Exploring the effects of quantity and location of pseudo-absences and sampling biases on the performance of distribution models with limited point occurrence data

SummaryIn the last decade, the application of predictive models of species distribution in ecology, evolution, and conservation biology has increased dramatically. However, limited available data and the lack of reliable absence data have become a major challenge to overcome. At least two approaches have been proposed to generate pseudo-absences; however it is not clear how the number of pseudo-absences created affect model performance. Moreover, the spatial bias in the collecting localities of a species (presence data) may add extra noise to the final distribution model. Here, we use a virtual species to assess the effects of spatial sampling bias, and number and location of pseudo-absences on model accuracy. We found that both number of pseudo-absences and spatial bias in sampling localities, as well as their interaction, significantly influence all accuracy measures (AUC, sensitivity, and specificity). However, location of pseudo-absences (either generated across the entire study area or only outside the environmental envelope of the species) does not affect model performance. These results provide some methodological guidelines for developing reliable distribution hypotheses when presence data are scarce.