Aim Predictive models of species occurrence have potential for prioritizing areas for competing land uses. Before widespread application, however, it is necessary to evaluate performance using independent data and effective accuracy measures. The objectives of this study were to (1) compare the effects of species occurrence rate on model accuracy, (2) assess the effects of spatial and temporal variation in occurrence rate on model accuracy, and (3) determine if the number of predictor variables affected model accuracy.
Location We predicted the distributions of breeding birds in three adjacent mountain ranges in the Great Basin (Nevada, USA).
Methods For each of 18 species, we developed separate models using five different data sets — one set for each of 2 years (to address the effects of temporal variation), and one set for each of three possible pairs of mountain ranges (to address the effects of spatial variation). We evaluated each model with an independent data set using four accuracy measures: discrimination ability [area under a receiver operating characteristic curve (AUC)], correct classification rate (CCR), proportion of presences correctly classified (sensitivity), and proportion of absences correctly classified (specificity).
Results Discrimination ability was not affected by occurrence rate, whereas the other three accuracy measures were significantly affected. CCR, sensitivity and specificity were affected by species occurrence rate in the evaluation data sets to a greater extent than in the model-building data sets. Discrimination ability was the only accuracy measure affected by the number of variables in a model.
Main conclusions Temporal variation in species occurrence appeared to have a greater impact than did spatial variation. When temporal variation in species distributions is great, the relative costs of omission and commission errors should be assessed and long-term census data should be examined before using predictive models of occurrence in a management setting.