Beyond a diagnostic tool: Validating standardized Mahalanobis distance as a species distribution model for invasive alien species in North America

Species distribution models (SDMs) are useful tools for predicting where new invasive species can establish within a country, supporting both preparatory and response activities. National Plant Protection Organizations use SDMs to inform risk assessment and surveillance activities for emerging plant pests. However, SDMs face multiple and difficult statistical challenges, including multi-collinearity of input variables, correlation structures in climate variables that vary through time and space, limited species observation data, and they often lack formal tests of model performance. We have implemented a previously-reported extrapolation-detection tool as an SDM, rather than a diagnostic tool of SDMs. This method characterizes the observed multivariate climate envelope by using Mahalanobis distance to take advantage of the correlation between climate variables, and identifies areas where the climatic conditions are outside the range of the observed climate envelope. Model outputs include climate suitability maps, and most-important covariate analyses to identify the environmental drivers of the results while assisting in variable reduction. We performed a formal test to assess the ability of the SDM to identify areas invaded by invasive plant pests in North America. Using a list of 23 species from the Canadian Food Inspection Agencys regulated plant pest list, we demonstrate that this method achieves a high level of accuracy (> 85%) for determining climate suitability for North American plant pest invasions, especially when combined with most-important covariate-guided variable reduction. This suggests that the model is suitable for identifying areas of North America that are susceptible to future invasions. We show that many of the errors occur at the edge of climate suitable areas, where we would expect greater uncertainty in model predictions due to potential over-constraining and geospatial averaging. We present additional analyses to support recommendations on the use and limitations of this SDM in a regulatory context.