Bayesian mediation analysis for zero-inflated mediator.

Area of Science:

• Biostatistics
• Causal Inference
• Statistical Modeling

Background:

• Mediation analysis is crucial for understanding causal pathways involving intermediate variables.
• Zero-inflated mediators, common in health research, pose challenges for standard mediation methods.
• Existing methods may lack flexibility in handling diverse outcome distributions and zero-inflated data.

Purpose of the Study:

• To develop a flexible Bayesian mediation analysis framework for zero-inflated mediators.
• To accommodate a wide range of outcome distributions in mediation analysis.
• To decompose mediation effects into components related to zero probability and non-zero distribution means.

Main Methods:

• A Bayesian framework employing Markov Chain Monte Carlo (MCMC) for parameter estimation.
• Joint modeling of zero-inflated mediators and various outcome distributions.
• Decomposition of total mediation effects into natural direct and indirect effects, and further into zero-inflation and non-zero mean components.

Main Results:

• The proposed Bayesian method demonstrates superior accuracy in point estimates and coverage probabilities compared to alternatives.
• The framework accurately decomposes mediation effects, providing more precise insights into causal pathways.
• The R package 'mediationBayes' facilitates practical implementation of the novel methodology.

Conclusions:

• The flexible Bayesian framework effectively handles zero-inflated mediators across diverse outcome distributions.
• The method offers improved accuracy and precision in mediation analysis, particularly for complex data structures.
• This approach enhances causal inference by providing a robust tool for dissecting mediation effects.