Disentangling trait and phylogenetic signals in species co-occurrence: a Bayesian dyadic regression approach

Kyle Rosenblad

Aim: Understanding the factors shaping species co-occurrence is a central objective in ecology. Progress is hindered by correlations among species traits and phylogenetic relationships, which confound each other’s signals. Regression-based approaches hold promise to ameliorate this issue, but the current toolkit lacks methods that combine suitable likelihoods for co-occurrence data with latent variables that account for non-independence, or pseudoreplication, among species pairs.

Innovation: I introduce an approach for estimating the effects of multiple trait and phylogenetic distances on species co-occurrence via Bayesian dyadic regression. Implemented in the compnet R package, this method disentangles the roles of correlated predictors by jointly estimating their conditional effects. compnet extends existing tools by incorporating latent variables that model dependence among species pairs, as well as response distributions suitable for co-occurrence data.

Main Conclusions: In mechanistic metacommunity simulations featuring trait correlations, compnet recovers variation in the strength of community assembly processes while avoiding false detections for ecologically irrelevant traits. In an empirical analysis of North American freshwater fish assemblages, the model reveals opposing effects on co-occurrence for two correlated traits, which appear weak or absent in models that do not account for the correlation. Simulation tests also show that the latent factor component, which is not included in other ecologically motivated dyadic regression tools, is important for avoiding pseudoreplication and associated risks of false detection. compnet offers a complementary new approach for analyzing co-occurrence data, revealing the roles of correlated traits while reducing risks of spurious inference.