Abstract
AbstractMendelian randomization (MR) is an approach to causal inference that utilises genetic variants to obtain estimates of the causal effect of an exposure on an outcome in the presence of unobserved confounding. MR relies on a set of assumptions to obtain unbiased effect estimates, one of these assumptions is that there is no pathway from the genetic variants to the outcome that does not act through the exposure. Increasing genome-wide association study (GWAS) sample sizes for the exposure enables discovery of instrumental variables with smaller effect sizes. We illustrate through simulations how smaller effect sizes could arise from genetic variants that act through traits that have greater liability to confound an exposure-outcome relationship. When such genetic variants are selected as instruments this can bias the MR effect estimate obtained from that instrument in the same direction as the confounded observational association but with larger magnitude. Through simulation we illustrate how the total bias of the MR estimates increases across a range of standard MR estimation methods increases as the proportion of the genetic instruments that are associated with the confounder increases. However, if such heritable confounders are known and can be instrumented, the confounder free effect estimate can be obtained through applying a pre-estimation filtering to standard MR methods, removing instruments that explain more variation in that confounder than the exposure, or by estimating effects through multivariable MR. We highlight the potential for SNPs identified in GWAS to be associated with potential confounders through examination of a recent GWAS of C-Reactive Protein. Finally, we illustrate our approach through estimation of the causal effect of age at menarche on type 2 diabetes, hypothesising that the MR effect estimate may be biased by confounding due to the inclusion of genetic variants associated with early life adiposity as instruments.
Publisher
Cold Spring Harbor Laboratory