Screening and Follow-Up Experimentation with Partial Replication

Robert D. Leonard and David J. Edwards
Virginia Commonwealth University

Optimal design of experiments can provide a useful way to create and augment screening designs when information concerning the system model can be harnessed to guide the choice of design points. Dependable model selection procedures are also needed to convert newly collected data into useful information for updating the criteria of interest. We present results from a simulation study focusing on the benefits of using partial replication in screening and follow-up experiments. A modified forward selection procedure is compared to the Dantzig Selector using scenarios that incorporate variation in model assumptions, signal/noise ratio, number of active effects, and increasing complexity of the truth model. In particular, a Bayesian D-Optimal approach is considered in a sequential follow-up procedure to improve the ability to detect active two-factor interactions.

2015 QPRC