Cluster Randomization Trials

7. Sample Size Assessment

Exercise 4

The goal of this exercise is to reinforce to the user how to determine the best strategies for cluster trials but asking them to think about what makes a strategy good or poor.

Review the strategies listed below for cluster trials and select whether it is a good or poor strategy. Upon selecting the answer, feedback will reinforce chapter content which defines why a certain strategy is good or poor.

The Strategies are:

1. Undertaking of a census of eligible clusters

2. To decide if it is necessary to adjust for the effect of clustering in the statistical analysis, the investigators consider testing the observed value of the intracluster correlation coefficient for statistical significance .In the presence of a null result it is decided to apply standard statistical methods in evaluating the effect of intervention, i.e., they will assume that substantial clustering effects are implausible if the test of significance reveals the observed value of the intracluster correlation can be attributed to chance.

3. For trials randomizing a small number of clusters (10 or less per group) the investigators decide to perform a cluster-level analysis in evaluating the effect of intervention.

Answers:

1. Undertaking of a census of eligible clusters
[Good strategy. This strategy would be particularly useful if larger clusters such as schools, worksites or cities are to be randomized. It would not only help to determine if enough clusters are available to achieve the desired power, but also help to establish the generalizability of the trial results.]

2. To decide if it is necessary to adjust for the effect of clustering in the statistical analysis, the investigators consider testing the observed value of the intracluster correlation coefficient for statistical significance .In the presence of a null result it is decided to apply standard statistical methods in evaluating the effect of intervention, i.e., they will assume that substantial clustering effects are implausible if the test of significance reveals the observed value of the intracluster correlation can be attributed to chance.
[Poor strategy. It is well-known that small values of the intracluster correlation coefficient combined with reasonably large cluster sizes can yield sizeable variance inflation factors (design effects). Moreover the power of a significance test for detecting small values of the intracluster correalation coefficient is generally very low. Therefore this strategy should be discouraged.]

3. For trials randomizing a small number of clusters (10 or less per group) the investigators decide to perform a cluster-level analysis in evaluating the effect of intervention.
[Good strategy. The large sample approximations underlying individual level analyses become questionable under these conditions.Therefore this is a very reasonable strategy.]