A Bayesian finite mixture of bivariate regression model for causal mediation analyses

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Causal mediation analysis is crucial for understanding complex relationships between exposures, mediators, and outcomes.
• Existing methods may struggle with confounding and unmeasured heterogeneity, limiting accurate effect estimation.
• Finite mixture models offer a flexible framework for handling complex data structures.

Purpose of the Study:

• To propose a novel Bayesian finite mixture of bivariate regression model for causal mediation analyses.
• To develop a method for estimating natural direct and indirect effects under confounding and heterogeneity.
• To apply the model to real-world data investigating the effects of inhaled corticosteroids on birthweight.

Main Methods:

• Development of a Bayesian finite mixture of bivariate regression model.
• Utilizing an identifiability condition within mixture components to define direct and indirect effects.
• Validation using simulated data under various confounding scenarios and application to a pregnancy cohort.

Main Results:

• The proposed model accurately estimates natural direct and indirect effects, even with confounding.
• The mixture model effectively accounts for unmeasured binary or categorical mediator-outcome confounders.
• Application to inhaled corticosteroid exposure in asthmatic women demonstrated the model's utility in estimating effects on birthweight.

Conclusions:

• The Bayesian finite mixture model provides a robust framework for causal mediation analysis.
• This approach enhances the ability to disentangle direct and indirect effects in the presence of confounding and heterogeneity.
• The model has significant implications for epidemiological research, particularly in perinatal health studies.