Hi,

You can easily use the command "reg3" in STATA, which also allows you to compare the 3SLS and 2SLS results. As regards which model is better...it depends. 3SLS models are useful if you expect some sort of cross-correlations in the residuals of the equations, and should result in better efficiency than 2SLS (provided that the model is well specified). In other words, 3SLS is useful if you want to test cross-equation restrictions which would otherwise be impossible to do using 2SLS.

If you plan to run a panel data model, in particular a fixed-effect model, I guess you could do the within-group de-meaning yourself, and then use "reg3" on the de-meaned data.

For more info on this, in particular for STATA,

see http://www.stata.com/statalist/archive/2007-08/msg00483.html

There is also a short comment from Kit Baum, so it's worth reading.

Hi Enrico,

Thnx.  Like we start in STATA by specifying 'xtreg' for panel regression followed by specifying 'fe' / 're' for effects, it seems not straight in case of 3SLS.  Should demeaning  be done for all equations in my case of panel? And if i have a dummy exogenous variable, the common demeaning holds true?

Thnx again.

I think if you go over to the book called Microeconometrics by Cameron and Trivedi, they mention that 3SLS results in more efficiency than 2SLS.  Given that you have 17 states and 10 years, have you thought also about panel time series methods? 

Hi Ali,

Of course, i wish to solve the equations considering panel data setup, through allowing fixed/random effects.  Just i wish to follow 3 SLS procedure, what i believe, is more efficient than the 2SLS procedure.  Unfortunately, STATA has no explicit commands for this.

Dear Sanjay,

you could try including dummy variables for the firm FE, the command reg3 should allow you to do that, e.g.:

xi: reg3 (endog1 endog2 exog1 i.year i.firm) (endog2 endog1 exog2 i.year i.firm)

However, depending on the number of firms, it may take a while for STATA to estimate the model.

Best regards

3 stage least squares are, under general conditions, asymptotically more efficient then 2 stage least squares. By iterating to higher stages until convergence the resulting estimates are identical with maximum likelihood estimates. See Oberhofer and Kmenta in Econometrica in the 1970s.

three stage least squares is simply a combination of sur (seemingly unrelated regression) and 2sls and provides more efficient estimates for linear regression models where the predictor variables are correlated with the error term