Overall AUC for Survival Models

• Activity Number: 87 - Survival and Longitudinal/Clustered Data Analysis
• Type: Contributed
• Date/Time: Monday, August 9, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Biometrics Section
• Author(s): Wenna Xi* and Yiyuan Wu and Hongzhe Zhang and Samprit Banerjee
• Companies: Weill Cornell Medicine and Weill Cornell Medicine and Weill Medical College, Cornell University and Weill Medical College of Cornell University
• Keywords: survival analysis; AUC; time-dependent AUC; time-dependent ROC; overall AUC; concordance index

Abstract

In survival analysis, time-dependent ROC (receiver operating characteristic) curve and subsequently time-dependent AUC (area under the ROC curve) are introduced to evaluate the performance of survival models at different time points. Since the first proposal of time-dependent AUC by Heagerty and Zheng in 2005, different versions of time-dependent AUCs and overall summary measures of time-dependent AUCs have been proposed. However, the differences (e.g., interpretations, strengths, limitations) between each version of time-dependent AUCs have not been well studied, resulting in the fact that certain R packages provide estimate of time-dependent AUCs without specifying the version of AUCs that is estimated, which may confuse users who are not familiar with the multiple versions of time-dependent AUCs. In this presentation, we aim to fill in the gap by summarizing and comparing different versions of time-dependent AUCs, and for some versions of time-dependent AUCs, we also provide a global measure named “overall AUC,” which includes the concordance index as a special case. We provide both simulations and real data analysis to illustrate the differences of each time-dependent AUCs.