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; } } }

SVEM with split/strip designs

ChristianStopp

A customer is curious about the use of SVEM in analysis of designs involving the blocking/split/strip families.

Any thoughts/comments on the topic I might offer? I'd say related resources but something tells we we're still too early in the game there.

Thanks in advance!

Accepted answers

All comments

DonMcCormack

Until I see research saying otherwise, I wouldn't try this. A big challenge with split plots is the estimation of whole plot error. AFAIK, the work done with SVEM has only considered MS error for testing effects. Given they're asking about SVEM would lead me to believe that the designs they're looking at are sample size limited, compounding the problem of split plot error estimation.

ChristianStopp

Thanks @DonMcCormack. Indeed about the small sample sizes. Makes sense to me, I appreciate the response!

Phil_Kay

I agree with @DonMcCormack . SVEM for split plots does not seem like a natural fit.

However, I think it might still be informative. SVEM takes a predictive/machine learning approach rather than a statistical/hypothesis testing approach. So you could say that SVEM doesn't really care about accurately estimating the different sources of variance.

For SVEM, you would have to specify the whole plot effect as a fixed effect (you can't have random effects in Gen Reg, of course), which seems a bit of a weird thing to do, and you would end up with a model that is conditional on whole plot - again, weird. But it could still tell you about the most important effects in the system. In that way it might help to guide your model selection for a split plot experiment, which is challenging because you can't include random effects in model selection platforms in JMP or Pro.

I would never trust SVEM to give me the definitive "best" model. But I will almost always use it now as part of my analysis of any experiment.

DonMcCormack

You could only specify the whole plot effect as fixed if it wasn't confounded with model terms, which is less likely as the sample size gets smaller. I'd ignore the factor(s) using the whole plot as its/their error term, too (e.g., the hard to change factor).

SamGardner

From my observation, to use SVEM effectively, you have to use entirely different types of DOEs with SVEM. You need designs similar to Space Filling designs, with several interior points across the design region. Creating a split-plot design with that structure is not easy to do. And SVEM is for prediction, not inference, so if you want to uncover what effects are truly important, you need more orthogonal designs that don't work well with SVEM.

Here are some papers that discuss Fast-Flexible Filling split plot designs: https://arxiv.org/pdf/2003.09507.pdf, https://www.researchgate.net/publication/340115623_Space_Filling_Split_Plot_Design_using_Fast_Flexible_Filling

SamGardner

If you think there is a need for this, please make a wishlist request, or ask the customer to use the public wishlist.

ChristianStopp

Thanks @SamGardner for the additional feedback. I shared a similar takeaway from the resources I'd seen about SVEM better suited for more complex (and small to medium) designs; relatedly, I've leaned into Adsurgo's Discovery talk accordingly (which includes a space filling mixture premise) in the past. I'll share the PDF resource as well and see about their wishlist desire.