Power for Cluster-Randomized Trials: Software, Web app, and Methods

总结

Summary Cluster-randomized trials are an increasingly common trial design that has unique advantages for some questions and is required for others. It has also spawned several related trial designs. Recent high-profile studies using these designs include trials of Ebola vaccines and for preventing the spread of resistant pathogens in hospitals. The key feature of cluster-randomized trials is that subjects are randomized in groups, rather than as individuals, so that all members of a community, hospital, or practice receive the same treatment. All trials must have accurate power and sample size calculation for moral and ethical reasons. It is wrong to randomize more persons than are needed for good power, as this exposes the excess persons to the risks of randomization unnecessarily. It is also wrong to randomize fewer persons than create good power, as then all are exposed to the risks of the study for naught—there is little hope of showing a benefit of any treatment-- and a lack of study effect may be due to poor power rather than ineffective treatments. Though it is a lesser concern, it is also unethical to have too small or large a sample size, as this wastes scarce resources such as the investigators’ time and the funder’s dollars. For some cluster-randomized trial designs, there exist analytic (closed-from) sample size formulae that rely on assumptions that can be unrealistic. For other designs, only approximate formulae exist. In general, these calculations can be found only in textbooks, scientific papers, and in software that is costly and can be difficult to understand and apply. There are very limited options for the most accurate calculations, which are based on simulations. Simulation-based power calculations can accommodate complex designs and realistic scenarios that are only awkwardly possible in formulae. We propose to generate a comprehensive free and open-source software suite to provide approximate, analytic, and simulation-based power assessment. In addition, we will develop a web app for the code to allow users who have less computing knowledge to make use of the software. Finally, we will make use of the software to answer outstanding questions in the design of cluster-randomized trials.