The Evaluation of Joint Predicted Risks from an Analysis of Semi-Competing Risks Data in the Presence of Censoring

• Activity Number: 168 - Risk analysis and related topics
• Type: Contributed
• Date/Time: Tuesday, August 10, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Section on Risk Analysis
• Author(s): Catherine Lee* and Sebastien Haneuse
• Companies: Kaiser Permanente, Division of Research and Harvard TH Chan School of Public Health
• Keywords: Semi-competing risks; multi-category outcomes; multi-category risk prediction; hypervolume under the manifold; extension of AUC

Abstract

We outline a unified approach to risk prediction and evaluation of predictive accuracy using semi-competing risks data (nonterminal event subject to dependent censoring by a terminal event). We propose to calculate and evaluate patient-specific absolute risk profiles for both events simultaneously. At any given point in time, a patient will have experienced: both events; the non-terminal event but not the terminal event; the terminal event without the non-terminal event; or neither event. We consider the task of prediction as being one where we seek to classify patients into one of four categories based on a vector of probabilities that add to 1. While the evaluation of predictions is well-established when the outcome is binary, less has been published on methods for multi-category outcomes and none specifically in the semi-competing risk setting. We propose a framework for evaluation of predictive performance for risk profiles based on the hypervolume under the manifold (HUM), an extension of the area-under-the-curve statistic for univariate binary outcomes. We propose a method for estimating the HUM statistic accounting for censoring when the true outcome category is unknown.