Predictive performance of plant species distribution models depends on species traits

Abstract

Abstract Predictive species distribution models are standard tools in ecological research and are used to address a variety of applied and conservation related issues. When making temporal or spatial predictions, uncertainty is inevitable and prediction errors may depend not only on data quality and the modelling algorithm used, but on species characteristics. Here, we applied a standard distribution modelling technique (generalized linear models) using European plant species distribution data and climatic parameters. Predictive performance was calculated using AUC, (Cohen’s) Kappa and true skill statistic (TSS), that were subsequently correlated with biological and life-history traits. After accounting for phylogenetic dependence among species, model performance was poorest for species having a short life span and occurring in human disturbed habitats. Our results clearly indicate that the performance of distribution models can be dependent on functional traits and provide further evidence that a species’ ecology is likely to affect the ability of models to predict its distribution. Biased and less reliable predictions could misguide policy decisions and the management and conservation of our natural heritage.