Seemingly Unrelated Regression and multivariate regression

[Ghazal Khan]

Hello. I am a beginner with stata, using it for the first time and doing a model for my dissertation where i have five bond indexes as my dependent variables (so 5 different y variables) and about 15 dependent or predictor variables. this was clearly a multivariate regression so i used the stata in-built 'mvreg' command to do the regression.. On my teacher's suggestion i also used the Seemingly unrelated regression to execute the 5 regressions together, getting almost the same results..
I want to run tests for checking model specification, for MVreg or SUREG, which tests should i be doing? can please someone share the commands for them too because the ones for simple linear regression that i saw for multicollinearity, autocorrelation or normality (such as vif, acf, swilk) do not seem to work for this kind of mvreg/sureg data. i will be grateful for some help.
Many thanks

[Jeff Wooldridge]

Ghazal: mvreg and sureg will give numerically identical estimates when you use the same explanatory variables in each equation. mvreg gives you no choice in the matter. With sureg, you have the option of exclusion restrictions. If your theory has no exclusion restrictions for the explanatory variables then you should just do OLS on each equation -- which is what mvreg does.

Unfortunately, mvreg doesn't allow robust variance matrix estimators -- an oversight that maybe can be fixed in Stata 17. For that matter, sureg should also allow robust options (to allow heteroskedasticity in the variances and covariances, and also neglected serial correlation in time series problems). You'll be better off estimating each equation separately and using, say, the -newey- command to obtain HAC standard errors.

[Ghazal Khan]

Thank you very much for your help Jeff Wooldridge, i will estimate each regression individually then, however, what do u suggest i do if i want to find out if the errors of the 5 individual regressions are correlated? if they are correlated, then that will support the use of SUR as the preferable estimation method, even if the results from OLS and SUR are the same, right? but if they are not, then i will have a good reason to use simple OLS rather than SUR as suggested by my teacher. many thanks

[Jeff Wooldridge]

Ghazal: The point is that there is only one estimator. It doesn't make sense to talk about using SUR when it reduces to OLS on each equation. The only reason to use SUR is to test restrictions across equations. But mvreg does that also, and you should get identical results. Regrettably, Stata does not have a way to make the test robust to general heteroskedasticity in the variance-covariance matrix.