Analyses of Timing and Type of Rearrest: Comparing Semi-Parametric and Machine Learning Survival and Classification Models

• In Event: Machine Learning Applications in Criminology & Criminal Justice

Abstract

Studies of recidivism often seek to balance including information on the time to and type of rearrest. This study follows a cohort of 87,000 active probationers and parolees in Georgia between 2016 and 2019. Using the National Corrections Reporting Program charge categories, we compared specifications of first rearrest by (1) any arrest and (2) rank order of the most serious charge as violent, property, drug, public order, or probation/parole arrests. This analysis tests the utility of static predictors, including time- and charge-specific criminal history measures, and dynamic protective and risk factors captured while under supervision. We examine arrest using survival models as well as through period-specific classification models to assess changes in risk profiles over time. Survival analyses compare semi-parametric statistical models to machine learning (ML) implementations (tree-based and deep survival learning models). Classification analyses compare logistic regression to ML classifiers to gauge differences in accuracy. This analysis also explores class weights and different sampling techniques to handle class imbalance. Results suggest minimal differences in AUC or concordance between semi-parametric models and ML techniques. Models incorporating dynamic measures perform better than static-only specifications. This study provides a framework for comparing different classification and survival implementations in the analysis of rearrest.

Authors

• Stephen Tueller, RTI International

• Christopher Inkpen, RTI International

• Pamela K. Lattimore, RTI International

• Rainer Hilscher, RTI International

• Caroline Kery, RTI International

• Emily Burtch, RTI International