12:51 a.m.
Hi all,
I've been using the modtest --autocorr option to test for autocorrelation in
a VAR model. I set the sample 1985 01 - 2009 04 (100 observations). The
automatic test and the manually specified LM test calculation do not yield
the same result because the former (automatic) uses observations (for the m
lags) outside the sample while the latter (manual) stays within
the specified sample period. I am of the opinion that the latter is more
accurate because the sample is restricted for a priori reasons that would be
invalid in the automatic autocorrelation testing option. Thoughts?
Here's the output from the script:
*AUTOMATIC: *
Breusch-Godfrey test for autocorrelation up to order 4
OLS, using observations 1985:1-2009:4 (T = 100)
Dependent variable: uhat
coefficient std. error t-ratio p-value
-------------------------------------------------------------
const -0.00580334 0.320830 -0.01809 0.9856
time -4.86302e-07 0.000692273 -0.0007025 0.9994
l_USGDP_1 -0.348853 0.657432 -0.5306 0.5970
l_USGDP_2 0.627695 0.609177 1.030 0.3057
l_USGDP_3 -0.123856 0.568652 -0.2178 0.8281
l_USGDP_4 -0.154690 0.333191 -0.4643 0.6436
l_COMP_1 0.00101458 0.0120923 0.08390 0.9333
l_COMP_2 -0.00117663 0.0211498 -0.05563 0.9558
l_COMP_3 -0.0101049 0.0314406 -0.3214 0.7487
l_COMP_4 0.0113482 0.0192114 0.5907 0.5563
uhat_1 0.377927 0.669859 0.5642 0.5741
uhat_2 -0.255529 0.496828 -0.5143 0.6083
uhat_3 -0.127548 0.238803 -0.5341 0.5946
uhat_4 -0.0859906 0.212572 -0.4045 0.6868
Unadjusted R-squared = 0.027148
Test statistic: LMF = 0.599973,
with p-value = P(F(4,86) > 0.599973) = 0.664
*Alternative statistic: TR^2 = 2.714813,*
*with p-value = P(Chi-square(4) > 2.71481) = 0.607*
Ljung-Box Q' = 0.955537,
with p-value = P(Chi-square(4) > 0.955537) = 0.916
*MANUAL*
Model 2: OLS, using observations 1986:1-2009:4 (T = 96)
Dependent variable: uhatUSGDP
coefficient std. error t-ratio p-value
-------------------------------------------------------------
const -0.0211644 0.397158 -0.05329 0.9576
uhatUSGDP_1 0.338611 0.910909 0.3717 0.7111
uhatUSGDP_2 -0.261190 0.584142 -0.4471 0.6560
uhatUSGDP_3 -0.149634 0.245107 -0.6105 0.5432
uhatUSGDP_4 -0.0877872 0.221146 -0.3970 0.6924
time -3.10909e-05 0.000855037 -0.03636 0.9711
l_USGDP_1 -0.309372 0.900280 -0.3436 0.7320
l_USGDP_2 0.580976 0.772587 0.7520 0.4542
l_USGDP_3 -0.107547 0.720912 -0.1492 0.8818
l_USGDP_4 -0.159781 0.485321 -0.3292 0.7428
l_COMP_1 0.00192274 0.0123714 0.1554 0.8769
l_COMP_2 -0.00188331 0.0216852 -0.08685 0.9310
l_COMP_3 -0.0100372 0.0391818 -0.2562 0.7985
l_COMP_4 0.0116237 0.0259833 0.4474 0.6558
Mean dependent var -0.000048 S.D. dependent var 0.004580
Sum squared resid 0.001943 S.E. of regression 0.004868
R-squared 0.024783 Adjusted R-squared -0.129825
F(13, 82) 0.160296 P-value(F) 0.999620
Log-likelihood 382.5531 Akaike criterion -737.1062
Schwarz criterion -701.2053 Hannan-Quinn -722.5945
rho -0.001343 Durbin-Watson 1.998754
Excluding the constant, p-value was highest for variable 52 (time)
Generated scalar T = 96
Generated scalar R = 0.024783
Generated scalar LM = 2.37917
Chi-square(4): area to the right of 2.37917 = 0.666394
(to the left: 0.333606)
Thanks!
Mj
3:04 a.m.
On Wed, 12 Oct 2011, Muheed Jamaldeen wrote:
I've been using the modtest --autocorr option to test for
autocorrelation in a VAR model. I set the sample 1985 01 - 2009 04
(100 observations). The automatic test and the manually specified
LM test calculation do not yield the same result because the
former (automatic) uses observations (for the m lags) outside the
sample while the latter (manual) stays within the specified sample
period. I am of the opinion that the latter is more accurate
because the sample is restricted for a priori reasons that would
be invalid in the automatic autocorrelation testing option.
