We use SUR (Seemingly unrelated regressions) in our paper to estimate the model specified with three equations.
The reviewers asks for justified the use, and what about when errors are correlated across equations? In which cases SUR is preferred over regular regression?
Thank you for your help!!!!
Hi
I expecte that you want to know When the suprior regression is? The answer is :if the Derbun Watson coefficient is less than R-Square and if the test of homogeniaty and correlated error shows acceptsnce of founding these problem then we can conclud that there is suprior regression.
Hi
I expecte that you want to know When the suprior regression is? The answer is :if the Derbun Watson coefficient is less than R-Square and if the test of homogeniaty and correlated error shows acceptsnce of founding these problem then we can conclud that there is suprior regression.
Palomino -
I am not an expert on SUR, but according to section 15.5, pages 579-580, in Maddala, G.S.(2001), Introduction to Econometrics, 3rd ed., Wiley, Zellner proposed SUR for panel data, where a set of regression equations have 'errors' which are correlated across cross-section, but independent over time. Maddala noted that this would be the case if all equations had common omitted variables.
I'm sure more has been done on this, and according to slide 50, in a set of slides by Alan Wan (see below this paragraph), the SUR method has been extended to cover heteroscedasticity, serial correlation, and varying numbers of explanatory variables across the equations, but notes those are advanced topics. But I'd say those are very important considerations.
(To find A. Wan's slides, search on the following: CHAPTER 4: Forecasting by Regression Prof. Alan Wan).
I think that if you cannot account for heteroscedasticity, etc., then you might possibly do better without using SUR.
Experimentation using test data might be helpful.
An important consideration appears to be: Do you have reason to expect "correlation across cross-section units" as noted on page 579 in Maddala(2001)?
Cheers - Jim
Bot OLS and SUR (you mean FGLS?) are consistent. FGLS is more efficient, with lower variance for the betas, when the error terms are correlated across the equations. You could:
1. Provide the referee with the significance levels for the correlation coefficients of the variance-covariance matrix for the error terms. This should be enough to justify FGLS.
2. Provide the referee with equation-by-equation OLS and show that it does not change your conclusions. If your policy coefficient is insignificant you can argue that OLS is not efficient in your setting.
Best,v
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression