An efficient, biogeographically constrained occupancy model in Stan

Occupancy models are a widespread tool for modeling species occurrence while accounting for imperfect detection during biological surveys. Multi-species occupancy models (MSOMs) extend the occupancy model to treat entire assemblages of species jointly via shared random effects. Recent applications of MSOMs have sought to extend the models to very large (e.g. continental) spatial scales in order to pool information on occupancy and detection probabilities over large species pools whose constituent species’ ranges may overlap only partially or not at all. However, such models typically employ relatively simple covariate structures that might fail to capture the spatial complexities of multiple species distributions. Here, we present an approach that leverages pre-existing range maps to constrain species’ distributions a priori, thereby enabling the pooling of information across biogeographically complex communities using minimally complex model structures. We apply our method to continental-scale survey data for 53 species of North American warblers (Parulidae) and to a large-scale survey of a hyperdiverse bird community from the western Amazon. Compared to biogeographically naïve alternatives, our approach yields markedly improved model performance, drastically different out-of-sample predictions, and substantially faster compute times.
In addition to presenting model results, this talk will discuss (in plain language) two Stan-specific coding challenges and their solutions:
1) How to marginalize over the (discrete) latent occupancy state, and
2) How to apply reduce_sum() to two-dimensional occupancy data in order to realize the speed-up associated with within-chain parallelization in Stan.

Presenter biography:
Jacob Socolar

Jacob Socolar (Norwegian University of Life Sciences) and Simon Mills (University of Sheffield) are postdoctoral researchers interested in avian responses to anthropogenic change. Both are currently working on biodiversity change at very large scales in Colombian tropical forests, and both have begun using Stan in earnest in the course of this project.