Interaction Effects Do Not Consistently Improve the Transferability of Species Distribution Models.

Species responses to the environment are often non-stationary, varying in space and time depending on the environmental context. This is one of the reasons why species distribution models (SDMs) often have low transferability. Context dependence may arise due to frequent and widespread interactive effects of environmental drivers on species occurrence. Thus, accounting for interactions in SDMs could in principle improve their transferability. Alternatively, incorporating driver interactions may lead to overfitting issues and reduce model transferability. We systematically assessed the extent to which interactions improve model transferability across taxa and biomes. We fitted MaxEnt and Boosted Regression Trees (BRT) models to presence data of 1158 species and evaluated their transferability to (1) independent native areas of plants, reptiles, birds and bats, (2) exotic ranges of plant species and (3) exotic ranges of bird species. While the performance of both model types (MaxEnt and BRT) was highly correlated, gains or losses in transferability when incorporating interactions were not consistent between model types. Incorporating interactions did not have any consistent impacts on transferability in 84% of the models according to a concordance (between model types) and significance criterion. In the remaining models, similar percentages of models had improved and worse transferability (~8% each). We did not detect any relevant effects of taxon, dataset, sample size, sampled area and environmental novelty on transferability gains when incorporating predictor interactions. Overall, uninformed predictor interactions modelled via widely used machine learning algorithms did not increase SDM transferability, suggesting that modelling routines should carefully consider their specification and usefulness for model extrapolation.