Optimal randomisation strategies – simple, stratified or minimised: lessons from the WHIST trial
KNIGHT R., PECKHAM N., COSTA ML., DUTTON SJ.
Background Random allocation to treatment groups is a key feature of clinical trial design, used to achieve balance between treatment groups on baseline characteristics. Several different randomisation techniques, such as simple randomisation, stratified randomisation, and randomisation by minimisation have been developed and are used in a variety of trial scenarios; however, it is not always clear what the optimal randomisation strategy is for a given trial. The optimal randomisation strategy depends on the planned sample size, and the number of key characteristics (stratification factors) on which we wish to ensure suitable balance. The Wound Healing in Surgical Trauma (WHIST) trial (n = 1629) used randomisation by minimisation with three stratification factors (open versus closed wound at presentation, Injury severity score ≤15 versus ISS ≥ 16, and recruitment centre). The optimal randomisation strategy in this trial is explored. Methods Anticipated baseline imbalance in the WHIST trial under different randomisation scenarios (simple randomisation, stratified randomisation, and randomisation by minimisation) was investigated. Simulations were conducted to explore the optimal randomisation strategy under a variety of different scenarios including varying numbers of stratification factors and overall sample sizes. Results In the WHIST trial, randomisation by minimisation led to increased balance on minimisation factors compared with simple randomisation; however, the benefit of this increased balance was small compared to the added complexity of the randomisation system required. Stratified randomisation resulted in decreased balance compared to simple randomisation. Conclusions In some instances more complex randomisation schemes than needed are used. At other times inappropriate randomisation schemes may lead to unacceptable imbalances in baseline characteristics. Further guidance on this is needed.