Species occurrence data are overwhelmingly presence-only, so pseudo-absences often serve as imperfect surrogates for true absences when training species distribution models. Unfortunately, this use of pseudo-absences is known to bias assessments of predictive skill. That is why MaxEnt is a notable exception. It avoids prediction bias by using pseudo-absences to sample the background distribution of environmental covariates. The most widely used metric to evaluate this difference is the area under the curve (AUC), which measures the ability of a model to discriminate presence from absence locations. Because the AUC assumes that pseudo-absences are true-absences, however, this method of estimating prediction bias is itself susceptible to bias. Here, we attempt to quantify bias in AUC estimation using simulation methods. We generate virtual worlds in which environmental suitability is defined by variation in the patchiness of a random Gaussian field and populate those worlds with virtual species defined by their functional response to each field. We then train a MaxEnt, RandomForest, and GLM (both Poisson and Binomial) for each species in each world using true and pseudo-absence points and evaluate them with the AUC, again using true and pseudo-absence points. We further compare AUC scores to other common metrics for evaluating models.
Results/Conclusions
Our results indicate that the use of pseudo-absences can introduce previously unrecognized bias in AUC, especially under certain environmental conditions. We, thus, recommend a complementary method to the AUC that accounts for this bias.