The distributions of many organisms are spatially autocorrelated, but it is unclear whether including spatial terms in species distribution models (SDMs) improves projections of future species distributions. We provide one of the first comparative evaluations of the ability of a purely spatial SDM, a purely non-spatial SDM, and a SDM that combines spatial and environmental information to project species distributions across eight millennia of climate change. To distinguish between the importance of climatic versus spatial explanatory variables, we fit three Bayesian models to modern occurrence data of Fagus and Tsuga, two tree genera whose distributions can be reliably inferred from fossil pollen in Eastern North America: a spatially-varying intercept model, a non-spatial model with climatic variables, and a full model with climate variables and spatial terms. Using high temporal resolution paleoclimate data, we hindcasted the models for 8000 years, and compared model projections with palynological data. For both taxa, spatial SDMs provided better fits to the calibration data, more accurate predictions of a hold-out validation dataset of modern trees, and when projected lower false positive rates at all time periods than non-spatial SDMs. Hindcasted projection of spatial SDMs had higher variance than those of non-spatial SDMs. Overall performance of non-spatial and spatial SDMs varied temporally and as a function of niche overlap.
Including spatial terms in projected SDMs reduced false positive rates, perhaps by accommodating serial dependence among the model residuals or by accounting for missing environmental factors and/or biological processes that determine species distributions. Therefore, spatially explicit SDMs, though computationally demanding, may lead to improved projections of species responses to climate change.