Interpretation of coefficients in SUR

[Martin B]

Hi all!

I want to do the following:

I have 3 time series of dependent variables (e1,e2,e3). I have 2 time series of explanatory variables (f1,f2). I have 6 additional time series of explanatory variables (m1,m2,m3 and r1, r2, r3). Hereby, r1 and m1 correspond to e1 and so on.

I want to estimate the following regression system:

e1 = c(1) + c(2)*f1 + c(3)*f2 + c(4)*m1 + c(5)*r1
e1 = c(6) + c(7)*f1 + c(8)*f2 + c(4)*m2 + c(5)*r2
e1 = c(9) + c(10)*f1 + c(11)*f2 + c(4)*m3 + c(5)*r3

The variables e1, e2 and e3 have common coefficients in the m and r variables and specific coefficients in the intercepts and the f1 and f2 variable.

I want to estimate this as a system of seemingly unrelated regressions with feasible generalized least squares. In addition, I want to apply the Newey-West estimator of the covariance matrix to correct for autocorrelations.

This is how I proceed: I create a systems object from object/new object/system. I type in the specification written above. Then I hit the button Estimate. Then I choose Seemingly Unrelated Regression.

Now here is my question: why can't I then choose the option generalized least squares? Why can't I choose to correct with the Newey-West method?

I thank you very much for your valuable time. An answer would be most helpful!

Kind regards

Martin

[EViews Gareth]

Now here is my question: why can't I then choose the option generalized least squares?

Not quite sure what you mean by this, aren't GLS and ML estimation the same for SUR, or am I misunderstanding what you want?

Why can't I choose to correct with the Newey-West method?

Simple answer is that we haven't implemented it. EViews 7 will have this feature.

[Martin B]

Thank you very much for your answer!

So when applying SUR there is no need to set for GLS. Am I right?

How can I then correct for autocorrelation, if Newey-West is not implemented yet?

Thank you very much!

Kind regards

Martin

[EViews Gareth]

Thank you very much for your answer!

So when applying SUR there is no need to set for GLS. Am I right?

right.

How can I then correct for autocorrelation, if Newey-West is not implemented yet?

Unfortunately, you can't...

[Martin B]

Ok, let me quickly review the system I posted:

e1 = c(1) + c(2)*f1 + c(3)*f2 + c(4)*m1 + c(5)*r1
e2 = c(6) + c(7)*f1 + c(8)*f2 + c(4)*m2 + c(5)*r2
e3 = c(9) + c(10)*f1 + c(11)*f2 + c(4)*m3 + c(5)*r3

Is there another option in Eviews to allow for correlation among the error terms of the three equations and to correct for autocorrelation in the error terms of a given equation (e.g. in the first equation)?

I have relatively strong autocorrelation. So this is kind of important...

Many thanks

Kind regards

M

[startz]

Ok, let me quickly review the system I posted:

e1 = c(1) + c(2)*f1 + c(3)*f2 + c(4)*m1 + c(5)*r1
e2 = c(6) + c(7)*f1 + c(8)*f2 + c(4)*m2 + c(5)*r2
e3 = c(9) + c(10)*f1 + c(11)*f2 + c(4)*m3 + c(5)*r3

Is there another option in Eviews to allow for correlation among the error terms of the three equations and to correct for autocorrelation in the error terms of a given equation (e.g. in the first equation)?

I have relatively strong autocorrelation. So this is kind of important...

Many thanks

Kind regards

M

Just write out the serial correlation by hand. For example, the first equation could be written

Code: Select all

e1 = c(1)*(1-c(20)) + c(2)*(f1-c(20)*f1(-1)) + c(3)*(f2-c(20)*f2(-1)) + c(4)*(m1-c(20)*m1(-1)) + c(5)*(r1-c(20)*r1(-1)) + c(20)*e1(-1)

where c(20) is the serial correlation coefficient.

[Martin B]

I thank you very, very much!

Your advice was most helpful. Now I have the Durbin-Watsons that I want;-)

Kind regards

M

[remuct]

Hi,
I would like to add to this subject on the issue of the number of equations and of variables per equation supported by SUR. I have a large model of 33 equations containing each 9 variables. I need to estimate this sytem by accounting for the contemporaneous correlation of residuals but also residual autocorrelation (because all of my residuals are autocorrelated). If I express all my equations by writing out the serial correlation, my SUR system will look like:
yi = a0i (1-bi) + bi yi(-1) + a1i (x1 - bi x1(-1)) + ... + a9i (x9 - bi x9(-1)), i=1,...,33
with 33x11=363 coefficients to be estimated.

With this equation specification Eviews returns me non-singular matrix message when the system exceeds 8 equations. Is there a size limit for system models, if not do you have any idea why I can't estimate the whole model?

Kind regards

[remuct]

Concerning the message, I meant "near-singuler matrix", sorry for the confusion.