Probability Models of Recidivism: An Exploration

• NCJ Number: 107806
• Author(s): R L Linster; E B Patterson
• Date Published: 1987
• Length: 98 pages
• Grant Number(s): 84-IJ-CX-K046
• Agencies: NIJ-Sponsored

Annotation

This paper explores empirically the notion that selected actuarial models can plausibly uncover individual recidivism probabilities on the basis of information about the offender and his or her history and that such models have predictive validity at least under a split-sample test.

Abstract

Focus is on a comparison of the static logit model with a time-dependent hazard model. These models are applied to data from a 2-year followup study of male inmates released from prison in North Carolina and a 10-year study of criminal justice contacts following parole among subjects from a California Youth Authority institution. Results of the analyses clearly support recidivism as a process in which chance plays an essential role. The variables and models used here assigned most subjects to failure probabilities that lie in a middle range, resulting in substantial prediction error. While the logit model was found to perform somewhat better in assigning individual probabilities of failure within some fixed period of time, hazard models generally possess practical and analytical advantages. Additional statistical data are appended. 26 tables and 23 references.