Estimating treatment effects in clinical crossover trials

Some current approaches to modeling crossover trials in two treatments are critically reviewed from the perspective of the practical requirements of the drug developer. Particular attention is paid to the AB/BA design, and the inadequacies of the once popular two-stage procedure are discussed in detail. The use of baseline data is also examined. Both frequentist and Bayesian alternatives to approaches currently advocated are considered and critically compared. It is concluded that it is crucial for the applied statistician working in this field to have an appreciation of the practical medical and pharmacological background.