PNNL

Abstract

The article “Response Surface Design Evaluation and Comparison” by C.M. Anderson-Cook, C.M. Borror, and D.C Montgomery is discussed. The discussion covers three issues. First, graphical methods for evaluating and comparing experimental designs should utilize the unscaled prediction variance (PV) rather than the scaled prediction variance (SPV). Using PV makes it easier for practitioners to see how prediction variance is decreased by designs with larger numbers of points, and to assess the “cost” of that decrease in a flexible manner. Using SPV confounds “prediction variance” and “cost” information in a pre-defined way, thus making design choice harder for practitioners. The second and third comments are that bias properties as well as variance properties of designs should be accounted for in constructing, evaluating, and comparing experimental designs.