Logit with interactions: report logits or margins?

[Hanna Kuschel]

Dear community,

I am working on several logit models for my thesis. Some of them include interaction terms, others don't. Now I am unsure whether to report the original b-coefficients (logits, log odds) of the model or whether to just report the associated margins (or combine and report both measures).

My concerns:
Logits/log odds:
Pro: used and reported in similar papers, Stata retuns values for interaction terms
Contra: I want to plot the interaction effects by using marginsplot. But if I understand and see correctly, this plot uses margins as a basis. Thus, I am not consistent in my measures if I report logits in the table, but graph interactions as margins, right? Or is it acceptable to do this?

Margins:
Contra: No values returned for interaction terms (Dummy#Continuous interaction) when I use the margins, dydx(_all) command after the logit command (but I learned that I can compute them manually by specifying the 0/1-categories of my Dummy#Continuous interaction separately within the margins, dydx(_all) command, right? I can then enter my results in my regression table).
Pro: Consistent measures for table and plot

Report both measures: I found that some independent variables are significant as logits but not significant as margins. This speaks against reporting both measures.

What is your recommendation?

I am looking forward to anyone who can help me with this!
Thanks a lot in advance,
Hanna

[Richard Williams]

I really don't see why you can't report both. Both have informative value. I pay much more attention to the significance of coefficients than I do to the significance of the marginal effects. To the extent that you rely on significance tests when choosing your model, the significance of the coefficients is going to guide what you include, not the significance of the marginal effects.

Vince Wiggins touches on this and other things in

https://www.stata.com/statalist/arch.../msg00293.html

Among other things he says "If you are trying to test whether an interaction term in your model is statistically significant, do that in the metric in which you estimated the model. That is to say, look at the test statistics on the interaction term."

Vince also explains why margins doesn't report interaction terms separately.

[Richard Williams]

Also, welcome to Statalist. If you haven't seen it already, my handout explaing (and praising) margins can be found at

https://www3.nd.edu/~rwilliam/stats3/Margins01.pdf

[daniel klein]

I want to plot the interaction effects by using marginsplot. But if I understand and see correctly, this plot uses margins as a basis.

I think you might be confusing (average) marginal effects with adjusted predictions here; margins can produce both and marginsplot can plot both. You probably need to be more specific about what exactly you want. I usually plot the predicted probabilities to visually present results form a regression model.

Best
Daniel

[Hanna Kuschel]

Thank you both, Daniel for pointing out my confusion - you were right here. Richard's great handout about margins helped me to clarify that. Your reponds really helped me a lot!