systemfit - SUR

Arne Henningsen <ahenningsen at email.uni-kiel.de> schrieb am 30.11.2004,
10:20:29:
On Tuesday 30 November 2004 09:34, contact at thomasalmer.com wrote:
Arne Henningsen  schrieb am 29.11.2004,

17:02:12:
On Monday 29 November 2004 16:42, contact at thomasalmer.com wrote:
Hello to everyone,

I have 2 problems and would be very pleased if anyone can help me:

1) When I use the package "systemfit" for SUR regressions, I get two
different variance-covariance matrices when I firstly do the SUR
regression ("The covariance matrix of the residuals used for
estimation") and secondly do the OLS regressions. In the manual for
"systemfit" on page 14 I see however, that the variance-covariance
matrix for SUR is obtained from OLS. How can this be explained?
Hi Thomas,
I get identical residual covariance matrices:

R> library(systemfit)
R> data( kmenta )
R> demand  supply  labels  system  fitols  fitols$rcov
         [,1]     [,2]
[1,] 3.725391 4.136963
[2,] 4.136963 5.784441
R> fitsur  fitsur$rcovest
         [,1]     [,2]
[1,] 3.725391 4.136963
[2,] 4.136963 5.784441
It is a pity, but my matrices are not as nice :-(
Please show how you obtained these results. 
I did not provide the steps, because the data is not public. But that is
what I did after defining the system and labels (like you & the
documentation):
fitols<-systemfit("OLS", system, labels)
fitsur<-systemfit("SUR", system, labels, maxit=100)

But after all I think your answers are sufficient. Thanks once again for
your support. R and the community are really excellent and a big thread
to competitors!
This is what I did:
R> data( kmenta )
R> demand  supply  labels  system 
R> # OLS estimation:
R> fitols  # (non-iterated) SUR estimation
R> fitsur  iterated SUR estimation
R> fitsurit 
R> fitols$rcov
         [,1]     [,2]
[1,] 3.725391 4.136963
[2,] 4.136963 5.784441
R> fitsur$rcovest
         [,1]     [,2]
[1,] 3.725391 4.136963
[2,] 4.136963 5.784441
R> fitsurit$rcovest
         [,1]     [,2]
[1,] 6.199071 7.493383
[2,] 7.493383 9.128547

An excerpt:
fitsur$rcovest
            [,1]         [,2]         [,3]       ...
 [1,] 0.015097517  0.018005050
 [2,] 0.018005050  0.276259834
 ...

fitols$rcov
               [,1]          [,2] 	[,3]	 ...
 [1,]  1.010326e-02  0.0096103837
 [2,]  9.610384e-03  0.2329884378
 ...

fitsur "The covariance matrix of the residuals used for estimation":
    eq1         eq2        eq3         ...
eq1 0.01317429  0.01504719 0.007981307
eq2 0.01504719  0.25233860
...

fitols "The covariance matrix of the residuals":
    eq1          eq2          eq3      ...
eq1 9.51154e-03  0.009137884  0.002648577
eq2 9.13788e-03  0.220435063
...

By the way: Why are the figures larger for SUR?
OLS minimizes the residuals and, thus, also the variance of the residuals 
(=diagonal of the residual covariance matrix).
Iterated SUR is equivalent to a maximum likelihood estimation. Maximizing the 
likelihood value is equivalent to minimizing the determinant of the residual 
covariance matrix. Thus, the determinant of the residual covariance matrix 
and not the residuals itself are minimized:

R> det(fitols$rcov)
[1] 4.434845
R> det(fitsurit$rcov)
[1] 0.4376941
R> det(fitsurit$rcovest)
[1] 0.4377184

Did you do _iterated_ SUR?
Yes:
"systemfit results
method: iterated SUR
convergence achieved after 30 iterations"
If you use iterated SUR, the SUR estimations are iterated. In the first SUR 
estimation the residual covariance matrix of the OLS estimation is used. In 
all following iterations the residual covariance matrix of the previous step 
SUR estimation is used. 

I do not know how to change that.
Please read the documentation. It says to set argument "maxit" to 1 - or do 
not provide this argument, since 1 is the default.

Best wishes,
Arne
THANKS A LOT FOR YOUR IMMEDIATE HELP!!!

2) Is there an easy possibility to test a) the OLS equations, and b)
the SUR system for SUR structures? In other words: Is the LM-Test from
Breusch and Pagan available in R?
I don't understand what you want to test. Does the hausman test what you are 
looking for (see ?hausman.systemfit). If you have questions regarding this 
test, you might ask my co-author of systemfit, Jeff Hamann.
Primarily I want to test if the variance covariance matrix of the OLS
residuals is diagonal. But this can be done manually, of course.
Best wishes,
Arne

Thanks for the attention!

Best Regards,
Thomas Almer

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
--
Arne Henningsen
Department of Agricultural Economics
University of Kiel
Olshausenstr. 40
D-24098 Kiel (Germany)
Tel: +49-431-880 4445
Fax: +49-431-880 1397
ahenningsen at agric-econ.uni-kiel.de
http://www.uni-kiel.de/agrarpol/ahenningsen/

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
-- 
Arne Henningsen
Department of Agricultural Economics
University of Kiel
Olshausenstr. 40
D-24098 Kiel (Germany)
Tel: +49-431-880 4445
Fax: +49-431-880 1397
ahenningsen at agric-econ.uni-kiel.de
http://www.uni-kiel.de/agrarpol/ahenningsen/
[ . . .]
Please show how you obtained these results.
I did not provide the steps, because the data is not public. 
A general remark:
In most cases a questioner can use the example data provided with a package or 
construct simple data, e.g. with matrix(1:9,3). See the posting guide, 
section "Examples". (If the surprising behavior does not appear with the 
sample data, the questioner has to take a deeper look in _his_ data.)
But that is 
what I did after defining the system and labels (like you & the
documentation):
fitols<-systemfit("OLS", system, labels)
fitsur<-systemfit("SUR", system, labels, maxit=100)

But after all I think your answers are sufficient. Thanks once again for
your support. R and the community are really excellent and a big thread
to competitors!

[ . . . ]
2) Is there an easy possibility to test a) the OLS equations, and
b) the SUR system for SUR structures? In other words: Is the
LM-Test from Breusch and Pagan available in R?
I don't understand what you want to test. Does the hausman test what you
are looking for (see ?hausman.systemfit). If you have questions regarding
this test, you might ask my co-author of systemfit, Jeff Hamann.
Primarily I want to test if the variance covariance matrix of the OLS
residuals is diagonal. But this can be done manually, of course.
Please provide the code and a reference to the test. Then, I will add this 
test to systemfit. This is how R works: useRs become developerR :-)

Best wishes,
Arne
Arne Henningsen
Department of Agricultural Economics
University of Kiel
Olshausenstr. 40
D-24098 Kiel (Germany)
Tel: +49-431-880 4445
Fax: +49-431-880 1397
ahenningsen at agric-econ.uni-kiel.de
http://www.uni-kiel.de/agrarpol/ahenningsen/