However, for the purpose of this study, we want to evaluate the probabilities generally rather than under specific conditions or under ranges of conditions. A subtle issue arises when evaluating the quality of the probabilities in our setting: although the trees are estimating probabilities of class membership, neither for the training data nor for the test data do we know the true probabilities. All we know is the true class of each example. For this paper, our task is simplified because all we address is how well the estimated probabilities rank cases