Subscribe Bookmark RSS Feed

Transformation of factor (in fit-model dialog) results in high VIF

joberi

Community Trekker

Joined:

Jun 23, 2011

Hello,

After execution of an experiment (CCD), I realized that it makes sense to log-transform one of the factors (I do it in the fit model dialog), which eventually gives a decent fit. However, the linear and square term have a high correlation then (VIF = 56). When manually transforming this factor, the VIF is low (VIF < 2). Can someone explain to me how this can happen and how I can avoid it?

Thanks!!!

Joachim

5 REPLIES
jiancao

Staff

Joined:

Jul 7, 2014

I tried with my data but didn't see this behavior. Could you share the data and the models you ran?

joberi

Community Trekker

Joined:

Jun 23, 2011

Attached now the file - should be self explanatory, just run the two models and check the VIFs. I am actually wondering, if it is related to the coding somehow. If using the "transform" option, I cannot code.

Thanks!

jiancao

Staff

Joined:

Jul 7, 2014

Yes, getting high VIFs is a result of the square term, Log(B)*log(B), not centered. (Center Polynomials is the default option when higher order terms are specified as model effects in Fit Model see Model Specification Options​)

12979_pastedImage_0.png12984_pastedImage_2.png

Somehow JMP won't automatically center the crossed terms on the transformed variables created temporarily within the Fit Model window.

joberi

Community Trekker

Joined:

Jun 23, 2011

Hmm - ok - and what can I do now? Any workaround?

Thanks!

jiancao

Staff

Joined:

Jul 7, 2014

The work around is just like what you already did: saving the transformed columns to your data table before you specify your models. Are you using this feature Transform Columns​? It lets you quickly create new columns  (e.g.,log) in Fit Model launch window (in the column list pane) and then use Add to Data Table option  to save them--done with two right mouse clicks.