The study examines a two-year recidivism rate among 6,172 offenders released in a U.S. state, finding it to be around 45 percent. A detailed analysis shows a significant age-related trend: those under 25 have a recidivism rate of 58 percent, while those 45 and older see a rate of 36 percent. The distribution of prior convictions is skewed to the right, with an average of 3.2, and nearly 25 percent of the sample had five or more previous convictions. Despite the sample being predominantly male (81 percent) and African-American (51%), data integrity is maintained with minimal missing values and outliers, providing a solid basis for further analysis. In comparing logistic regression, decision tree, and polynomial logistic regression, all methods achieve predictive accuracies ranging from 66.7 to 68.8 percent on the test set, with an AUC of approximately 0.72, which is notably higher than the 55 percent baseline. The classification tree, optimized with a maximum depth of five, entropy splitting, and 70 percent random feature sampling, strikes a balance between accuracy and interpretability. Both logistic and polynomial models provide consistent coefficients for the number of priors, age categories, and their interactions. Notably, the impact of prior convictions increases nonlinearly; offenders with five or more priors have an odds ratio six times that of first-time offenders. The findings consistently show a positive effect for the under-25 age group and a negative effect for the over-45 age group, supporting life-course theory. Economically, the study suggests that recidivism risk is largely influenced by the accumulation of crime-specific human capital and the varying opportunity costs associated with age, rather than demographic labels alone. By integrating age and prior convictions into sentencing support tools, high-risk cases can be identified without relying on race. This approach offers judges transparent and auditable probability estimates, directing limited correctional resource.
Lu et al. (Wed,) studied this question.