On Thu, 28 Mar 2019, Sven Schreiber wrote:
Am 28.03.2019 um 02:56 schrieb Allin Cottrell:
> There's a colleague in my department who uses gretl for all his
> empirical work, and does a lot of logistic regressions. He's a
> political economy guy and in many cases the dependent variable is the
> percentage vote in some legislative body for some bill or other.
> We have logistic regression in gretl but it hasn't had much attention
> in a long time. At my colleague's prompting I've recently enabled the
> --robust and --cluster options for "logistic". But in addition he also
> wants to use logistic regression in a panel context. We have fixed
> effects (binary) logit already -- any thoughts on what it would take
> to extend that to cover fixed effects logistic?
After looking at the help for 'logistic', spontaneously I'd say it's
much simpler than felogit. The LHS of the regression is some
transformation of a variable, but the same is true for example for a
simple log-transform. So I don't see why in terms of estimation a
standard FE specification with a transformed regressand wouldn't work --
of course always sticking to the (sometimes overlooked) assumption that
the regressors are strictly exogenous, so in particular no lagged
endogenous terms and so on.
Thanks, so that sounds quite promising for a --logistic option
for the "panel" command.
(Note that for nonlinear transformations there's also a
getting predicted values properly, because the expected transformed
value will not be the same as the transformed expectation. This could
also apply here AFAICS, but this would already affect the 'logistic'
command as-is, without panel fixed effects.)
Hmm, good point. It turns out, so far as I can see, that there's no
closed-form solution for the expected value of a sigmoid function of
an r.v. X, even if X is normal, but there are some standard
approximations. We can probably do a bit better than we have done up