CFE 2019: Start Registration
View Submission - CMStatistics
Title: Bayesian model averaging for two-sample summary data Mendelian randomization in the presence of pleiotropy Authors:  Jack Bowden - University of Exeter (United Kingdom) [presenting]
Abstract: Mendelian randomization (MR) uses genetic variants as instrumental variables to estimate the causal effect of a modifiable health exposure on a downstream health outcome. The technique can be implemented using only summary data estimates of genetic association from genome wide association studies, which is referred to as `two sample summary data MR. Typically many hundreds of genetic are used in such an analysis, but it is highly likely that a sizable number of them are, in fact, invalid instruments. A main cause of instrument invalidity is when a genetic variant exerts a direct effect on an outcome not through the exposure of interest. This is a violation of the exclusion restriction, and is referred to in the MR field as `pleiotropy'. We develop a Bayesian Model Averaging approach that intelligently searches the space of all $2^{L}$ models ($L$ being the number of genetic variants) to obtain a more robust and reliable causal estimate. Our algorithm favours models with large numbers of variants, but down-weights sets of variants that lead to heterogeneous causal effect estimates. It also naturally accounts for weak instrument bias via the use of a a posterior profile likelihood function. We illustrate our approach first for a basic one parameter causal model, and then show how it can be extended to more complex modelling frameworks.