Local Causal Discovery for Structural Evidence of Direct Discrimination
Jacqueline Maasch, Kyra Gan, Violet Chen, Agni Orfanoudaki, Nil-Jana Akpinar, Fei Wang
TL;DR
The paper tackles the challenge of identifying direct discrimination in complex domains without full causal graphs by introducing LD3, a local causal discovery method that targets the causal parents of the outcome and yields a valid adjustment set for the weighted controlled direct effect. It embeds LD3 within a CFA framework using a graphical interpretation of causal partitions and proves asymptotic correctness under mild assumptions, while maintaining computational efficiency with $O(|oldsymbol{Z}|)$ CI tests. The authors introduce a graphical criterion for WCDE, enabling direct discrimination assessment through the derived adjustment set $oldsymbol{A}_{ ext{DE}}$, and demonstrate that LD3 provides more stable and interpretable results than global baselines across synthetic benchmarks and two real-world cases (COMPAS recidivism and liver transplant allocation). The practical impact lies in offering a scalable, interpretable, and statistically sound tool for policy analysis and algorithmic fairness, enabling targeted interventions based on the detected direct mechanisms of unfairness.
Abstract
Identifying the causal pathways of unfairness is a critical objective for improving policy design and algorithmic decision-making. Prior work in causal fairness analysis often requires knowledge of the causal graph, hindering practical applications in complex or low-knowledge domains. Moreover, global discovery methods that learn causal structure from data can display unstable performance on finite samples, preventing robust fairness conclusions. To mitigate these challenges, we introduce local discovery for direct discrimination (LD3): a method that uncovers structural evidence of direct unfairness by identifying the causal parents of an outcome variable. LD3 performs a linear number of conditional independence tests relative to variable set size, and allows for latent confounding under the sufficient condition that all parents of the outcome are observed. We show that LD3 returns a valid adjustment set (VAS) under a new graphical criterion for the weighted controlled direct effect, a qualitative indicator of direct discrimination. LD3 limits unnecessary adjustment, providing interpretable VAS for assessing unfairness. We use LD3 to analyze causal fairness in two complex decision systems: criminal recidivism prediction and liver transplant allocation. LD3 was more time-efficient and returned more plausible results on real-world data than baselines, which took 46$\times$ to 5870$\times$ longer to execute.