Re: [Gretl-users] ok to edit .po files offline now?
On Sun, May 28, 2006 22:40, Sven Schreiber wrote:
1) Failed to initalize print struct for loop\n
apart from typo in English, what's struct?
Hmm. It's an internal construct. This is a message we issue when, for some
reason, printing things within loop fails. Allin, maybe this needs to go to
stderr or to be edited to be more transparet to the end user, no?
10) PWE
Prais-Winsten estimator
12) Rescaled-range plot for
what's the context here again?
The rescaled range plot is used in long memory models to compute the (R/S)
statistic (see below). R/S is the name of the statistic, we don't need to
translate that.
18) Data set is not recognized as a panel.\n
Please use \"Sample/Set frequency, startobs\".
-- does this really still reflect the current menu structure?
This is probably going to change in a few days.
21) log(R/S)
R? S?
See (12)
22)\n
Hausman test matrix is not positive definite (this result may be treated
as\n
\"fail to reject\" the random effects specification).\n
-- This is different: I perfectly understand it, but I refuse to
translate that ;-) given that I have written an entire (though short)
paper entitled "The Hausman test statistic can be negative even
asymptotically", and it turns out it can only happen when H0 is false.
So we shouldn't advise people to accept H0 in those cases.
Oh, yes, I remember your paper. Nice piece of work. But, as you say in the
paper, the trouble comes when you use two different estimators for the
variance of the two estimators you compare. Under the null, they are
asymptotically equivalent, so a negative HT is indeed only a finite-sample
phenomenon. Under the alternative, one of them is inconsistent, so really
anything can happen.
However, if you use the same estimator for the variance (which is ok under the
null) the problem disappears. This is why, for example, I prefer computing the
HT for OLS vs IV via the "variable addition" approach rather than by going
through the quadratic form. Allin and I talked the thing over to some length,
and eventually decided that in some cases (eg HT for random effects vs fixed
effects in panels) we present both forms in order to make it easier to compare
our results with those from other packages (eg Stata).
Riccardo "Jack" Lucchetti
Dipartimento di Economia
Facoltà di Economia "G. Fuà"
Ancona