9:27 p.m.
Hello,
either this post contains a relatively stupid question or I've stumbled
upon a strange bug in gretl. this is with the released 1.7.4 on windows
(no snapshot). Since you would have asked me for it anyway, I'm
providing a testcase right away.
The attached script contains two gmm blocks, which AFAICS should be
independent and not affecting each other. The funny thing is, if I
comment out the first gmm block, the results of the second gmm
estimation change. As I'm writing this, tired and everything, I feel
sure that somebody is going to point out a programming mistake and make
me feel embarrassed, but well, maybe it's a gretl bug after all.
good night,
sven
9:47 p.m.
On Fri, 30 May 2008, Sven Schreiber wrote:
Hello,
either this post contains a relatively stupid question or I've stumbled
upon a strange bug in gretl. this is with the released 1.7.4 on windows
(no snapshot). Since you would have asked me for it anyway, I'm
providing a testcase right away.
The attached script contains two gmm blocks, which AFAICS should be
independent and not affecting each other. The funny thing is, if I
comment out the first gmm block, the results of the second gmm
estimation change. As I'm writing this, tired and everything, I feel
sure that somebody is going to point out a programming mistake and make
me feel embarrassed, but well, maybe it's a gretl bug after all.
Fortunately, it isn't.
When you do 2 or more steps in GMM, the weights matrix gets updated.
Hence, in your script the second GMM blocks starts with the weights matrix
out of the first one. If you insert the statement
matrix dnew2weights = inv(dnew2mat'dnew2mat)
between the two gmm blocks, you get that the second estimate is precisely
equal to what would have been if you had skipped the first gmm block.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
12:13 p.m.
Riccardo (Jack) Lucchetti schrieb:
Fortunately, it isn't.
When you do 2 or more steps in GMM, the weights matrix gets updated.
Hence, in your script the second GMM blocks starts with the weights
matrix out of the first one. If you insert the statement
matrix dnew2weights = inv(dnew2mat'dnew2mat)
between the two gmm blocks, you get that the second estimate is
precisely equal to what would have been if you had skipped the first gmm
block.
First of all, thanks a lot for the quick explanation!
However, I'm wondering whether the updated weight matrix shouldn't
rather be stored in some internal gretl variable instead of what the
user passed as the *initial* weights. Are there any other cases where a
user-defined variable is implicitly altered by gretl?
IMHO at least it should be stated explicitly in the manual -- but I
confess I haven't checked and I can very well imagine that it is already
mentioned there.
thanks,
sven
1:18 p.m.
On Sat, 31 May 2008, Sven Schreiber wrote:
However, I'm wondering whether the updated weight matrix
shouldn't
rather be stored in some internal gretl variable instead of what the
user passed as the *initial* weights.
If you need something like that, you can do it "by hand":
matrix W0 = (whatever)
matrix W = W0
gmm
...
weights W
...
end gmm --iterate
Are there any other cases where a user-defined variable is implicitly
altered by gretl?
Lots. For example,
scalar a = 0
mle ...
params a
end mle
Same goes for gmm and nls.
IMHO at least it should be stated explicitly in the manual --
but I confess I haven't checked and I can very well imagine that
it is already mentioned there.
Thanks for the idea. I'd already planned to do some work on the manual
this weekend. I'll throw this in.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
3:48 p.m.
Riccardo (Jack) Lucchetti schrieb:
On Sat, 31 May 2008, Sven Schreiber wrote:
> Are there any other cases where a user-defined variable is implicitly
> altered by gretl?
Lots. For example,
scalar a = 0
mle ...
params a
end mle
Same goes for gmm and nls.
well of course, there it's very obviuos -- so I meant _except_ with
'params' :-)
BTW, this got me thinking: in some sense it might be more consistent to
use the accessor $coeff also for gmm, mle, nls, and so forth. But since
I'm not in favor of breaking backwards-compatibility (as you know), I
guess I'm not proposing it seriously.
> IMHO at least it should be stated explicitly in the manual -- but I
> confess I haven't checked and I can very well imagine that it is
> already mentioned there.
Thanks for the idea. I'd already planned to do some work on the manual
this weekend. I'll throw this in.
thanks for the work!
-sven
5802
days inactive
5803
days old
4 comments
2 participants
participants (2)
-
Riccardo (Jack) Lucchetti -
Sven Schreiber