topic Re: Interactions without principal effects in SAS Data Science

Interactions without principal effects

frak — Sat, 28 Nov 2015 15:59:04 GMT

Hi all, i've a problem:

i'm using a node of regression in SAS EM WORKSTATION 12.1; I'm using a cross-validation method and my aim is getting a logistic regression with the possible insertion of significant interactions.

So my settings are:
principal effect--> YES
2 factors interactions -->yes
polynomial terms -->yes
maximum degree of polynomial terms = 2
regression kind = logistic regression
selection method= stepwise
selection criterion= crossvalidation error (this bcs i'm using cross-validation)

The final model proposed is:

Intercept Gravidanze*Pedigree IMP_BMI*IMP_Diastolica IMP_BMI*IMP_Glucosio IMP_Insulina*Pedigree
(for the full output see the attachment stattalk.txt)

So all my effects are interactions of quantitative variable ad no principal effect seems to be significant.
I'm trying to find an explanation to this output.. can someone help me?

But that's not all, because if i modify settings and sat up "2 factor interactions-->NO" (I do not modify others settings)
and i run the node i obtain the same output.. ..so i think there is something wrong, but can someone tell me what?

i wait for some advice Smiley Happy
thank you in advance!

Re: Interactions without principal effects

WendyCzika — Tue, 01 Dec 2015 14:39:10 GMT

I'm guessing those terms you are seeing that include 2 inputs are actually polynomial terms (those are "interactions" of the interval inputs), so you need to set Polynomial Terms to No as well if you don't want any interaction effects. The Two-Factor Interactions property applies to interactions involving the class inputs only.

Re: Interactions without principal effects

frak — Tue, 01 Dec 2015 17:40:28 GMT

Ok 🙂 thanks a lot, now i understand why using the setting "two Factor Interaction" or not my output was the same! 🙂

Althought my problem is that that a model with only interactions (in my opininion) has no sense: interaction has sense only if also principal effects are significant in the model. So now my new question is:

how could i tell EM to use a hierarchic procedure while it puts variables into the model?

I mean how can i tell it FIRST to enter significant principal effects and only THEN interactions (between princ. eff. wihch had entered as significant)?

I imagine that i should modify the settings in "use default selection values" in model selection panel but, what should i modify in particular?

Again, thank you in advance!

Re: Interactions without principal effects

WendyCzika — Tue, 01 Dec 2015 17:52:27 GMT

In the Regression node, you can set Use Selection Defaults to No, then click on the ellipsis for Selection Options. There you can change Hierarchy Effects to All then select Single for the Moving Effect Rule property. Hope that does the trick!

Re: Interactions without principal effects

frak — Tue, 01 Dec 2015 18:26:30 GMT

Perfect! thank you again!