Using Algorithms to Address Trade-Offs Inherent in Predicting Recidivism

Abstract

Although risk assessment has increasingly been used as a tool to help reform the criminal justice system, some stakeholders are adamantly opposed to using algorithms. The principal concern is that any benefits achieved by safely reducing rates of incarceration will be offset by costs to racial justice claimed to be inherent in the algorithms themselves. But fairness tradeoffs are inherent to the task of predicting recidivism, whether the prediction is made by an algorithm or human. Based on a matched sample of 67,784 Black and White federal supervisees assessed with the Post Conviction Risk Assessment (PCRA), we compare how three alternative strategies for “debiasing” algorithms affect these tradeoffs, using arrest for a violent crime as the criterion. These candidate algorithms all strongly predict violent re-offending (AUCs=.71-72), but vary in their association with race (r= .00-.21) and shift tradeoffs between balance in positive predictive value and false positive rates. Providing algorithms with access to race (rather than omitting race or ‘blinding’ its effects) can maximize calibration and minimize imbalanced error rates. Implications for policymakers with value preferences for efficiency vs. equity are discussed.

Keywords: risk assessment, algorithm, race, fairness, bias