Bayesian Sparse Mediation Analysis with Targeted Penalization of Natural Indirect Effects

Area of Science:

• Biostatistics
• Genomics
• Epidemiology

Background:

• Causal mediation analysis quantifies exposure effects via mediators.
• High-dimensional data (e.g., epigenome, microbiome) requires advanced statistical methods.
• Identifying active mediators is crucial for understanding biological pathways.

Purpose of the Study:

• Develop novel Bayesian prior models for high-dimensional mediation analysis.
• Enable targeted penalization of the natural indirect effect (NIE) for active mediator identification.
• Improve selection and estimation accuracy in identifying active mediators.

Main Methods:

• Proposed two novel prior models: a four-component Gaussian mixture prior and a product threshold Gaussian prior.
• Specified joint prior distributions on exposure-mediator and mediator-outcome effects.
• Enabled targeted penalization of the product of these effects within a Bayesian framework.

Main Results:

• Simulations demonstrated improved selection and estimation accuracy compared to competing methods.
• Applied methods to the Multi-Ethnic Study of Atherosclerosis (MESA) and LIFECODES birth cohort.
• Identified active mediators revealing important biological pathways for disease mechanisms.

Conclusions:

• The novel Bayesian methods effectively identify active mediators in high-dimensional settings.
• These approaches enhance understanding of complex disease mechanisms through biological pathways.
• The methods offer improved accuracy for mediation analysis with large numbers of potential mediators.