An Open Source Replication of a Winning Recidivism Prediction Model

Area of Science:

• Criminology
• Machine Learning
• Data Science

Background:

• Recidivism forecasting is crucial for criminal justice.
• Ensuring fairness in predictive models across racial groups is a significant challenge.
• The National Institute of Justice (NIJ) sponsored a challenge to address these issues.

Purpose of the Study:

• To present the winning solution for the NIJ recidivism forecasting challenge.
• To evaluate model accuracy using the Brier score and fairness using racial false positive rates.
• To explore different machine learning model specifications and fairness metric adjustments.

Main Methods:

• Utilized XGBoost, a non-linear machine learning model.
• Investigated various model specifications to assess predictive performance.
• Implemented a simple bias strategy to equalize false positive rates across racial groups.

Main Results:

• Achieved high accuracy as measured by the Brier score.
• The proposed method resulted in zero false positive rates for both White and Black parolees.
• Demonstrated that different models often yield similar predictive performance.

Conclusions:

• The winning solution balanced accuracy and fairness criteria effectively.
• A trivial bias adjustment can eliminate racial disparities in false positive rates.
• Open-source materials are provided to enable replication and further research in recidivism forecasting.