Common scale minimal sufficient balance: An improved method for covariate‐adaptive randomization based on the Wilcoxon‐Mann‐Whitney odds ratio statistic

Abstract

Minimal sufficient balance (MSB) is a recently suggested method for adaptively controlling covariate imbalance in randomized controlled trials in a manner which reduces the impact on randomness of allocation over other approaches by only intervening when the imbalance is sufficiently significant. Despite its improvements, the approach is unable to consider the relative clinical importance or magnitude of imbalance in each covariate weight, and ignores any imbalance which is not statistically significant, even when these imbalances may collectively justify intervention. We propose the common scale MSB (CS-MSB) method which addresses these limitations, and present simulation studies comparing our proposed method to MSB. We demonstrate that CS-MSB requires less intervention than MSB to achieve the same level of covariate balance, and does not adversely impact either statistical power or Type-I error.

Adaptive Randomization Method to Prevent Extreme Instances of Group Size and Covariate Imbalance in Stroke Trials

“…The newly proposed CSSize-MSB algorithm will incorporate a measure of group size balance control into the existing CS-MSB algorithm. CSSize-MSB uses the previously demonstrated mechanisms of CS-MSB 22 to estimate covariate imbalance with the generalized odds ratio. 25 Additionally, CSSize-MSB will estimate the group size imbalance between the treatment groups after each individual patient randomization as the proportion of allocated patients that have been randomized to the control group.…”

“…2 This type of randomization approach has not only been used in stroke research but also in Parkinson disease 20 and COVID-19 research. 21 An extension of this randomization method called common-scale minimal sufficient balance (CS-MSB) 22 has been proposed that provides improved covariate balance with the same level of treatment allocation randomness compared with MSB. Additionally, it has been demonstrated that CS-MSB has identical type 1 error and power control compared with MSB.…”

“…Additionally, it has been demonstrated that CS-MSB has identical type 1 error and power control compared with MSB. 22 CS-MSB relies on 2 key parameters: the strength of the biased coin used to intervene in randomization 22 (as opposed to using a fair coin to randomize participants, a biased coin is used where a higher probability of assignment corresponds to a stronger coin; for example, a biased coin with a strength of 0.7 corresponds to a 70% chance of assignment to a particular group), and the threshold for determining whether the existing covariate imbalance is substantial enough to intervene in an individual randomization 22 (a P value corresponding to a single measure of imbalance across all covariates). However, while CS-MSB effectively controls for covariate balance and allocation randomness, it is not capable of controlling for group size balance.…”

“…However, while CS-MSB effectively controls for covariate balance and allocation randomness, it is not capable of controlling for group size balance. 22 As adaptive randomization methods become more technically sophisticated, the underlying complexity makes it harder to clearly communicate the essence of the method to clinicians. Lack of clarity may lead to reduced trust in the proposed method, thus becoming a barrier to the adoption of these useful adaptive randomization methods.…”

Adaptive trials in stroke: Current use & future directions

“…In this trial, the two-arm, two-stage design was complemented by an adaptive sample size reestimation within the second stage using the Mehta and Pocock method described above 30 . The randomisation for this trial uses the covariate adaptive common scale minimum sufficient balance algorithm 38 that improves the balance of key prognostic variables while maximizing the randomness of allocation.…”