Comparing Meters to Yards: A Nationally Representative Evaluation of Gender Bias in Risk Assessment

Risk-needs assessments (RNAs) assist correctional staff in assigning supervision and programming. While gender is a well-known predictor of crime, for decades contemporary RNAs have claimed “gender-neutrality” or risk prediction equality for males and females. Unfortunately, females are frequently overclassified, relegated to a category higher than their risks indicate. Using ridge and mixed effects regression methods, we sought to enhance the predictive accuracy of the Modified Positive Achievement Change Tool (MPACT) comparing three common assessment development methods via a 10-state sample of youth (<i>N</i> = 241,596) across multiple justice settings. Findings demonstrated recidivism rates vary substantially by gender and bias/overclassification is likely for assessments using a “gender-neutral” approach. Tools that oversample females, or attempt “equally weighting,” have similar issues. However, “gender-specific” methods create prediction parity, rooting out overclassification. Study takeaways include the need to evaluate tools for overclassification, methods of eliminating gender bias, while achieving strong predictive validity in development of the MPACT.

风险需求评估（Risk-needs assessments，RNAs）可协助惩教人员制定监管方案与处遇计划。尽管性别是公认的犯罪预测因子，但数十年来，当代RNAs均宣称自身具备‘性别中立性’，或可实现针对男性与女性的风险预测平等性。但遗憾的是，女性群体常被分类过高，即被划入高于其实际风险等级的类别中。本研究采用岭回归与混合效应回归方法，旨在提升改良式积极成就变化量表（Modified Positive Achievement Change Tool，MPACT）的预测准确性；我们通过覆盖多司法场景的10个州的青少年样本（N=241,596），对三种常见的评估开发方法进行了比较。研究结果显示，累犯率因性别存在显著差异，采用‘性别中立’方法的评估工具大概率存在偏差与分类过高问题。过度抽取女性样本或试图采用‘均等加权’的工具也存在类似问题。但‘性别特定’方法可实现预测均等性，彻底消除分类过高的问题。本研究的核心启示包括：在开发MPACT的过程中，需评估工具是否存在分类过高问题，探索消除性别偏差的方法，同时保障较强的预测效度。