Thoughts?
I think the basic point here is what one means by "restricting the
sample to 1985:1-2009:4" in this context (i.e. estimating a model
with lags).
This has been discussed on the list before, and opinions may differ.
I have maintained that the most intuitive interpretation is that
_estimation_ should run from 1985:1-2009:4, with prior lags being
accessed if possible. An alternative view is that no data earlier
than 1985:1 should be accessed in any way, in which case estimation
of a 4-lag VAR would have to start in 1986:1.
I guess my view is that so long as it's clear what gretl in fact
does, the user should be able to achieve what he or she wants. That
is, knowing that gretl tries to start estimation at the top of the
sample range if possible, if you want a 4-lag VAR that doesn't
access any data before 1985:1, you'll have to do
smpl 1986:1 2009:4
But as for your point about the autocorrelation test, I don't really
get it. The automatic test uses the same sample range as the VAR,
which seems to me right. Your manual test uses a shorter sample, and
that seems to me less pertinent. (Gretl follows the Breusch-Godfrey
and Kiviet approach in setting pre-sample residuals to zero in the
auxiliary regression.)
Besides, the results are substantively the same eiither way (fail to
reject H0, as one would expect in a reasonably specified VAR).
Here's the output from the script:
*AUTOMATIC: *
Breusch-Godfrey test for autocorrelation up to order 4
OLS, using observations 1985:1-2009:4 (T = 100)
Dependent variable: uhat
coefficient std. error t-ratio p-value
-------------------------------------------------------------
const -0.00580334 0.320830 -0.01809 0.9856
time -4.86302e-07 0.000692273 -0.0007025 0.9994
l_USGDP_1 -0.348853 0.657432 -0.5306 0.5970
l_USGDP_2 0.627695 0.609177 1.030 0.3057
l_USGDP_3 -0.123856 0.568652 -0.2178 0.8281
l_USGDP_4 -0.154690 0.333191 -0.4643 0.6436
l_COMP_1 0.00101458 0.0120923 0.08390 0.9333
l_COMP_2 -0.00117663 0.0211498 -0.05563 0.9558
l_COMP_3 -0.0101049 0.0314406 -0.3214 0.7487
l_COMP_4 0.0113482 0.0192114 0.5907 0.5563
uhat_1 0.377927 0.669859 0.5642 0.5741
uhat_2 -0.255529 0.496828 -0.5143 0.6083
uhat_3 -0.127548 0.238803 -0.5341 0.5946
uhat_4 -0.0859906 0.212572 -0.4045 0.6868
Unadjusted R-squared = 0.027148
Test statistic: LMF = 0.599973,
with p-value = P(F(4,86) > 0.599973) = 0.664
*Alternative statistic: TR^2 = 2.714813,*
*with p-value = P(Chi-square(4) > 2.71481) = 0.607*
Ljung-Box Q' = 0.955537,
with p-value = P(Chi-square(4) > 0.955537) = 0.916
*MANUAL*
Model 2: OLS, using observations 1986:1-2009:4 (T = 96)
Dependent variable: uhatUSGDP
coefficient std. error t-ratio p-value
-------------------------------------------------------------
const -0.0211644 0.397158 -0.05329 0.9576
uhatUSGDP_1 0.338611 0.910909 0.3717 0.7111
uhatUSGDP_2 -0.261190 0.584142 -0.4471 0.6560
uhatUSGDP_3 -0.149634 0.245107 -0.6105 0.5432
uhatUSGDP_4 -0.0877872 0.221146 -0.3970 0.6924
time -3.10909e-05 0.000855037 -0.03636 0.9711
l_USGDP_1 -0.309372 0.900280 -0.3436 0.7320
l_USGDP_2 0.580976 0.772587 0.7520 0.4542
l_USGDP_3 -0.107547 0.720912 -0.1492 0.8818
l_USGDP_4 -0.159781 0.485321 -0.3292 0.7428
l_COMP_1 0.00192274 0.0123714 0.1554 0.8769
l_COMP_2 -0.00188331 0.0216852 -0.08685 0.9310
l_COMP_3 -0.0100372 0.0391818 -0.2562 0.7985
l_COMP_4 0.0116237 0.0259833 0.4474 0.6558
Mean dependent var -0.000048 S.D. dependent var 0.004580
Sum squared resid 0.001943 S.E. of regression 0.004868
R-squared 0.024783 Adjusted R-squared -0.129825
F(13, 82) 0.160296 P-value(F) 0.999620
Log-likelihood 382.5531 Akaike criterion -737.1062
Schwarz criterion -701.2053 Hannan-Quinn -722.5945
rho -0.001343 Durbin-Watson 1.998754
Excluding the constant, p-value was highest for variable 52 (time)
Generated scalar T = 96
Generated scalar R = 0.024783
Generated scalar LM = 2.37917
Chi-square(4): area to the right of 2.37917 = 0.666394
(to the left: 0.333606)
Thanks!
