Conservation biologists need to be able to predict species distributions ba
sed 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 train
ing-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 model
s. The predicted probability of species presence from the second portion of
the data was compared to actual presence-absence values to assess model pe
rformance. Overall, differences in average performance among the parameteri
zed models were small. Vegetation, climate, and spatial models each predict
ed approximately 60% of the presences correctly. Models employing a combina
tion of these factors consistently improved model performance, but only sli
ghtly (an approximate 4% improvement). In contrast, the null model correctl
y predicted just 35% of the presences. Our results suggest that (1) paramet
erized models are a substantial improvement over a null model but still mak
e 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, availa
ble and relatively inexpensive climate data or incomplete distributional in
formation may be the preferred data option.