Cluster randomised trials: time for improvementBMJ 1998; 317 doi: https://doi.org/10.1136/bmj.317.7167.1171 (Published 31 October 1998) Cite this as: BMJ 1998;317:1171
The implications of adopting a cluster design are still largely being ignored
- Marion K Campbell (), Senior statistician,
- Jeremy M Grimshaw, Programme director
Cluster randomised trials, where groups of patients rather than individuals are randomised, are increasingly being used in health services research. Randomisation by individual is inappropriate for evaluating some interventions, such as organisational changes, where it may not be feasible to randomise at the patient level. In such cases cluster randomisation at the level of the health professional or organisation is necessary. Such randomisation can also minimise the potential for contamination between treatments when trial patients are managed within the same setting.
The main consequence of adopting a cluster design is that the outcome for each patient can no longer be assumed to be independent of that for any other patient (which is the case in an individually randomised trial). Patients within any one cluster are more likely to have similar outcomes. For example, the management of patients within a single general practice is more likely to be consistent than management across several practices.
This lack of independence has implications for the design and analysis of these trials.1 The statistical power of a cluster randomised trial is greatly reduced in comparison with a similar sized individually randomised trial. Therefore standard sample size estimates have to be inflated to take account of the cluster design.
The impact on sample size can be substantial and depends on the size of the clustering effect and the number of clusters available. The clustering effect would be high if, for example, management of patients within individual hospitals was very consistent but there was wide variation across hospitals. Consider a trial of an educational intervention to implement a clinical guideline. A patient randomised trial would require 194 patients to detect a change from 40% to 60% in the proportion of patients who are managed appropriately (with 80% power and 5% significance). However, this design would be inappropriate because of the potential for contamination. For a cluster randomised trial with a moderate clustering effect and 10 available patients per cluster the equivalent sample size adjusting for clustering is 38 clusters or 380 patients—that is, almost double.1
The analysis of cluster randomised trials must also take into account the clustered nature of the data. Standard statistical techniques are no longer appropriate, unless an aggregated analysis is performed at the level of the cluster,2 as they require data to be independent. If the clustering effect is ignored P values will be artificially extreme, and confidence intervals will be over-narrow, increasing the chances of spuriously significant findings and misleading conclusions.
Although an aggregated analysis can be performed at the cluster level using standard statistical tests, this approach is statistically inefficient. Furthermore, it does not allow variation at the patient level to be explored—for example, it cannot take account of patient characteristics such as disease severity. More advanced techniques have now been developed to analyse patient level data arising from a clustered design, which allow the hierarchical nature of the data to be modelled appropriately.3 They essentially allow variation to be modelled at each level of the data—for example, at both the practice and the patient level.
Despite the increased use of cluster randomised trials, the implications of adopting such a design continue to be largely ignored. For example, a review by Devine et al which examined studies of physicians' patient care practices observed that 70% of studies identified had not appropriately accounted for the clustered nature of their study data.4
Many trials do not take cluster randomisation into account when calculating the required sample size, resulting in studies which are underpowered. This may, in part, be explained by the lack of published information on the likely size of the clustering effect, known as the intracluster correlation coefficient.5 Reliable estimates of this coefficient are required to ensure robust sample size calculations, yet publication of intracluster correlation coefficients in trial reports is rare.
We need a standardised approach to the reporting of all aspects of cluster randomised trials, including intracluster correlation coefficients. The CONSORT statement for standards of reporting of patient randomised trials6 has improved the reporting of such trials, and a similar approach for the reporting of cluster randomised trials would be beneficial. A letter to the BMJ highlighted the need for such an approach,7 and a proposed amendment to the CONSORT statement is being formulated. We believe that this will aid the appraisal of published trial reports and the planning of future research. It deserves the full support of the medical research community.