Letters

Cluster randomised trials

BMJ 1999; 318 doi: https://doi.org/10.1136/bmj.318.7193.1286b (Published 08 May 1999) Cite this as: BMJ 1999;318:1286

Standardised approach to analysing and reporting these trials is misguided

  1. Nick Freemantle, Senior research fellow,
  2. John Wood, Principal statistician
  1. Medicines Evaluation Group, Centre for Health Economics, University of York, York YO10 5DD
  2. Statistical Sciences, SmithKline Beecham Pharmaceuticals, Harlow, Essex CM19 5AW
  3. Health Services Research Unit, University of Aberdeen, Aberdeen AB25 2ZD

    EDITOR—In their editorial on cluster randomised trials Campbell and Grimshaw correctly identify the importance of the appropriate choice of the unit of analysis.1 Unfortunately, they display a disappointingly poor grasp of the basis of estimation, and the specific recommendations that they make are inappropriate.

    The authors' plea for a standardised approach to the analysis and reporting of cluster randomised trials is misguided. The framework that they describe has the patient as the principal unit of the experiment, with the correlation between patients in the same cluster having to be taken into account in the analysis. Following the recommendations in the editorial may, however, make investigators prey to the same errors that the authors are counselling them to guard against.

    Although some trials fall naturally into the framework that the authors describe, others do not. It is not helpful in their own example of an educational intervention to implement a clinical guideline. There, the clinician (whose behaviour is the target of the intervention) is clearly the principal unit of the experiment. We are interested in the change in his or her behaviour as a result of the intervention, the average size of the change, and how the size varied between clinicians. Focusing on the estimation of these quantities of direct interest should lead to a sensible and interpretable analysis. Recognising that such estimates should be fit for purpose should lead to a sensible design.2

    Adjusting the analysis on the basis of the intracluster correlation coefficient assumes the observed covariance structure to be the true structure, without taking into account measurement error. It is robust only where a reasonable estimate of variability at the clinician level is available. Thus to suggest that such methods are more efficient than aggregated analyses at the level at which the intervention is targeted is incorrect. The authors' quest for information on intracluster correlation coefficients is a search for an answer to the wrong problem. Thoughtless focus on correlation coefficients could lead to the use of values estimated in one study for calculations of sample size in quite a different area, which is probably worse than doing no such calculation at all.

    The authors also suggest incorrectly that covariates at the patient level may not be examined in aggregated analyses. Generalised linear modelling provides the opportunity of specifying covariates at the patient level while analysing at the level at which the intervention is targeted.3

    References

    Authors' reply

    1. Marion K Campbell, Senior statistician (m.k.campbell{at}abdn.ac.uk),
    2. Jeremy M Grimshaw, Programme director
    1. Medicines Evaluation Group, Centre for Health Economics, University of York, York YO10 5DD
    2. Statistical Sciences, SmithKline Beecham Pharmaceuticals, Harlow, Essex CM19 5AW
    3. Health Services Research Unit, University of Aberdeen, Aberdeen AB25 2ZD

      EDITOR—We did not suggest that a standardised approach be adopted for the analysis of cluster trials. Our primary plea is that researchers account for the clustering in their data. Several methods are available to do this, and considerable debate surrounds the choice of method. We emphasised the use of hierarchical models; they allow all the information and variation at each level of the data to be explored while retaining the validity of the analysis. An aggregated approach may, however, be more appropriate for some analyses. Considerable debate also surrounds the choice of the unit of analysis. As Murray argues, however, this may be misplaced, and attention should rather be focused on the appropriate specification of the model for the analysis; the model selected should be well matched to the underlying structure of the data.1

      We were surprised that Freemantle and Wood considered a standardised approach to the reporting of cluster randomised trials to be misguided. As Freemantle himself has indicated, implementation of the CONSORT guidelines for randomised trials should minimise the serious deficiencies in the design, analysis, and reporting of such trials.2 Extension of this statement to include the issues specific to cluster trials should yield similar benefits.

      We also cannot agree with the suggestion that we should refrain from seeking further information on the intracluster correlation coefficient. Reliable estimates of this coefficient are required if robust calculations of sample size are to be made. Naturally, the gold standard approach is to measure the correlation coefficient directly within the study setting. However, the estimate of this coefficient, like other assumptions, is often informed by previous research or published reports. Because of the paucity of published information on the potential size of correlation coefficients in specific settings and on factors that might affect their magnitude, we (with other authors 3 4) are promoting their routine publication. However, to suggest that using an external estimate, even if it is the best available, is worse than not calculating sample size at all is absurd. The adoption of such an approach would invalidate most calculations of sample size ever undertaken.

      References

      View Abstract

      Log in

      Log in through your institution

      Subscribe

      * For online subscription