Interpreting interaction from proc genmod

[Bluey44]

Apologies if this is a very basic question.  But it has been awhile since I've run a model with an interaction term and I want to make sure I am interpreting things correctly.  I saw a similar question was asked recently so apologies if this is duplicative but I still wasn't sure I 100% understood the interpretations there.

This is from a proc genmod model.  In the first two tables it shows the interaction term is non-significant.  But in the  differences table there do appear to be some strata that are statistically significant while others are not.  Would I still use the main maximum likelihood tables to conclude there was no interaction?  If there was interaction would I expect to see more differences between the strata effect estimates in the least square model (i.e. 95% CIs not overlap)?  Or am I supposed to look at the differences table and conclude there is some kind of interaction going on?

Also is there way to get just an overall relative risk measure for the interaction term?  Or is that not of any use?  In my write up of testing for significance of the interaction do I just cite the p-value?  (the other effect estimates included for other analyses are relative risks).

[StatDave]

The main effect tests in the Parameter Estimates and Type3 tables show that ANYBPZ is strongly significant but HEALTH is not. So, it is not surprising to see that the two comparisons in the LS-means differences table that are significant differ on ANYBPZ. You might find it more useful to use a SLICE statement to analyze the interaction. The following will give tests of the ANYBPZ difference in each level of HEALTH:

   slice anybpz*health / sliceby=health;

[Ksharp]

Check this for interact effect explanation:
https://support.sas.com/kb/38/384.html

Check this for relative risk:
https://support.sas.com/kb/23/003.html

[StatDave]

Regarding the overall relative risk, in the model with interaction it would be the average of the two differences of ANYBPZ in the HEALTH levels (or vice versa if preferred). You could do that with an LSMESTIMATE statement like this:

lsmestimate anybpz*health 'anybpz RR' 0.5 0.5 -0.5 -0.5 / exp cl;

But since there is no strong evidence of an interaction, that estimate would probably be pretty close to the estimate form a model without interaction which you would get from your LSMEANS statement with the DIFF and EXP options. I'm assuming you are fitting a model to a binary response with DIST=POISSON LINK=LOG and an offset so that you are modeling the log of the event probability.