Venturing Into the Unknown: The Importance of Variable Selection When Modelling Alien Species Under Non-Analogue Climatic Conditions

Species distribution models (SDMs) are widely used to address species' responses to bioclimatic conditions in the fields of ecology, biogeography and conservation. Among studies that have addressed reasons for model prediction variability, the impact of climatic variable selection has received limited attention and is rarely assessed in sensitivity analyses. Here, we tested the assumption that this aspect of model design is a major source of uncertainty, especially when projections are made to non-analogue climates. As a study system, we used 142 alien plant species introduced to the sub-Antarctic islands. Based on global occurrence data, we fitted SDMs as functions of seven bioclimatic variable sets that only differed in the identity of two temperature variables. Moreover, we calculated the overlap between the island's climatic conditions and the niches the species have realised outside of the islands. Despite comparable internal evaluation metrics, projections of these models were in sharp contrast with each other, with some models predicting the sub-Antarctic islands' climate to be almost ubiquitously suitable to most species and others unsuitable to almost all species. In particular, the mean temperature of the warmest month led to strong underpredictions of the SDMs, while its replacement by the mean temperature of the coldest month led to massive overpredictions. Partitioning the variance in projections demonstrated that predictor identity was its most important source, followed by island and species identity. The size of area projected to be suitable was also related to the overlap in predictor values realised in the global range of species (outside of the islands) and on the islands. Our findings emphasise the importance of bioclimatic variable selection in SDMs, especially when making projections to non-analogue climates. Such extrapolations are often required, especially when using SDMs to assess invasion risk under both current and future climates.