Species distribution models generally fail to account for the multiple processes driving species distributions, the different spatial scales at which these act, and the inherent ambiguity of species absence data. This has led to criticism of species distribution models as possessing weak statistical power and producing unrealistically narrow estimates of potential range. This is an important problem because such models are widely used to design reserves and predict species responses to climate change. We present hierarchical Bayesian regression models that improve on single-level models by distinguishing, at different model levels, among three distinct factors influencing presence/absence and abundance: local environment, land use, and neighborhood or spatial processes. Our goal is to integrate the increasingly available large data sets on species distributions to build down from the broad regional spatial scale of species distribution data down toward population-level processes that control the local performance of species.

Results/Conclusions

We show that by using hierarchical models to analyze data from South Africa’s Cape Floristic Region, we can assimilate large data sets from multiple sources (including volunteer-collected survey data, remote-sensed data, and traditional environmental GIS layers) at different scales, and partition the explanatory power of the models among different processes. The models produce more robust estimates of the environmental tolerances of species, while also predicting their potential distributions in the absence of land use and/or dispersal limitation. Generally, the models predict larger potential ranges than those typically used to project climate change impacts. Comparison with data from field experiments provides an important validation of these predictions. We also find that generally, spatial neighborhood effects play a more important role in explaining the distributions of localized endemic species than those of widespread species.  Finally, we find that the hierarchical models generally outperform single-level distribution models when fed the same data.