A Unified Approach to Nonparametric Comparison of Receiver Operating Characteristic Curves for Longitudinal and Clustered Data

• Activity Number: 100
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2007 : 8:30 AM to 10:20 AM
• Sponsor: WNAR
• Author(s): Gang Li and Kefei Zhou*+
• Companies: University of California, Los Angeles and Amgen Inc.
• Keywords: ROC curve ; Nonparametric comparison ; clustered data ; quantile process ; partial area under the curve ; confidence band

Abstract

This article presents a unified nonparametric method for comparisons of ROC curves with clustered data. Most previous methods are developed for a single index, and the asymptotic variance usually involves unknown densities whose nonparametric estimates are unstable. We derive the joint asymptotic distribution of correlated empirical ROC curves. We then develop a Monte Carlo method to approximate their joint distribution without involving density estimation. We prove that the Monte Carlo method is consistent. The theory is applied to derive inferential procedures for comparing sensitivities at a fixed specificity and weighted areas under ROC curves, construct confidence bands for the difference of ROC curves, estimate the set of specificities at which a diagnostic test is more sensitive than the other. Simulations showed that our methods work well for small samples.