Effect Measure Modification by Covariates in Mediation: Extending Regression-based Causal Mediation Analysis.

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Traditional mediation analysis assumes constant effects across patient characteristics.
• Investigational insights often suggest effect modification by baseline factors.
• Modification of mediation effects like natural direct effect (NDE) and natural indirect effect (NIE) is of clinical interest.

Purpose of the Study:

• To extend regression-based causal mediation analysis to incorporate effect measure modification (EMM).
• To assess how baseline characteristics modify NDE, NIE, and proportion mediated.
• To provide a practical tool for implementing EMM in mediation analysis.

Main Methods:

• Extended closed-form regression-based causal mediation analysis to include EMM.
• Utilized a simulated numerical example to demonstrate method's performance.
• Applied the extended method to an empirical case study of anti-interleukin-1 therapy for anemia reduction.

Main Results:

• Naive mediation analysis ignoring EMM can yield biased NDE and NIE estimates.
• Baseline characteristics significantly influence the presence and magnitude of EMM for NDE and NIE.
• The proportion mediated via an inflammatory biomarker was higher in younger, nondiabetic patients with lower baseline inflammation.

Conclusions:

• Considering EMM in causal mediation analysis is crucial for accurate effect estimation.
• The findings highlight differential utility of biomarkers based on patient characteristics.
• An open-source R package, regmedint, is provided to facilitate EMM adoption in research.