Am 08.07.2016 um 23:30 schrieb Allin Cottrell:
On Fri, 8 Jul 2016, Sven Schreiber wrote:
> The errors go away when I precede the estimation and omit commands
> with a "smpl ... --no-missing" line, that's why I think it has to do
> with the missings.
However, what actually happened was that gretl kept the original n=50
sample as established by the initial OLS: it's not fooled so easily. I
tried setting INCOME to NA also, for good measure, but it's the same
story: gretl sticks with the sample defined by the original model.
So what you're describing must be something more subtle.
More subtle indeed; my script has been restructured (for other reasons)
and when I now comment out the "smpl ... --no-missing" line I don't get
errors anymore. So can't construct the example anymore, sorry.
What I seem to remember is that first gretl kicked out some variables
that were all zeros (those were some constructed interaction terms),
then estimated the "full" model, then started the "omit --auto"
removing something like 21 variables, and then the "Datenfehler" (= data
When I manually estimated the semi-restricted model without the 21
removed variables so far, I think the estimation sample was longer (at
the current edge of the time series) than that of the full model.
But I realize that this information is not really symptoms that would
help to pinpoint the error.