Heterogeneous Mediation Analysis for Cox Proportional Hazards Model With Multiple Mediators

Area of Science:

• Biostatistics
• Causal Inference
• Survival Analysis

Background:

• Mediation analysis in survival data is complex, especially with multiple mediators and sparse predictors.
• Existing methods often struggle to account for overlapping confounders and effect modifiers across different causal pathways.

Purpose of the Study:

• To propose a robust heterogeneous mediation analysis framework for survival data.
• To develop a joint modeling approach integrating mediation and survival models using Bayesian additive regression trees.
• To enable causal discovery and quantify heterogeneity in survival outcomes.

Main Methods:

• A joint modeling approach linking mediation regression and proportional hazards models via Bayesian additive regression trees with shared typologies.
• Incorporation of a sparsity-inducing prior to identify relevant confounders and effect modifiers.
• Derivation of individual-specific interventional direct and indirect effects on the log-hazard and survival function scales.
• Utilizing a Bayesian approach with Markov chain Monte Carlo (MCMC) for effect estimation.

Main Results:

• The proposed method effectively handles multiple mediators and predictor sparsity in survival data.
• Simulation studies confirm the empirical performance and validity of the Bayesian approach.
• The method successfully quantifies causal effects and heterogeneity in survival data.

Conclusions:

• The developed heterogeneous mediation analysis provides a powerful tool for understanding complex causal relationships in survival data.
• This approach facilitates improved causal discovery and heterogeneity quantification, particularly in the presence of multiple mediators.
• The application to the ACTG175 study highlights the practical utility of the method in real-world biomedical research.