Predicting Recidivism Fairly: A Machine Learning Application Using Contextual and Individual Data

• In Event: Lightning Talk: Findings and Insights from the National Institute of Justice Recidivism Forecasting Challenge

Abstract

Community corrections agencies across the U.S. face daunting challenges. Too often, caseloads are large and resources are thin. Services provided must balance surveillance and support programming. Too much of a poorly targeted intervention or too little of an urgently needed program will do little to improve reintegration and reduce recidivism, and may even make matters worse. Predicting recidivism risk accurately and fairly is therefore critical to effective community corrections, release, and programming decisions. In an effort to improve the accuracy and fairness of risk assessments, we responded to NIJ’s Recidivism Forecasting Challenge by utilizing additional contextual factors characterizing the communities where supervised individuals live. To accomplish this, we integrated a range of contextual metrics with the individual-level Georgia parolee dataset supplied by NIJ, including community-level measures of environment and health, housing, economic opportunity, quality of life, and social and government services. We then assessed and validated various machine learning algorithms using both individual and contextual factors to predict recidivism outcomes. In our validation efforts, we focused on maximizing predictive accuracy across gender and race according to NIJ’s judgement criteria. We found that a lasso model performed best overall. In this presentation, we describe our methodological approach and results.

Authors

• Eric Sevigny, Georgia State University

• Thaddeus Johnson, Georgia State University

• Jared Greathouse, Georgia State University