Paper Abstract
85. Alex Albright, If You Give a Judge a Risk Score: Evidence from Kentucky Bail Decisions, 05/2019.
Abstract: High-stakes decisions are increasingly informed by predictive tools. Many assume that these tools should reduce disparities across groups by limiting human discretion, but empirical evidence on this is lacking. In this paper, I outline how interactions between prediction tool recommendations and human discretion can actually exacerbate disparities across groups. In particular, I discuss a policy change in Kentucky that set a recommended default for judge bail decisions based on risk scores. Counter to expectations, the policy caused an increase in raw racial disparities in initial bond, first illustrated by Stevenson (2017). Using case-level data, I show that this increase was not simply a consequence of different risk scores by race. Rather, the recommended default was also more likely to be overridden (in favor of harsher bond conditions) for black defendants than similar white defendants. I discuss two forces behind this result. First, judges varied in their policy responsiveness; judges in whiter counties responded more to the new default than judges in blacker counties. Second, even within judge and time, judges were more likely to override the recommended default for moderate risk black defendants than similar moderate risk white defendants. This result suggests that interaction with the same predictive score may lead to different predictions by race.