There are a variety of statistical methods to predict the presence or absence of disease in medical research. When a new factor is added to the existing prediction model, the improvement of the predictive power can be evaluated using c-index or net re...
There are a variety of statistical methods to predict the presence or absence of disease in medical research. When a new factor is added to the existing prediction model, the improvement of the predictive power can be evaluated using c-index or net reclassification index (NRI). For binary outcome data, a number of studies show that NRI is more sensitive to detect an improvement in predictive ability of the added factor. For survival outcome data, however, such an evaluation has not been done thoroughly. In this study, we compare several methods estimating the improvement of the added predictive power for survival outcome data through a simulation study under various scenarios. We consider c-index (Harrell, 1982), iAUC (integrated area under the ROC curve) (Heagerty, 2005), c-index (Uno, 2011), NRI (Pencina, 2011), and NRI (Uno, 2013). The simulation study results demonstrate that there is an inflation of the type I error rate for c-index and iAUC, while NRIs are very conservative. We also compare the results of the five methods applied to a real data example.