Differential Prediction by Race in IRAS-PAT Assessments: An Application of Debiasing Strategies

Abstract

Abstract There remain serious concerns about the potential for pretrial risk assessments to exacerbate racial disparities. Yet, current evidence on differential prediction in pretrial risk assessments is limited. The present investigation tests for differential prediction by race as an indication of bias in Indiana Risk Assessment System–Pretrial Assessment Tool (IRAS-PAT) assessments. Using pooled data drawn from a five-county IRAS-PAT validation, which included 689 Black and 2,850 White defendants, we primarily used a hierarchical regression-based approach to test between-group differences in the slopes of regression lines. Where slope-based bias was present, differential prediction was reevaluated once algorithmic corrections were applied. Findings showed IRAS-PAT assessments produced less accurate predictions of pretrial misconduct for Black defendants relative to White defendants. Only one debiasing strategy—which accounted for item-level differences across subgroups—corrected differential prediction. Debiasing strategies can mitigate differential prediction but may have limited utility for local jurisdictions under current legal frameworks.