Print PageClassification and Regression Tree (CART) models provide a flexible way to relate species presence/absence to habitat characteristics, but standard CART algorithms ignore imperfect detection. When imperfect detection is ignored, present but not seen is treated as absent. We describe a new tree splitting algorithm based on a likelihood ratio test. The likelihood function accounts for imperfect detection. We describe four possible likelihood functions that correspond to different assumptions about how detection changes over the classification tree.
Results/Conclusions
We illustrate the analysis with data on mountain plovers. Three of the methods produce the same tree structure as the naive tree that ignores detection; the fourth does not. Estimated occupancy probabilities are consistently lower in a detection-adjusted tree.
