Unbiased causal inference with Mendelian randomization and covariate-adjusted GWAS data

Area of Science:

• Genetic Epidemiology
• Statistical Genetics
• Causal Inference

Background:

• Mendelian randomization (MR) uses genome-wide association study (GWAS) data for causal inference from observational data.
• Adjusting for covariates in GWAS can introduce collider bias, affecting SNP-trait associations and downstream MR analyses.
• Previous numerical studies indicated bias in univariable MR (UVMR) with covariate-adjusted GWAS data, suggesting multivariable MR (MVMR) as a solution.

Purpose of the Study:

• To provide a rigorous theoretical explanation for why and how multivariable Mendelian randomization (MVMR) mitigates bias introduced by covariate adjustment in GWAS.
• To derive analytical results supporting empirical observations on bias in UVMR and its mitigation by MVMR.
• To investigate the application of UVMR and MVMR methods, including after collider-bias correction.

Main Methods:

• Derivation of analytical results for scenarios with multiple covariates adjusted in GWAS for exposure and/or outcome.
• Conducting extensive simulations to compare UVMR and MVMR performance with covariate-adjusted GWAS summary data.
• Application of UVMR and MVMR methods to real GWAS data, adjusting for body mass index (BMI) and metabolomic principal components.

Main Results:

• Analytical results explain the mechanisms of bias in UVMR and its avoidance in MVMR when covariates are adjusted, irrespective of collider bias presence.
• Simulations demonstrate MVMR's advantage over UVMR in producing nearly unbiased causal estimates with covariate-adjusted GWAS data.
• In specific situations, applying UVMR after bias correction was found to be advantageous.

Conclusions:

• MVMR provides a robust method to avoid bias in Mendelian randomization analyses stemming from covariate adjustment in GWAS.
• The derived analytical results offer theoretical substantiation for the empirical benefits of MVMR.
• The study confirms MVMR's utility in real-world genetic epidemiology, exemplified by the analysis of BMI's causal effect on blood pressure.