I am currently working on presenting the results of a logistic regression that includes a dummy*dummy interaction in AME format. I am working with multiply imputed data, and Stata SE version 16.01.
When I had previously calculated the log odds, I found that the interaction term was negative.
However, when looking at AMEs, the same interaction becomes positive.
It should be noted that the interaction is not statistically significant, but I would still like to comment on the direction of the effect nonetheless.
Does anyone know why these results are so different, and whether I should "trust" the AME results over the log odds?
Thank you!
When I had previously calculated the log odds, I found that the interaction term was negative.
Code:
//LOG ODDS mimrgns dummy1, dydx(dummy2) pwcompare // dy/dx= -0.56, se=0.65
Code:
//AME mimrgns dummy1, dydx(dummy2) predict(pr) pwcompare // dy/dx= 0.06, se=0.12
Does anyone know why these results are so different, and whether I should "trust" the AME results over the log odds?
Thank you!
