Sensitivity of species habitat-relationship model performance to factors of scale

Researchers have come to different conclusions about the usefulness of habi tat-relationship models for predicting species presence or absence. This di fference frequently stems from a failure to recognize the effects of spatia l scales at which the models are applied. We examined the effects of model complexity, spatial data resolution, and scale of application on the perfor mance of bird habitat relationship (BHR) models on the Craig Mountain Wildl ife Management Area and on the Idaho portion of the U.S. Forest Service's N orthern Region. We constructed and tested BHR models for 60 bird species de tected on the study areas. The models varied by three levels of complexity (amount of habitat information) and three spatial data resolutions (0.09 ha , 4 ha, 10 ha). We tested these models at two levels of analysis: the site level (a homogeneous area <0.5 ha) and cover-type level tan aggregation of many similar sites of a similar (and-cover type), using correspondence betw een model predictions and species detections to calculate kappa coefficient s of agreement. Model performance initially increased as models became more complex until a point was reached where omission errors increased at a rat e greater than the rate at which commission errors were decreasing. Heterog eneity of the study areas appeared to influence the effect of model complex ity. Changes in model complexity resulted in a greater decrease in commissi on error than increase in omission error. The effect of spatial data resolu tion on the performance of BHR models was influenced by the variability of the study area. BHR models performed better at cover-type levels of analysi s than at the site level for both study areas. Correct-presence estimates ( 1 - minus percentage omission error) decreased slightly as number of specie s detections increased on each study area. Correct-absence estimates (1 - p ercentage commission error) increased as number of species detections incre ased on each study area. This suggests that a large number of detections ma y be necessary to achieve reliable estimates of model accuracy.