Interpretation of interaction in post hoc probing

[Mike Rouven]

Hello,

I (using STATA 14.2) have the following hypothesis: Geographic proximity (continuous) moderates the relationship between firm type (binary) and firm cooperation success (binary).

I answer this with the following regression:

xtlogit coop type c.center_geo type#c.center_geo c.log_size c.exp age, re vce(cluster pid) or nolog

The interaction was significant, so I did a post hoc probing for better interpretation. First, I created two new variables by adding and subtracting the standard derivation from the moderator "geo", which is centered:

sum(center_geo)
gen High_center_geo=center_geo-(.3458912)
gen Low_center_geo=center_geo-(-.3458912)

Subsequently, I re-performed the interaction with each of the two values in a separate regression:

xtlogit coop type c.High_center_geo type#c.High_center_geo c.log_size c.exp age, re vce(cluster pid) or nolog
estimates store post_hoc_1
xtlogit coop type c.Low_center_geo type#c.Low_center_geo c.log_size c.exp age, re vce(cluster pid) or nolog
estimates store post_hoc_2

The result of the post hoc probing looks like this:

My question is now: how do I interpret the result? The odds ratios for both constellations of the interaction are identical (odds ratio: 1.492), which I understand that both high geographic proximity and low geographic proximity moderate the relationship between firm type and firm cooperation success. However, this does not make sense to me, since this would mean, in reverse, that the moderator, formulated in extreme terms, has the same influence on the relationship both when it exists and when it is absent. Am I interpreting the results correctly? If so, what can I do to make differences visible? I have already tried double and quadruple standard derivation, but the results above do not change at all in terms of odds ratio.

Thank you very much for your support!

[Dick Campbell]

Check you gen statements. The two terms you have created are identical.

[Mike Rouven]

Thank you very much for your reply.
As I understand it, the two gen statements differ in the use of the standard derivation , since on the one hand the standard derivation is subtracted (High_center_geo) and on the other hand the standard derivation is added to the centered moderator (Low_center_geo). The command -sum- also displays different mean values, maxima and minima:

[Tim Meier]

Is there anyone who can kindly help me with my problem? I would be so grateful if someone would take a moment.

[Mike Rouven]

Is there anyone who can kindly help me with my problem? I would be so grateful if someone would take a moment.

I think this was the wrong thread, but nevertheless I would of course also be very grateful if someone could help me.

[Dick Campbell]

I misread your original code. I apologize for the error. Look at the correlation between high_center_geo and low_center_geo. It should be 1.0 because all you have done is added or subtracted a constant from the original value.

Here might be a better way to think about this. The interaction says that the slope of coop on center_geo varies (is conditional on) by the binary variable type and, conversely that the slope of coop on type varies (is conditional on) the value of center_geo. I would have have explored the interaction by looking at the conditional slopes in the first case, getting separate estimates for the model within the two values of type, holding the remaining variables in the model at some reasonable value. Actually, Stata margins and marginsplot commands will do the heavy lifting for you. You decided to go the other way, looking at the slope of coop on type conditional on specific values of center_geo. But it doesn't look to me that you actually looked at conditional models. I note that the output you show use has 3000 cases for each model suggesting that you used the entire sample in running both models. Don't you want to restrict model 1 to cases where you have high proximity and model 2 to cases where you have low proximity?