The Effects of Multicalibration on Fairness Violations in Personalized Prediction Models

Abstract

Personalized clinical prediction models can introduce preference-based fairness violations: some groups may perform worse under personalization than under a generic model (rationality violations), or benefit from misreporting their group membership (envy-freeness violations). This study investigates whether multicalibration post-processing can serve as a proactive mechanism for reducing such violations, across three clinical datasets, two calibration holdout strategies, three personalization strategies, three model architectures, and four training pipeline configurations. The answer is conditional. Multicalibration succeeds only under a Mixed balancing configuration (training with joint group-label balance while calibrating on natural class prevalences), which gives the HKRR algorithm an unconfounded, group-specific miscalibration signal. Under this configuration, ECE and error rationality violations are eliminated on the Mortality dataset and envy-freeness violations decline substantially on Apnea. Under other configurations, label skew either confounds the calibration signal or has already been removed by construction, leaving nothing for HKRR to correct. Training balance is a precondition for multicalibration to function, not an interchangeable alternative; this sensitivity has not been documented in prior empirical evaluations. Beyond pipeline configuration, architecture matters: logistic regression and neural networks both benefit, while random forests do not improve on Mortality, where severe class imbalance leaves insufficient signal for HKRR, exacerbated by RF’s discrete outputs reducing correction precision. Random forests do improve on more balanced datasets. Across metrics, ECE improves most directly; accuracy improves only when corrections cross the decision threshold. AUC exhibits an unexpected divergence: multicalibration degrades AUC rationality because HKRR’s bin-level corrections disturb within-group rankings, while AUC envy-freeness improves because the same corrections compress between-group score gaps. This divergence has not been previously characterized.