Pleiotropy robust Mendelian randomization

Across the world, scientists are trying to uncover the causes of disease and ill-health, so that resources can be intelligently targeted to improve the lives of the general public. However, ethical considerations often remove the possibility of a randomized controlled trial (RCT), which is the most straightforward and reliable way of assessing causality, and decisions must be made using non-experimental data.
Since our genes are randomly assigned from parents to their children at the point of conception, they should not be related to factors such as diet and other lifestyle choices that impact an individual’s health in later life. We can therefore think of them as creating a natural randomized trial. By comparing people’s health across different genetic sub-groups, we can assess whether a risk factor (e.g. cholesterol) causally influences the risk of a disease (e.g. stroke). This is technique is called Mendelian randomization. My research focuses on developing the statistical methods required to perform this analysis, so that it is reliable and efficient.

The explosion in publicly available data from genome-wide association studies is accelerating the use of Mendelian randomization (MR) in bio-medicine. Summary data estimates of genetic association from large numbers of variants are now being synthesised for causal inference within the two-sample framework. Over the coming years, cohort studies such as UK Biobank will also provide a rich individual-level data resource for MR investigations. In a short time, the field has seen, and will continue to see, a dramatic rise in the power for testing causal hypotheses. There is a justified concern that when large numbers of genetic variants are included in MR analyses, with many lacking a firm biological basis for their association with the exposure, a sizeable proportion of these variants are likely to be invalid instrumental variables.
A chief concern is that genetic variants may exert an effect on the outcome not through the exposure of interest - which is referred to as horizontal pleiotropy. This programme will focus on the development of statistical methods for MR in the presence of pleiotropy and other associated biases. Building upon recent methodological advancements in this field, we will develop analysis tools that:
• Provide natural robustness to pleiotropy;
• Adjust for non-random selection into (or out of) cohort studies used for MR;
• Novel methods for data visualisation and analysis will be incorporated into statistical software platforms such as MR-Base
This programme links key players in methodological research around the world to leading biomedical scientists at the Integrative Epidemiology Unit. It will develop the tools to enable epidemiological researchers to fully capitalise on emerging large-scale data sources whilst preserving the principles and rigour of causal inference.