ABSTRACT
This chapter provides a framework for conceptualizing randomization in clinical trials and for linking randomization to a statistical test. The aim of this chapter is to demonstrate how randomization works, how it does not work, and why a discussion of randomization without reference to a randomization test is incomplete. Randomization works because, on average, the differences between the treatment averages of any potentially confounding variable are zero. Randomization does not work because of some magic equating potion that spirits away all biases for any particular clinical trial. Furthermore, randomization is closely linked to statistical inference by the concept of a randomization test. This test is formally defined and its main features are highlighted: The randomization test is valid and powerful by construction, it can be used with any design and any test statistic, without random sampling and without assuming specific population distributions, and it results in a frequentist, conditional, and causal inference. The randomization test derives its statistical validity by virtue of the actual randomization, and conversely a randomized design is complemented with a calculation of the randomization test p-value because it provides a quantification of the probability of the outcome (or a more extreme one) if the null hypothesis is true.
