Design Augmentation

Augmenting a design adds runs to improve the properties of the design. Stat-Ease offers powerful tools for augmenting an existing data set. Access this feature by clicking on the Design node to get back to the design layout spreadsheet and selecting Design Tools > Augment Design from the menu.

It is always more efficient to design for the appropriate model in the first place, but if such knowledge is not available yet, default designs can be augmented to fit the true response surface better. The available options for augmenting a design depend on the initial design. Use the Help button on the augment dialogs if you are not sure what to do.

Response Surface and Mixture designs:

With the aid of optimal selection, the program provides additional experiments needed to fit a higher order model. If the lack-of-fit test shows significant, or the fit summary shows a higher order (but aliased) model fits better, and/or subject matter knowledge indicates a higher order model is needed, response surface and mixture designs can be augmented.

When using optimal augmentation to augment a response surface polynomial, Stat-Ease will add two new design points for each coefficient to be estimated. Two are needed because if only one were added, the new coefficients often end up with leverages of one and very large standard errors.

Factorial designs:

While using Stat-Ease, two-level factorial designs with center points are commonly augmented to a central composite designs.

Resolution III designs, including the regular two-level (red), and Plackett-Burman designs, augment best with a full-factor foldover. Resolution IV designs, including regular two-level (yellow), and Min Run Screening, benefit most from a single-factor foldover, or a semi-foldover. All factorial designs can be augmented with an optimal set of points.

Optimal augments are most commonly used to augment a response surface design to handle higher-order model terms. The new model is specified by clicking the Edit Model button on the dialog. The algorithm calculates the required number of runs to fit the new model. By default, the required runs are doubled in the additional model points. The extra runs do two things. First, they provide greater precision for estimating the new coefficients. Second, they correct the original design’s model misspecification problems. The additional model points can be set to zero, but it is not recommended. Extra lack-of-fit and replicate points are used to provide a check for the model and a more realistic estimate for the variation in the process that is not coming from the factors changing. If the additional model points must be reduced, replace some of them with replicates.