Web-Bootstrap Estimate of Area Under ROC Curve

Abstract

The accuracy of binary discrimination models (discrimination between cases with and without any condition) is usually summarized by classification matrix (also called a confusion, assignment, or prediction matrix). Receiver operating characteristic (ROC) curve can visualize the association between probabilities of incorrect classification of cases from the group without condition (False Positives) versus the probabilities of correct classification of cases from the group with condition (True Positives) across all the possible cut-point values of discrimination score. Area under ROC curve (AUC) is one of summary measures. This article describes the possibility of AUC estimate with the use of web based application of bootstrap (resampling). Bootstrap is useful mainly to data for which any distributional assumptions are not appropriate. The quality of the bootstrap application was evaluated with the use of a special programme written in C#.NET language that allows to automate the process of repeating different experiments. Estimates of AUC and confidence limits given by bootstrap method were compared with bi-normal and nonparametric estimates. Results indicate that usually bootstrap confidence intervals are narrower than nonparametric one, mainly for small data samples.