Comparative study of statistical methods for clustered ROC data: nonparametric methods and multiple outputation methods

Abstract

In clustered receiver operating characteristic (ROC) data each patient has several normal and abnormal observations. Within the same cluster, observations are naturally correlated. Several nonparametric methods have been proposed in the literature to handle clustered data structure, but their performances on simulated and real datasets have not been compared. Recently, a multiple outputation method has been considered for clustered data in areas other than diagnostic accuracy to account for within-cluster correlation. The multiple outputation method offers a resampling-based alternative for one sample clustered data with or without covariates, or for hypothesis testing in two sample clustered data. The method does not require a specific within-cluster correlation structure and yields a valid estimator while accounting for the within-cluster correlations. This paper contributes to the literature by introducing the multiple outputation method to the ROC setting, and empirically comparing the performance of these clustered ROC curve methods. The performance of these methods is also evaluated through two real examples.