Participating in the NIJ Recidivism Forecast Challenge enabled SAS Institute to showcase the performance of its analytical subject-matter expertise in helping to identify the key components of recidivism, notably in providing strategies to account for racial bias in recidivism calculations.
In this report, SAS describes its process to account for racial bias in recidivism forecasting in the Challenge’s year 3 male parolees. SAS built models using a data-mining approach with both SAS and open-source software. SAS trained several machine learning algorithms to derive measures of variable importance. SAS found leakage in the data, which impacted its ability to make significant conclusions. Suggestions are provided for ensuring model accuracy by group (e.g., race, ethnicity, educational attainment, and gender), in which metrics could include, but are not limited to, equalized odds, false positive/negative parity, and truer positive/negative parity. This report advises that an approach for creating meaningful bias-assessment threshold is to use the marginal recidivism rate for a given year as a classification cut-off value. The practical interpretation of these threshold values is to classify a case as “risky” when there is a higher-than- average chance to recidivate.