Improving range shift predictions: enhancing the power of traits

Accurately predicting species range shifts in response to environmental change is a central ecological objective and applied imperative. In synthetic analyses, traits emerge as significant but weak predictors of species range shifts across recent climate change. These studies assume linearity in the relationship between a trait and its function, while detailed empirical work often reveals unimodal relationships, thresholds, and other nonlinearities in many trait-function relationships. We hypothesize that the use of linear modeling approaches fails to capture these nonlinearities and therefore may be under-powering traits to predict range shifts. We evaluate the predictive performance of four different machine learning approaches that can capture nonlinear relationships (ridge-regularized linear regression, ridge-regularized kernel regression, support vector regression, and random forests). We validate our models using four multi-decadal range shift datasets in montane plants, montane small mammals, and marine fish. We show that nonlinear approaches perform substantially better than least-squares linear modeling in reproducing historical range shifts. In addition, using novel model observation and interrogation techniques, the trait classes (e.g. dispersal-or diet-related traits) that we identify as primary drivers of model predictions are consistent with expectations. However, disagreements among models in the directionality of trait predictors suggests limits to trait-based statistical predictive frameworks.