A0317
Title: A generalization of the Dirichlet-multinomial regression model for microbiome counts
Authors: Sonia Migliorati - University of Milano Bicocca (Italy) [presenting]
Andrea Ongaro - University of Milano-Bicocca (Italy)
Roberto Ascari - University of Milano-Bicocca (Italy)
Abstract: Multinomial regression is a widespread tool to model microbiome counts as a function of environmental and biological covariates, whereas the Dirichlet-multinomial model represents an enhancement to cope with overdispersion. Though, both models often show a poor fit to real data due to their rigid dependence structure, which rules out the possibility of modelling positive associations among bacterial taxa and poor parameterization. A new regression model based on a mixture of Dirichlet-multinomial distributions is proposed. The model is a compound multinomial model, which is obtained by considering a (conditional) multinomial response and assigning an extended flexible Dirichlet distribution to its parameters. This new model succeeds in clearly identifying and interpreting relationships between taxa counts and covariates, allowing for possible positive associations among taxa too. Moreover, the mixture structure of the model naturally enables the identification of clusters of bacterial genera sharing similar biota compositions, which, in turn, can be associated with enterotypes. The analysis of a human gut microbiome dataset confirms the better performance of the new model concerning competitors. The analysis has been performed via bayesian inference, resorting to the Hamiltonian Monte Carlo algorithm, with a spike and slab approach to variable selection.
