Consideration of functional traits is thought to offer promise in the quest to predict climate-induced geographic range shifts across taxonomic groups. Previous studies using linear models have found traits to be significant, but weak, predictors of the magnitude of range shifts. Functional traits such as thermal sensitivity, life history, diet specialization, and dispersal ability mediate biological responses that are nonlinear and involve thresholds. Since these responses are poorly described by linear models, we investigate nonlinear predictive modeling methods including multivariate adaptive regression splines, random forests, support vector regression, and neural networks. We evaluate individual traits’ contributions to range shifts when feasible for a method. Our evaluation framework leverages permutation testing and cross-validation to compare prediction errors on historic climate-induced range shift data for linear and nonlinear methods. We apply these approaches to a survey of elevational range shifts for 139 Swiss alpine plant species.
All nonlinear predictive methods offer increased predictive performance when compared to linear modeling. In particular, support vector regression techniques (when coupled with parameter tuning) most improved the potential to use traits to forecast range shifts in our Swiss alpine plants data. We find, using random forest analysis, that dispersal-related traits such as breeding mode offer the greatest predictive power. Our study highlights the benefits of nonlinear modeling techniques for capturing the influence of traits on responses to changing environments and suggests their potential for applications such as species distribution modeling and vulnerability analyses.