nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

DSD with a hard-to-change categorical

Phil_Kay

Hi,

I have an interesting DOE challenge from a customer.

They want to try a DSD for an experiment. They have 7 continuous factors and a single 2-level categorical. Simple.

However, the categorical is hard-to-change. It is the type of liner in the system. Changing the liner means shutting down the whole system, which is time-consuming and could lead to a change in the "baseline" performance of the system. So they want to minimise the number of changes.

I have had a few goes at creating a DSD for the 7 factors with 8 additional runs and 3 additional centre-points and then using these runs as covariates in a custom split plot with 4 whole plots. The designs are okay but I am wondering if anyone has other ideas.

Ultimately, I think it would be useful to have something as close as possible to a DSD. They are not too concerned about the number of runs (30 would be okay). They would probably be okay changing the liner every 3 runs if required.

BTW, they would be okay with us writing this up as a blog if it is sufficiently interesting.

Thanks,

Phil

Accepted answers

DonMcCormack

Don't use the blocking from the DSD platform, manually create the blocks based on the categorical factor values from the original DSD. You're not changing the design, only the run order. It will retain the properties of the original DSD.

All comments

MikeD_Anderson

Hi, Phil!

I think you want to use blocking variables. Since you’re basically trying to get to a split-plot design, I’m pretty sure that adding a blocking scheme of 2 and then treating that as your liner variable will do what you want. The output table will be sorted by block but randomized within the blocks.

A couple of caveats: I’d keep an eye on the correlations. And, you’re probably violating some of the assumptions in the DSD, so I’m not sure the Fit DSD will be appropriate. Maybe Brad or a DOE dev will chime in on that last point. As long as the hard to change factor drops out during the modeling process, I expect you’ll be okay – but again, you might want a second opinion there.

Just a thought.

Best,

M

Phil_Kay

Thanks, Mike.
I looked at using 4 DSD blocks to create 4 whole plots for the liner factor (2 whole plots per level of liner). The result is pretty good. However, we have a handful of correlations between Main Effects and 2FIs of ~0.5. Not very definitive. I wonder if we can do better?
And you are right that Fit DSD will not be possible with the added factor as the runs are no longer fold-overs. However, Fit DSD can still be used to screen the 7 continuous factors.
Thanks,
Phil

DonMcCormack

Phil,

Create four random block, two for each level, then alternately assign the runs to the appropriate random block according to the categorical factor value. To get the analysis, you can copy the scripts created when you generate a RSM split plot for the same factor configuration. You could do this a differnt number of block, if you want.

Don

DonMcCormack

Should have added that you'll need one column to ID the random block (name it Whole Plot so it matches what's in the analysis script) and one for the hard to change factor.

Phil_Kay

Thanks, Don. Yes, I tried exactly that (see above in thread). Not bad but not great.

Mark_Bailey

Custom design with alias D-optimal criterion?

Phil_Kay

I found a way that produces a better design in terms of Main effects being clear of 2nd order:

Create DSD for 7 continuous and 1 categorical. No additional runs. 18 runs.

Then use continuous factors as covariates in custom design. Ignore the original categorical factor and add a new one that is hard-to-change with 6 whole plots.

DonMcCormack