Am 17.06.2010 13:46, schrieb Riccardo (Jack) Lucchetti:
On Thu, 17 Jun 2010, Sven Schreiber wrote:
> I'm trying to get the estimates from a restricted logit model.
> Here's roughly what I mean to do (except it's not working the way I
> open greene12_1
> logit selfemp age income const
> matrix unrestricted = $coeff
> end restrict
> matrix restricted = $coeff
> I also tried the "--full" option for the restrict block, but that just
> gives me errors.
> Is there a way in gretl to get restricted logit estimates? (without
> manually coding it, I mean)
There isn't. With a bit of work, the --full option could be extended to
logit/probit models. However, why should we? Let me explain: in the case
of logit/probit models, the situation is very different if your
constraint is homogeneous or not. (Just to establish notation: you would
like to estimate a vector \beta under the constraint R\beta = d. If d
If d is NOT zero (like in your example), then it isn't clear to me why
you would want to test such a constraint, since the \beta vector in
logit/probit models is identified up to a scale factor, which is why
everyone uses the arbitrary convention of setting \sigma^2 to 1.
In other terms, in this context it makes sense to compare the relative
magnitude of coefficients, but (if I'm not mistaken) a hypothesis on
their absolute value does not.
Thanks for your explanations, jack. I will try various ways of answering:
First, I actually want to restrict more than one coefficient, I gave the
example with one coefficient because I thought I was simplifying, but
apparently too much so. So it is about the relative magnitudes. As I
understand you,that's not possible?
Second, the background is that we are actually estimating a panel logit
model (with Stata). Since we are using conditional fixed effects, the
unit (or group) effects are not estimated. My idea was to plug in the
beta estimates from Stata into gretl in separate logit models unit per
unit and then back out the constant terms. Of course I could do that by
explicitly writing down the likelihood conditional on the given beta and
maximize that, but I thought maybe there was a quicker way.