Re: [Gretl-users] additional .pot info
On Mon, May 29, 2006 21:13, Sven Schreiber wrote:
Ok, but I don't find it anywhere. Any hints? (cli only?)
Variable -> Hurst exponent
In fact, I think we should reorganise our (still a bit limited) offering wrt
long-memory models. Currently, you can compute the R/S statistic there, plus
the GPH and LW estimators for d as a by-product of the periodogram. Maybe we
could put everything in one plce.
And finally re the Hausman test and Jack's remarks: First,
I'm amazed
that you knew that paper of mine, and I won't venture any further
speculation about the reasons for that... But to the point, so you agree
that gretl is partially using methods where "anything can happen", but
still advises to accept H0 in those cases? I still think gretl shouldn't
do that.
You're right, IMO. The situation, the way I see it, is: there are two ways to
compute a Hausman test. Under the null, both are ok and any discrepancies
should vanish in large samples. Under the alternative, one is still ok (the
"regression" approach), while the other (the "quadratic form"
approach, from
the definition) may go crazy. The problem is: what do you do when your
"definition" HT falls into the acceptance region? How do you know whether the
null is true, or rather you're seeing a perverse effect of the alternative?
Technically speaking, the "qf" version is not always a consistent test
statistic (nice one-line abstract of your paper, no? :-D).
On these grounds, one should obviously prefer the regression approach. The
trouble is that there's a whole lot of econometric software out there that
goes for the other one, so we should report that as well (for the same reason
everyone reports the Durbin-Watson statistic even if it's nearly useless
nowadays).
But in short, yes, in my opinion we should avoid commenting and leave that to
the user.
Riccardo "Jack" Lucchetti
Dipartimento di Economia
Facoltà di Economia "G. Fuà"
Ancona