7:45 a.m.
New subject: About Pooled Mean Group estimation
On Tue, 12 Jan 2021, Sven Schreiber wrote:
I'm also noticing, however, that the parameters of the
mean-group
estimator {0.89, -0.46, -1.8} lead us to the (presumably) correct PMG
result again. Maybe we should routinely also try MG as an
initialization. Jack?
Excellent idea. We could offer several initialisation options. Another one
I briefly experimented with is adding lagged diiferences of X to the FE
model we use for initialisation (but didn't help very much).
On Wed, 13 Jan 2021, Reynaldo Senra wrote:
I would like to add that this PMG package seems to put in advantage
Gretl compared to Eviews and Stata. Eviews
doesn't allow for MG and Hausman test for the comparison between PMG and MG. Stata
does, but when I used 2 lags
with the xtpmg package, all the outputs yielded totally irrational coefficients.
As the outputs of Eviews and Gretl were identical in most cases, I combined them. I may
be wrong, but the only
limitation of PMG Gretl package is the output limited to report the long run and loading
coefficients. I think
it would be great to include the short run coefficients in the output but of course, I
know time is limited and
the developers do a great contributions with this and other packages.
Thanks, this is nice to know. As for the short-run coefficients, I don't
quite understand what Eviews reports in this case, since to my
understanding with the PMG estimator the short-run coefficients are
unit-varying, aren't they?
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
11:18 a.m.
New subject: About Pooled Mean Group estimation
Am 13.01.2021 um 08:45 schrieb Riccardo (Jack) Lucchetti:
On Tue, 12 Jan 2021, Sven Schreiber wrote:
On Wed, 13 Jan 2021, Reynaldo Senra wrote:
> I would like to add that this PMG package seems to put in advantage
> Gretl compared to Eviews and Stata. Eviews
> doesn't allow for MG and Hausman test for the comparison between PMG
> and MG. Stata does, but when I used 2 lags
> with the xtpmg package, all the outputs yielded totally irrational
> coefficients.
Aha? I find that surprising. (With Stata, I mean.)
> long run and loading coefficients. I think
> it would be great to include the short run coefficients in the output
> but of course, I know time is limited and
> the developers do a great contributions with this and other packages.
Thanks, this is nice to know. As for the short-run coefficients, I don't
quite understand what Eviews reports in this case, since to my
understanding with the PMG estimator the short-run coefficients are
unit-varying, aren't they?
Eviews reports the MG-style estimates of the short-run coeffs, so simple
cross-unit averages.
If I remember correctly the way our code is currently set up, the
short-run coefficients are not directly estimated, the terms are
partialled out. So it's not directly available. Note that it also
wouldn't make much sense to try to calculate common dynamic multipliers,
precisely because the short-run coefficients are not common but
heterogeneous.
If you want to retrieve the full dynamic specification for a certain
unit i, but under the PMG restriction, you can:
1. Estimate the long-run coefficients via PMG.
2. Grab the unit-specific loading in the bundle member b.phis[i] and the
longrun coefficients b.lrcoeff.
3. Restrict the sample to unit i and then calculate the new LHS variable as:
diff(y) - b.phis[i] * (y(-1) - lincomb(X, b.lrcoeff))
4. Regress the LHS from step 3 on the differenced terms (plus exogenous)
to get the short-run coefficients for unit i.
I guess we could think about adding this algorithm to the package.
cheers
sven
1439
days inactive
1439
days old
2 comments
3 participants
participants (3)
-
Reynaldo Senra -
Riccardo (Jack) Lucchetti -
Sven Schreiber