I Have Randomized by Clinic; Now What? Four Decades After Cornfield

It has been nearly 40 years since the publication of Jerome Cornfield’s 1978¹ paper calling for the appropriate treatment of the analysis of data from group (or cluster) randomized trials (GRTs). Up until that point, researchers were left with two equally unsatisfactory analytic approaches. First, they typically ignored the induced correlation between outcome measures resulting from the GRT approach, and analyzed the data at the level of the individual, likely exposing themselves to committing a type I error—that is, rejecting the null hypothesis when the null is actually true. Second, they analyzed a cluster-level summary measure of the outcome, thereby eliminating the correlated individual responses, but had to forgo the ability to make individual-level covariate adjustments, and this usually resulted in an enormous loss of power. As few were the wiser, and no good analytic solutions were available, most researchers opted for the former approach.

METHODS FOR CORRELATED DATA

But Cornfield’s statistical bell could not be unrung, and we were no longer able to ignore that unwelcome consequence of the GRT in our analyses. Fortunately, necessity is the mother of invention, and fueled by statistical innovation and rapid advances in computing power, the subsequent decades produced myriad methods for correlated data, from generalized estimating equations to random effects and mixed models. So it was clearly time to gather, assess, and report on the available approaches for the design and analysis of GRTs. Three papers in particular were commissioned for and published in the March 2004 AJPH Evaluation Issue (its cover sporting the iconic cartoon by Jim Borgman, to remind researchers to endeavor not to contribute to the scientific noise) to summarize the state of the art for GRTs.^2–4

Thirteen years later, it is time for an update, and two articles by Turner et al. in this and the subsequent issue of AJPH provide reviews and assessments of various GRT-related designs⁵ and GRT-related analyses.⁶ These sister articles provide a compendium of information on GRTs, and serve as the go-to comprehensive resource on the topic for beginners and experts alike, summarizing nicely a vast array of topics surrounding GRTs.

PRIMER ON GROUP RANDOMIZED TRIALS

Part 1, on design,⁵ serves as an excellent primer on GRTs, proceeding from the simple (i.e., definition of an individual and group randomized trials) to the complex. At its most basic, the motivation for conducting a group rather than an individual randomized trial is twofold: when the threat for intervention contamination is high or when administratively it becomes extremely difficult to randomize individuals. They continue the discussion around the problematic correlation produced as a function of the increasingly sophisticated group randomized design features (to be solved in part 2), and describe the various design types that form the GRT and alternative GRT family of designs. As in many fields, the “solution” to one issue generates challenges in another, and Turner et al. do a thorough job of articulating the solution parts and illuminating the challenges. In this contribution (as in the second article), the authors provide a glossary of terms at the end of the article, to help demystify the GRT jargon.

In part 2, on analysis,⁶ Turner et al. begin by describing the “problem” that arises in the analytic phase of a GRT that we are trying to cure. Although randomization by group or cluster (e.g., clinic, school, hospital, village) now ensures that the groups are independent, the dependent variable measured on participants (usually people) within the group tend to be more alike (i.e., correlated) than responses from participants in other groups. Although the measure of this correlation, the intraclass (or intracluster) correlation coefficient, tends to produce deceptively tiny values of this correlation (i.e., usually less than 0.05), the intraclass correlation coefficient is small but mighty, and can wreak havoc on the unsuspecting (or even the seasoned) researchers.⁷ Turner et al. proceed to review the array of analytic methods, careful to note the pros and cons, provide a resource for how to implement these methods with a variety of software, and remind us of appropriate reporting standards in publications for GRTs.

ONE-STOP-SHOPPING SET

As the wisdom, appropriateness, and use of GRTs expand, it is extremely helpful to have a one-stop-shopping set of articles to refer to, to help guide both the design and the analysis. I have no doubt these articles will be well-dog-eared in both paper and electronic versions.