Optimizing predictive performance of criminal recidivism models using registration data with binary and survival...

Area of Science:

• Criminology
• Statistics
• Machine Learning

Background:

• No consensus exists on optimal modeling strategies for recidivism prediction.
• Previous studies benchmarked various techniques but underexplored tree ensemble methods like gradient boosting and random forests.

Purpose of the Study:

• To evaluate gradient boosting and random forests for binary criminal prediction.
• To assess statistical and machine learning methods for censored time-to-event recidivism data.
• To compare model performance across Dutch and American datasets.

Main Methods:

• Fitted manually specified statistical and semi-automatic machine learning models.
• Applied models to Dutch recidivism data and two American datasets (North Carolina prison data).
• Evaluated models for both binary and censored time-to-event outcomes.

Main Results:

• Semi-automatic modeling for binary outcomes showed no improvement over traditional statistical models.
• Gradient boosting demonstrated slightly improved performance for survival data (time-to-event).
• Model comparison across datasets suggests a need to explore both statistical and machine learning approaches.

Conclusions:

• Flexible machine learning models can validate traditional models by identifying potential misspecifications or missing interactions.
• Gradient boosting shows promise for survival analysis in recidivism prediction.
• A combined approach of statistical and machine learning methods is advisable for robust recidivism prediction.