Cluster-randomised trial

Some interventions cannot be given to one person and withheld from their neighbour: a screening campaign run by community health workers, a change in how a clinic organises follow-up, a training programme for doctors. A cluster-randomised trial randomises the unit that receives the intervention, whether a village, a general practice, a hospital or a district, and measures outcomes in the individuals inside each cluster. Because people in the same cluster resemble each other, each cluster contributes less information than the same number of unrelated individuals, so the trial needs more participants than an individually randomised one (the design effect), and the analysis must account for the clustering or it will overstate its certainty.

The corpus holds three of the most consequential cluster-randomised trials in cancer, all from India. The Kerala oral cancer trial randomised 13 clusters in Trivandrum district, more than 190,000 adults, to repeated visual inspection of the mouth by trained health workers or usual care, and cut oral cancer deaths among tobacco and alcohol users, by a third after three rounds and by four-fifths in users who attended every round at fifteen years; it remains the only randomised evidence that oral cancer screening saves lives. The Mumbai trial randomised 20 slum clusters, more than 150,000 women, to visual inspection of the cervix with acetic acid every two years or health education and reduced cervical cancer deaths by 31 percent. The Osmanabad trial randomised 52 villages, more than 130,000 women, to one round of HPV testing, cytology, visual inspection or usual care, and only HPV testing reduced advanced cancers and deaths, the result behind the World Health Organization's HPV-first screening recommendation. SANO, in oesophageal cancer, used a stepped-wedge cluster design to roll active surveillance out across Dutch hospitals.

The design's hazards are recruitment bias, when the people who come forward in intervention clusters differ from those in control clusters because they know which arm their village is in, and small numbers of clusters, which no number of individuals can compensate for. Consent is also layered: a community or institution agrees to be randomised, and individuals still consent to the intervention and to data collection. The CONSORT extension for cluster trials sets out what a report must include.