When to specify an interaction effect

noetsi

Fortran must die
#1
I have a model with about 20 variables in it (6 of them are dummies from a categorical variable with 7 levels). All of these are suggested by the literature. However, the literature makes no suggestion about possible interaction effects and I have no reason to suspect interaction (although I am not an expert in this field admittedly).

So do you just specify all two way interaction effects and see which ones matter? :p With 20 variables that would be a lot of interactions. Is there a diagnostic to suspect interaction?
 

hlsmith

Omega Contributor
#2
So you let literature guide you, though you don't want to miss something. I always think about it as a variables slope differing if controlling for another variable in the model (stratifying, it into two different slopes). You can always sleuth for them, but when do you stop (second order, third order, etc.) and how do you interpret them if you didn't even see them theoretically before hand.
 

hlsmith

Omega Contributor
#3
I know you may not completely be a content expert for data you are asked to run. So you can always think about an EXPLORATORY stepwise approach to release the feelers. you just have to then really think about how you want to treat results if you find something.
 

noetsi

Fortran must die
#4
THOUS HAST COMMITTED BLASPHEMY BY SUGGESTING STEPWISE REGRESSION :MAD:

Seriously I have never seen stepwise done with interaction. Wouldn't this cause a problem if the stepwise failed to add the main effects before the interaction term? I will have to look for literature on stepwise with interaction terms. I think of interaction much as you suggested, but the literature on this topic simply never brings that up. The field is not particularly methodologically sophisticated [which is obvious since I am making this comment and I am not particularly sophisticated]:p
 

hlsmith

Omega Contributor
#5
I think in SAS you can add a @2 or some other option. Yes, there are inherent issues in doing this, but it could help you take a quick cursory glance. Do you have continuous variables in your model?