Abstract
BACKGROUND: Four main types of cluster-randomised trial (CRT) are well known: parallel-group (PG), factorial, stepped-wedge and crossover designs. This established typology relates to how clusters are exposed to intervention(s) or control(s) during the trial. Published guidance is lacking on how to link design features to how individuals within clusters may be exposed and measured. Thus, the aim of this paper was to develop a classification system for different types of cluster membership in CRTs, focussing on PG designs and building on our experiences of delivering a care home trial. METHODS: The classification system was developed in seven stages: (i) a scoping review was conducted to explore the use of open-cohort PG-CRTs in a range of settings; (ii) a version of the classification system was developed, using the stepped-wedge CRT typology; (iii) this was tested using a sample of published trials from the scoping review; (iv) a second version was developed, reviewed and further amendments made to aid clarity; (v) 15 trialists with experience of CRTs in a range of settings provided feedback in a 1-day, face-to-face user engagement workshop; (vi) a wider group of 39 trialists completed an online survey, providing examples and additional feedback; and (vii) all authors reviewed and approved the final version. RESULTS: Six types of cluster membership in PG-CRTs are proposed: the closed-cohort and cross-sectional designs already established, a new-admission-continuous-recruitment, open-cohort with discrete-recruitment, open-cohort with continuous-recruitment, and a non-standard closed-cohort design. The final classification system is made up of six core design features and five additional design considerations. Diagrams of each type of cluster membership are introduced and used to illustrate examples. CONCLUSIONS: Implications of distinctions between the six types of cluster membership for the statistical analysis require further research. CONSORT guidance needs updating to include specific guidance on reporting the type of cluster membership alongside the description of how design features apply to clusters. Further methodological research is required into both the statistical and the practical implications of adopting previously unlabelled but frequently used types of cluster membership.