The Impact of (Not) Stratifying Analyses by Site When Randomization Was Stratified

• Activity Number: 637
• Type: Topic Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): John Seaman*+ and Brian Wiens
• Companies: Alcon and Alcon
• Keywords: non-inferiority ; clinical trials

Abstract

We consider analyses of clinical trials of novel medical devices when the randomized assignment of treatment to subject was stratified by investigative site. In pharmaceutical studies, it is common to stratify the analysis by the same factor(s) used to stratify the randomization. For medical devices, advice from regulators has often been to report an unstratified analysis as primary. To evaluate the impact, we report simulations of a clinical trial with a dichotomous primary endpoint and a non-inferiority hypothesis. When the outcome differs by site (due to differing surgeon skill, heterogeneous study subjects or other differences) the stratified analysis produced superior control of the type I error rate. Power was maintained or improved. When outcome did not differ by site, the stratified analysis maintained the type I error rate with minor loss of power. Assessment of consistency of results among sites is important to properly interpret the stratified analysis. We conclude that the stratified analysis has few drawbacks while providing important advantages in both size and power.