GLS estimation

[seg_b]

Hi !

I'm trying to estimate an equation of world wheat consumption. I did an LS regression but the Durbin-Watson statistic is not satisfying. So I want to do a GLS regression to correct this autocorrelation problem.
But there isn't any GLS option on Eviews 7. Is that because LS already means GLS?
I searched in eviews help, but I found no command or options...

Thanks for answering

[startz]

Look up AR(1) in the help system.

[seg_b]

Thanks a lot! I did.
But I don't understand :

my first equation is :
conso = argentina + canada + corn + ratio + wheat + pop

There is correlation in my residuals. I tested this serie and it follows a AR(1).

So, with AR(1) in the help system, my new equation is :
conso = argentina + canada + corn + ratio + wheat + pop + [ar(1)=c(1)]
right?

But the result is just concerning c(1).
I tested the stationnarity of the residuals, and the serie is!
I don't understand what is the point with this method.
How can I get the new coefficients of this new regression ?
And this specification : AR(1) = c(1) is really about the residuals ?

thanks for answering

[startz]

Why don't you show us exactly what command you used so we can see where the coefficients are.

[seg_b]

I just did :
Quick/estimate equation/
I wrote : CONSO = ARGENTINA_ENDING_STOCK + CANADA_ENDING_STOCK +
CORN___T + RATIO_EXPORT_WORLD + WHEAT___T +
WORLD_POP + [AR(1)=C(1)]

And this is the result I get.
I try to estimate this equation with GLS method and I don't understand what c(1) means and where are the new coefficient I should get with this new estimation.

[startz]

You probably want coefficients on all those other variables....

[EViews Glenn]

Easier to specify by list

Code: Select all

CONSO C ARGENTINA_ENDING_STOCK CANADA_ENDING_STOCK CORN___T RATIO_EXPORT_WORLD WHEAT___T WORLD_POP AR(1)

Note that I've added a constant

[seg_b]

I tried Cochrane-Orcutt method by this way :

http://www.getproxy.org/view/index.php? ... YzcyMDY%3D

but I don't understand, R² is lower and lower at every new equation and, at the end, just the constant is significant...
It makes me crazy!

I tried to use your method Glenn, but my result is :

http://www.getproxy.org/view/index.php? ... YzcyMDY%3D

So if AR(1) is not significant, my equation is false, isn't it?
And, anyway, Durbin-Watson is too low...

[startz]

Including AR(1) effectively does iterated Cochrane-Orcutt. The odd thing in your regression is that the AR(1) coefficient comes out close to zero while the Durbin-Watson indicates remaining serial correlation. You might try posting your workfile, including the estimated equation.

[seg_b]

Yes, I don't understand why I get those results.
I join my workfile and the equation I use...

thanks for your help :)

[startz]

There are two things going on I believe.
(1) The estimation algorithm is having trouble finding a global maximum. You can see that by providing different starting values for the coefficients. Maybe someone at EViews has a suggestion for getting around this.
(2) Someone has given you "fake" data. The same value is repeated for several months in a row, which is unlikely to have really happened. (It may be that this the data nonetheless represents the best guess as to what happened.)

[seg_b]

For your second hypothesis, the problem is that population and gdp are annual series.
But I try to modelise consumption which is a monthly serie...
I tried with 2 solutions :
1/ repeat the annual value for every month of the year
2/ GDP/12 and POP/12 and I put those values at every month of the year
...

And I'm sure those variables are important to regress consumption.