EVOP/Mixture Design/Frankenstein Design Question

Question

I don't have lots of experience with mixture designs so I'm looking for some personal guidance for 'next time this happens' with a prospect/customer. I've got a prospect that uses extreme vertices mixture designs quite often. Their thinking was, if one region of the design factor space looks promising, but it's on the edge of the space, what about moving the design space ala an EVOP kind of approach in the direction of promise for a second design. So second design overlaps a relatively large portion of the first design space. First analyze the second design as it's own distinct single design. Then bolt the two designs together and analyze the entire mess.

My first thought is if nothing else the second design can serve as a means to confirm predictions coming from the first design since the space overlaps. Are there other 'gotchas' I just haven't considered from an analysis point of view for the Frankenstein design as it builds...maybe into EVOPlike iterations, 3 and 4?

Something tells me the Fit Model/Standard Least Squares platform might not be the best place to analyze the Frankenstein design? Maybe Generalized Regression, penalized methods?

I'd love to hear people's thougths.

bradleyjones · Accepted Answer

I agree with Mike that the Augment Designer is the tool to use here. Once you click the Augment button in that designer, you can change the ranges of the factors to move to a new region. You should also click the check box to make the new runs a separate block. If that effect is significant, it is an indication that the intercept has shifted between the two sets of runs.

bradleyjones · Accepted Answer

Mixture designs have built-in correlation due to the fact that the percentages of the ingredients must add to 1. Further constraining the ingredients by adding upper and lower bounds or linear inequality constraints makes for still more correlation in the parameter estimates. Finally, adding more and more model terms (2nd and 3rd order effects) makes for still more correlation.

As a result, I believe that mixture designs should almost always be viewed as response surface designs. The most important design diagnostic is the variance of prediction within the region of experimentation.

For model selection, I would recommend removing higher order terms in groups to see whether each group makes substantial differences in the predicted responses. You can also add up the sums of squares for a group and do a multiple degree of freedom hypothesis test to see the significance of including that group of terms.