Speaker: Anna Plantinga, Graduate Student, UW Biostatistics
Abstract: The human microbiome is intricately related to human health and disease. To better understand the roles it plays, there is significant interest in association testing and prediction using the microbiome. However, microbiome data provides several analytic challenges: they are compositional, high-dimensional, and have external structure that can be leveraged to improve statistical power. Two high-dimensional regression-based approaches that may address these challenges will be discussed. I will first introduce a kernel machine regression-based test of association for entire microbial communities with survival outcomes. The test has proper type I error control, high power, and is efficient to compute. I will then discuss a penalized regression method that utilizes a novel multilevel model to enable prediction and estimation with hierarchically structured, compositional data. These approaches will be applied to studies of the gut microbiome in graft-versus-host disease and obesity.