Mj
3:37 a.m.
Thanks for that Allin.
I thought I'd clarify the point regarding the autocorrelation test which was
unclear in my previous email. The sample period 1985:01 to 2009:04 is chosen
for specific reasons relating to stability and floating of the exchange
rate. If the autocorrelation test (automatic) auxiliary equation uses
information prior to 1985:01 then the information in that test is not the
same as that in a VAR (4) estimated from 1986:01. Since Gretl's VAR or OLS
options use 1985:01 as the start there isn't a problem of inconsistency.
It's just that I noticed the difference when I used J-MULTI to estimate the
VAR and it uses 1986:01 as the estimation period even though the sample is
specified 1985:01 to 2009:04. For my purposes of doing the autocorrelation
tests on Gretl, I'll change the sample period to 1986:01 and that should
resolve it.
While we're on the subject, is there a script for a portmanteau test for
auto correlation?
And using the var option in Gretl what's the best way to specify
heterogeneous (in variables) equation structure? (e.g. to impose block
exogeniety for a small open economy VAR)
eg.
Y = Y-1 + X-1 + Z-1
X= X-1 + Z-1
Z=X-1 + Z-1
Thanks!
On Wed, Oct 12, 2011 at 2:04 PM, Allin Cottrell <cottrell(a)wfu.edu> wrote:
On Wed, 12 Oct 2011, Muheed Jamaldeen wrote:
> I've been using the modtest --autocorr option to test for
> autocorrelation in a VAR model. I set the sample 1985 01 - 2009 04
> (100 observations). The automatic test and the manually specified
> LM test calculation do not yield the same result because the
> former (automatic) uses observations (for the m lags) outside the
> sample while the latter (manual) stays within the specified sample
> period. I am of the opinion that the latter is more accurate
> because the sample is restricted for a priori reasons that would
> be invalid in the automatic autocorrelation testing option.
> Thoughts?
I think the basic point here is what one means by "restricting the
sample to 1985:1-2009:4" in this context (i.e. estimating a model
with lags).
This has been discussed on the list before, and opinions may differ.
I have maintained that the most intuitive interpretation is that
_estimation_ should run from 1985:1-2009:4, with prior lags being
accessed if possible. An alternative view is that no data earlier
than 1985:1 should be accessed in any way, in which case estimation
of a 4-lag VAR would have to start in 1986:1.
I guess my view is that so long as it's clear what gretl in fact
does, the user should be able to achieve what he or she wants. That
is, knowing that gretl tries to start estimation at the top of the
sample range if possible, if you want a 4-lag VAR that doesn't
access any data before 1985:1, you'll have to do
smpl 1986:1 2009:4
But as for your point about the autocorrelation test, I don't really
get it. The automatic test uses the same sample range as the VAR,
which seems to me right. Your manual test uses a shorter sample, and
that seems to me less pertinent. (Gretl follows the Breusch-Godfrey
and Kiviet approach in setting pre-sample residuals to zero in the
auxiliary regression.)
Besides, the results are substantively the same eiither way (fail to
reject H0, as one would expect in a reasonably specified VAR).
