Multivariable MR can mitigate bias in two-sample MR using covariable-adjusted summary associations

Abstract Genome-Wide Association studies (GWAS) are hypothesis free studies that survey the whole genome for polymorphisms associated with a trait of interest. To increase power and to estimate the direct effects of these single nucleotide polymorphisms (SNPs) on a trait GWAS are often conditioned on a covariate (such as body mass index (BMI) or smoking status). This adjustment can introduce bias in the estimated effect of the SNP on the trait. Mendelian randomisation (MR) studies use summary statistics from GWAS estimate the causal effect of a risk factor (or exposure) on an outcome. Covariate adjustment in GWAS can bias the effect estimates obtained from MR studies using the GWAS data. Multivariable MR (MVMR) is an extension of MR that includes multiple traits as exposures. Using simulations we show that MVMR can recover unbiased estimates of the direct effect of the exposure of interest by including the covariate used to adjust the GWAS within the analysis. We show that this method provides consistent effect estimates when either the exposure or outcome of interest has been adjusted for a covariate. We apply this method to estimate the effect of systolic blood pressure (SBP) on type-2 diabetes (T2D) and waist circumference on systolic blood pressure both adjusted and unadjusted for BMI.