Title: Multivariate logistic mixture regression for microbiome analysis
Authors: Jack OBrien - Bowdoin College (United States) [presenting]
Abstract: Two standard approaches to understanding microbiome data are mixture models and regression techniques. Mixture models attempt to capture the structure and number of components of the data, with the aim of associating these with environmental or experimental conditions. Regression techniques seek to directly assess the connection between covariates and the specific features of the data (compositional, excess zeros, temporal or spatial structure). We show how recent advances in inference for multivariate logistic regression can be extended to a mixture context, determining the component structure as part of the regression. This translates the regression to a classification procedure and shows how the Voronoi tessellation can be used to understand microbial dynamics. A compositional approach is covered and details are given about how it may be extended to a non-compositional framework that may have wider applicability in marine ecology.