Estimating dispersion parameters in the analysis of data from crossover trials

Abstract

Continuous data from a crossover experiment are likely to be correlated within each subject and an efficient analysis may have to estimate the dependence in the error term. Assuming that the error term follows a first-order autoregression, and that subject effects are taken to be fixed, two methods of estimating the autocorrelation coefficient are considered. One method applies restricted maximum likelihood, the other uses conditional profile likelihood. Simulation studies for three-period and four-period designs and a numerical example are presented.