Hey Commissioner... Hand Me That Predictive Tool

Can predictive tools help control the costs of incarceration?

Posted May 01, 2009

People worry about crime - about becoming a victim, about its costs, and about a leaky parole system. Are we doing all we can to control those risks?

Is it any wonder that Senators Jim Webb and Arlen Specter's crime bill, The National Criminal Justice Commission Act of 2009, won bipartisan support? Economic collapse has a way of focusing everyone's attention on fiscal responsibility, and the U.S. corrections system has been on an unaccountable spending binge for the last three decades. Though not mentioned in the bill, there are predictive tools we could use at parole time that would help us to control this addiction to incarceration.

It took a new administration and financial desperation before states eyed our gluttonous and wayward corrections system as a place to trim. Facing shrinking budgets, states across the country started looking for ways to save money, and they soon found that the U.S. had become The Incarcerated Society. In 2007, the U.S. had over 2.3 million persons incarcerated. We now lead the world in incarceration rate - the U.S. hovers at five to eight times higher than the countries of Western Europe and twelve times higher than Japan. Senators Webb and Specter came to the obvious conclusion: There are huge inefficiencies in the corrections system and, not being prisoners ourselves, we continue to pay the price because we adapted to those inefficiencies. In the last ten years, parole boards have been using woefully outmoded methods for deciding which prisoners to release, and legislators have been hoping that everything will be ok as violent criminals are released to make room for the mandatory sentences of nonviolent offenders.

This costly embarrassment was a long time coming. From 1985 to 2005, the number of people in local jails rose threefold, from 256,615 to 747,529. This dramatic rise was not some kind of unanticipated side-effect; it was the direct result of policy decisions. These policies freshly criminalized behavior by reclassifying it. From 1974 and 2001, the average American became 3 times more likely to go to prison sometime in his or her lifetime (about 6.6%). Were people getting less compliant? It wouldn't seem so. The crime rate was dropping for much of that period.

The real explanation is much simpler. Well-meaning but poorly-tested offender-treatment programs couldn't stem the escalating crime rates in the 1970s, so new policies resolved to control crime by locking up more people for longer periods. Those policies are still with us, along with the problems they caused: overcrowded prisons, enormous prison administration costs, and higher recidivism rates. We couldn't build prisons fast enough.

The American public has lost its taste for the tough-talking, high-spending, "lock ‘em up at any cost" attitude of prior administrations. At the same time, we aren't prepared to flood the streets with unreformed criminals. So now we may be ready to exercise some science-based governing. As a result, it is all the more disappointing that so promising a bill gives scant attention to the savings that could be harvested by fairer and more accurate parole decision-making.

In states that still have discretionary parole, a parole-eligible prisoner comes before a board that has the inmate's file for review. In addition to that file, the parole board can direct questions to the prisoner about, for example, their adjustment to prison life, their remorse, and their positive activities while incarcerated. They then form a judgment about whether the prisoner merits parole. We want to release parole-eligible inmates, but keep incarcerated those who are likely to commit again. Satisfying these twin goals is usually the job of a parole board. But at what cost?

At least 175,000 parolees on the streets now were deposited by the unsteady, inferior judgment of parole boards. At the same time, prisons are full of parole-eligible, expensive inmates whose low risk was missed by parole boards. According to a National Research Council report, parole is at least $18,000 per year less than housing a parole-eligible prisoner. How much money can be saved depends on how many low-risk inmates denied by parole boards would be identified for release by predictive tools.

These predictive tools have a long pedigree, and by using them the Commission that Senators Webb and Specter propose could save money while finding a more accurate balance of community safety and proper punishment. For decades we have known that much behavior - from college success and cancer treatment outcomes to post parole violence - is better predicted by simple formula than by human "experts". (Pioneered by philosopher and psychologist Paul Meehl, these techniques were made famous in the bestseller Moneyball. My colleague Michael Bishop and I explored their potential contribution to philosophy, in Epistemology and the Psychology of Human Judgment, and I work out their promise for public policy in The Empathy Gap - in a chapter called "Stat vs. Gut".) During this time, crime researchers developed tools to identify parole-eligible inmates at risk to commit again if released. Their accuracy consists in distinguishing the likely recidivists from the harmless inmate, by creating a model that includes variables known to be correlated with recidivism for particular crimes - like behavior while incarcerated, age at first crime, and gender of victim, to name a few.

These tools amount to microsurgery in social science, but for some reason the parole system prefers mangling patients with triage. We have superior predictive tools, so why don't we insist that parole boards use them? The best guess is that everyone thinks the "personal touch" of the parole board is a kind of ineffable expertise. Like the winetaster's unchallengable skill, the parole board member merits expert status as well. Unfortunately, we just aren't very good at identifying genuine experts. Just as wine experts fall to these predictive models, so too do parole boards. When pitted against expert judges, predictive models win hands down.

In the next post, I will complete the story of how these predictive tools win, and what we could expect if we actually used them.