Pleiotropic bias correction in Mendelian randomization analyses

: novel approaches and applications to causal inference leveraging large-scale genetic data

Abstract

Background-
The proliferation of large-scale studies with genetic data, coupled with technological advances in genetics has served as a catalyst for the development of causal inference approaches leveraging genetic data. One such approach, Mendelian randomization (MR), uses genetic variants as instrumental variables to estimate causal relationships while mitigating bias due to residual confounding. MR methods can be broadly be categorised based on the type of data utilised,
specifically approaches using either individual level data or summary data from genome-wide association studies (GWAS).

Objectives-
To evaluate and contribute to the existing MR methods literature, introducing novel statistical frameworks for use with individual and summary level data. In this work radial multivariable Mendelian randomization (RMVMR) is presented as an approach for visualising summary MR analyses, while MR using Gene-by-Environment interactions (MR-GxE) allows for the validity of individual instruments to be evaluated. Analyses of the causal determinants of stroke are
subsequently conducted within an evidence triangulation framework, highlighting the utility of employing individual and summary level MR methods in combination.

Results-
An appraisal of existing MR methods is conducted in chapter 1, after which RMVMR is presented as a novel approach for visualising key aspects of multivariable MR analyses in chapter 2. Interaction MR is described in chapter 3, providing an overview of how causal effects can be estimated using a single instrument in the presence of measured or unmeasured gene-by-covariate
interactions. Simulation studies are performed for each MR method, highlighting the relative strengths and limitations of each approach. After conducting a preliminary review of stroke in chapter 4, applied analyses are performed
in chapter 5, identifying a range of modifiable risk factors with respect to stroke.

Conclusion-
The statistical methods presented in this work have the potential to overcome common limitations of existing MR methods, and highlight their application to stroke as a key area of global health research.

Date of Award	27 Sept 2022
Original language	English
Awarding Institution	University of Bristol
Sponsors	Wellcome Trust
Supervisor	George Davey Smith (Supervisor), Eleanor C M Sanderson (Supervisor) & Jack Bowden (Supervisor)