Comparative evaluation of balancing properties of stratified randomization procedures

If in a clinical trial prognostic factors are known in advance to be associated with the outcome of a patient it is often recommended that the randomization for a clinical trial should be stratified on these factors, particularly in a multicenter trial. Unfortunately, stratified or covariate-adaptive randomization does not always promote greater balance between the numbers of treatment assignments to A and B within each stratum and thus overall. Because such designs have numerous parameters that must be specified, simulation is a good tool to investigate the impact of these parameters on balance.

We investigate and discuss in more detail the difference in balancing performance of three stratified randomization procedures. The permuted-block randomization within strata, the "minimization" method and "self-adjusting" design are assessed overall, within levels of prognostic factors, and within strata.

We show the superior performance of "self-adjusting" design and the extent of balancing losses occurring with permuted-block randomization within levels of factors and with "minimization" within strata. A summary of principal conclusions regarding the balancing properties of stratified randomization procedures is presented and general recommendations are offered.

The relative merits of each procedure should be weighted carefully in relation to the characteristics of the trial. Considering the likelihood of imbalances, the sample size and values of parameters of stratified randomization procedures (number of prognostic factors, number of factor levels, block size) are important when choosing a randomization procedure.