Bayesian Estimation of Cooccurrence Affinity with Dyadic Regression

Abstract

O_LIEstimating underlying co-occurrence relationships between pairs of species has long been a challenging task in ecology as the extent to which species co-occur is partially dependent on their prevalence. While recent work has taken large steps towards solving this problem, the next question is how to assess the factors that influence co-occurrence. C_LIO_LIHere, I show that a recently proposed co-occurrence metric can be improved upon by assigning Bayesian priors to the latent co-occurrence relationships being estimated. In the context of analysing the factors that affect co-occurrence relationships, I demonstrate the need for a generalised linear model (GLM) that takes raw data (co-occurrences and species prevalence) not derived quantities (co-occurrence metrics) as its data. Next, I show the form that such a GLM should take in order to perform Bayesian inference while accounting for non-independence of dyadic matrix data (e.g. distance and co-occurrence matrices). C_LIO_LII then present 3 example analyses to highlight the types of scientific questions these methods can help answer, using existing data sets - measuring the effects of trait dissimilarity among dung beetle species and relatedness between ant species on co-occurrence, and constructing co-occurrence networks of bacteria found in cystic fibrosis patient sputum samples. C_LIO_LIFinally, I present the software package CooccurrenceRegression.jl, which provides a straightforward interface for researchers to put these methods into practice. C_LI