Re: [Gretl-users] Information criteria in VARs
On Sat, 23 Jul 2005, Allin Cottrell wrote:
On Sun, 24 Jul 2005, jack wrote:
> Sorry Allin, I must admit I wasn't clear at all. If we really want to go
> with the others, the following patch should be applied to CVS:
>
> int k = var->ncoeff;
>
> var->ll = -(g * n / 2.0) * (LN_2_PI + 1) - (n / 2.0) * var->ldet;
> - var->AIC = (-2.0 / g) * var->ll + 2.0 * k;
> - var->BIC = (-2.0 / g) * var->ll + k * log(n);
> + var->AIC = (-2.0 * var->ll + 2.0 * k * g) / n;
> + var->BIC = (-2.0 * var->ll + log(n) * k * g) / n;
I'm willing to do that, but it seems a bit odd to divide the whole
thing by the number of observations. Is there an obvious rationale?
Not really. Information criteria are ordinal, so multiplication by a
constant is immaterial. As I said in an earlier post, the only thing we
gain is replicating what other packages do. Not a big deal, if you ask me,
but it doesn't hurt either.
I'd like to know the opinion of other list members, but at the end of the
day whatever you (Allin) decide is fine by me.
cheers,
Riccardo `Jack' Lucchetti
Dipartimento di Economia
Università di Ancona
jack(a)dea.unian.it
http://www.econ.unian.it/lucchetti