Yuying Jin
The receiver operating characteristic (ROC) curve is a popular diagnostic tool assessing the diagnostic accuracy of a putative biomarker. When a biomarker is used for prognosis, often the marker is not measured concurrently with disease outcome. Further, the disease onset time can be censored. Therefore the classic concept of ROC curves needs to be extended to integrate both the time-varying nature of the marker and the clinical onset time of the disease. Naive estimates of predictive accuracy ignore the censoring can be biased. Incorporating the time dimension in ROC curve analysis has recently been an area of active research. We contrast the forms and attributes of various time-dependent ROC estimators discussed in the literature. The bias and efficiency of these estimators are investigated in simulation studies. Recommendations for their uses in practical situations are given.
