4:50 a.m.
Dataset:
https://raw.githubusercontent.com/srk7774/data/refs/heads/master/table10....
Hi,
Is there anyway we can get newey-west standard errors in gretl? HAC robust errors are
differing slightly with textbook result.
Text book result:
Dependent Variable: LCONSUMPTION Method: Least Squares
Sample: 1947-2000 Included observations: 54
Newey-West HAC Standard Errors & Covariance (lag truncation = 3)
Variable Coefficient Std. Error t-Statistic Prob.
CONSTAN -0.467714 0.043937 -10.64516 0.0000
LINCOME 0.804871 0.017117 47.02132 0.0000
LWEALTH 0.201272 0.015447 13.02988 0.0000
INTEREST -0.002689 0.000880 -3.056306 0.0036
Statistic Value Statistic Value
R-squared 0.999560 Mean dependent var. 7.826093
Adjusted R-squared 0.999533 S.D. dependent var. 0.552368
S.E. of regression 0.011934 F-statistic 37832.71
Sum squared resid. 0.007121 Prob. (F-statistic) 0.000000
Durbin-Watson stat. 1.289237
Gretl result:
Model 10: OLS, using observations 1947-2000 (T = 54)
Dependent variable: l_C_expenditure
HAC standard errors, bandwidth 2, Bartlett kernel
coefficient std. error z p-value
-------------------------------------------------------------
const −0.467711 0.0458627 −10.20 2.02e-024 ***
Interest −0.00268906 0.000879808 −3.056 0.0022 ***
l_Yd_income 0.804873 0.0171749 46.86 0.0000 ***
l_W_wealth 0.201270 0.0160382 12.55 4.00e-036 ***
Mean dependent var 7.826093 S.D. dependent var 0.552368
Sum squared resid 0.007121 S.E. of regression 0.011934
R-squared 0.999560 Adjusted R-squared 0.999533
F(3, 50) 23576.95 P-value(F) 9.64e-79
Log-likelihood 164.5880 Akaike criterion −321.1760
Schwarz criterion −313.2201 Hannan-Quinn −318.1077
rho 0.324678 Durbin-Watson 1.289219
Log-likelihood for C_expenditure = −258.021
11:10 a.m.
On 15/03/2026 05:50, Ramki S wrote:
Dataset:
https://raw.githubusercontent.com/srk7774/data/refs/heads/master/table10....
Hi,
Is there anyway we can get newey-west standard errors in gretl? HAC robust errors are
differing slightly with textbook result.
Hi, thanks for testing.
First of all, it must be said that, when it comes to robust estimation,
there are tons of slightly different ways to do things and a single
different detail can affect the results visibly (especially in a small
sample like the one you're considering). Take a look at Chapter 22 of
the gretl User's Guide. Judging by the formatting of the output, it
looks as if Gujarati and Porter are using Eviews; if that was true, I
don't know what the precise formulae Eviews used, so... However, they
seem to use a lag truncation parameter equal to 3, rather than the
automatic value used by gretl, which is 2 in this case. You may want to
use the "set hac_lag" command. But even in that case, result don't match
exactly, although they qualitatively do.
That said, I believe the results reported in the book are a little
weird: the OLS coefficients reported on page 451 (the ones you posted)
are different from the ones reported on page 336 (example 10.2), albeit
only at the sixth decimal digit. That makes me thing that Gujarati and
Porter may have used for the two tables two slightly different versions
of the same dataset (decimal digits, perhaps?), and the table on page
451 may be impossible to replicate with publicly available data.
-------------------------------------------------------
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
-------------------------------------------------------
11:21 a.m.
Am 15.03.2026 um 12:10 schrieb Riccardo (Jack) Lucchetti:
On 15/03/2026 05:50, Ramki S wrote:
> Dataset:
>
https://raw.githubusercontent.com/srk7774/data/refs/heads/master/table10....
>
>
>
> Hi,
>
> Is there anyway we can get newey-west standard errors in gretl? HAC
> robust errors are differing slightly with textbook result.
>
Hi, thanks for testing.
