A0325
Title: Fast fitting of phylogenetic random effect models
Authors: Bert van der Veen - Norwegian University of Science and Technology (Norway) [presenting]
Robert OHara - Norwegian University of Science and Technology (Norway)
Abstract: Ecologists survey locations in space or time to collect data on the presence or abundance of species. The models fitted to such data are used to assess the impact of changes in environmental conditions on species and potentially make recommendations for conservation purposes. Often, there are few non-zero observations for many of the species, so adding an extra source of information would help to successfully estimate the parameters in the models. One way to do that is to add information on the relatedness of species via a Phylogenetic tree. This allows us to borrow information from species that occur more frequently in order to determine the response of less abundant species that share a similar evolutionary ancestry. We present an implementation of Phylogenetic random effects models in the gllvm R package, which uses Variational Approximations (VA) for estimation. The key to a fast approximation is to reduce the number of VA parameters as much as possible, which we do by applying a matrix normal structure for the VA distribution and combining it with a reduced rank approximation. We also apply a Nearest Neighbor approximation to the inverse for the Phylogenetic covariance matrix, and use Template Model Builder for parallel computations.
