9.5. Summary

In this chapter we describe randomization procedures used by pharmaceutical industry. We devote most attention to the popular permuted block design and its variations brought in by variable block size, constrained randomization and randomization balanced across the centers. For each of the described approaches, we provide SAS code to create a randomization schedule.

We also discuss covariate-adaptive allocation procedures. When there is a need to balance the treatment groups with respect to baseline covariates, the stratified permuted block randomization is typically used. However, if the number of baseline covariates to balance upon is large, stratified randomization might not be feasible. In this case, one of the dynamic covariate-adaptive allocation procedures can be used to achieve the required balance. We describe the popular minimization procedure—a largely deterministic Taves (1974) version and a version by Pocock and Simon (1975) where a random element is added at each allocation step. Both algorithms are implemented in the SAS macro. An example that illustrates the use of the macro to randomize the patients dynamically is provided.

The randomization designs not covered in this chapter include biased coin designs (Efron, 1971), the maximal procedure (Berger et al, 2003), and designs based on urn models (see the review in Wei and Lachin, 1988). The maximal procedure (Berger et al, 2003) was proposed as an alternative to a sequence of permuted blocks of small sizes. The maximal procedure maintains the imbalance no larger than a prespecified positive integer number b throughout the enrollment and provides less potential for selection bias than the set of randomized blocks of size 2b. Although originally developed for trials with two treatments and a 1:1 treatment ratio, the maximal procedure can be easily expanded for more than two treatments. Each of these designs achieves a certain trade-off between the treatment balance and the amount of randomization the design provides.

Randomization designs can be viewed as adaptive allocations that use previous treatment assignments to modify the probability of the next assignment to ensure good balance in treatment assignments and to provide some amount of randomization. A completely different class of randomization designs are response adaptive allocation procedures. These designs use responses observed so far in the trial to change allocation away from perfectly balanced allocation based on a certain objective. We refer the reader to Hu and Ivanova (2004) for the most recent review of response adaptive designs.