Panel Paper: Power Calculations for Moderator Effects in Cluster Randomized Trials

Saturday, November 9, 2013 : 3:50 PM
3017 Monroe (Washington Marriott)

*Names in bold indicate Presenter

Jessaca Spybrook, Western Michigan University
Cluster randomized trials (CRTs) have become increasingly common as a means for evaluating the effectiveness of social programs. For example, CRTs have been conducted to examine the effectiveness of programs designed to prevent drug abuse among youth, reduce violent behavior among students, enhance parenting skills, and improve student achievement. CRTs have become more widespread in evaluations of the effectiveness of social programs and policies for two primary reasons. First, when they are feasible and if they are well designed and implemented, randomized trials are the best way to establish causal relationships. Second, the natural clustering in many organizations and the fact that programs are often implemented at the group level makes CRT’s a logical choice. For example, in the U.S. education system, students are nested within classrooms within schools within districts, and educational interventions are typically delivered at the classroom, school, or district level.            

However, the sheer presence of a CRT to evaluate a program or policy is not enough to generate rigorous evidence of the effectiveness of a program. As noted above, the trials must be well-designed and implemented in order to generate high-quality evidence of program effectiveness. In the past 15 years, the field has made substantial progress in terms of how to design CRTs and how to calculate the statistical power for the main effect of treatment. However, designing a study to detect the main effect of treatment may not be sufficient. It is quite reasonable that context matters in these studies and thus designing studies to examine for whom and under what conditions a program is effective is critical.

The purpose of this paper is to provide a resource for researchers designing studies that not only test whether or not a program works, but under what conditions and for whom. Specifically, I provide power calculations for dichotomous student and cluster level moderators for the following types of CRTs: 2-level CRT, 3-level CRT, 3-level multisite cluster randomized trial (MSCRT), and 4-level MSCRT. To make the calculations accessible for researchers planning CRTs, I also include a tool, R code in this case, for researchers conducting these calculations. The calculations and code in this paper is a next step towards building capacity in field for researchers to design CRTs that move beyond the main effect of treatment.