Am 31.12.2013 15:38, schrieb Riccardo (Jack) Lucchetti:
On Mon, 30 Dec 2013, Sven Schreiber wrote:
> Dear devels,
> hope you've had nice holidays so far.
> I have started using the new random-effects probit in gretl, and I will
> also try out the fixed-effects logit (conditional) function package
> soon. I was wondering, however, why (instead of RE-probit or in addition
> to it) Jack didn't implement the RE-logit. It would seem more natural to
> have the two logits together (RE and FE), and not an incomplete mix of
> probit and logit. (I'm aware of the impossibility of FE-probit.)
> Is it something technical which makes RE-logit more difficult to
> program? (I really have no idea.)
Not really. As long as individual effects are assumed to be normal, we
have a relatively efficient implemetation of gaussian quadrature which
should make in relatively easy. The only reason why we started from
RE-probit and FE-logit is that those are the "traditional" panel
estimators for binary dependent variables.
In fact, even people with little familiarity with C but proficient in
coding with tools like eg Python (hint, hint) should be able to use the
plugin/reprobit.c as a template and implement relogit without suffering
unendurable torments. Maybe it could be a new year resolution for these
people (hint, hint again). ;)
Or perhaps, somebody could try coding relogit in hansl throught the
ghq() function and make a nice function package out of it.
Ok thanks, good to know, I'll have a look when I'm back home.
In the meantime, here's some feedback on re-probit:
1) It seems that robust standard errors are not available for panel
probit; at least in the output it says "based on inverse Hessian" even
though the option '--robust' is passed. That's ok, but perhaps gretl
should flag an extra warning, not ignore silently.
2) also RE-probit: The reporting of the “lnsigma2” estimate in the
column of the slope coefficients is a bit confusing IMHO. For example,
it's not clear what the p-value is supposed to tell me; if the
log-variance is zero, then the variance was unity, and that's the
implied test for that?
3) In my case RE-probit leaves out td94 and td11 (those are manual time
dummies), but only says so for td11.
4) Speaking of time-dummies, it would be nice if "probit
--random-effects" also allowed the --time-dummies option. (or does it
already? at least it's not documented)
5) And it seems that panel probit only appears in the online help, not
in the guide (dated "November 2013").