When randomisation is not good enough: Matching groups in intervention studies

Abstract

AbstractRandomised assignment of individuals to treatment and controls groups is often considered the gold standard to draw valid conclusions about the efficacy of an intervention. In practice, randomisation can lead to accidental differences due to chance. Researchers have offered alternatives to reduce such differences, but these methods are not used frequently due to the requirement of advanced statistical methods. Here, we recommend a simple assignment procedure based on variance minimisation (VM), which assigns incoming participants automatically to the condition that minimises differences between groups in relevant measures. As an example of its application in the research context, we simulated an intervention study whereby a researcher used the VM procedure on a covariate to assign participants to a control and intervention group rather than controlling for the covariate at the analysis stage. Among other features of the simulated study, such as effect size and sample size, we manipulated the correlation between the matching covariate and the outcome variable and the presence of imbalance between groups in the covariate. Our results highlighted the advantages of VM over prevalent random assignment procedure in terms of reducing the Type I error rate and providing accurate estimates of the effect of the group on the outcome variable. The VM procedure is valuable in situations whereby the intervention to an individual begins before the recruitment of the entire sample size is completed. We provide an Excel spreadsheet, as well as scripts in R, MATLAB, and Python to ease and foster the implementation of the VM procedure.