3SLS and GMM (error correlation)??

[Thaer Alhalabi]

Dear Statalist users

I am having a problem in my research between 2-step GMM and 3sls. I have a system of 3 equations that are profit, capital and growth. Each of them is a dependent and a regressor when it's not dependent
(Profit=l.Profit+l.captial+l.growth+l.control+eit) (capital=l.captial+l.Profit+l.growth+l.control+eit ) (growth=l.growth+l.captial+l.profit+l.control+eit) .

So what I know is 3SLS assumes that there is a correlation in the error term between the equations. I have done lmcovreg3 and lmhreg3 tests for autocorrelation and heteroscedasticiy and the null hypothesis was rejected. However, when I do the 3SLS regression through cmp command it gives me unrealistic results as one of the dependent has no any significant regressors in its equation and another barely have any significant regressors. I even demeaned (centered) all the variables hoping that might solve the problem but still the results are too bad. I did the three equations using 2-step GMM each equation independently and and gave much better results and the postestimations were great.

So my question is does GMM work between those three equations when the error term between them is correlated as per the lmcovreg3 results?? Plus, do article that use GMM when they have two equations(for example) assume that the error between the equations is not correlated??

Thanks

[Phil Bromiley]

You didn't get a quick answer. You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

As I understand it, GMM is a general approach so it could be applied to simultaneous equation systems but whether the routine you used for GMM did this is not clear. At least in non-GMM estimator, you can do one equation at a time and get consistent results (assuming you have sufficient exogenous variables to instrument the endogenous variables) - the estimates are just less efficient than 3SLS (but don't depend on all three equations being correctly specified). You might look at xtdpdml which uses a GSEM/SEM approach to this problem and produces SEM code that you can then modify as you please.