made everything worse. But it is, in any case, well to remember the Economist's warning about what kinds of crime get plugged in to a predictive model. Is it the break-ins alone--i.e. what supposedly prompted Zimmerman to be driving around that night? Or is it the act of a Zimmerman taking the law into his own hands? And if that isn't defined as a crime, what happens?

No comments:

Post a Comment