Clustering by health professional in individually randomised trialsBMJ 2005; 330 doi: http://dx.doi.org/10.1136/bmj.330.7483.142 (Published 13 January 2005) Cite this as: BMJ 2005;330:142
- Katherine J Lee, PhD student (firstname.lastname@example.org)1,
- Simon G Thompson, director1
- 1MRC Biostatistics Unit, Institute of Public Health, Cambridge CB2 2SR
- Correspondence to: K J Lee
- Accepted 8 November 2004
Patient outcomes in many randomised trials depend crucially on the health professional delivering the intervention, but the resulting clustering is rarely considered in the analysis
Almost all trials that randomise individuals assume that the observed outcomes of participants are independent. The validity of this assumption is doubtful, however, in some situations. One example is when more than one health professional (such as surgeons, nurses, general practitioners, or therapists) delivers a non-pharmaceutical intervention to participants. Because health professionals may vary in their effectiveness, observations on participants treated by the same professional may be somewhat similar or clustered. Clustering of outcomes may also appear less obviously (such as in clustering by centre in a multicentre trial) or in a more dominant form (as in cluster randomised trials). In each of these situations the assumption of independence is violated, which means that standard statistical methods are invalid and may give misleading conclusions. The presence of clustering in a trial inflates standard errors and reduces the effective sample size, thus reducing the power of the trial. We examine the prevalence and importance of potential clustering in individually randomised trials and present an example of the effect it can have on the overall results and conclusions of a trial.
Types of clustering
In a trial comparing a new one stop clinic with a dedicated breast clinic for breast cancer screening,1 patients were randomised to a clinic, where they attended an appointment with one of several consultants. The main outcome was patient anxiety, which is likely to be influenced by the consultant treating the patient, yielding potential clustering by consultant. In this trial the clustering is imposed by the design of the trial because of the interventions being compared and is nested within treatment groups since each consultant participates in one treatment arm only (fig 1).