Randomisation by cluster and the problem of social class bias.

For randomised population trials the unit of randomisation is normally the individual person. In some situations, however, investigators take other groups as basic unit and one such design is cluster randomisation. Considerable attention has been given to this design recently in statistical and epidemiological literature. The Edinburgh randomised trial of breast cancer screening is an example which takes general practices as clusters of patients. The experience of this trial is reported here. Mortality from all causes, cardiovascular disease and lung cancer over the first 5 year period of follow up are examined. We found that spurious mortality differences were present in the analyses, which do not allow for socio-economic status. From examination of methods of adjusting for this, we conclude that allowance can be made at the analysis stage, and it is intended that this approach will be adopted when breast cancer mortality is analysed in the Edinburgh trial. Nevertheless, we recommend that for future studies with outcome related to socio-economic status, randomisations which use this design be stratified by socio-economic criteria where this is feasible.