Cluster Randomization Trials

8. Factors Influencing Power

Exercise 5

The goal of this exercise is to help the user understand the benefits and perils when considering subsampling. Feedback would include why subsampling was used appropriately, and if not, what specifically was improper.

Instructions to user:
Review the examples and select whether use of subsampling was appropriate or inappropriate.

Examples:

1. A trial is being planned in which an information package is to be mailed out to individuals in the intervention communities, Since the average population of each community is in excess of 1000, the investigators conclude that selecting a random subsample of individuals within each community will lead to a negligible loss in power.

2. In a hospital randomized trial of a new approach for managing palliative care , patients are recruited by physicians or their staff prospectively over time after randomization.

3. An investigator elects to select a random subsample of subjects from each cluster to be enrolled in a trial. The size of the subsample to be selected from a given cluster is to be proportional to the size of that cluster since this should maximize the power for detecting a significant intervention effect.

4. To avoid the risk of contamination in a community intervention trial in which the communities are geographically proximate, it is suggested that subsampling should be restricted to the geographic center of each community.

Answers:

1. A trial is being planned in which an information package is to be mailed out to individuals in the intervention communities. Since the average population of each community is in excess of 1000, the investigators conclude that selecting a random subsample of individuals within each community will lead to a negligible loss in power.
[Appropriate. Since the power of a cluster randomized trial depends much more on the number of clusters randomized rather than on their size, this assertion is correct. The principal benefit of subsampling in this case could be substantial savings in cost and effort with little accompanying loss in precision.]

2. An investigator elects to select a random subsample of subjects from each cluster to be enrolled in a trial. The size of the subsample to be selected from a given cluster is to be proportional to the size of that cluster since this should maximize the power for detecting a significant intervention effect.
[Inappropriate. This is not true. For a given number of clusters and overall sample size, maximum power is obtained if the subsamples selected are of equal size in all clusters. The proposed strategy would be correct if the primary aim of the trial were to estimate the overall population mean, as it often is in survey samples. However this is not the case in cluster randomized trials.]

3. To avoid the risk of contamination in a community intervention trial in which the communities are geographically proximate, it is suggested that subsampling should be restricted to the geographic center of each community.
[Appropriate. This is a reasonable strategy for the purposes of trial planning although the ability to generalize the trial results may be affected if residents in the center of the community are expected to respond differently to the intervention than residents of more outlying areas.]

4. In a hospital randomized trial of a new approach for managing palliative care, patients are recruited by physicians or their staff prospectively over time after randomization.
[Either could be correct based on the following. The risk here is of intervention-related selection bias. For example, hospital personnel who are responsible for recruiting participants to receive the new intervention may be more selective or motivated in their efforts than personnel who are assigned to recruit participants into the control group. This could result in substantial imbalance between intervention groups with respect to both the number of the subjects recruited and their characteristics. Similar imbalances could result if willingness to participate differed by intervention group.

If all eligible patients cannot be identified at baseline, then an alternative strategy that is less likely subject to less bias would be to have someone independent of the trial be responsible for recruiting.]