Cluster Unit Randomized Trials
9. Cluster Level Replication
Some of the earlier community intervention trials addressing cardiovascular risk factors enrolled only two clusters, one allocated to the experimental intervention group and the other to a control (e.g., Turpeinen et al., 1979), with justification resting largely on cost and logistical considerations. This two-cluster design, which can still be seen in the literature today, may be very useful for exploratory purposes as a prelude to a more definitive trial that adopts formal power considerations.
Yet the ability to secure a valid estimate of ρ depends on the ability to obtain an accurate estimate of such variation, and hence the design itself is invalid. It is only under the unlikely (and untestable) assumption that ρ is zero that a valid test of the intervention effect can be conducted. Otherwise the results are subject to the same problems of interpretation that would arise in an individually randomized trial that assigns exactly one patient to each of two treatments. Taking a series of repeated measures in each of the two clusters improves the value of this design as an exploratory tool, but does not remove the basic problem.
The main issue here is not power but rather the threat to trial validity, since the effect of intervention is inevitably confounded with the natural variation that exists between the two clusters.
This is not to say that trials randomizing, say, three or four clusters to each group should be encouraged, since, although technically valid, they will almost surely lack the ability to detect important intervention effects. It is only by adopting a formal probabilistic approach to sample size estimation (discussed earlier) that this problem can be avoided.