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

Background. A central objective among health researchers across disciplines is to identify modifiable factors that can reduce health disparities. Three common methods–difference-in-coefficients (DIC), Kitagawa-Oaxaca-Blinder (KOB), and causal decomposition analysis (CDA)–share the same goal to identify such contributors but can produce divergent results depending on confounding and model assumptions. Despite these challenges, applied researchers lack clear guidance on selecting appropriate methods for different scenarios. Methods. We start with a brief review of each method, assuming no unmeasured confounders. We then move to two more realistic scenarios: 1) unmeasured confounders affect the relationship between intermediate confounders and the mediator, and 2) unmeasured confounders affect the relationship between the mediator and the outcome. For each scenario, we generate simulated data, apply all three methods, compare their estimates, and interpret the results using Directed Acyclic Graphs. Results. Under the assumption of no unmeasured confounders, the DIC approach is a simplistic method suitable only when no intermediate confounders are present. The KOB decomposition is appropriate unless adjustment for baseline covariates is necessary. When unmeasured confounding exists, the DIC method yields biased estimates in both scenarios, and all three methods produce biased results in the second scenario. However, CDA, when paired with sensitivity analysis can help assess the robustness of its estimates. Conclusions. We advise against using the DIC method, particularly, in observational studies, as the assumptions of no intermediate confounders is often unrealistic. When unmeasured confounders are anticipated, CDA combined with sensitivity analysis offers a more robust approach for identifying mediators over other methods.