Predictive Bias in Juvenile Risk Assessment: Considering Race/Ethnicity and Sex

Abstract

This study examines the joint effects of race/ethnicity and sex on the classification quality of the Positive Achievement Change Tool (PACT). We assess predictive parity, calibration, and error rate balance using a statewide sample of youth who completed probation from 2007 to 2014. Similar rates of predictive parity were observed across groups. Calibration was observed consistently across groups, but only when low risk youth were used as the reference category. Black and Hispanic boys revealed the highest base rates of re-arrest and inaccurate classification (i.e., scored low risk but were rearrested). Our findings highlight the difficulty in making conclusions about test bias, especially when base rates of recidivism are not equal across groups. Implications for juvenile justice policy and practice are discussed.