Cliff-Aware Fairness for Algorithmic Public-Benefit Decisions

Machine learning and automated decision-making systems are increasingly used in public administration for high-impact decisions on access to public benefits. These programs often rely on sharp eligibility thresholds that create “benefit cliffs,” where small changes in applicant characteristics can trigger substantial losses in assistance. Because errors near these policy boundaries can carry greater economic consequences than errors elsewhere, population-level fairness metrics can fail to capture localized disparities. We introduce Cliff-Aware Fairness (CAF), a threshold-local framework for evaluating subgroup-specific errors near public-benefit eligibility boundaries. We further propose the Cliff-Weighted Fairness Gap (CWFG), which weights near-threshold errors by potential benefit loss. Using childcare-benefit eligibility as a case study, we evaluate multiple predictive models and fairness-aware methods across conventional and threshold-local fairness metrics. Our analysis tests whether models that appear fair globally exhibit larger subgroup disparities near benefit cliffs and whether accounting for economic consequence changes model fairness assessments.