Comparing models for predicting species' potential distributions: a case study using correlative and mechanistic predictive modelling techniques

Models used to predict species' potential distributions have been described as either correlative or mechanistic. We attempted to determine whether correlative models could perform as well as mechanistic models for predicting species potential distributions, using a case study. We compared potential distribution predictions made for a coastal dune plant (Scaevola plumieri) along the coast of South Africa, using a mechanistic model based on summer water balance (SWB), and two correlative models (a profile and a group discrimination technique). The profile technique was based on principal components analysis (PCA) and the group-discrimination technique was based on multiple logistic regression (LR). Kappa (kappa) statistics were used to objectively assess model performance and model agreement. Model performance was calculated by measuring the levels of agreement (using kappa) between a set of testing localities (distribution records not used for model building) and each of the model predictions. Using published interpretive guidelines for the kappa statistic, model performance was "excellent" for the SWB model (kappa = 0.852), perfect for the LR model (kappa = 1.000), and "very good" for the PCA model (kappa = 0.721). Model agreement was calculated by measuring the level of agreement between the mechanistic model and the two correlative models. There was "good" model agreement between the SWB and PCA models (kappa = 0.679) and "very good" agreement between the SWB And LR models (kappa = 0.786). The results suggest that correlative models can perform as well as or better than simple mechanistic models. The predictions generated from these three modelling designs are likely to generate different insights into the potential distribution and biology of the target organism and may be appropriate in different situations. The choice of model is likely to be influenced by the aims of the study, the biology of the target organism, the level of knowledge the target organism's biology, and data quality. (C) 2003 Elsevier Science B.V. All rights reserved.