AI R&D to Support Community Supervision: Integrated Dynamic Risk Assessment for Community Supervision (IDRACS), Final Report

• NCJ Number: 309339
• Author(s): Pamela K. Lattimore; Christopher Inkpen
• Date Published: 2024
• Length: 140 pages
• Grant Number(s): 2019-75-CX-0012
• Agencies: NIJ-Sponsored

Annotation

This publication provides results from the Integrated Dynamic Risk Assessment for Community Supervision (IDRACS) project.

Abstract

This report describes findings from the Integrated Dynamic Risk Assessment for Community Supervision (IDRACS) project, a study led by RTI International in collaboration with Applied Research Services, Inc. (ARS) and the Georgia Department of Community Supervision (DCS). The outcome of this project was a suite of predictive algorithms and data management processes that will supplement DCS’s supervision practices, allowing for accurate and time-specific predictions of the risk of felony or violent rearrest or revocation. The authors produced separate models stratified by biological sex and supervision type (straight or split probation and parole). The project team found that including detail on the nature and timing of the underlying criminal history produced more accurate results compared to models that used broad lifetime criminal histories. Furthermore, applying feature selection algorithms suggested that omitting arrests that occurred 5 years before the start of supervision did not worsen model accuracy. Tests of including dynamic measures revealed substantial gains in model accuracy. Additionally, period-specific models proved most accurate for predictions in the first year of supervision. Applying machine learning techniques revealed that while these models sometimes produced modest improvements in accuracy, they were often not significantly or substantively different in contrast to the tradeoffs in model interpretation and ease of implementation when compared to traditional statistical models (i.e., logistic regression). The RTI team also developed and implemented a process that entailed bootstrapping predictions to create confidence intervals around individual predictions, incorporating uncertainty into one’s predicted probability of rearrest.