A Within-Subjects Permutation Approach for Comparing Predictive Ability of Two Continuous Predictors for a Survival Outcome

• Activity Number: 363
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Ming-Wen An* and Gwenael Le Teuff and Stefan Michiels
• Companies: Vassar College and Université Paris Sud and Université Paris Sud
• Keywords: predictive ability ; non-nested ; c-index ; resampling ; permutation ; Cox

Abstract

We consider the question of how to compare the predictive ability of two candidate continuous predictors of a survival outcome. This is distinct from a related question to assess added predictive ability of a predictor to an existing set of predictors, which has been well-studied. Resampling-based methods have been proposed for testing an association between a predictor and an outcome. However, the null sampling distribution generated with such methods corresponds to no association between a predictor and the outcome. In contrast, the null hypothesis for our question is that the two predictors have the same predictive ability for a survival outcome. We thus propose a within-subjects permutation approach to generate the null sampling distribution. We conducted a simulation study for a variety of hypotheses and used different predictive ability measures (c-index, Brier score and iAUC) with the Cox proportional hazards model. As opposed to standard approaches (e.g. jackknife or bootstrap), our test showed valid type I error rates and power. We illustrate this approach with competitive gene signature data for predicting survival in early breast cancer.