MRC Integrative Epidemiology Unit (IEU) Seminar Series

Title: Bayesian Multivariate Regression and Path Analysis Models with Variable Selection for high-dimensional and longitudinal Genetic Epidemiology

Abstract: We present Bayesian seemingly unrelated regressions and path analysis models for integrating high-dimensional data sets in genetic epidemiology applications. A causal model encoded in a Directed Acyclic Graph (DAG) is proposed between blocks of variables, and automatic feature selection on the set of arrows in the DAG is used to find parsimonious models of the data. The Bayesian analysis produces model-averaged probabilities of associations between pairs of variables. We can thus obtain Bayesian estimates of indirect, direct and total effects between variables, conditional on the space of models. Further, we obtain full posterior uncertainty on these effects and on the space of DAGs, so we can explore probabilistically the space of different causal models.

We present two applications. The first is an analysis of a dataset of 158 NMR spectroscopy measured metabolites and over 9000 Single Nucleotide Polymorphisms, measured in a cohort of more than 5000 people. Here the residual correlations between metabolites are modelled simultaneously with the associations with genetic variants. The second application models epidemiological cohort data from seven time points along the life course, including genetics and environmental risk factors for adult obesity.

Biography: 

Alex Lewin's research lies at the interface between statistics and machine learning, developing highly-structured, high-dimensional Bayesian models for statistical genomics and genetic epidemiology.

Her background is in maths and physics, and she has worked in Biostatistics for several years at Imperial College, developing new Bayesian models and software for gene expression data and other 'omics' data. She joined LSHTM in 2018, to work with people across the School on methods and applications using omics data in epidemiology.

All welcome