Discussion
The designs presented in this paper provide flexible choices for 12-, 16-, 20- and 24-run orthogonal split-plot experiments under five realistic design scenarios. We proposed an approach for assigning word lengths to the factorial effects under each of the five scenarios and then used the resulting EWLPs to rank the designs. For identical values of N,n1 ,n2 and S we showed that the generalized MA designs in our catalog fare well when compared to designs constructed via other approaches. This is attributable to the fact that for a given run-size and number of columns (factors) our construction approach examines all non-isomorphic OAs.
Undoubtedly other design scenarios and ranking schemes could be proposed for ranking such designs. We are neither wed to our five proposed scenarios nor to the ranking schemes; however, we believe our approach encapsulates many of the circumstances and objectives encountered by experimenters.
One deficiency of the MA criterion and its generalizations is that it is difficult to incorporate certain design principles such as effect sparsity and effect hierarchy. In the nonregular factorial setting, Bingham and Chipman (2007) use a Bayesian approach for incorporating prior information in optimal design selection. Having pre-specified which models are most plausible, their methodology allows the user to select designs (possibly nonregular) that discriminate between competing models. A natural extension of this research would be to adapt such techniques to the split-plot setting. This topic is presently under consideration.
The proposed approach in this article focuses on orthogonal split-plot designs. Alternatively, one could choose to construct a D-optimal split-plot design, which has been implemented in JMPs . The D-criterion used in JMPs does not distinguish between the five scenarios. A cursory inspection of several of our designs in comparison to the ‘‘D-optimal’’ designs found in JMPs suggested that our designs generally have high relative efficiencies. Although this is a desirable characteristic, it is not the intention of this article to compare directly orthogonal designs with D-optimal designs. While the optimal designs can be constructed more easily using a computer algorithm, such designs are usually nonorthogonal. Orthogonal split-plot designs can be more preferable in practice as important effects can be more easily chosen to be free of aliasing with other effects. If one wishes to further de-alias low-order effects, common methods for doing so are readily available in the literature (e.g., foldover and semifoldover designs).