Help with categorical interaction terms in Poisson model

Thanks in advance for any assistance you have to offer.

I'm working with a Poisson model of weighted survey data. My outcome is ACCESS to a weapon (0/1). I have several predictors, including GENDER (0 female /1 male) and a RISK category (0 = no mental health risk factors / 1 = intermediate risk factors / 2 = severe risk factors). I'd like to test an interaction between GENDER (0/1) and the RISK (0/1/2) categories.

My stata command is: svy, subpop(varname): poisson ACCESS AGE i.RISK i.GENDER i.RISK#i.GENDER varname3, irr

My specific questions are:
1) Is this the proper way to create the interaction term?
2) Why does the stata output only show results for the 1/1 and 1/2 interaction levels (attached below)?
3) Can I appropriately interpret each of the interaction IRR's as the IRR relative to the 0/0 interaction level?

View attachment 4266

I'm unsure why; but if I test the interaction term without the main effect terms, I get results for each interaction level (attached).
My stata command is: svy, subpop(varname): poisson ACCESS AGE i.RISK#i.GENDER varname3, irr

View attachment 4267


TS Contributor
1) Both are appropriate, but different ways of specifying the interaction term.

2) because it takes the reference category into acount

3) not in your first model, they are "ratios of risk ratios" instead of "risk ratios" (since you have a binary dependent variable, the exponentiated coefficients are risk ratios and not incidence rate ratios)

There is a discussion of these different ways of specifying the interaction term in this Stata tip, and this Stata tip could also be useful for interpreting interaction terms in this type of model.
Thank you, that's really helpful. One quick follow up questions re: your response to #2.

I think the total number of interactions from GENDER (0/1) and RISK (0/1/2) would be:
0 0 (reference)
0 1
1 0
1 1
2 0
2 1

Why would the output eliminate all but 1/1 and 1/2 (attached below)?

View attachment 4270

thanks again.


TS Contributor
That would be true if you included only one of the main effects. In your model you, correctly, included two main effects. I suggest you go through the articles I linked to before, and work out exactly what is being compared and who is compared with who for each parameter. Then you will understand your model.