DOE/DSD: References talking about effect of runs with no results or removed runs

Question

Hi,

a customer asked me for good references about the effect it has to remove e.g. center runs or any other run from an DSD, or if a run does not profide a response value (e.g. an experiment exploded).

Of course we can check for the effect, ... but is there any (good) scientific paper or book talking about this in more detail (e.g. effect on degrees of freedom, orthogonality, ...)

Thanks,

Martin

louv · Accepted Answer

Martin,

From an applied perspective the degrees of freedom and orthogonality are the "lawyer" issues.

The detective persona makes me think if the treatment combination exploded that would be a zero rather than a non-entry. I would certainly not be afraid to penalize an explosion.

Ryan_Lekivetz · Accepted Answer

Hi Martin,

I'm not aware of anything in the literature.

The biggest issue is if you remove a non-center point, you break the foldover structure. You'll still have a "nice" design, that will be close for orthogonality, but the Fit Definitive Screening platform would not fit the response. The pure center runs being removed won't cause the same problem, as you maintain the foldover structure.

As Lou said, it wouldn't concern me, other than the realization that you won't be able to use Fit DSD for analysis.

martindemel · Accepted Answer

Thanks, so I can conclude: (?)

* using two stage forward selection in GenReg (or double lasso) could be used then instead of Fit DSD (Of course if one has JMP Pro).
* removing non-true center points effect the analysis more than removing true centerpoints

Of course taking into account all detective approaches to confirm the model as best as possible.

Ryan_Lekivetz · Accepted Answer

Yes, that looks reasonable to me. My typical approach is to use GenReg in conjunction with Fit DSD to see how the models match, so it really is just losing the ability to use Fit DSD.

bradleyjones · Accepted Answer

Martin,

I do not think there have been published articles investigating the effect of missing runs in a DSD. Here is what I know:

1) a missing center run will not affect the orthogonality of the main effects or the mutual orthogonality of main effects and 2nd order effects. The only negative is that one of the quadratic effects will no longer be estimable.

2) a missing run that is not a center run will destroy both the orthogonality of the main effects and the mutual orthogonality of main effects and second order effects.However, the factor with a value of zero for the missing run will be orthogonal to the other factors and all the 2nd order effects. The two-factor interactions involving this factor will also be orthogonal to all the other main effects. Small correlations among the other main effects will exist and will increase the variance of these main effects by roughly 10%.

I hope this helps.