Robust split-plot designs for model misspecification

Abstract

Many existing methods for constructing optimal split-plot designs, such as D-optimal or A-optimal designs, focus only on minimizing the variance of the parameter estimates for the fitted model. However, the true model is usually more complicated; hence, the fitted model is often misspecified. If significant effects not included in the model exist, then the estimates could be highly biased. Therefore a good split-plot design should be able to simultaneously control the variance and the bias of the estimates. In this article, I propose a new method for constructing optimal split-plot designs that are robust under model misspecification. Four examples are provided to demonstrate that my method can produce efficient split-plot designs with smaller overall aliasing. Simulation studies are performed to verify that the robust designs I construct have high power, low false discovery rate, and small mean squared error.