Parallel Hypothesis Testing for Science and Industry in a Large Randomized Clinical Trial
The Carotid Revascularization Endarterectomy/Stenting Trial (CREST) will assess the relative efficacy of carotid endarterectomy (CEA) and Carotid Angioplasty with Stenting (CAS) in the prevention of stroke. As other trials, data from CREST are being used to meet different aims. The NIH and scientific community are interested in establishing if CEA and CAS differ in their long-term efficacy to prevent strokes, while the FDA and industry are interested in establishing if CAS is as safe or safer than CEA at one year of follow-up. The CREST investigators have chosen the novel approach of addressing these needs by having two “primary” analyses. For the scientific community and NIH, a null hypothesis of no difference between two treatments in the average hazard over 4 years of follow-up is contrasted with an alternative hypothesis that one treatment is better (a two-tailed test). This hypothesis will be addressed using proportional hazards analysis. For the FDA, an “equivalency” analysis where the null hypothesis that CAS has an event rate marginally higher than CEA (by δ of 2.6%) at the one-year point in the follow-up is contrasted against the alternative that CAS event rate is as low (within δ of 2.6%) or lower than CEA (a one-tailed test). This hypothesis will be addressed by comparing Kaplan-Meier estimates of survival at one year after the procedure. Having two primary hypotheses that are analyzed using different methods opens the possibility that “discordant” study results may occur. For example, the NIH analysis could conclude that CEA has a lower event rate than CAS, while the FDA analysis could conclude CAS is as good or better than CEA. Such a result would be confusing to the scientific community and lower the confidence placed in CREST study results. The likelihood of discordant findings was evaluated using a simulation approach that considered a wide range of outcomes ranging from CEA being better, to CEA being worse, than CAS. The outcome of the “NIH” and “FDA” analysis were never discordant in 3500 simulations. These findings suggest that the data from CREST can be used to address the different needs of the NIH and FDA with a low likelihood of introducing conflicting results.