MRC Integrative Epidemiology Unit (IEU) Seminar Series

Title: Bayesian network analysis offers an alternative to Mendelian randomization for exploratory analysis of causal relationships in complex data

Abstract: Mendelian randomization (MR) is an increasingly popular causal inference tool used in genetic epidemiology. But it has limitations for evaluating simultaneous causal relationships in complex data sets that include, for example, multiple genetic predictors and multiple potential risk factors associated with the same genetic variant.

We use computer simulations to investigate Bayesian network analysis (BN) as an alternative approach.  Firstly, we compare MR and BN using simple four-variable models where the causal assumptions required for MR or BN are violated. Secondly, we consider a more complex scenario that may arise when analysing biomarker data as generated from modern “omics” technologies, whereby multiple genetic variants are associated with multiple potential intermediate variables (biomarkers) that mediate an outcome of interest. This highly pleiotropic scenario violates one of the required assumptions of MR. As well as BN and MR we evaluate several other recently-proposed causal inference methods: multivariable MR based on Bayesian model averaging (MR-BMA), a multi-SNP mediation intersection-union test (SMUT) and a latent causal variable (LCV) test. We show that BN is a useful complementary approach to existing methods for performing causal inference in complex data sets.

 We also consider the missing data problem when performing BN, where the standard approach is to remove every individual with missing data beforehand. This can be wasteful and undesirable when there are many individuals with missing data, perhaps with only one variable missing, making imputation a natural choice. We present a new imputation method designed to increase the power to detect causal relationships whilst accounting for model uncertainty. This method uses a version of nearest neighbour imputation, whereby missing data from one individual is replaced with data from another individual, the nearest neighbour. An important feature of this approach is that it can be used with both discrete and continuous data. For each individual with missing data, subsets of variables that can be used to find the nearest neighbour are chosen by bootstrapping the complete data to estimate a Bayesian network. We show that, with the use of our imputation method, the power to detect the correct model can be increased by as much as 50%. Thus, this approach has great potential to boost the power of BN to identify possible causal relationships.

Biography: Richard Howey is a Research Associate at Newcastle University working in the Institute of Genetic Medicine with Prof. Heather Cordell. His research focuses on the development of statistical genetics methodology and associated software. Richard has a diverse research background earning his PhD in pure mathematics at Newcastle University under the supervision of Prof. Barry Johnson in the field of functional analysis. His subsequent research has included developing methods and software in artificial intelligence at Durham and Strathclyde Universities and developing mathematical models in Epidemiology at Edinburgh University.

His current research is focussed on developing causal inference methods and software for use with complex genetic and biological data with the use of Bayesian Networks.

All welcome