Dealing with "incorrect" or out-of-range runs

JL_LZH

Does someone have a hint for me on how to deal with out-of-range-runs (maybe there is a better name for it)? 

For example:

When running a test plan with the kerf roughness as response in a cutting process, the plan unfortunately includes some parameter sets where the energy is too low, resulting in just some scratches on the surface. I can't use the results for my response, but the outcome is still interesting and I want to include the information about the bad parameter set.

I know there is the option to set limits in the design phase, but maybe I just don't know about it exactly before or the combination of parameters is complex. 

Is there a way to include those results somehow when fitting a model? I believe that I read or watched something about it, but I can't remember what media it was. I'm pretty sure the post or video was also about informative missing..

@martindemel 

martindemel

It is not clear to me what the issue exactly is. I expect the following:

Some parameter settings result in no cut (just scratches on the surface due to too low energy), so no roughness value. These runs have therefore a roughness value which is set to missing, correct? (That would explain why you cannot use it in the modeling as you described, as Standard least squares ignores rows with missing values in parameter used for the model).

 

Assuming my expectation is correct, you might want to use in the model dialog "informative missing" to at least see if the missing value runs behave differently than the mean response of the other runs. 

In case you can calculate or measure the energy level, and assuming you know the minimum energy necessary to have a cut instead of just scratches on the surface, you might want to consider also using censoring to reflect the limit of detection for energy. 

(an old but still appropriate exampple on how to deal with that kind of data can be found here: GenReg for DOEs  (page 36ff of the PDF document)