Abstract: | In certain areas of medical research, the two-period crossover design is a frequent choice for comparing treatments A and B in a randomized clinical trial. Earlier work by Grizzle and by Brown was based upon a parametric theory linear model. Recently, the present authors employed D. R. Cox's additive randomization models and, for the case of zero residual effect, found a discrepancy between it and the parametric model with respect to the precision of period effects. In the present note, this divergence is accounted for by allowing for the possibility of non-additivity through the use of a completely general randomization model. It is concluded that the structure of the crossover design is such that use of the parametric theory linear model is required if a single, consistent model is desired. |