Classifying Adult Probationers by Forecasting Future Offending

• NCJ Number: 238082
• Author(s): Geoffrey C. Barnes Ph.D.; Jordan M. Hyatt, J.D., M.S.
• Date Published: March 2012
• Length: 64 pages
• Grant Number(s): 2008-IJ-CX-0024
• Agencies: NIJ-Sponsored
• Publication Type: Report (Study/Research)
• Award: This project will test, with a randomized controlled trial, an intensive community supervision strategy for probationers and parolees who are at extremely high risk of committing homicidal acts.

Annotation

This report presents results from a demonstration project focused on the development of a risk-prediction model for a local probation and parole department in a large urban city.

Abstract

Study results demonstrate that the project's risk-forecasting models were able to increase the Philadelphia Adult Probation and Parole Department's (APPD's) ability to predict recidivism, leading to the restructuring of agency supervision protocols. The project resulted in the construction of three different prediction models based on a statistical process known as "random forest." One benefit of random forest modeling is that there is no theoretical limit on the number of predictors that can be included in the model. Throughout the duration of the project, hundreds of different predictor variables were drawn from electronically available administrative records, and they were tested for possible use in these models. The most recent version of APPD's model produced an accurate forecast for 79,299 of the 119,935 probation case starts in the construction sample. These estimates suggest that this model can be correct nearly two-thirds (66.1 percent) of the time; however, a more reasonable method of measuring the model's accuracy is to re-examine forecasted and actual outcomes separately, focusing on each of the three different outcome categories. The power and promise of the random forest forecasting methods is clear in Philadelphia; their introduction has allowed the agency to stratify offenders by the risk they pose, to tailor supervision requirements, to focus resources in accordance with policy directives, and to balance caseload sizes in the face of budgetary constraints. Twelve recommendations for building random forest prediction models in any jurisdiction are offered. Figures, tables, and references