12:17 p.m.
Hi,
Not sure if this is a gretl issue or an econometric one actually.
I'm trying to come up with a script that replicates Stata's "rreg"
command
which is basically a robust estimation for outliers correction. It is based
on an iterative WLS resulting in an optimal weight vector.
Now, for testing purposes, I'm running OLS and WLS models successively. I
note that the WLS' "Statistics based on the original data" SSR and S.E of
regression are slightly different from the respective OLS' output (*in bold
*below). Should these parts of outputs be the same?
If not, hhat is the difference coming from ?
Best,
Artur
? ols crime X
Model 1: OLS, using observations 1-50
Dependent variable: crime
...
Mean dependent var 566.6600 S.D. dependent var 295.8773
*Sum squared resid 2442333 S.E. of regression 227.9573*
...
? wls W crime X
Model 2: WLS, using observations 1-50
Dependent variable: crime
Variable used as weight: W
...
Statistics based on the weighted data:
Sum squared resid 1925024 S.E. of regression 202.3807
...
Statistics based on the original data:
Mean dependent var 566.6600 S.D. dependent var 295.8773
*Sum squared resid 2469623 S.E. of regression 229.2273*
12:39 p.m.
Hi, This is an econometric issue. In the presence of Heteroskedasticity
under OLS , to improve the model in terms of test of significance WLS is
used.
On Mon, 21 Dec 2020 at 5:48 PM, Artur Bala <artur.bala.tn(a)gmail.com> wrote:
Hi,
Not sure if this is a gretl issue or an econometric one actually.
I'm trying to come up with a script that replicates Stata's "rreg"
command
which is basically a robust estimation for outliers correction. It is based
on an iterative WLS resulting in an optimal weight vector.
Now, for testing purposes, I'm running OLS and WLS models successively. I
note that the WLS' "Statistics based on the original data" SSR and S.E of
regression are slightly different from the respective OLS' output (*in
bold *below). Should these parts of outputs be the same?
If not, hhat is the difference coming from ?
Best,
Artur
? ols crime X
Model 1: OLS, using observations 1-50
Dependent variable: crime
...
Mean dependent var 566.6600 S.D. dependent var 295.8773
*Sum squared resid 2442333 S.E. of regression 227.9573*
...
? wls W crime X
Model 2: WLS, using observations 1-50
Dependent variable: crime
Variable used as weight: W
...
Statistics based on the weighted data:
Sum squared resid 1925024 S.E. of regression 202.3807
...
Statistics based on the original data:
Mean dependent var 566.6600 S.D. dependent var 295.8773
*Sum squared resid 2469623 S.E. of regression 229.2273*
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website:
https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
1:23 p.m.
Am 21.12.2020 um 13:17 schrieb Artur Bala:
Hi,
Not sure if this is a gretl issue or an econometric one actually.
I'm trying to come up with a script that replicates Stata's "rreg"
command which is basically a robust estimation for outliers
correction. It is based on an iterative WLS resulting in an optimal
weight vector.
Now, for testing purposes, I'm running OLS and WLS models
successively. I note that the WLS' "Statistics based on the original
data" SSR and S.E of regression are slightly different from the
respective OLS' output (*in bold *below). Should these parts of
outputs be the same?
If not, hhat is the difference coming from ?
Well, it should be the different estimators, no? If I understand the
output correctly, WLS reports what's the outcome if the unweighted data
are used, but the coefficient values from WLS are applied, in order to
have a meaningful comparison with the OLS results. And if there were no
difference, WLS wouldn't have made any difference.
cheers
sven
1:39 p.m.
Le lun. 21 déc. 2020 à 14:23, Sven Schreiber <svetosch(a)gmx.net> a écrit :
Am 21.12.2020 um 13:17 schrieb Artur Bala:
Well, it should be the different estimators, no?
Hhhhhhh!!!! Of course IT IS!!!!! And you're right. This is a new set of
estimators indeed!!!
That's not my day definitely. Sorry !!!! Thank you Sven (and Kuppusamy) for
the reaction!!!
Best,
Artur
(PS. Indeed my mind was stuck to gretl output with --vcv correction which
produces no change on estimators... this is why it sounded strange to me
that SSR may change...but is was WLS instead :( )
1875
days inactive
1875
days old
3 comments
3 participants
participants (3)
-
Artur Bala -
Dhanasekaran Kuppusamy -
Sven Schreiber