On Sat, July 22, 2006 01:26, Allin Cottrell wrote:
On Fri, 21 Jul 2006, Talha Yalta wrote:
>> You're making sense, if you mean to request that
>> transformations of variables could be written into a model
>> specification. However, I would dispute that this is a "GUI"
>> feature: it requires typing the model specification. The
>> gretl GUI model dialog doesn't require (or allow) any typing
>> of the model spec.
> Yes this is not really a GUI feature but it is indirectly
> related with the GUI as well. Currently, you run a regression
> and end up with a lot of newly generated variables, which
> really don't have to be permanently in the variable list.
I agree, this can be an issue. It raises a larger design
Is it, then, worth considering a bifurcation of the model
specification dialog? A "user-friendly" option (which we have
already) and an "expert" version (which sticks a vamped-up
version of the current command-line model specification into a
text-entry dialog box)?
IMO the dialog is ok the way it is now. In the long run, there's no way any
dialog box can match the power and flexibility of the command line. Playing
catch-up would only lead to a monstrously complex dialog. The dialog box
serves its purpose just right: a quick interface for simple tasks. It
shouldn't aim to completeness: simply, big jobs aren't its jobs. Besides, the
"expert" version is there already, in a way: press Alt-X and there you go.
A different question (which makes more sense) is: should we just drop
variables created "on the fly", after estimation is done? This would make a
lot of sense, but then, it becomes very difficult to handle cases where you
need those variables after estimation. Imagine this scenario: you estimate
model 1, which contains a "volatile" variable, say log(x); you save it and the
"volatile" variable is dropped; you estimate other stuff, and/or possibly
re-define x. The you go back to model 1 and ask for a variable deletion test,
which doesn't involve log(x). You can compute the test statistic, but you
can't compute the restricted model, because log(x) is gone forever (ok,
strictly speaking you can compute coefficients and standard errors, but not --
for instance -- the residuals). If you sort this kind of issue out, then I'd
drop volatile variables any time.
Riccardo "Jack" Lucchetti
Dipartimento di Economia
Facoltà di Economia "G. Fuà"