Re: [Gretl-devel] logit problem
Am 12.11.2014 um 23:30 schrieb Allin Cottrell:
On Wed, 12 Nov 2014, Sven Schreiber wrote:
> Hi,
>
> when 'logit' is automatically dropping variables (or observations?)
> because of perfect prediction, there is a problem it seems:
>
> <hansl>
> open abdata.gdt
> series binary = (INDOUTPT > INDOUTPT(-1))
> series bincheck = binary logit binary const EMP bincheck
> </hansl>
>
> produces this error message here:
> <error>
> Note: bincheck != 0 predicts success perfectly
> 338 observations not used
> Not a Number geschah bei Berechnung
>
> Fehler bei Skriptausführung: Stopp
> > logit binary const EMP bincheck
> </error>
>
> (BTW, note the non-translated strings, perhaps they still need to be
> marked for translation?)
Well, at least gretl didn't crash! But seriously, how do you reckon
gretl ought to respond when fed a nonsensical specification such as
this? Should we check every regressor to verify that it's not identical
to the regressand? That would enable us to emit a more precise error
message but I'm not sure it's worthwhile. (And it would still miss cases
where a linear combination of the regressors is exactly equal to the
regressand.)
No, the point is that according to what I've seen and AFAICS also what's
in the docs, when gretl is able to find the "offending" variable it
drops it and then carries on with a reduced specification.
The example above was just to provide a showcase for a failure of this;
I thought that the message "bincheck != 0 predicts success perfectly"
implies that gretl knows what's going on and thus I expected that it
would drop 'bincheck' and then estimate a logit model with the remaining
variables.
Which part of my assumptions was incorrect?
thanks,
sven