Inference on Correlated Discrimination Measures in Survival Analysis

The area under the receiver operating characteristic curve (overall C) is a widely used measure of a prognostic model discrimination. In this paper, we develop a nonparametric test for the comparison of two correlated C indexes of two different models when applied to the same population. We extend the DeLong's approach and use the theory of the jackknife methodology applied to correlated one sample generalized U-statistics. We derive the distribution of correlated estimators of C indexes as well as a consistent estimate of the asymptotic variance, leading to an asymptotically normal test.