Background
Cluster randomization trials in which intact social units are randomly assigned to different intervention groups have become very popular in recent years, particularly for the evaluation of innovations in the delivery of health care. An extensive literature dealing with the associated methodological challenges has also appeared. Although the monitoring of such trials using formal stopping rules is clearly indicated when the outcomes are irreversible and individual-level data are available sequentially, simple and reliable statistical methods that may be used for this purpose are currently not available.
Purpose
To investigate the validity of standard group sequential methods when applied to cluster randomization trials having binary outcomes.
Methods
The large sample distributions for each of five test statistics computed from sequentially accumulated data are derived. A simulation study is performed to evaluate the finite sample properties of these statistics when applied to the interim analysis of cluster randomization trials. Data from the World Health Organization antenatal care trial [1] are used to illustrate the methods.
Results
Each of the joint distributions is shown to be characterized by a covariance structure that asymptotically satisfies an independent increments structure, a foundation that simplifies group sequential methods [2]. The simulation study reveals that four of the five test statistics evaluated provide satisfactory performance with as few as 10 clusters allocated to each of two interventions.
Limitations
The applicability of our results to effect estimation following a group sequential cluster randomization trial is not investigated, although a theoretical foundation which may be used for this purpose is presented.
Conclusions
Standard group sequential methods can be applied to cluster randomization trials when interim analyses are warranted.
