CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: A Bayesian approach to overlapping-sample Mendelian randomization Authors:  Hui Guo - University of Manchester (United Kingdom) [presenting]
Abstract: Mendelian randomization (MR) is a popular approach to causal inference in medical research. It uses exposure associated genetic variants, or most commonly, single nucleotide polymorphisms (SNPs) as instruments to investigate causal relationships between exposures and outcomes. Existing MR methods and their platforms have mainly focused on using publicly available summary statistics of the SNP-exposure association and SNP-outcome association obtained from two independent studies (namely two-sample MR). However, it has been brought to our attention that there are cases where a subgroup of participants was included in both of the studies, especially when data were collected at the population level. To enable a two-sample MR analysis, it is common practice that one study is discarded and a third study used instead of such that the studies are independent. Or, if data are available at the individual level, summary statistics are estimated from association analysis using data only from the non-overlapping participants. We will introduce a Bayesian MR approach that converts a two-sample or overlapping sample case into a one-sample setting, where the unmeasured data are treated as unknown parameters in the model that can be imputed in Markov chain Monte Carlo. It allows us to use all the available data without removing any participants or study. By its nature, Bayesian approach offers a more flexible way of dealing with complex models (e.g., multiple exposures, pleiotropy).