Crossover designs and repeated measurements.

In crossover or changeover designs, the different treatments are allocated to each experimental unit (e.g. patient in clinical trial) in a randomized order. To analyze the results of such experiments, a mixed analysis of variance model is usually assumed. But such a model implies unrealistic and unnecessary restrictions on the variance-co-variance structure. These restrictions can be avoided by assuming the measurements obtained from a unit (patient) as repeated measurements from a multivariate distributed random vector. The experimental effects are primarily characterized by the mean of this vector. The usually defined treatment, period and interaction (residual) effects are estimable functions of these mean values. In this paper, the general approach is discussed and parametric as well as nonparametric tests for the various hypotheses are presented. This approach is developed in detail for the two-period crossover design and demonstrated with an example.