First of all, it must be said that, when it comes to robust
estimation, there are tons of slightly different ways to do things and
a single different detail can affect the results visibly (especially
in a small sample like the one you're considering). Take a look at
Chapter 22 of the gretl User's Guide. Judging by the formatting of the
output, it looks as if Gujarati and Porter are using Eviews; if that
was true, I don't know what the precise formulae Eviews used, so...
However, they seem to use a lag truncation parameter equal to 3,
rather than the automatic value used by gretl, which is 2 in this
case. You may want to use the "set hac_lag" command. But even in that
case, result don't match exactly, although they qualitatively do.
Just to add to that: there are several other settings in gretl that
affect those calculations, see the help for the "set" command -
hac_kernel, hac_prewhiten, hac_missvals. For example, it's not obvious
from the Eviews output which kernel was used there. For standard usage
those settings can be left at their defaults, but if you want to do
comparisons and replications, then it may be necessary to adjust them.
cheers
sven
11:33 a.m.
Thanks a lot. I know that the values are pretty close. But for cross-verification, I
posted.
12:13 p.m.
Hi Ramki,
The textbook standard errors and the ones reported by gretl are both
Newey-West, they just differ in this way: the textbook values have a
degrees-of-freedom adjustment applied while gretl uses the maximum
likelihood estimator of the covariance matrix. You can replicate the
textbook values thus:
ols l_C_expenditure const l_Yd_income l_W_wealth Interest --robust
matrix V = $vcv * $T/($T - $ncoeff)
se = sqrt(diag(V))
print se
Which version is "right"? Well, the rationale for the Newey-West
variance estimator is asymptotic, and opinions vary on whether df
adjustment should be made in such cases.
Allin Cottrell
On Sun, Mar 15, 2026 at 12:50 AM Ramki S <ramakrishnasalagrama(a)gmail.com> wrote:
>
> Dataset:
>
https://raw.githubusercontent.com/srk7774/data/refs/heads/master/table10....
>
>
> Hi,
>
> Is there anyway we can get newey-west standard errors in gretl? HAC robust errors are
differing slightly with textbook result.
>
> Text book result:
> Dependent Variable: LCONSUMPTION Method: Least Squares
> Sample: 1947-2000 Included observations: 54
> Newey-West HAC Standard Errors & Covariance (lag truncation = 3)
>
> Variable Coefficient Std. Error t-Statistic Prob.
> CONSTAN -0.467714 0.043937 -10.64516 0.0000
> LINCOME 0.804871 0.017117 47.02132 0.0000
> LWEALTH 0.201272 0.015447 13.02988 0.0000
> INTEREST -0.002689 0.000880 -3.056306 0.0036
>
> Statistic Value Statistic Value
> R-squared 0.999560 Mean dependent var. 7.826093
> Adjusted R-squared 0.999533 S.D. dependent var. 0.552368
> S.E. of regression 0.011934 F-statistic 37832.71
> Sum squared resid. 0.007121 Prob. (F-statistic) 0.000000
> Durbin-Watson stat. 1.289237
>
> Gretl result:
> Model 10: OLS, using observations 1947-2000 (T = 54)
> Dependent variable: l_C_expenditure
> HAC standard errors, bandwidth 2, Bartlett kernel
>
> coefficient std. error z p-value
> -------------------------------------------------------------
> const −0.467711 0.0458627 −10.20 2.02e-024 ***
> Interest −0.00268906 0.000879808 −3.056 0.0022 ***
> l_Yd_income 0.804873 0.0171749 46.86 0.0000 ***
> l_W_wealth 0.201270 0.0160382 12.55 4.00e-036 ***
>
> Mean dependent var 7.826093 S.D. dependent var 0.552368
> Sum squared resid 0.007121 S.E. of regression 0.011934
> R-squared 0.999560 Adjusted R-squared 0.999533
> F(3, 50) 23576.95 P-value(F) 9.64e-79
> Log-likelihood 164.5880 Akaike criterion −321.1760
> Schwarz criterion −313.2201 Hannan-Quinn −318.1077
> rho 0.324678 Durbin-Watson 1.289219
>
> Log-likelihood for C_expenditure = −258.021
> _______________________________________________
> 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/
6
days inactive
6
days old
4 comments
4 participants
participants (4)
-
Cottrell, Allin -
Ramki S -
Riccardo (Jack) Lucchetti -
Sven Schreiber