Data-driven decision making is increasingly being used in policy and business applications. Prediction models are helping determine who to release on bail or parole, who to hire or recruit, and who is eligible for a loan. But just because a statistical model is guiding the decision, doesn't guarantee that the outcome will not discriminate against certain demographic groups. It's important to understand how discriminatory forms of model bias can arise, how they can be detected, and how we can develop methods that are fair by design.

This talk explores questions of bias and disparate impact in the context of recidivism risk prediction. Recidivism prediction instruments provide decision makers with an assessment of the likelihood that a criminal defendant will reoffend at a future point in time. While such instruments are gaining increasing popularity across the country, their use is attracting tremendous controversy. Much of the controversy concerns potential discriminatory bias in the risk assessments that are produced. In this talk we describe several fairness criteria that have recently been applied to assess the fairness of recidivism prediction instruments. We demonstrate that the criteria cannot all be simultaneously satisfied when recidivism prevalence differs across groups. We then show how disparate impact can arise when a recidivism prediction instrument fails to satisfy the criterion of error rate balance.