Modeling the occurrence of bird species: Are the errors predictable?

Typical assessments of models where many species occurrences are predicted (e.g., from species-habitat matrices or Gap Analyses) report overall omission and commission errors. Yet species' attributes suggest that we may predict a priori that some species are more likely to be modeled correctly than others. Because the likelihood of modeling species correctly is related to species incidences in surveys, a method was created that ranked the 183 avian species known to be breeding in Maine as to how likely they would be to occur in surveys. Attributes (e.g., population level, niche width, aggregation) were used to model 79% of the variation in incidence within the Maine Breeding Bird Atlas. Likelihood of Occurrence Ranks (LOORs) were assigned to each species based upon the modeled incidences to reflect how likely the species are to be observed in future surveys. The occurrence of birds on areas with species checklists were then modeled and compared to the LOORs. For five of six areas, the number of species correctly modeled using species-habitat associations was highly correlated with LOORs: species judged a priori to be likely to be modeled correctly actually were. For one large area (9172 ha) with a checklist covering 52 years, the number of species correctly modeled was not correlated with LOORs, evidence that the checklist is essentially complete. In general, sites with checklists from many years (e.g., >10 yr) and from large areas (e.g., >1000 ha) yielded the lowest commission error. These results demonstrate that the confidence assigned to results where the occurrences of species are modeled (e.g., Gap Analysis) is highly dependent on the test sets and the species modeled.