• Thomas Grischott.
• Institute of Primary Care, University and University Hospital of Zurich, Pestalozzistrasse 24, CH-8091, Zurich, Switzerland. thomas.grischott@usz.ch.
• Bmc Med Res Methodol. 2018 Oct 16; 18 (1): 108.

BackgroundIn randomised controlled trials with only few randomisation units, treatment allocation may be challenging if balanced distributions of many covariates or baseline outcome measures are desired across all treatment groups. Both traditional approaches, stratified randomisation and allocation by minimisation, have their own limitations. A third method for achieving balance consists of randomly choosing from a preselected list of sufficiently balanced allocations. As with minimisation, this method requires that heterogeneity between treatment groups is measured by specified imbalance metrics. Although certain imbalance measures are more commonly used than others, to the author's knowledge there is no generally accepted "gold standard", neither for categorical and even less so for continuous variables.MethodsAn intuitive and easily accessible web-based software tool was developed which allows for balancing multiple variables of different types and using various imbalance metrics. Different metrics were compared in a simulation study.ResultsUsing simulated data, it could be shown that for categorical variables, χ2-based imbalance measures seem to be viable alternatives to the established "quadratic imbalance" metric. For continuous variables, using the area between the empirical cumulative distribution functions or the largest difference in the three pairs of quartiles is recommended to measure imbalance. Another imbalance metric suggested in the literature for continuous variables, the (symmetrised) Kullback-Leibler divergence, should be used with caution.ConclusionThe Shiny Balancer offers the possibility to visually explore the balancing properties of several well established or newly suggested imbalance metrics, and its use is particularly advocated in clinical studies with few randomisation units, as it is typically the case in cluster randomised trials.

19 keywords...

hide…