Defining the scale of habitat availability for models of habitat selection.

Statistical models of habitat preference and species distribution (e.g., Resource Selection Functions and Maximum Entropy approaches) perform a quantitative comparison of the use of space with the availability of all habitats in an animal's environment. However, not all of space is accessible all of the time to all individuals, so availability is in fact determined by limitations in animal perception and mobility. Therefore, measuring habitat availability at biologically relevant scales is essential for understanding preference, but herein lies a trade-off: Models fitted at large spatial scales, will tend to average across the responses of different individuals that happen to be in regions with contrasting habitat compositions. We suggest that such models may fail to capture local extremes (hotspots and coldspots) in animal usage and call this potential problem, homogenization. In contrast, models fitted at smaller scales will vary stochastically depending on the particular habitat composition of their narrow spatial neighborhood, and hence fail to describe responses when predicting for different sampling instances. This is the now well-documented issue of non-transferability of habitat models. We illustrate this tradeoff, using a range of simulated experiments, incorporating variations in environmental gradients, richness and fragmentation. We propose diagnostics for detecting the two issues of homogenization and non-transferability and show that these scale-related symptoms are likely to be more pronounced in highly fragmented or steeply graded landscapes. Further, we address these problems by treating the neighborhood of each cell in the landscape grid as an individual sampling instance (with its own neighborhood), hence allowing coefficients to respond to the local expectations of environmental variables according to a Generalized Functional Response (GFR). Under simulation this approach is consistently better at estimating robust (i.e., transferable) habitat models at smaller scales, and less susceptible to homogenization at larger scales. At the same time, it represents the first application of a GFR to continuous space (rather than multiple, spatially distinct datasets), allowing the predictive advantages of this extension of species distribution models to become available to data from large-scale but single-site field studies.