Categorical variables and interaction in MLR

[Myron Tsisinskiy]

Would it be correct, to add in multiple linear regrression interaction between two categorical variables (i transformed them - dummy variavles) with coding 0 and 1?

 

[noetsi]

It depends on whether your theory supports doing so or not. Dummy variable interaction is not different than any other interaction.

 

[Dason]

Dummy variable interaction is not different than any other interaction.

It's a little bit easier to get your head around a dummy variable interaction than an interaction between continuous covariates though!

 

[noetsi]

It's a little bit easier to get your head around a dummy variable interaction than an interaction between continuous covariates though!

True, but theoretically (I think) it's the same thing. Which is what I meant.

 

[spunky]

but as Jake corrected in a post of mine sometime earlier, the interpretation of dummy-coded interactions can be pretty confusing. effects coding is what i think people usually go for when they're dealing with interactions between categorical variables...

 

[Dason]

Is effects coding the {-1, 1} way of coding things?

Edit: And by that I mean it's either 0 or 1 for all groups except one group that is -1 for all of the dummy variables...

 

[spunky]

yep... that's how we know it in the social sciences.

how would you call it?

 

[Dason]

Meh - I don't really use that very often. But if I think about it when I learned it I think it was "factor effects coding" - so... yeah ... I guess effects coding makes sense.

 

[noetsi]

Strange I always use reference coding...although the only thing I usually care about an interaction variable is if its sig or not. When it is I try to find out why it is substantively and look at graphs and run the regression at different levels of the interacting variables...

I am not sure what it matters if you use reference or effect coding if your doing that

 

[Jake]

We call them "contrast" codes over here