Seemingly unrelated estimation vs seemingly unrelated regression

[Belinda Foster]

Hello. Stata has two separate commands: sureg for seemingly unrelated regression and suest for seemingly unrelated estimation. I have consulted the manual but it is still unclear to me what is the difference between the two. My understanding is that in the former case two linear regressions are estimated jointly whereas in the latter the results of two difference regressions (of any kind) are simply "combined" together but estimation is separate and thus not as efficient.

Am i correct? If not, can someone please explain the difference between the two? And when is appropriate to use one over the other?

[Carlo Lazzaro]

Belinda:
this quite old Stata thread might be interesting: http://www.stata.com/statalist/archi.../msg01157.html

[Belinda Foster]

Thanks Carlo, i am aware of this post but i was hoping that some other users (or people from Stata - Jeff Pitblado (StataCorp)) would provide some more detail. For example, for two linear regression models should the resulting V matrix be the same from sureg and suest? Does suest also take into account the covariance of the residuals? I am asking because sureg offers cross-equation error correlation.

[Xiaogeng Xu]

Hi,
I saw this thread as I have the same question. Yes, the difference is that sureg considers the correlation of error terms and compute the variance-covariance matrix when joint estimating the models, while suest combines the models and give a robust covariance matrix (which sureg cannot). I think whether to use sureg or suest depends on whether one needs the jointly estimates of the coefficients or not. The demand for the joint estimates comes from the belief that the error terms from the individual models are correlated with each other. This can happen when there are two different dependent variables in a given sample. However, if one only needs to compare the coefficients from the individual models, then suest is a better choice since it doesn't change the estimate of the coefficient but gives a robust s.e. (robust to cluster or heteroskedasticity). Notice that when suest is used for the cross-model testing and the models are based on same/dependent samples, the coefficients can be correlated so one needs to stack the sample and use cluster robust s.e.

[Xiaogeng Xu]

(Tried to delete a duplicate post but failed) Typo: *gives a robust covariance matrix