Overrating Classifier Performance in ROC Analysis in the Absence of a Test Set: Evidence from Simulation and Italian CARATkids Validation

Abstract

Background The use of receiver operating characteristic curves, or “ROC analysis,” has become quite common in biomedical research to support decisions. However, sensitivity, specificity, and misclassification rates are still often estimated using the training sample, overlooking the risk of overrating the test performance. Methods A simulation study was performed to highlight the inferential implications of splitting (or not) the dataset into training and test set. The normality assumption was made for the classifier given the disease status, and the Youden's criterion considered for the detection of the optimal cutoff. Then, an ROC analysis with sample split was applied to assess the discriminant validity of the Italian version of the Control of Allergic Rhinitis and Asthma Test (CARATkids) questionnaire for children with asthma and rhinitis, for which recent studies may have reported liberal performance estimates. Results The simulation study showed that both single split and cross-validation (CV) provided unbiased estimators of sensitivity, specificity, and misclassification rate, therefore allowing computation of confidence intervals. For the Italian CARATkids questionnaire, the misclassification rate estimated by fivefold CV was 0.22, with 95% confidence interval 0.14 to 0.30, indicating an acceptable discriminant validity. Conclusions Splitting into training and test set avoids overrating the test performance in ROC analysis. Validated through this method, the Italian CARATkids is valid for assessing disease control in children with asthma and rhinitis.