Orateur
Description
Randomization is a key step in clinical trials to ensure a valid estimation of treatment effect. Most popular randomization method is stratification with blocks. This method can cause serious imbalances which makes this method unworkable in case of small sample size trials or incorporation of several prognostic factors. Minimization can overcome these problems, by accounting for many factors at the design stage, even when the number of patients is low.
The origin of this work is a future study with Ceva Santé Animale, about testing a new medicine on dogs. Stratification had to be applied to allocate 150 individuals to three treatment arms, depending on four qualitative factors and a continuous variable. Continuous criteria are often imbalanced at baseline without any clinically relevant classes to categorize the factor. A new randomization method accounting for both continuous and categorical factors with many strata was developed. This method is derived from classical Pocock and Simon’s minimization and accounting for continuous factors directly without transformation into categorical variables while maintaining randomness.
The main objective of this work was to compare different randomization methods, including stratification with blocks, classical minimization and the new randomization method. Simulations were run on randomly generated sets of individuals, to estimate the impact of the randomization method, the number of factors used and the number of individuals on the imbalance between treatment arms.
The method proposed, i.e. non-deterministic neighborhood-based minimization, allows to consider continuous covariate in a way which is at least as efficient as categorized-based usual methods. Minimization is preferable in terms of balance when the number of patients decreases. In the case of this study, the final decision was to use the new randomization method, which creates similar imbalance as stratification with blocks and classical minimization for qualitative factors, but a smaller one for the quantitative variable.
