The utility of modeling hazard rates associated with recidivis

This study measures recidivism rates through the use of arrest data. The focus is to model the arrest process of a large cohort of New Jersey offenders using hazard rates. Nonparametric estimating techniques and parametric distributional forms were used to predict the timing of arrests related to a group of offenders. Ordinary Least Squares and an Accelerated Failure Time model (i.e., Lifereg procedure) were used to consider the effects of individual covariates on the timing of arrest. Participation in offending behavior was also measured by arrest and was studied through multivariate Logistic and Probit regressions.;The effects of model specification, sample size, the effects of censored data, duration of follow-up time required, and whether past criminal histories can be used to predict future behavior were studied. This study shows that estimates of the timing of recidivism, for the population as a whole, were consistent over different time periods using the Lifereg model but not the OLS model. Further, the Logistic and the Probit models did not produce estimates of participation that were consistent for both retrospective and prospective time periods. Since event history analysis has been advocated as preferable to other methods (Hagan and Palloni, 1988), clarification of issues studied in this dissertation has important statistical and practical implications. This study shows that ten percent samples may seriously inflate the variability of parameter estimates, even when the estimates are based on samples of nearly one-thousand individuals. Additionally, censored data significantly affect distributional predictions of the cumulative density function when using lognormal and exponential distributions. A method to correct for censored data is offered for the lognormal distribution with mixed results.