Causal mediation analysis with latent subgroups

Area of Science:

• Biostatistics
• Genetics
• Epidemiology

Background:

• Causal mediation effects in biomedical studies can vary significantly among individuals.
• Understanding this heterogeneity is crucial for accurate interpretation of biological pathways.

Purpose of the Study:

• To develop a statistical method for identifying latent subgroups with distinct mediation effects.
• To propose a mixture modeling approach for estimating subgroup-specific causal mediation effects.
• To provide a robust method for selecting the number of subgroups and computing confidence intervals.

Main Methods:

• Proposed a mixture modeling approach within a linear structural equation model framework.
• Utilized the expectation-maximization (EM) algorithm for parameter estimation.
• Employed the singular Bayesian information criterion (BIC) for subgroup selection and non-parametric bootstrap for confidence intervals.

Main Results:

• Demonstrated that subgroup-specific mediation effects can be identified under group-specific sequential ignorability assumptions.
• Simulation studies confirmed the empirical performance of the proposed iMed method.
• Reanalysis of DNA methylation data showed heterogeneous mediation effects of CpG sites in the smoking-lung function pathway across two latent subgroups.

Conclusions:

• The iMed method effectively accounts for latent subgroup structures in heterogeneous mediation analysis.
• Identified distinct subgroups with differing mediation effects in a real-world dataset.
• The developed R package 'iMed' facilitates the application of this novel methodology in biomedical research.