|
Abstract:
|
Randomized clinical trials use either stratified or unstratified randomization. For the former, the stratification factors are typically categorical baseline covariates (region, age group, ECOG status, etc.) that are presumed to influence the clinical endpoint of interest. We caution that uncertainty at the trial design stage can contribute to "ineffective" stratification and the corresponding stratified analysis can lead to an adversely biased or imprecisely estimated treatment effect, especially for trials designed to assess whether a test treatment prolongs survival relative to a control treatment. To mitigate this non-trivial risk, we show how “effective” stratification can be achieved using a pre-specified treatment-blinded algorithm applied to the clinical trial outcomes, followed by a novel power-boosting stratified analysis after treatment unblinding. We illustrate the utility of our proposal vs. current practice using a graphical summary of p-values and hazard ratio estimates from 23 real data examples. We also discuss alignment of our proposal with FDA guidance on covariate-adjusted analyses, and with related publications by John Tukey, Stuart Pocock, and others.
|
