Beyond Outcome Disparities: A Sensitivity Framework for Auditing Algorithmic Fairness

Algorithmic decision systems increasingly govern high-stakes domains, yet traditional fairness metrics reveal disparities without explaining why or which features drive them. We introduce optimal-transportbased global sensitivity analysis (OT-GSA) as a structural diagnostic framework that bridges quantitative sensitivity analysis and critical algorithm studies. Our four-dimensional ethical taxonomy-magnitude, asymmetry, justification, and stability-translates sensitivity indices into actionable fairness interventions.

Analysis of Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) recidivism prediction reveals criminal history features exhibit structural amplification (OT index = 0.34), while age demonstrates extreme asymmetry (∆S = 0.25) across racial groups-indicating fundamentally different decision rules violating equal treatment principles. Counterfactual analysis shows bias propagates through correlated feature families, requiring holistic redesign beyond simple feature removal. Longitudinal analysis of Center for Medicare & Medicaid Services (CMS) Hospital Readmissions Reduction Program (HRRP) reveals a novel fairness violation: measurement system shock.

During the 21st Century Cures Act implementation, all features simultaneously inflated in importance (+270% to +4,200%), with geographic region exhibiting 170-fold variation across policy eras (temporal variance σ 2 = 0.07). Counterfactual validation confirmed removing epistemically fragile features reduced temporal variance by 17%, while stabilizing feedback loops paradoxically increased instability by 13%demonstrating structural policy redesign, not engineering fixes, is required. Our taxonomy demonstrates algorithmic fairness requires multi-dimensional assessment: epistemic robustness across contexts constitutes an independent fairness requirement beyond demographic parity or calibration.

When measurement systems become unpredictable during policy transitions, even wellcalibrated algorithms violate procedural fairness. The framework provides tools to move from detecting disparities to diagnosing transmission mechanisms and prescribing targeted interventions-essential for accountability as automated systems increasingly govern consequential domains.