Greetings.

Browsing ResearchGate, I find that Plamen Akaliyski asked a similar question and Daniel J. Tulloch gave a very useful answer.  With credit to these, Professor Tulloch recommended two YouTube videos, which may help you also:

Specifying interaction:

http://www.youtube.com/watch?v=f-tLLX8v11c

Interpreting interactions:

http://www.youtube.com/watch?v=ScKL40dp8M4

The videos use the statistical software Stata.

Let me reframe the question also because I have another problem. Given that i cannot use factor notation for interaction effects (I am using Mlwin environment within stata with runMlwin) I have to create all the combination by hand, But this seems meaningless if i am interested in only one of these combinations. If I can use only it (unemp*weekly=1, all the other combinations=0) and interpret the coefficient as the differential with all the other combinations together it is fine for me.

best,

Francesco

If your variables are dummy coded; it wouldn't be interpreted as the odds of being religious for that combination compared to all other combinations; it would be the odds of being religious for that combination compared to the odds if there were no interaction and the effects of Unemployed and Weekly were merely additive. i.e., you have an effect of Unemployed (showing the odds of being religious for Unemployed vs. baseline) and an effect of Weekly, and combining those in the regression equation gives you a predicted log-odds of being religious for unemployed-weekly people; the interaction shows how much different the actual predicted log-odds are from what that prediction would have been.

Thanks Stephen! This is exactly what I was looking for. So, it is proper tu use only that interaction term? I do not have to use all the combinations?

bests and thanks!

Whether or not it's proper is probably debatable. Mathematically I believe it's fine; it's really not much difference than if you had coded each of these variables as 2-level instead of 4- and 5-level (i.e., as "unemployed" vs. "everything else", and as "weekly" vs. "everything else"). Conceptually, on the other hand, people might argue that you are potentially ignoring meaningful variance and therefore not getting the best possible model. Personally, I feel like it's probably justifiable if the other interactions are irrelevant to your hypothesis, but I can't guarantee that others wouldn't challenge it.

Perfect. If mathematically is fine I have a lot of theoretical reasons to do that,

Thank you so much, really

bests,

Francesco

Careful: not everything that is "mathematically fine" makes sense in "real life" or is useful in "real life".

Examples: you can calculate the ratio of fuel consumed per distance traveled. Mathematically fine. Useful to compare engines, for instance. Absolutely useless to investigate the effect of carpooling. You can calculate the the ratio of fuel consumption and car weight. Completely useless as a measure of fuel efficiency. You can approximate a relationship of weight and age in a particular age interval well by some linear function, and it is mathematically fine to extrapolate this - but such extrapolations may be entirely insensible in real-life (e.g. to calculate the weigt you had 2 years before you were born)

How is unemployed coded? Is it a binary variable? If it is a separate binary variable than Job status then its interaction with ParentsWeekly is meaningful otherwise not. I suppose ParentsWeekly has 5 levels.

dear Jochen, I completely agree with your statement. In fact, I have a lot of a priori theorectical reasons for testing ONLY that interaction term but I had some doubt about its mathematical plausibility and thus about the interpretation of that coefficient.

Given that mathemaically seems proper, I will go with that model which is the most useful in my "academic real life"...

bests,

Francesco

dear Mehmet, bot varables are categorical and re-coded as dummy. So I have a dummy for unemployed (=1, 0 otherwise) and another one for weekly attendance (=1, 0 otherwise)