Species distribution models (SDMs) project the outcome of community assembly processes - dispersal, the abiotic environment, and biotic interactions - onto geographic space. Recent advances in SDMs account for these processes by simultaneously modeling the species that comprise a community in a multivariate statistical framework or by incorporating residual spatial autocorrelation in SDMs. However, the effects of combining both multivariate and spatially-explicit model structures on the ecological inferences and the predictive abilities of a model are largely unknown. We used data on eastern hemlock  (Tsuga canadensisL.) and five additional co-occurring overstory tree species in 35,569 forest stands across Michigan, USA to evaluate how the choice of model structure, including spatial and non-spatial forms of univariate and multivariate models, affects ecological inference about the processes that shape community composition as well as model predictive ability.

            Incorporating residual spatial autocorrelation via spatial random effects did not improve out-of-sample prediction for the six tree species, although in-sample model fit was higher in the spatial models. Spatial models attributed less variation in occurrence probability to environmental covariates than the non-spatial models for all six tree species, and estimated higher (more positive) residual co-occurrence values for most species pairs. The non-spatial multivariate model was better suited for evaluating habitat suitability and hypotheses about the processes that shape community composition.  Environmental correlations and residual correlations among species pairs were positively related, perhaps indicating that residual correlations were due to shared responses to unmeasured environmental covariates. This work highlights the importance of choosing a non-spatial model formulation to address research questions about the species-environment relationship or residual co-occurrence patterns, and a spatial model formulation when within-sample prediction accuracy is the main goal.

Please see README.txt in the data package.