[Gretl-users] model selection criteria
cottrell at wfu.edu
Tue Sep 30 16:25:42 EDT 2014
On Mon, 29 Sep 2014, Summers, Peter wrote:
> Thanks for the clarification!
> Maybe it's because I'm primarily a time series guy, but my
> impression is that the info criteria are primarily used in time
> series -- before this semester (tomorrow, actually), I'd never
> taught them in a cross-section or panel context. Of course there's
> no reason not to...
> At any rate, I think it'd be better to either a) eliminate that
> message or b) report the changes in all the relevant stats
> (adj-R^2, log-likelihood, & info criteria, assuming this is
> feasible). Option a has the advantage (imho) of forcing the user
> to confront issues of model specification directly, rather than
> relying on his/her software to make the decision. Of course
> there's always c) go back to referring to "information criteria"
For the moment (without prejudice to later amendments) I've gone for
the easy option, (c). In CVS and snapshots, the message refers to
"information criteria" rather than "model selection statistics".
>> On Sep 29, 2014, at 6:49 PM, Allin Cottrell <cottrell at wfu.edu> wrote:
>>> On Mon, 29 Sep 2014, Sven Schreiber wrote:
>>>> Am 29.09.2014 um 16:24 schrieb Summers, Peter:
>>>> What are the 3 “model selection statistics” that Gretl mentions when
>>>> doing tests omitting variables? I’m guessing adj. R-squared,
>>>> log-likelihood & AIC, but what about SC & HQ? Or is she referring to the
>>>> 3 info criteria?
>>> Yes I would think the latter, AIC and SC and HQ. But I'm not 100% "in
>>> the know".
>> Yes, it's the information criteria. But maybe this should be revisited, in terms of what exact statement is offered and/or which criteria are included (or whether the statement should just be dropped).
>> At one time (long ago, and before we calculated the HQC) the Akaike and Schwartz criteria were explicitly labeled as "Model selection criteria". Now, as Peter's question implies, it's not so clear what we're talking about in that message.
More information about the Gretl-users