CMStatistics 2020: Start Registration
View Submission - CMStatistics
B0475
Title: Design selection and analysis for two-level supersaturated designs Authors:  Rakhi Singh - UNC Greensboro (United States) [presenting]
John Stufken - George Mason University (United States)
Abstract: An extensive literature is available on design selection criteria and analysis techniques for 2-level supersaturated designs. The most notable design selection criteria are the popular $E(s^2)$-criterion, $UE(s^2)$-criterion, and Bayes $D$-optimality criterion, while the most notable analysis technique is the Gauss-Dantzig Selector. It has been observed that while the Gauss-Dantzig Selector is the superior analysis technique, differences in screening performance of different designs are not captured by any of the common design selection criteria. We will consider new design selection criteria inspired by the Gauss-Dantzig Selector. We will establish that designs that are better under these criteria also tend to perform better as screening designs. In addition, most supersaturated designs are studied under the main effects model because the number of runs is small to study even the main effects, let alone be studying the main effects as well as the two-factor interactions. However, in practice, there is no guarantee that the presence of two-factor interactions does not influence the response. We will also consider a random forest-inspired analysis technique that we have developed to circumvent this problem. We will see that this technique outperforms the existing analysis techniques.