Linear high-dimensional mediation models adjusting for confounders using propensity score method

Area of Science:

• Genetics and Bioinformatics
• Epidemiology
• Statistical Genetics

Background:

• High-dimensional mediation analysis (HIMA) investigates epigenetic mediation of exposure-disease pathways.
• Existing HIMA models often assume no confounders, which is unrealistic in observational studies.
• Confounder bias is a significant challenge in applying HIMA to real-world data.

Purpose of the Study:

• To develop and evaluate methods for adjusting confounder bias in high-dimensional mediation analysis.
• To compare propensity score-based approaches with traditional covariate regression.
• To identify epigenetic mediators in the smoking-lung cancer pathway.

Main Methods:

• Proposed three propensity score-related methods: propensity score regression (PSR), propensity score weighting (PSW), and propensity score union (PSU).
• Integrated propensity score calculation with sure independence screening, minimax concave penalty (MCP) variable selection, and joint-significance testing.
• Compared the performance of PSR, PSW, and PSU against traditional covariate regression using simulations.

Main Results:

• Simulation results indicated that the propensity score union (PSU) model performed best in adjusting for confounder bias.
• Application to the TCGA lung cancer dataset revealed specific DNA methylation sites mediating the effect of smoking on lung disease.
• Identified key mediating sites, including Cg24480765 (gene RP11-347H15.2) and Cg22051776 (gene KLF3).

Conclusions:

• Propensity score methods, particularly PSU, effectively adjust for confounder bias in high-dimensional mediation analysis.
• The findings highlight the role of specific DNA methylation sites as mediators in the causal pathway from smoking to lung cancer.
• This approach enhances the reliability of HIMA in observational epidemiological research.