More accurate and unbiased associations in genetic association studies with applications to Mendelian Randomization

A genome-wide association study (GWAS) is a research approach which compares the genomes of a large group of people in order to locate genomic variants statistically associated with a specific trait or disease. In general, it has been observed that estimated variant-trait associations tend to be lower in magnitude in follow-up replication studies than in the initial study which discovered these associations. A phenomenon known as Winner’s Curse has been recognised as largely responsible for this bias impacting discovery effect size estimates. This bias can negatively affect statistical applications that rely on GWAS variant-trait association estimates. One such example is Mendelian randomization (MR) analyses. In observational epidemiology, MR is a popular analytical tool which exploits genetic variation to assess the causal effect of a modifiable exposure on an outcome. If the same GWAS sample is used to both select genetic variants for the MR study and provide variant-exposure or variant-outcome association estimates, the MR exposure-outcome causal effect estimate will be biased due to Winner’s Curse. This thesis focuses on mitigating Winner’s Curse bias in the summary-level data setting. It presents a comprehensive review of Winner’s Curse correction methods, considering methods which solely use the initial discovery study as well as those that also incorporate replication information. Several modifications to improve existing ‘discovery-only’ methods are suggested, while an approach which uses the parametric bootstrap, adapted from the existing literature, is also proposed. With respect to the discovery and replication setting, the original likelihood-based method is modified, making it more suitable for application to modern studies. In addition, the use of another method, originally published for the purpose of estimating treatment effects by combining randomised and observational evidence, is explored in this context. Comparative method performance, in terms of both bias and mean square error, is assessed using a wide range of simulated data sets as well as real data sets from the UK Biobank. This thesis demonstrates the benefit of adjusting for Winner’s Curse bias by employing methods which jointly consider the effect sizes of all variants, as opposed to implementing corrections independently. The final portion of this thesis introduces a novel summary-level MR method, namely MR Simulated Sample Splitting (MR-SimSS), which seeks to provide exposure-outcome causal effect estimates void of Winner’s Curse bias. By conducting a simulation study and several same-trait empirical analyses, the ability of MR-SimSS combined with Robust Adjusted Profile Score (MR-RAPS) to consistently yield accurate causal effect estimates is demonstrated. With an adequate number of instrument variants, it is shown that this form of MR-SimSS can successfully eliminate bias due to both Winner’s Curse and weak instruments. Most notably, MR-SimSS continues to yield unbiased estimates in settings in which the variant-exposure and variant-outcome association estimates have been generated by GWASs performed with a single, fully overlapping sample.

Publisher

University of Galway

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International

cbn

Collections

University of Galway Theses (PhD Theses)