This is “Team Sherrill’s” report in its participation in the National Institute of Justice’s (NIJ’s) Recidivism Forecasting Challenge, which sought to identify promising approaches to predicting recidivism for a sample of people on parole in Georgia.
The performance metrics for the competition prioritized both accuracy and reduced racial bias for awarding winners. Team Sherrill used deep neural network (DNN) machine learning to build models for reducing recidivism with the provided data. This approach was selected because neural networks learn from the data provided in contrast to being limited to the team’s existing knowledge, and it can be trained to minimize cross-entropy and other selected loss functions, enabling the modeling of complex and nonlinear relationships. They also facilitate building models with interaction effects such as those related to gender and racial differences. To assess model performance, the team used a few different metrics. Validity was measured using the Brier score. Racial bias was assessed through a “fair and accurate” method that accounted for disparate false positive rates between races, as well as overall model accuracy.
Downloads
Similar Publications
- Social Support, Victimization, and Stress in a Women’s Prison: The Role of in-Prison Friendship for Reducing Perceptions of Stress
- Racial Politics in the Contemporary Prison Society: The Importance of Race and Ethnicity to Prison Social Organization
- Relations between youths’ community violence exposure and their physical aggression: The protective role of adults