DOE - Split Plot Designs: Does one ever include center points?

[#1]

Greetings:  I'm new to split plot designs, understand the basic structure and application.  Does one ever include center points in such a design to detect curvature?  I've searched and searched, but haven't been able to find an answer.  I suspect the answer is "No", but if anyone out there can refer a document or has an answer, I'd love to know the reason.

Cheers, Bruce

[#2]

Bruce, there would be nothing wrong with adding center points to a split-plot design.  You should read "Split-plot designs for robust product experimentation", Box and Jones if you really want to take advantage of the capabilities of split-plot designs.  They are NOT just for hard-to-change factors.

[#3]

Awesome Statman, thanks for the info. I'll get the text.

Happy experimenting!

Bruce

[#4]

Hi,

Came across your post, and I'm in same predicament. How do you set this up in JMP?

Joe.

[#5]

To add center points, the requirement is all of the factors must be continuous (there are workarounds, but none are ideal).  Code the factors levels -1, 1. Then all you do is add rows to the data table populated with 0's.  If you run center points randomly throughout the experiment, you can not only test for curvature, but also stability over the design space.

[#6]

Stability over the design space is an underrated use case for center points in my humble opinion. That "stability" is as important as testing for a curvature effect, because it gives the experimenter an indication of whether or not their measurement process is stable to begin with.  Without a stable measurement process, how can we trust the inferences we make from the results of our experiment on the whole?  This goes hand in hand with running properly design MSA studies that confirm that the measurement system meets certain pre specified criteria (e.g. for a continuous measurement process, we might want to see that %GRR  is less than or equal to a certain value, and/or that P/T Ratio is less than or equal to a certain value). 

[#7]

The only thing I'll add is maybe if your budget allows, just build a x-optimal design, that aims at estimating curvilinear effects in the first place. Depending on the factors and the nature of the split plot structure, you might end with a similarly sized overall design when adding center points to detect curvilinearity vs. estimating the effects outright. Build some candidate designs as @PatrickGiuliano and @statman suggest then use the Evaluate Design platform to see how they compare for power, correlation of effects, prediction variance etc. Worth a try at the least I think?