> Here's the output from the script:
> *AUTOMATIC: *
>
> Breusch-Godfrey test for autocorrelation up to order 4
> OLS, using observations 1985:1-2009:4 (T = 100)
> Dependent variable: uhat
>
> coefficient std. error t-ratio p-value
> -------------------------------------------------------------
> const -0.00580334 0.320830 -0.01809 0.9856
> time -4.86302e-07 0.000692273 -0.0007025 0.9994
> l_USGDP_1 -0.348853 0.657432 -0.5306 0.5970
> l_USGDP_2 0.627695 0.609177 1.030 0.3057
> l_USGDP_3 -0.123856 0.568652 -0.2178 0.8281
> l_USGDP_4 -0.154690 0.333191 -0.4643 0.6436
> l_COMP_1 0.00101458 0.0120923 0.08390 0.9333
> l_COMP_2 -0.00117663 0.0211498 -0.05563 0.9558
> l_COMP_3 -0.0101049 0.0314406 -0.3214 0.7487
> l_COMP_4 0.0113482 0.0192114 0.5907 0.5563
> uhat_1 0.377927 0.669859 0.5642 0.5741
> uhat_2 -0.255529 0.496828 -0.5143 0.6083
> uhat_3 -0.127548 0.238803 -0.5341 0.5946
> uhat_4 -0.0859906 0.212572 -0.4045 0.6868
>
> Unadjusted R-squared = 0.027148
>
> Test statistic: LMF = 0.599973,
> with p-value = P(F(4,86) > 0.599973) = 0.664
>
> *Alternative statistic: TR^2 = 2.714813,*
> *with p-value = P(Chi-square(4) > 2.71481) = 0.607*
>
> Ljung-Box Q' = 0.955537,
> with p-value = P(Chi-square(4) > 0.955537) = 0.916
>
> *MANUAL*
>
> Model 2: OLS, using observations 1986:1-2009:4 (T = 96)
> Dependent variable: uhatUSGDP
>
> coefficient std. error t-ratio p-value
> -------------------------------------------------------------
> const -0.0211644 0.397158 -0.05329 0.9576
> uhatUSGDP_1 0.338611 0.910909 0.3717 0.7111
> uhatUSGDP_2 -0.261190 0.584142 -0.4471 0.6560
> uhatUSGDP_3 -0.149634 0.245107 -0.6105 0.5432
> uhatUSGDP_4 -0.0877872 0.221146 -0.3970 0.6924
> time -3.10909e-05 0.000855037 -0.03636 0.9711
> l_USGDP_1 -0.309372 0.900280 -0.3436 0.7320
> l_USGDP_2 0.580976 0.772587 0.7520 0.4542
> l_USGDP_3 -0.107547 0.720912 -0.1492 0.8818
> l_USGDP_4 -0.159781 0.485321 -0.3292 0.7428
> l_COMP_1 0.00192274 0.0123714 0.1554 0.8769
> l_COMP_2 -0.00188331 0.0216852 -0.08685 0.9310
> l_COMP_3 -0.0100372 0.0391818 -0.2562 0.7985
> l_COMP_4 0.0116237 0.0259833 0.4474 0.6558
>
> Mean dependent var -0.000048 S.D. dependent var 0.004580
> Sum squared resid 0.001943 S.E. of regression 0.004868
> R-squared 0.024783 Adjusted R-squared -0.129825
> F(13, 82) 0.160296 P-value(F) 0.999620
> Log-likelihood 382.5531 Akaike criterion -737.1062
> Schwarz criterion -701.2053 Hannan-Quinn -722.5945
> rho -0.001343 Durbin-Watson 1.998754
>
> Excluding the constant, p-value was highest for variable 52 (time)
>
> Generated scalar T = 96
> Generated scalar R = 0.024783
> Generated scalar LM = 2.37917
> Chi-square(4): area to the right of 2.37917 = 0.666394
> (to the left: 0.333606)
>
> Thanks!
>
> Mj
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
8:23 a.m.
Am 12.10.2011 05:37, schrieb Muheed Jamaldeen:
And using the var option in Gretl what's the best way to specify
heterogeneous (in variables) equation structure? (e.g. to impose block
exogeniety for a small open economy VAR)
eg.
Y = Y-1 + X-1 + Z-1
X= X-1 + Z-1
Z=X-1 + Z-1
Thanks!
You would need to specify this equation system "manually" and estimate
it via SUR.
-sven
9:23 a.m.
El 12/10/11 05:37, Muheed Jamaldeen escribió:
Thanks for that Allin.
[...]
While we're on the subject, is there a script for a portmanteau test for
auto correlation?
We have the Ljung-Box Q statistic and p-value in the text window of the
correlogram command 'corrgm'.
--
Ignacio Diaz-Emparanza
DEPARTAMENTO DE ECONOMÍA APLICADA III (ECONOMETRÍA Y ESTADÍSTICA)
UPV/EHU Avda. Lehendakari Aguirre, 83 | 48015 BILBAO
T.: +34 946013732 | F.: +34 946013754
www.ea3.ehu.es
12:51 p.m.
On Thu, 13 Oct 2011, Ignacio Diaz-Emparanza wrote:
El 12/10/11 05:37, Muheed Jamaldeen escribió:
> Thanks for that Allin.
>
>
> [...]
> While we're on the subject, is there a script for a portmanteau test for
> auto correlation?
We have the Ljung-Box Q statistic and p-value in the text window of the
correlogram command 'corrgm'.
There's also the corresponding ljungbox() function.
Allin Cottrell
4824
days inactive
4825
days old
5 comments
4 participants
participants (4)
-
Allin Cottrell -
Ignacio Diaz-Emparanza -
Muheed Jamaldeen -
Sven Schreiber