Identifying Causally-Robust Mediators of Health Disparities: A Review and Simulation Studies With Directed Acyclic Graphs

2025-06-23

AI Summary

This study addresses the causal identification of mediation effects in health disparities, targeting estimation bias in conventional linear decomposition methods (e.g., Oaxaca–Blinder, Kitagawa decompositions) under unmeasured confounding. We propose a novel framework integrating causal decomposition analysis with sensitivity analysis, grounded in directed acyclic graphs (DAGs) to simulate diverse confounding structures and systematically evaluate method robustness. Results demonstrate that traditional approaches exhibit substantial bias under complex unmeasured confounding, whereas our proposed causal decomposition—augmented by formal sensitivity assessment—enables quantitative evaluation of unmeasured confounding’s impact on mediation effect estimates. This enhances the reliability and interpretability of causal inferences. The framework advances health equity research by offering a theoretically rigorous yet empirically implementable paradigm for mediation analysis in observational studies of health inequality.

Technology Category

• Reasoning under Uncertainty: Causality
• Machine Learning: Causal Learning
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Background Traditionally researchers have used linear approaches such as difference in coefficients DIC and Kitagawa Oaxaca Blinder KOB decomposition to identify risk factors or resources referred to as mediators underlying health disparities Recently causal decomposition analysis CDA has gained popularity by defining clear causal effects of interest and estimating them without modeling restrictions Methods We begin with a brief review of each method under the assumption of no unmeasured confounders followed by two realistic scenarios where unmeasured confounders affect first the relationship between intermediate confounders and the mediator and second the relationship between the mediator and the outcome For each scenario we generate simulated data apply all three methods compare estimates and interpret results using directed acyclic graphs Results The DIC approach performs well only when no intermediate confounders are present a condition unlikely in real world health disparities that arise from many factors over the life course The KOB decomposition is appropriate only when baseline covariates such as age need not be controlled When unmeasured confounding exists DIC yields biased estimates in both scenarios while both KOB and CDA yield biased estimates in the second scenario however CDA supplemented by sensitivity analysis can reveal how robust its estimates are to unmeasured confounding Conclusions We recommend against using DIC for investigating drivers of health disparities and instead advise applying CDA combined with sensitivity analysis as a robust strategy for identifying mediators of health disparities

Problem

• Identifying robust mediators of health disparities using causal methods
• Comparing traditional vs. causal approaches under unmeasured confounding
• Recommending causal decomposition with sensitivity analysis for mediator identification

Innovation

• Uses causal decomposition analysis (CDA)
• Simulates unmeasured confounder scenarios
• Recommends CDA with sensitivity analysis