|
Abstract:
|
Longitudinal data are frequently analyzed by linear mixed models. Optimal designs for these analyses, however, depend on the specified model. Practitioners seldom know the model structure before data collection. When the underlying true model differs from the model that is used to specify the optimal design, this design becomes inefficient. We apply a D-optimal maximin criterion to find a set of designs that remain highly efficient whenever the model is incorrectly specified. Based on different model and parameter spaces, the relative efficiency of each chosen optimal design with respect to the true model is computed. Thereafter, the minimum relative efficiencies as function of the true models are derived. Finally, designs with the maximum efficiencies are identified. We show that these robust designs have a larger efficiency than the conventional equally spaced designs.
|
