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