Ewan Carr
Department of Biostatistics and Health Informatics, King's College London
September 5, 2019
This script (randomisation.R) implements covariate-constrained randomisation,
as described in Carter and Hood (2008). It is intended for use in
cluster-randomised trials where blocks of clusters (of varying size; 4, 6, 8)
are allocated sequentially. Please see example.R for details.
This work will be presented at the 5th International Clinical Trials Methodology Conference in Brighton, UK (6-9 October, 2019):
Covariate constrained block randomisation for a cluster randomised trial
Kirsty James, Sabine Landau, Ewan Carr, Ben Carter (2019)Introduction Cluster randomised controlled trials require randomisation at the level of the cluster as opposed to the level of the participant. As there are fewer units being randomised than in an individually randomised trial the risk of baseline covariate imbalance is high. Standard methods of stratified randomisation can be employed but are limited to categorical covariates. In an ongoing trial we used stratified covariate constrained randomisation in order to accommodate continuous covariates.
Methods Clusters were identified within catchment areas, 4-6 within each. We required balance in the trial arms for characteristics of the area's service user populations hence the randomisation was stratified by catchment area. In addition, we balanced trial arms for two continuous cluster level covariates; surgery quality and deprivation. The randomisation algorithm, adapted from the work of Carter and Hood, balanced trial arms within and across catchment areas for these two covariates.
Results We randomised 28 clusters from 7 catchment areas (strata). All clusters within a stratum were supplied as a set over the course of the randomisation period. Every time the covariate information on clusters of a stratum became available the algorithm worked out all possible cluster assignments within the stratum and constructed a balancing index based on the clusters that have been randomised so far. An assignment is then chosen at random from the best performing allocations in terms of the balancing index to avoid the algorithm becoming deterministic.
Discussion There were several added complexities in using this randomisation technique in terms of performing the allocations as it was a bespoke algorithm executed by the statisticians. This method does require all cluster information within a stratum to be provided at once which could be a limitation. Outside of this the algorithm allowed the flexibility that was required to balance on continuous covariates in a reliable way.
This software was developed for a single trial and was not intended for wider use. You are welcome to adapt/reuse this code. However, if you are thinking of using covariate-constrained randomisation we recommend you refer to one of the following existing packages, which are easier to use and offer many more features:
- R
- Stata
Carter BR, Hood K. Balance algorithm for cluster randomized trials. BMC Medical Research Methodology 2008;8. doi:10.1186/1471-2288-8-65