[Some methodological questions about randomizing collective units (author's transl)].

Abstract

Trials often consist in comparing two samples, or more, which have been differently treated. Often these two samples are obtained by choosing individuals at random. Yet it may be easier, or even unavoidable, to randomize groups rather than individuals. It this case, the accuracy of estimates, that is the significance level of tests is different. In order to evaluate them, it is necessary to take into account a possible "cluster effect". On the basis of the two-stage sampling theory, it is possible to build a Student-like test. Nevertheless, the convergence towards a Student variable is slower than in the case of an individual randomization. Thus, the approximation cannot always be accepted. Aside from this question of how to build a correct test, the fact is that randomization by group most often means a loss of efficiency. The loss may result from the homogeneity of the clusters, from their unequal sizes and from the fact that the degrees of freedom are fewer. Sometimes, one can limit the loss by using a sampling technique which at least takes into account the sizes of the clusters.