BAYESIAN MENDELIAN RANDOMISATION USING INDIVIDUAL-LEVEL DATA

Abstract

Mendelian randomisation (MR) is an effective tool to identify the causal relationship between different variables. This method mimics randomised controlled trials (RCTs) based on the random segregation of people's alleles, i.e. Mendel's genetic law. Traditional MR analysis is limited by a number of methodological challenges: one of which is overlapping sample analysis. Classic MR methods only focus on one- or two-sample MR settings, which often involves a waste of information. In this thesis, a Bayesian approach that is applicable to any of the three sample settings, i.e. one-, two- and overlapping-sample, has been developed based on a series of different MR models. The flexibility of this novel approach has been demonstrated through analysis based on the various models with multi-instrument/exposure/outcome, pleiotropy, interaction and random effects. Through a number of simulation experiments, this method has been shown to possess better performance in terms of precision and statistical power, compared to classic methods. In addition, MR is also an area that needs good foundations for hypothesis testing. However, the frequentist hypothesis test can only lead to binary decision logic and is handicapped by its inability to accommodate decision of uncertainty. Such black-and-white thinking is more likely to lead to misinterpretation and confusion, especially in MR, in which some of the assumptions are untestable. This thesis extends the classic binary decision logic (``acceptance'' and ``rejection'') to a ternary logic ("acceptance", "rejection" and "uncertain") by replacing the point null hypothesis with a region of practical equivalence which consists of values of negligible magnitude for the effect of interest, while exploiting the ability of Bayesian analysis to quantify evidence of the effect that falls inside/outside the region. The proposed method has also been calibrated via loss function. In the Bayesian framework, one common downside is that Bayesian approach is often limited by its computational expensiveness based on large studies. Therefore, introducing some strategies to handle big data Bayesian analysis is also significant. For this purpose, a data partitioning method (evenly divide the full dataset into some subsets, and then combine the posteriors obtained from these subsets) has been incorporated into our Bayesian MR method in order to improve the efficiency of Bayesian computation with large studies.

Date of Award	1 Aug 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Carlo Berzuini (Supervisor) & Hui Guo (Supervisor)