Testing the effect of sample prevalence and sampling methods on probability- and favourability-based SDMs

Predicting the occurrence probability of species is intrinsically dependent on the quality of the training dataset and, in particular, on the sample prevalence (i.e., the ratio between presences and absences). Whenever the number of presences and absences is not equal within the training dataset, the predictions deviate towards higher values as the sample prevalence increases and vice versa. As a result, probability models of species occurrence with different sample prevalence cannot be directly compared. The favourability concept was introduced to amend this limitation. Indeed, the favourability - i.e., the variation in the probability of occurrence regardless the sample prevalence - could reduce the degree of uncertainty when comparing species distributions despite different sample prevalences. To test this hypothesis, we simulated 50 virtual species and compared the predictive performance of four probability-based and favourability-based Species Distribution Models (GLM, GAM, RF, BRT) under a set of different prevalence values and sampling strategies (i.e, random and stratified sampling). Favourability-based models performed slightly better than probability-based models in predicting the species distribution over geographic space, confirming also their capability to reduce the variability of the predictions across different degrees of sample prevalence.