A Bayesian nonparametric approach for multiple mediators with applications in mental health studies

Area of Science:

• Causal inference
• Statistical modeling
• Bayesian nonparametrics

Background:

• Mediation analysis with multiple mediators is crucial for understanding complex causal pathways.
• Existing parametric methods risk model misspecification and often neglect mediator interactions.
• Current approaches may only estimate joint mediation or sum individual effects, ignoring synergistic or antagonistic interactions.

Purpose of the Study:

• To propose a novel Bayesian nonparametric method for mediation analysis with contemporaneously observed multiple mediators.
• To overcome limitations of existing parametric models, including model misspecification and the inability to capture mediator interactions.
• To enable flexible estimation of individual, joint, and interaction mediation effects.

Main Methods:

• Developed a flexible Bayesian nonparametric model using a three-level enriched Dirichlet process mixture.
• Modeled the joint distribution of outcome, multiple mediators, treatment, and confounders.
• Employed standardization (g-computation) to compute all possible mediation effects, including pairwise and higher-order interactions.

Main Results:

• The proposed method successfully identified significant individual mediators.
• Significant pairwise interaction effects among mediators were also detected.
• Application to mental health data revealed complex mediation pathways from unintended pregnancies to maternal depression.

Conclusions:

• The novel Bayesian nonparametric approach offers a flexible and robust alternative for mediation analysis with multiple mediators.
• This method effectively captures complex mediation structures, including interactions, providing deeper insights into causal relationships.
• The findings highlight the utility of advanced statistical methods in uncovering nuanced pathways in health research.