Peer led programme for asthma education in adolescentsBMJ 2001; 323 doi: https://doi.org/10.1136/bmj.323.7304.110 (Published 14 July 2001) Cite this as: BMJ 2001;323:110
Issues of design and analysis are crucial in cluster randomised trials
- Tim J Peters (firstname.lastname@example.org), reader in medical statistics,
- Anna Graham, clinical research fellow,
- Chris Salisbury, senior lecturer in general practice,
- Laurence Moore, senior research fellow
- Department of Social Medicine, University of Bristol, Bristol BS8 2PR
- Division of Primary Health Care, University of Bristol
- Cardiff School of Social Sciences, Cardiff University, Cardiff CF10 3WT
- Barts and the London, Queen Mary's School of Medicine and Dentistry, London E1 4NS
- Airway Research Centre, Department of Respiratory and Sleep Medicine, John Hunter Hospital, Newcastle 2310, Australia
- Department of Public Health and Community Medicine, Westmead 2145, Australia
- Children's Hospital at Westmead, Westmead 2145, Australia
- School of Paediatrics, University of New South Wales, Randwick 2031, Australia
EDITOR—We have concerns about the design and analysis of Shah et al's cluster randomised trial of a peer led education programme for asthma.1 Neither the printed nor the (longer) electronic version mentioned how clustering was accounted for in the trial design. The sample size was not justified—neither the number of clusters (six) nor numbers of children in them. This may seem unimportant since confidence intervals were provided for the comparisons between arms, but the omission is crucial.
The authors did not specify the magnitude of differences considered in advance as clinically important. The small intracluster correlations observed could just be fortuitous. With so few clusters, any estimate of between-cluster variance (and hence intracluster correlation) will be extremely imprecise. Without proper details of trial design, the danger of publication bias remains, where a study with low power is more likely to be published when significance is attained. The widths of the comparative confidence intervals are not reassuring here.
It is unlikely that a fully considered trial design would include only six clusters, with apparently no attention being paid to stratification, given the sex and year imbalances.2 Important information was also omitted regarding the analysis plan, with between-arm comparisons in table 2 for quality of life as a total score and three subdomains. Only the electronic version states that the total score was the primary outcome.
More fundamentally, comparisons for each …