To assess the efficacy of potential new drugs in the initial phase of clinical research, one must use an efficient design that satisfies conditions to guarantee the safety of the subjects. For a parallel design, a two-period crossover design, two three-period crossover designs, and a Latin square design with three periods, we compared variances of estimators based on a mixed analysis of variance model. The proposed three-period crossover designs turned out to be only slightly less efficient than the Latin square design, which is not capable of satisfying the necessary safety conditions. The analysis of data from the crossover design poses several problems, including nonconstant variances for all observations and the possibility of carryover effects. To resolve these issues, we generalized the Box-Cox transformations to the mixed model at hand and, using simulation, investigated the sensitivity of the analysis to the presence of (first-order) carryover effects. This showed that results from the model without carryover are reliable for only very small carryover effects.