Species distribution models (SDMs) can be thought of as projecting the outcome of community assembly processes - dispersal, the abiotic environment, and biotic interactions - onto geographic space. Many current improvements to SDMs focus on accounting for these processes by simultaneously modeling the species that comprise a community in a multivariate statistical framework or by accounting for residual spatial autocorrelation in SDMs. However, the effects of combining both multivariate and spatially-explicit model structures on both the ecological inference that can be made and model predictive ability are largely unknown. We evaluated how the choice of model structure affects ecological inference about the processes that shape community composition, as well as model predictive ability, using data on eastern hemlock (Tsuga canadensis) and five additional canopy tree species in 35,569 forest stands across Michigan, USA.
Results/Conclusions
We found that incorporating residual spatial autocorrelation via spatial random effects dramatically improved model predictive ability for all tree species, but the spatial models attributed substantially less variation in occurrence probability to environmental covariates for five of the six species. The non-spatial multivariate model was better suited for evaluating hypotheses about the processes that shape community composition. We found considerable variation in shared environmental responses and residual correlations among species. Environmental correlations and residual correlations among species-pairs were positively related, perhaps indicating shared responses to unmeasured environmental covariates.