A0318
Title: Smooth time-dependent ROC curve estimators
Authors: Juan-Carlos Pardo-Fernandez - Universidade de Vigo (Spain) [presenting]
Pablo Martinez-Camblor - University of Oviedo (Spain)
Abstract: The receiver operating characteristic (ROC) curve is a popular graphical method often used to study the diagnostic capacity of continuous (bio)markers. When the considered outcome is a time-dependent variable, two main extensions have been proposed: the cumulative/dynamic ROC curve and the incident/dynamic ROC curve. In both cases, the main problem for developing appropriate estimators is the estimation of the joint distribution of the time-to-event variable and the marker. As usual, different approximations lead to different estimators. We explore the use of a bivariate kernel density estimator which accounts for censored observations in the sample and produces smooth estimators of the time-dependent ROC curves. The performance of the resulting cumulative/dynamic and incident/dynamic ROC curves is studied by means of Monte Carlo simulations. Additionally, the influence of the choice of the required bandwidth is explored. The obtained results suggest that the smooth estimators provide good approximations, specially when the area under the ROC curve is not too large.
