Using simulated data to explore the effects of spatial structure, sampling strategy, and statistical methods on species distribution models

The ability to map and monitor animal and plant species distributions has become more important in the context of awareness of environmental change and its effects on biodiversity (Franklin 1995; Guisan and Zimmermann 2000; Guisan and Thuiller 2005). Species distribution models (SDM) are based on the quantification of speciesenvironment relationships and have seen increasing use as the availability of geospatial technologies and spatial analysis tools expands. SDM require digital maps of important environmental variables, such as topography and climate, as well as spatial information on the species attribute of interest (e.g. presence/absence, type, abundance), usually from a sample of locations. Although used quite widely in ecological applications with often very sophisticated statistical techniques, most SDM are still developed without considering the spatial autocorrelation that is inherent in biogeographical data (Miller et al 2007). More traditional statistical methods used in SDM, such as generalized linear models (GLM) are typically based on the implicit assumption that the distribution of species is random and, therefore, each observation is independent. This assumption violates one of the basic tenets of geography, the direct relationship between distance and likeness, as well as of ecological theory, that elements of an ecosystem close to one another are more likely to be influenced by the same generating process and will therefore be similar. Ignoring spatial autocorrelation in biogeographical data can lead to poorly specified models in general and inflated significance estimates for predictor variables in particular (Legendre 1993).