Stratified Time-to-Event Trials: Is There a Best Analysis?

• Activity Number: 89
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Devan V. Mehrotra*+ and Shu-Chih Su and Xiaoming Li
• Companies: Merck Research Laboratories and Merck & Co., Inc. and Merck Research Laboratories
• Keywords: Cox Model ; Interaction ; Minimum Risk Weights ; Stratification ; Survival Analysis ; Time-to-Event

Abstract

Consider a typical two-treatment randomized clinical trial involving a time-to-event endpoint, with randomization stratified by a categorical covariate (e.g., gender). At the design stage, it is often assumed that the treatment hazard ratio (HR) is constant across the strata, and the data are commonly analyzed using the stratified Cox proportional hazards model. In this talk, we caution that this common approach is needlessly risky because departures from the constant HR assumption can result in a costly loss in power. We propose an alternate approach in which the [log] HR is estimated separately for each stratum using an unstratified Cox model, and the stratum-specific estimates are then combined to obtain an overall estimate using either sample size or "minimum risk" (Mehrotra and Railkar, 2000) stratum weights. We quantify the potential efficiency gains of the proposed approach versus the common Cox model analysis using realistic simulations. The design of a vaccine mega-trial is used to motivate the problem.