Reducing Incorrect Classification: Risk Assessment from a Bayesian Perspective

• In Event: Criminal Justice Policy and Institutionalized Racial/Ethnic Discrimination

Abstract

Actuarial assessment has become commonplace in the criminal legal system, with risk classifications having wide-ranging implications for the institutional placement of offenders, allocation of services, and safety of the public. Despite the importance of making reliable predictions about future recidivism risk, these instruments (e.g. LSI-R) tend to perform only slightly better than chance, fail to account for gender differences in offending, and systematically over-classify risk among people of color—calling into question the widely noted “objectivity” of actuarial tools. Furthermore, ideal instruments should be flexible to varying influences of risk factors on recidivism across time and place, so classification can reflect local characteristics of offending. To address these concerns, Washington State DOC records are used to construct the first actuarial assessment tool based on Bayesian, rather than Frequentist, methods. Briefly, these models make use of historically informative priors, taken from earlier cohorts of recidivists, to “learn” which factors influenced actual re-offending and desistance behaviors. The resulting posterior estimates of gender-specific models are then fed into, and updated, using data from a newer cohort of offenders to construct the final instrument. Findings highlight the utility of Bayesian methods for prediction, with particular emphasis on reducing incorrect classification by race and gender.

Author

• Philip Pendergast, University of Colorado Boulder