Department of Clinical Research, University of Basel
3 December 2024
main reasons for randomized treatment allocation:
unpredictability, to avoid patient/treatment selection biases
(balance of prognostic baseline covariates, but better balance can be achieved without randomization)
(possibility of randomization tests, but permutation tests are possible anyway)
improving balance to allow simpler analyses and increase the perceived scientific validity of conclusions
comparison with simple randomization and stratified block randomization:
better balance, no problems with large number of prognostic baseline covariates (also continuous) or small sample size
worse unpredictability, but never deterministic and unproblematic in blinded studies
comparable power, with choice of adjustment factors independent of stratification
include center as factor, in case centers need to be excluded
include constant factor, to improve balance in group sizes
secuTrial:
stochastic minimization
range or variance (slightly better)
no weighting (however, it should be possible to use copies of factors to practically create weights)
algorithm for unequal allocation ratios unclear (however, virtual groups can be balanced and then joined)
no continuous prognostic baseline covariates
cannot be reproduced in R (seed unknown)
distribution of test statistic for randomization test can be estimated by simulating new allocations, while keeping everything else unchanged
if there is no random component in the allocation, permutation tests can be used instead, randomizing e.g. the order in which the participants entered the trial in given periods of time (assuming no important effect of time over such periods)
randomization test for minimization implemented in R package carat
some regulatory agencies prefer randomization tests when minimization is used, but standard tests give comparable results
reasons for randomization tests are valid also for any allocation except simple randomization
estimates and confidence intervals corresponding to randomization tests may be possible, but less standard (hence potentially less credible for many readers)