Improved status and trend estimates from the North American Breeding Bird Survey using a hierarchical Bayesian generalized additive model

The status and trend estimates derived from the North American Breeding Bird Survey (BBS), are critical sources of information for bird conservation. However, many of the varied uses of these estimates are note well supported by the current suite of models. For example, inferences about population recovery require modeling approaches that are more sensitive than the standard model to changes in the rates of population change through time, such as change points and cycles. In addition, regional status assessments would benefit from models that allow for the sharing of information across the species’ range, to improve regional estimates. Here we describe hierarchical Bayesian generalized additive mixed-models (GAM) that fit these criteria, generating status and trend estimates optimized for many common uses related to conservation assessments. We demonstrate the models and their benefits using data for Barn Swallow Wood Thrush and a selection of other species, and we run a full cross-validation of the GAM against two other BBS models to compare predictive fit. We use a 15-fold cross-validation approach that provides a practical alternative to assessing predictive fit across the entire BBS dataset, while accounting for the spatial and temporal imbalances in the data. The GAMs have better predictive fit than the standard model for all species studied here, and better or comparable predictive fit compared to an alternative first difference model. In addition, one version of the GAM described here estimates a population trajectory that can be decomposed into a smooth component and the annual fluctuations around that smooth. Trajectories from this “GAMYE” model can be visualized either with or without the annual fluctuations, to suit particular research needs (e.g., visualizing patterns that may follow climatological cycles vs patterns that relate more to annual precipitation). This decomposition also allows trend estimates that are more stable between subsequent years because they remove the variation of the annual fluctuations and are therefore more useful for trend-based status assessments, such as those by the lUCN.