OB Brownbag - Jennifer Doleac, Texas A&M University
Algorithmic Risk Assessment in the Hands of Humans
Thursday 3 October 2019 (12h00 - 13h00) - Extranef - 126
We evaluate how adopting risk assessment tools (algorithmic predictions of future offending) affects sentencing, recidivism and race/age disparities for felony offenders. We find that scoring right above the “low-risk" cutoff increases the likelihood of incarceration by 6-7 percentage points and increases the sentence length by approximately 23-34%, depending on the crime of conviction. Using a difference-in-difference framework, with defendants who were ineligible for risk assessment as a control, we find no evidence that the adoption of risk assessment affected average sentencing for nonviolent offenders: to whatever extent sentences decreased for lower risk defendants this was counterbalanced by an increase for those with higher risk scores -- in particular, for young defendants. We find little evidence that risk assessment led to a decline in recidivism, and explore several potential reasons why not. Our results on racial disparities in sentencing are mixed. Statewide, we find no evidence that risk assessment affected racial disparities. However, racial disparities increased after the adoption of risk assessment in the subset of judicial circuits that appear to use risk assessment most.