Model selection for ecological community data using tree shrinkage priors

Abstract

Researchers and managers model ecological communities to infer the biotic and abiotic variables that shape species' ranges, habitat use, and co-occurrence which, in turn, are used to support management decisions and test ecological theories. Recently, species distribution models were developed for and applied to data from ecological communities. Model development and selection for ecological community data is difficult because a high level of complexity is desired and achieved by including numerous parameters, which can degrade predictive accuracy and be challenging to interpret and communicate. Like other statistical models, multi-species distribution models can be overparameterized. Regularization is a technique that optimizes predictive accuracy by shrinking or eliminating model parameters. For Bayesian models, the prior distribution automatically regularizes parameters. We propose a tree shrinkage prior for Bayesian multi-species distributions models that performs regularization and reduces the number of regression coefficients associated with predictor variables. Using this prior, the number of regression coefficients in multi-species distributions models is reduced by estimation of unique regression coefficients for a smaller number of guilds rather than a larger number of species. We demonstrated our tree shrinkage prior using examples of presence-absence data for six species of aquatic vegetation and relative abundance data for 15 species of fish. Our results show that the tree shrinkage prior can increase the predictive accuracy of multi-species distribution models and enable researchers to infer the number and species composition of guilds from ecological community data.