Spatial and temporal non‐stationarity in long‐term population dynamics of over‐wintering birds of North America

Understanding population changes across long time scales and at fine spatiotemporal resolutions is important for confronting a broad suite of conservation challenges. However, this task is hampered by a lack of quality long‐term census data for multiple species collected across large geographic regions. Here, we used century‐long (1919–2018) data from the Audubon Christmas Bird Count (CBC) survey to assess population changes in over 300 avian species in North America and evaluate their temporal non‐stationarity. To estimate population sizes across the entire century, we employed a Bayesian hierarchical model that accounts for species detection probabilities, variable sampling effort, and missing data. We evaluated population trends using generalized additive models (GAMs) and assessed temporal non‐stationarity in the rate of population change by extracting the first derivatives from the fitted GAM functions. We then summarized the population dynamics across species, space, and time using a non‐parametric clustering algorithm that categorized individual population trends into four distinct trend clusters. We found that species varied widely in their population trajectories, with over 90% of species showing a considerable degree of spatial and/or temporal non‐stationarity, and many showing strong shifts in the direction and magnitude of population trends throughout the past century. Species were roughly equally distributed across the four clusters of population trajectories, although grassland, forest, and desert specialists more commonly showed declining trends. Interestingly, for many species, region‐wide population trends often differed from those observed at individual sites, suggesting that conservation decisions need to be tailored to fine spatial scales. Together, our results highlight the importance of considering spatial and temporal non‐stationarity when assessing long‐term population changes. More generally, we demonstrate the promise of novel statistical techniques for improving the utility and extending the temporal scope of existing citizen science datasets.