Apostolos Gkatzionis, Paul J. Newcombe and Stephen Burgess Seminar

AUEB STATISTICS SEMINAR SERIES MARCH 2020

Apostolos Gkatzionis, Paul J. Newcombe and Stephen Burgess, MRC Biostatistics Unit, University of Cambridge

Mendelian Randomization

ROOMΤ105, NEW AUEB BUILDING, (TROIAS 2)

ABSTRACT

Mendelian randomization is the use of genetic information to assess the existence of a causal relationship between a risk factor X and a disease outcome Y. It is an application of instrumental variables analysis in the field of statistical genetics, using genetic variants (SNPs) as instruments. SNPs that are included in a Mendelian randomization study should satisfy the instrumental variable (IV) assumptions: they should be strongly associated with the risk factor and not associated with the outcome in any other way except through their association with the risk factor.

Recent Mendelian randomization analyses utilize data from large Genome-wide Association Studies. Individual-level data from such studies are typically not available for ethical reasons, therefore Mendelian randomization analyses often rely on summarized data (univariate SNP-risk factor and SNP-outcome association estimates and corresponding standard errors).

After providing a brief introduction to Mendelian randomization, we discuss approaches for estimating the X-Y causal effect and assessing the validity of the IV assumptions using only summarized data. We focus particularly on “robust” Mendelian randomization methods, which can yield unbiased causal effect estimates even under violations of the IV assumptions. We then propose a Bayesian variable selection algorithm for Mendelian randomization using summarized data. Our algorithm uses a general Bayesian setting where the likelihood from a linear regression model for the SNP-risk factor associations is augmented with a loss function that penalizes the SNPs’ direct effects on the outcome. Variable selection is done using reversible-jump MCMC, and the X-Y causal effect is estimated by model averaging.        

The performance of the new algorithm is compared against established Mendelian randomization methods in simulations. We conclude with a real-data application, studying the effect of blood pressure on the risk of suffering from coronary heart disease.