In using the 30 new cases to assess the performance of the models on an independent dataset, the highest probability was used to assign each case to a class. When the models were used to determine the exact severity class of the new cases, CART correctly classified 17 of the 30 validation cases. Nine of erroneous classifications were due to overestimation and four to underestimation of disease severity class. When used to estimate the probability of having disease seventy greater than 20% (class >1), 73.3% of the cases were correctly classified (Table
2.3), whereas four cases each were misclassified as being in class 1 and in a class other than 1. The PO model correctly classified 66.7% of the cases when used to predict the exact class to which a case belonged, and 73.3% of the cases when predicting the probability of being in a class higher than 1. In both instances, most of the misclassified cases were predicted as being in a class lower than their actual class, eight and six, respectively, and two cases were misclassified as being in a higher class. Similarly, the ECR model correctly classified 18 of the 30 cases when used to estimate the exact class and 21 when used to assign cases to a class above class 1. Most of the misclassification was due to underestimation.