Comparing covariate selection strategies for predictive dynamic occupancy models

Predicting how species will redistribute under global environmental change is a prerequisite for effective conservation. Dynamic occupancy models (DOMs) are increasingly popular tools for modelling range dynamics. DOMs estimate local colonisation and extinction probabilities while accounting for imperfect detection to capture the processes behind changing patterns of species occurrence. Covariate selection is particularly challenging for DOMs because four key parameters may be modelled as a function of covariates. Many techniques have been used for covariate selection with DOMs, but the relative merits of these for predictive performance have not been examined. We used long-term breeding bird survey data from Switzerland for 10 contrasting species to compare different covariate selection strategies in terms of the predictive performance of resulting models, computational efficiency, and consistency of estimated relationships. Using a shared set of candidate covariates, we attempted 39 alternative covariate selection workflows for each species, predicted site occupancy to hold-out years, and compared the outputs of each workflow. Covariate selection strategy affected predictive performance, although there was no overall best approach and differences tended to be relatively small. The largest drops in performance occurred when preliminary filtering stages of selection were used to reduce the size of the covariate pool. Selection strategies also resulted in different spatial predictions of site occupancy for some species, and methods were not consistent in which covariates were identified as important. In cases where DOMs are used to inform decision-making, covariate selection strategies should be applied in ways that capture the uncertainty around which model structure is best, and estimated ecological relationships should always be interpreted with caution. We encourage users to avoid unnecessarily filtering covariates prior to conducting full covariate selection to permit sufficiently complex models if supported by the available data.