NOTE: This is very much a work in progress. It often changes daily. I am happy to hear suggestions and questions.
A spatial iCAR model to estimate route-level trends from BBS data. One goal is to generate route-level trend estimates that could be used as data in a subsequent exploration of possible covariates of trends and abundances. Another goal is to integrate some covariates into the model directly.
The model is written in Stan. It is based on the bbsBayes, slope model (e.g., Sauer and Link 2011), but without the random year-effects, and with no particular stratification.
I've relied strongly on elements from these two excellent case studies in Stan:
-
This intrinsic CAR model structure for the BYM model from Mitzi Morris and co-authors.
-
I'm working at some elaborations of the model that would incorporate a route-level covariate based on this exact sparse CAR model from Maxwell Joseph.
It's effectively an over-dispersed Poisson regression model with random slopes representing the trends at each route, random intercepts representing the mean abundance (mean count) on each route, as well as the among and within observers effects (true observers, see below)
I've applied this model to the BBS for ~400 species. Estimated slopes, trends, and intercepts, for every route and species are available to explore in a personal Google Drive. The drive includes a readme file that explains the contents as well as trend maps for all species.
The observer effects here are random effects for each observer, not the observer-route combinations, used in the bbsBayes models. This is a new thing for the hierarchical Bayesian BBS models. The MCMC algorithm of JAGS and BUGS has a great deal of trouble separately estimating observer effects from route-level intercepts. The HMC algorithm in Stan, plus the spatially explicit estimates of the route-level intercepts, appears to have much more success.
Baird's Sparrow trends in the Great Plains. Dots represent BBS routes on which Baird's Sparrow has been observed between 2004 and 2019. Over this time period, there's an interesting spatial pattern: declines in the Northeast, increases in the Southwest.
Chipping Sparrow trends over the last 15 years suggest the species has been increasing in the St-Lawrence River valley, in the southeastern portion of its range, and through some parts of the Great Plains.
Wood Thrush trends over the last 15 years suggest the species has recently increased in the eastern parts of its range, and some portions of the Appalachians, but generally decreased in the regions where it is most abundant.
