Status: Resolved (Dave Doehlert)
Are the D-Optimal designs, in particular those generated with Gosset, just as good as I-Optimal designs?
No. The Optimal (D) algorithm focuses on getting the coefficients in the model estimated well (it minimizes the volume of the confidence ellipsoid for the b's). This is not sufficiently relevant to most projects for which industrial researchers need designs. Researchers need to run a few trials and then predict the trials (in the region of interest) they did not run.
So a researcher needs a model that predicts well. This means that the design should provide predictions with confidence limits as tight as possible. Optimal (I) designs have the tightest possible confidence limits. This is "best precision".
Mathematical Note: Optimal (I) designs minimize the average variance of prediction, averaged over the region of interest.
03-FEB-97
STRATEGY for Windows Frequently Asked Questions