Effectiveness of predicting breeding bird distributions using probabilistic models

Conservation biologists need to be able to predict species distributions based on easily collected ata available at regional scales. We quantified the effectiveness of different types of data for predicting bird distributions for the state of Idaho. We developed probabilistic models to evaluate the ability of vegetation, climate, and spatial autocorrelation data to predict the presence or absence of 40 bird species. We determined the probability of correctly predicting presence and absence for each species using a training-testing sample methodology. This method involves splitting the data set into two portions: one portion was used to "fit" the models using maximum likelihood, and the second portion was then predicted from the fitted models. The predicted probability of species presence from the second portion of the data was compared to actual presence-absence values to assess model performance. Overall, differences in average performance among the parameterized models were small. Vegetation, climate, and spatial models each predicted approximately 60% of the presences correctly. Models employing a combination of these factors consistently improved model performance, but only slightly (an approximate 4% improvement). In contrast, the null model correctly predicted just 35% of the presences. Our results suggest that (1) parameterized models are a substantial improvement over a null model but still make frequent mistakes in predicting the presence or absence of species, and (2) data availability may be the most important factor in determining which variables to use to predict species presence-absence. In some cases, available and relatively inexpensive climate data or incomplete distributional information may be the preferred data option.