Which Demographic Features Are Relevant for Individual Fairness Evaluation of U.S. Recidivism Risk Assessment Tools?

TL;DR

The paper investigates how to operationalize individual fairness in U.S. recidivism risk assessment tools by identifying which demographic features should be included in the individual similarity function. It employs a between-subjects human-subject study with four conditions to test whether including Race, Sex, or Age in the similarity function affects lay fairness judgments, formalizing the notion that pairs with $S=high$ and $O=diff$ should yield $J=unfair$ if the similarity function is good. Results indicate that Race and Sex are often ignored by participants, while Age yields nuanced effects; no statistically significant differences emerge across conditions after corrections, though qualitative analyses show Age prompting more detailed justifications. The study informs legal and policy considerations for fairness in RRA tools, arguing for including Age and Sex but omitting Race in the demographic inputs used by the individual similarity function, with practical implications for regulatory criteria and tool design.

Abstract

Despite its constitutional relevance, the technical ``individual fairness'' criterion has not been operationalized in U.S. state or federal statutes/regulations. We conduct a human subjects experiment to address this gap, evaluating which demographic features are relevant for individual fairness evaluation of recidivism risk assessment (RRA) tools. Our analyses conclude that the individual similarity function should consider age and sex, but it should ignore race.

Figures (1)

• Figure 1: Example interface for the main human study task (Condition: Treatment with Race)