Mendelian randomization analysis using mixture models for robust and efficient estimation of causal effects

Guanghao Qi,Nilanjan Chatterjee

doi:10.1038/s41467-019-09432-2

Guanghao Qi, Nilanjan Chatterjee

Open Access

https://doi.org/10.1038/s41467-019-09432-2

Copy DOI

Abstract

Mendelian randomization (MR) has emerged as a major tool for the investigation of causal relationship among traits, utilizing results from large-scale genome-wide association studies. Bias due to horizontal pleiotropy, however, remains a major concern. We propose a novel approach for robust and efficient MR analysis using large number of genetic instruments, based on a novel spike-detection algorithm under a normal-mixture model for underlying effect-size distributions. Simulations show that the new method, MRMix, provides nearly unbiased or/and less biased estimates of causal effects compared to alternative methods and can achieve higher efficiency than comparably robust estimators. Application of MRMix to publicly available datasets leads to notable observations, including identification of causal effects of BMI and age-at-menarche on the risk of breast cancer; no causal effect of HDL and triglycerides on the risk of coronary artery disease; a strong detrimental effect of BMI on the risk of major depressive disorder.

Highlights

Simulation studies show that MRMix can be far more robust compared to existing alternatives in a wide range of scenarios (Figs 2 and 3, Supplementary Figure 1)
When genetic correlation due to the causal relationship and pleiotropic effects are in the same direction (Fig. 2), MRMix generally produced nearly unbiased estimates of causal effects as long as the sample size for GWAS for the exposure (X) and the corresponding number of instruments reached a minimum threshold
We develop a novel and powerful method for conducting Mendelian randomization (MR) analysis using a large number of genetic instruments based on normal-mixture models for effect-size distribution where distinct mixture components are incorporated to allow genetic correlations to arise both from causal and non-causal relationships

Summary

Methods

If the assumption is satisfied, the two sets of regression coefficients will satisfy a proportional relationship in the form βy = θβx, where θ is the causal effect of X on Y. For both simulations and real data applications, we compare MRMix with existing popularly used MR methods that allow the estimation of causal effects. The estimate is given by θ 1⁄4 ρg=h2x, where ρg is the estimated genetic covariance of the pair of traits and h2x is the estimated heritability of X This estimator is nearly equivalent to Egger regression using the same set of SNPs (see Supplementary Notes for details). As a benchmark for comparison, in real data analysis, we report estimates of causal effect based on the LD score regression

Results

Discussion

Conclusion