Odd interaction results, what does this mean?

[Robert Oost]

Dear statalist members,

First of all, I am new to this forum and I hope to learn as much as possible while using stata. I appreciate such forums are available and we can help each other out. I am a beginner at statistics, therefore this question might sound simplistic to you who have more experience.

My first hypothesis stated that my independent variable would positively affect my dependent variable. So, the first regression testing my main relationship showed a significant and positive effect of X on Y. Therefore, I confirmed my hypothesis.

My second hypothesis argued that my moderating variable (or interaction term) would positively influence the relationship between my independent variable and dependent variable. Now, the regression showed some odd results. My independent variable became negative instead of positive (while insignificant) while the interaction term itself showed to be positive and significant (p<0,05).

Thus, the interaction term itself is positive and signficant as predicted. However, my independent variable is negative and insignificant. Therefore, my question is as follows: how should I interpret this result?

First I tried to find some information regarding interaction effects online, but it left me more confused and often these were quite sophisticated. I hope to get a simple answer in order to understand the basics first.

Thank you,

Robert

[Carlo Lazzaro]

Robert:
welcome to the list.
As per FAQ, the best approach to get helpful replies is to post what you typed and what Stata gave you back, so that interested listers can see what's the matter with your data. Thanks.

[Robert Oost]

Robert:
welcome to the list.
As per FAQ, the best approach to get helpful replies is to post what you typed and what Stata gave you back, so that interested listers can see what's the matter with your data. Thanks.

I did the following: regress Patchange (DV) Fhetero Fsize TMTSize Indus FAge Acqexp (controls) RDINT (IV) which gave me a significat and positive result for my IV (RDINT).

Then I did the following: regress Patchange (DV) Fhetero Fsize TMTSize Indus FAge Acqexp (controls) RDINT (IV) c.RDINT##c.Gendiv. This resulted in a significant and positive effect for the interaction effect Gendiv (moderator), while non significant effect for the RDINT (main relationship) and the coeffficient became negative (-6.3).

I am simply interested how I should interpret this. Should I reject my hypothesis or should I conclude that my hypothesis is confirmed, since interaction is significant? I predicted a positive effect of the moderating variable on the effect of RDINT on Patchange.

[Marcos Almeida]

Just as a side note to Carlo's excellent advice:

You may use - margins - followed by - marginsplot - so as to get an insightful view of the interaction effects.

Also, since it seems both variables are continuous, I wonder whether you centered them, otherwise the zero value can be meaningless, and that may be happening in your case.

[Maarten Buis]

Remember that when you include an interaction term between variables A and B, the main effect of variable A is the effect of A when B is 0. By including that interaction term you allowed the effect of A to change as B changes, so as a consequence there is no longer 1 effect of A, but as many effects as there are values in B. You can get weird values for the main effects, especially if the value 0 is extreme, e.g. if B was year of birth without centering at some meaningful value in the range of the data, then that would be the effect of A for someone born in the year 0, which is usually quite an extrapolation... The solution is to center your variables at some meaningful value in the range of the data, e.g. for year of birth I typically subtract 1960, so now 0 revers to people born in 1960.