Small-Sample Behavior of Robust Score and Wald Tests Arising from a Misspecifed Cox Proportional Hazards Model

• Activity Number: 180
• Type: Contributed
• Date/Time: Monday, August 4, 2003 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Eric S. Leifer*+ and Gregory DiRienzo and Stephen W. Lagakos and Eric V. Slud and David Hoberman
• Companies: National Heart, Lung & Blood Institute and Harvard School of Public Health and Harvard School of Public Health and University of Maryland and Food and Drug Administration
• Keywords: clinical trial ; Cox model ; robust tests

Abstract

In a randomized clinical trial, inference concerning treatment effect can be biased when based on a misspecified Cox model (Lagakos and Schoenfeld 1984). So-called robust variance estimators have been derived for settings where the source of such bias stems solely from the use of inconsistent model-based variance estimators (Lin and Wei 1989, Kong and Slud 1997, DiRienzo and Lagakos 2001). The resulting score and Wald tests are robust in the sense that they asymptotically preserve the nominal Type I error rate when the working Cox model is misspecified. It has been shown in simulations involving one or two covariates that such robust tests perform well for moderately sized clinical trials. Recently, there has been interest in the behavior of such tests when several covariates (e.g., 8-10) are included in the working Cox model, e.g., in investigational new drug (IND) submissions to the U.S. FDA. Using covariate data from actual clinical studies, we present simulation results that study the effects on such tests of covariate omission, adaptive variable-selection procedures, lagged treatment effects, and censoring that can depend on treatment group (of covariates).