Home | Publications | BB26a

Rank-Preserving Causal Fairness Under Unobserved Mediators

MCML Authors

Abstract

Rankings in admissions, lending, and hiring often inherit historical inequalities through causal pathways that are only partly observed. Existing counterfactual-based fairness methods can mitigate discriminatory mechanisms while preserving individual task-specific merit, assuming all variables are observed. However, assuming full visibility is fragile since mediating variables like opportunity, preparation, or task-specific skill are usually latent. We thus study the estimation of unit-level counterfactuals under unobserved mediators. We show that fairness methods that adapt the outcome while bypassing the mediator can remove group-level disparities while failing to recover the unit-level counterfactual, which may lead to unfair ranking of individuals. We propose Causal Mediator Flows (CMF), an autoencoder-style normalizing-flow model that infers a latent mediator representation from available proxies and uses it to emulate edge-wise counterfactual transport. Preliminary simulations and a law-school ranking study suggest that explicitly modeling the hidden mediator improves the reconstruction of the unit-level counterfactual<br>fair ranking and provides useful uncertainty diagnostics for practitioners.

inproceedings BB26a

ECAF 2026

5th European Conference on Algorithmic Fairness. Ghent, Belgium, Sep 02-04, 2026. Spotlight Presentation. To be published.

Authors

P. A. Boustani • L. Bothmann

Research Area

 A1 | Statistical Foundations & Explainability

BibTeXKey: BB26a

#p-bischl
Back to Top