Infering Epidemiological Causality using Mendelian Randomisation

It is very difficult to find out whether potential risk factors for disease, such as alcohol consumption or aspects of people s diet, really have causal effects on diseases such as heart disease and cancer. Rapid developments in genetics mean that we now know about genetic variants that influence these factors. This can happen either because your genes influence the amount of a substance such as vitamin D that circulates in the blood, or because your genes influence your behaviour, such as the amount of alcohol you consume. Because your genes are a random sample of your parents genes ( Mendelian randomization ), it is possible to use genetic variants associated with modifiable risk factors of interest to study whether these risk factors really have causal effects on diseases.
The statistical methods used to analyse these studies are called instrumental variables methods. These methods require development in order to make them useful in Mendelian randomization studies, including finding ways to check the assumptions made in these studies. This project aims to develop instrumental variables methods, and to apply them in particular examples of importance in medical research.

It has recently been proposed that Mendelian randomization ? the distribution of genotypes randomly amongst offspring when inherited from their parents via the process of meiosis ? may be used to infer causal effects of potentially modifiable exposures on health-related outcomes. Mendelian randomization studies have the potential to overcome the problems of confounding, reverse causation and measurement errors that affect observational epidemiological studies, and may provide inferences about causal effects in situations where randomized controlled trials are unethical or impracticable. In statistical terms, Mendelian randomization studies exploit the fact that genotype is an instrumental variable (IV) for the effect of modifiable exposure on outcome. However, standard IV methods make restrictive assumptions that may limit their use in this context. Extensions to these methods are therefore urgently required. This project aims to develop IV methods for the analysis of Mendelian randomization studies, and to assess their applicability using both simulation studies and real data. After reviewing published statistical methods and software, we will propose a formal causal framework that deals with a range of causal effects including risk ratios and odds ratios, document the assumptions required in different settings, and assess their relevance together with the extent to which they are likely to be violated in datasets used for Mendelian randomization studies. We will develop and apply methods that can test for and/or allow for potential violations of IV assumptions (population stratification, linkage disequilibrium, pleiotropy, developmental compensation). Without extra assumptions, bounds on the causal effect can be calculated. We will extend these to different causal parameters and investigate the sampling behaviour and conditions under which they are informative. The relevance of the local causal effect, estimable under certain conditions in the RCT scenario, will also be studied. Mendelian randomization can require very large sample sizes, and so we will extend current meta-analytic methods for genetic association studies to Mendelian randomization applications, including evaluation of the impact of taking genotype-outcome data and genotype-exposure data from different studies. Methods will be evaluated using simulation studies and in practical applications. The first two of these will be to use ALDH2 as an instrument to determine the causal effects of alcohol consumption on a wide range of health outcomes, and to use CRP and other genetic variants to determine the causal effect of C-reactive protein on cardiovascular disease.