Re: [Gretl-users] Variable selection
Am 17.06.2018 um 09:03 schrieb Riccardo (Jack) Lucchetti:
On Sat, 16 Jun 2018, John C Frain wrote:
> Can you give me one good example in which a stepwise regression program
> produces results that are not subject to the objections mentioned by
> Clive. OLS, VARS etc may be abused by some but such routines are the
> basis
> of much good work. Stepwise regression is not.
On a personal level, I agree 100%. As Clive remarked, the statistical
issues implied by automatic model selection are indeed thorny. The
only convincing framework in which you can (to some extent) substitute
human judgement with CPU cycles is BMA, but apart from that I myself
would never base any critical statistical procedure on tools like
stepwise regression.
Let me just remember everybody that this discussion is not only about
the stepwise method. For example, the latest code examples here on the
list were about what you might call best subset, by trying out every
combination (brute force, not elegant).
There are a lot of more recent estimation methods out there that tackle
shrinkage / penalized estimation / model reduction. There are
well-established reasons and a large literature out there justifying
shrinkage methods - where justifying does _not_ mean in all situations,
nor that they cannot be abused of course, and yes proper inference is at
least tricky and quite often unknown.
Having said that, however, I believe that the purpose of the gretl
project is not to tell the world how econometrics should be done.
research context, you may want to replicate other people results, and
do so by using free tools. Gretl must be up to the job, and therefore
I'm all in favour of implementing statistical techniques I would never
trust myself.
Very well put I think. Because if you think it through, otherwise we
couldn't offer OLS because it is biased in many situations.
cheers,
sven