Share

Cluster Unit Randomized Trials

15. Interim Analyses

The Role of Interim Analyses

Interim analyses are now a standard feature of individually randomized trials, particularly those with long-term follow-up and life-threatening outcomes. Although such analyses may have several objectives, the primary one is usually based on the need to detect unexpected differences in treatment effectiveness that may warrant early termination of subject accrual and follow-up.

There is no reason in principle that these factors should fail to apply to trials randomizing intact social units rather than individuals, as in, for example, nutritional supplementation trials having subject morbidity and mortality as the primary response variables. Yet formally planned interim analyses have not tended to play an important role in such trials.

A number of reasons may be responsible for this, including:

  • The relatively long lag time needed for an intervention to “settle in”;
  • The perception that the intervention in question is fairly benign, as in the case of lifestyle modification or behavioral trials; or
  • The likely belief that the assumptions underlying the stopping rules most frequently adopted for individually randomized trials, such as that developed by O'Brien and Fleming, 1979, may not hold in trials randomizing clusters.

But for trials in which cluster accrual occurs gradually over time it has now been shown (Zou et al., 2005) that such rules may in fact be safely applied under very general conditions. Of course, as in the case of individually randomized trials, it is vitally important that the treatment-related results be transmitted only to members of an independent data monitoring committee, and otherwise kept confidential.

A secondary aim of an interim analysis may be to reassess the values of some of the parameters used to estimate the required trial size. In cluster randomization trials, such parameters may not only include the standard deviation (for continuous outcomes) or the event rate in the control group (for dichotomous outcomes), but also the intracluster correlation coefficient. An example of sample size re-assessment for a cluster randomization trial is provided by Lake et al., 2002.