Optimal crossover designs under premature stopping

Abstract

Crossover designs are designed for experimentation over a certain number of periods but sometimes, for practical reasons, the experiment has to be terminated prematurely. This may lead to a problem since the resulting design after premature stopping may not allow efficient inference on the parameters of interest. So, as an insurance against such premature stopping, we need designs which are optimal if the entire experiment may be performed and which also remain optimal if the experiment has to be stopped at some intermediate stage. We show that the locally balanced designs of Anderson and Preece (2002) are good in this respect as they remain optimal if we stop any time after 3 periods. However, these designs require a large number of experimental units. As an alternative, we construct some designs which have substantially fewer units and which remain optimal if the experiment is terminated halfway.