Measuring and selecting scales of effect for landscape predictors in species-habitat models

Wildlife managers often use habitat models to determine species habitat requirements and to identify locations for conservation efforts, uses which depend on accurate specification of species-habitat relationships. Prior study suggests that model performance may be influenced by the way we measure environmental predictors. We hypothesized that species responses to landscape predictors are best represented by landscape composition-based measurements, rather than distance-based measurements. We also hypothesized that models using empirical data to select an appropriate scale of effect for each habitat predictor (multi-scale models) should perform better than models that assume a common scale of effect for all predictors (single-scale models). To test these hypotheses we constructed habitat models for three mammal species, Mephitis mephitis, Mustela erminea, and Procyon lotor, based on surveys conducted in 80 landscapes in southeastern Ontario, Canada. For each species we compared the performance of distance- and composition-based measurements, and we compared the performance of single- and multi-scale models. The composition-based measurement, measured at its empirically determined scale of effect, had greater explanatory power than the distance-based measurement of a given predictor more often than expected by chance, supporting our first hypothesis. Contrary to expectation, multi-scale models did not have better explanatory power or predictive performance relative to single-scale models. We identified and evaluated four potential mechanisms to explain this, and, depending on the species, we found that the best explanation was either that predictors have significant effects at a common scale or that, although the modeled effects were at multiple scales, they were of similar magnitude and direction at the scales modeled in single- and multi-scale models. Our results suggest that habitat modeling based on distance-based measurements could be improved by including composition-based measurements of landscape predictor variables, but that inclusion of predictor-specific scales of effect for composition-based measurements does not necessarily improve performance over models including composition-based measurements at a single scale. Conservation and wildlife management may be simplified when single-scale models perform as well as multi-scale models, as this suggests actions conducted at a single scale may address management objectives as well as actions taken at different scales for different landscape features.