I wasn't familiar with that article. Just skimming through it, I
notice it is using Stata 10 and the old -mfx- command. Part of the
problem with -mfx- is that it doesn't realize when variables are
dependent on each other, e.g. it doesn't know that if x1x2 = x1*x2,
then if x1 = 0 x1x2 must equal 0 as well. Factor variables and the
margins command addressed those concerns in Stata 11.
I must confess, I get confused by all these discussions about the
marginal effect of an interaction term. The margins command doesn't
even report a separate marginal effect for the interaction term --
instead it just reports marginal effects for the terms used to
compute the interaction. That makes sense to me, since an interaction
term can't change without either or both of its component parts
changing. So, I am not quite clear on what people want or expect when
they ask for the marginal effect of an interaction term. Maybe I just
need to break down and read these articles more carefully!