Factorial and response surface designs robust to missing observations

Compound optimum design criteria which allow pure error degrees of freedom may produce designs that break down when even a single run is missing, if the number of experimental units is small. The inclusion, in the compound criteria, of a measure of leverage uniformity is proposed in order to produce designs that are more robust to missing observations. By appropriately choosing the weights of each part of the criterion, robust designs are obtained that are also highly efficient in terms of other properties. Applications to various experimental setups show the advantages of the new methods.