Interpreting Interaction Effect in Proc Genmod (Categorical vs. Categorical with multiple levels)

[varatt90]

Hi all, 

I was wondering if someone can help me interpret the interaction variables presented in the output below. 

If we pretend that the interaction term in the Type 3 table is significant, my understanding is that:

- interaction between being a boy (vs. Girl) and being Depressed (vs. Not depressed) - significant

- interaction between being other (vs. Girl) and being depressed (vs. not depressed) - significant

The values for girl is not provided as it is the reference group that I have chosen. 

Is this correct? Any help would be appreciated!

 

  

[StatDave]

When there is interaction in the model, you interpret the effect of each variable at each level of the other variable it interacts with. To do that, use the SLICE statement. For example:

slice sex*depression/ sliceby=depression;

or

slice sex*depression/ sliceby=sex;

[varatt90]

Here, I used spliceby = sex

So from the Analysis of GEE Parameter Estimates Tables, I can identify the significance of whether or not interaction exists. And I can use this Simple Differences Table to identify significance between sex groups that have depression. Is that correct?

[StatDave]

That appears to be the results of a statement like
slice sex*depression / sliceby=depression diff;
But, yes, that table lets you compare the sex levels when depression=yes.

[varatt90]

Yes! My apologies! I did: slice sex*depression / sliceby(depression) diff oddsratio CL ;

[varatt90]

You mentioned "When there is interaction in the model, you interpret the effect of each variable at each level of the other variable it interacts with. To do that, use the SLICE statement. "

Just to confirm, if we didn't assume the interaction effect in the Type 3 table was significant and went by actually what we see in the output (i.e., the interaction term in the Type 3 table was not significant.), then I would not be able to interpret the effect of each variable at each level of the other variable it interacts with. Is this correct?

[StatDave]

If there is no interaction then that means that the effect of each variable is the same at all levels of the other variable. So, that greatly simplifies the situation. Each variable just has an overall effect.

[varatt90]

Great! Thank you so much for clarifying that!

[varatt90]

Are there any resources that you think would be helpful reading regarding interactions between categorical variables (esp when the outcome variable is also categorical)?

[StatDave]

See this note that discusses models with interaction in general and testing hypotheses in such models. In particular, Example 3 illustrates and discusses an interaction model in a binary response model.

[varatt90]

Thank you so much!!!

[PaigeMiller]

In addition to the good advice from @StatDave you should also take advantage of the excellent plots available in PROC GENMOD via the EFFECTPLOT statement to help you interpret the interaction. There are interaction plots, SLICEFIT plots, and SLICEBY plots.

[varatt90]

When I specify "effectplot interaction (x=depression)" where Y = smoking status into my code, this is what I see. Is this what you meant?

 

[StatDave]

Yes, that provides a graphical depiction of the comparison results from your SLICE statement.

[SteveDenham]

Yes, those plots look exactly like the ones I used to illustrate an interaction years ago when I taught an introductory stats class.

SteveDenham