Reporting probit/logit marginal effects for interaction terms in regression table

[Stijn Loon]

Hi all,

As is well known by many on this forum, there are some difficulties with calculating marginal effects for interaction terms. One way it has been suggested to deal with this problem is to calculate marginal effects for a variable at representative values. For instance, if I want to calculate the marginal effects for the interaction between x and y, one way to do so is to write:

margins, dydx(x) at(y =(0 1))

This will give me two marginal effect values, one for x=1 at y=0, another for x=1 at y=1.

Here is my question. Anytime a probit/logit model with interaction terms is used in an article, the marginal effects are reported as a single value, as though they are normal regression coefficients. This may just be because one of the representative values at which a marginal effect is calculated is implicitly taken to be most interesting or relevant for the purposes of the article (e.g. x=1 at y=1). Or is it just that I am missing something?

[Richard Williams]

Actually, I am not familiar with papers like you are citing. Can you give a publicly internet-available link or two? If what you say is correct, I imagine the papers are doing something different than what Stata does. Or they may even be wrong. If you compute the interactions yourself, even Stata wouldn’t know that the main effects and the interactions are related to each other. With Stata’s old mfx command, interactions were not handled correctly. The margins command is only around 10 years old, so it wouldn’t surprise me if older papers had mistakes.