We consider a three-period three-treatment crossover experiment to compare the effects of two drugs with a placebo. Subjects within each successive block of six are randomly assigned to the six possible treatment sequences. A full rank linear model is developed for this design. The experiment is carried out and several response variables are measured. Missing data disturb the balance, however, and normality assumptions appear untenable. The significance level of the results is therefore determined by random reassignment of treatment sequences to subjects, in accordance with the randomization actually performed in the experiment.