Rapid Responses to:
|
|
Rapid Responses: Rapid Responses published:
-
![[Read Rapid Response]](/icons/misc/sectionDOWN.gif)
Cluster randomised trials: can empirical estimates be used for sample size calculations
- Petra Denig (23 July 1999)
|
|
|||
|
Petra Denig, Senior researcher Dept.Clinical Pharmacology/NCH/GUIDE, University of Groningen, The Netherlands
Send response to journal:
|
Editor - Health care interventions attempting to influence practice behaviour are often randomised by cluster. Campbell and Grimshaw have argued that to calculate the required sample size it is best to use estimates for clustering effect from other studies, whereas Freemantle and Wood have pointed out the risks of using empirical estimates.1 2 3 We wish to shed some light on this matter. Using data from a large intervention study in which groups of doctors were randomised, we have calculated the intracluster correlation coefficients (ICC) for a number of outcomes.4 5 In this trial, an educational programme was tested in 24 groups of general practitioners in The Netherlands. The group size ranged from 4 to 13. ICCs were based on the variance between and within these groups.5 Baseline prescribing data were used to calculate three different prescribing indicators for evaluating treatment of asthma at GP level, i.e. the ratio of prophylactic to bronchodilator drugs, the proportion of asthma patients receiving inhaled corticosteroids, and the proportion of patients on continuous use of bronchodilators without inhaled corticosteroids.4 Surprisingly, ICCs differed largely depending on the indicator chosen. For the ratio of prophylactic to bronchodilator drug, the ICC was 0.10. For the proportion of patients on inhaled corticosteroids, it was 0.26. For the proportion of patients on continuous mono-use of bronchodilators, it was 0.04. This difference cannot be explained by a difference in setting, since the same data from the same groups were used. Why is it then that the number of patients on inhaled steroids is greatly affected by the clustering effect and the number of patients on continuous mono-use of bronchodilators is not? We can only speculate. The groups in our study meet 6 to 10 times per year to discuss rational pharmacotherapy, with most groups working towards a consensus. One would therefore expect GPs within one group to have more similar prescribing patterns than GPs from different groups. Apparently this is the case when looking at general indicators of prescribing, such as drug ratios or numbers of patients receiving certain drugs, but not for specific indicators that take into account detailed prescribing schedules or combinations of drugs at patient level. The variation in this latter indicator appears more dependent on individual doctors than on the group to which they belong. Researchers should thus be very cautious when using empirical ICCs, and calculate the ICC after conducting the trial to check whether the study might have been underpowered. Petra Denig, Senior researcher. Lisa Pont, Research fellow. Department of Clinical Pharmacology, Northern Centre for Healthcase Research, University of Groningen, 9713 AV Groningen, The Netherlands 1. Campbell MK, Grimshaw JM. Cluster randomised trials: time for improvement. BMJ 1998; 317: 1171-1172 2. Freemantle N, Wood J. Cluster randomised trials. Standardised approach to analysing and reporting these trials is misguided. BMJ 1999; 318: 1286 3. Campbell MK, Grimshaw JM. Cluster randomised trials. Authors'reply. BMJ 1999; 318: 1286 4. Veninga CCM, Lagerlov P, Wahlström R, Muskova M, Denig P, Berkhof J, Kochen MM, Haaijer-Ruskamp FM, Drug Education Project Group. Evaluating an educational intervention to improve the treatment of asthma in four European countries. Am J Respir Crit Care Med 1999 (in press) 5. Donner A. The estimation of intraclass correlation in the analysis of family data. Biometrics 1980; 36: 19-25 |
|||