gretl, like stata, incorrectly computes the z-statistics of interaction dummies in logit and probit output. the computation leaves out essential cross-partial derivatives of y-hat, and thus incorrectly computes the z-statistic and possibly reverses the coefficient sign. Norton et al (2004) corrected this problem in stata by creating the command "inteff." It would be nice to see a similar command available to gretl users.
OK, we'll look into this.
This sounds like a bug description to me, so I'm moving this over to the bug tracker.
thanks,
sven
make summary more precise --sven
setting priority -- BTW, my previous comment wasn't meant as imperative, but was just supposed to be a description of what I did to the title. --sven
I would suggest re-filing this as a feature request. It's a bug only if you
expect your econometrics program to be in psychic communion with
the user. Note that the "fix" in stata is a user-contributed special, not
a change to stata's basic behavior. (I agree that having something
similar in gretl would be desirable.)
Yep I guess you're right Allin, so moving this back.
-sven
p.s.: and resetting priority to normal
-sven
Ok, I'm closing this now because it seems clear that this will not happen in the core gretl program. It should be doable in principle to contribute a function package that does the same thing as Stata's 'inteff'.
Sorry,
sven