B1267
Title: Analysis of compositional benthic data via regularized Dirichlet-multinomial regression
Authors: Alysha Cooper - University of Guelph (Canada) [presenting]
Ayesha Ali - University of Guelph (Canada)
Zeny Feng - University of Guelph (Canada)
Abstract: Compositional data, measured by taxa counts at a specified taxonomic rank, are prevalent in many fields like microbiology and ecology. In ecology, measuring the relative abundances of benthic macroinvertebrate taxa can provide insight into the health of an aquatic ecosystem. Identifying environmental variables that influence each taxon's abundance is an important biological question. Analyzing such compositional data is challenging due to the high dimensionality and overdispersion of these multinomial counts. Dirichlet-multinomial (DM) regression treats proportion parameters as random variables from a Dirichlet distribution and easily accommodates overdispersion. However, estimation of model coefficients within the DM framework is difficult due to its non-concave log-likelihood. Moreover, variable selection becomes necessary when the number of variables and/or number of taxa is large. A novel algorithm is proposed to optimize a regularized DM regression model via Majorization-Minimization (MM). The proposed regularized DM-regression is applied to study compositional benthic macroinvertebrate counts in Canada's oil sands region.
