Risk-assessment has become all the rage in American criminal justice. In jurisdictions across the country, criminal-justice officials are utilizing increasingly sophisticated risk-assessment tools, which can be used to predict a given offender’s likelihood to reoffend based on his criminal history and a number of other variables. These predictive evaluations can be brought to bear at several important decisional points in the criminal process: pretrial release, diversion into treatment, sentencing, and others.
Although risk assessment has been widely applauded for its potential to support increased efficiency in the use of scarce criminal-justice resources, a recurring criticism has been that leading risk-assessment tools have built-in racial biases. A particular concern has been the heavy reliance on criminal history; to the extent that criminal history reflects biased actions by police or others in the past, then predictions based on that history may tend to overestimate the relative risk posed by minority defendants. Thus, for instance, a black defendant and a white defendant whose actual risk levels are identical could potentially receive quite different risk scores, leading to quite different bail or sentencing decisions.
Such concerns find some support in the empirical research.
A new study, however, reaches more reassuring conclusions, at least with respect to one risk-assessment tool used in federal court.
In their paper, Jennifer Skeem and Christopher Lowenkamp assess how successfully the Post Conviction Risk Assessment (PCRA) tool predicted the future offending of a sample of more than 30,000 federal offenders. Of course, no prediction about something so mercurial as human behavior will ever be 100% accurate. Still, Skeem and Lowenkamp found substantial differences in the rearrest rates of offenders in each the four different risk categories. For instance, only 11% of those categorized as low risk were rearrested, as compared to 64% of those categorized as high risk.
Nor did these results change much when offenders were broken out by race. Low-risk blacks, for instance, were rearrested 12% of the time, while the figure for low-risk whites was 10%. Similarly, the rearrest rate for high-risk blacks was 62%, while the rate for high-risk whites was 66%.
Skeem and Lowenkamp also checked more narrowly for violent rearrests — presumably the most pressing sort of recidivism risk. High-risk blacks were rearrested 23% of the time for violent offenses, while high-risk whites were rearrested for such offenses only 19% of the time.
There is no clear pattern of racial bias in any of this. For instance, based on the data on any rearrest, it does appear that there was a slightly higher likelihood of a black offender being erroneously classified as high risk. However, based on the violent rearrest data, it was the white offenders who were slightly more likely to be misclassified as high risk.
Does this mean risk assessment is vindicated?
A few questions come to mind. First, to what extent are these results specific to the PCRA tool and/or the federal offender population? One should note that the federal criminal docket is quite distinctive, weighted much more heavily to certain offenses (drugs, guns, and immigration) than are state systems. This should make us cautious about generalizing from studies focused on the federal system.
Second, how well do rearrest rates correlate with actual reoffending? Skeem and Lowenkamp use rearrest rates as a proxy for reoffending, but the sort of criminal history that lands one into a high-risk category is also likely to make one a “usual suspect” for the police. Biases against the individuals in the high-risk category may increase their risk of mistaken arrest. Race may exacerbate that risk for some. For these reasons, we should be careful about assuming that rearrest necessarily implies reoffense.
Third, to what extent (if any) does higher reoffending by those classified as high risk result from that very classification? The whole point of classification is to treat offenders differently on the basis of their assessed risk. Yet, the different treatment given to high-risk offenders could conceivably make them more dangerous. For instance, it is possible that more intensive monitoring by probation agents could be stigmatizing and disruptive to the offender’s efforts to obtain stable housing and employment, and to develop positive social relationships. The high-risk label could, in effect, become a self-fulfilling prophecy for some offenders.
Fourth, how much of the reoffending involves the sort of major, predatory crime that really drives the whole risk-control project? I am glad that Skeem and Lowenkamp, unlike many other recidivism researchers, focus on rearrests for violent crime. Many other studies report higher, and much more alarming, recidivism rates, but without any indication of what proportion of the rearrests were for nuisance-type crimes. Distinguishing violent from nonviolent crimes is helpful, but there remains considerable variation within the “violent” category. Ideally, we would further distinguish low-level scuffles and the like from the much more serious offenses.
Finally, even assuming reliable, unbiased risk assessment, what are the ethical limits on the use of necessarily imperfect predictive tools? Consider the violent rearrest rate of those categorized by the PCRA as high risk: 21%. How much loss of liberty is justified by that risk level? Most of the time — indeed, arguably more than three-quarters of the time, if one takes the data at face value — our fear of the high-risk offender seems misplaced. Such odds do not come close to the level of certainty (beyond a reasonable doubt) required to take a person’s liberty away in the first instance. Once a person has been convicted, though, does a 21% certainty as to future violence become a sufficient basis to add time to a person’s prison term? If so, how much time? And if not, could those odds at least justify a significantly longer and more onerous period of community supervision?
Such ethical questions are not easy, but we need to be grappling with them — and also extending and refining the sort of empirical research exemplified by the Skeem-Lowenkamp paper — as the criminal-justice system increasingly reorients itself around the new risk-assessment tools.
The new paper is “Risk, Race, and Recidivism: Predictive Bias and Disparate Impact,” 54 Criminology 680 (2016).
Cross posted at Life Sentences.