ROC CURVES FOR CLASSIFICATION TREES

A common problem in medical diagnosis is to combine information from s everal tests or patient characteristics into a decision rule to distin guish diseased from healthy patients. Among the statistical procedures proposed to solve this problem, recursive partitioning is appealing f or the easily-used and intuitive nature of the rules it produces. The rules have the form of classification trees, in which each node of the tree represents a simple question about one of the predictor variable s, and the branch taken depends on the answer. The authors consider th e role of misclassification costs in developing classification trees. By varying the ratio of costs assigned to false negatives and false po sitives, a series of classification trees are generated, each optimal for some range of cost ratios, and each with a different sensitivity a nd specificity. The set of sensitivity-specificity combinations define a curve that can be used like an ROC curve.