|
Abstract:
|
I will describe a Bayesian joint model for compositional microbiome data that allows for the identification of mediation effects. To accommodate overdispersion, the microbial abundance data are assumed to follow a Dirichlet-multinomial distribution, given the treatment assignment and a set of observed covariates. A compositional linear regression model relates the taxa proportions, transformed via balances, to the outcome. Spike-and-slab priors are imposed on the regression coefficients to provide direct inference on the presence of overall and marginal mediation effects, treatment effects, and covariate effects. I will compare the method's performance versus comparative approaches on simulated data and show an application to a benchmark study investigating the meditation effects of the gut microbiome on the relation between sub-therapeutic antibiotic treatment and body weight in early-life mice.
|
