Abstract

Experimental randomization is defended as a procedure to select an allocation of treatments in comparative experiments. To be convincing, a comparative experiment should allow many alternative allocations of treatments over experimental units, that look equally informative "on paper," but may have underlying differences. Randomization prevents these underlying differences to be causally influential when an allocation is chosen for implementation. Randomization helps prevent bias. ;Both Bayesian criticisms of randomization and what seems to be the opinion of the early Fisher are rejected. The Bayesian decision theoretic argument against randomization is shown to rest on a faulty analysis of experimental randomization that goes back, at least, to Abraham Wald. An improved decision theoretic framework is formulated in which randomization can indeed give higher expected utility. Against the early Fisher it is argued that the choice of a randomization scheme depends on the identification of equally informative, alternative allocations. Randomization alone can never validate a statistical analysis because what allocation is actually implemented, is relevant information that has to be taken into account in the statistical analysis. This does not, however, turn randomization into a useless ritual. The manner in which an allocation is obtained is equally relevant information