st: Heteroskedasticity in FE model

[Giovanni Canitano <giovanni.canitano@cce.unifi.it>]

Hy friends,

 

I have estimated my FE model with the command xtreg varlist, fe.

 

1) Obviously, Stata has dropped the variables time invariant. To obtain the coefficient of these variables I have retained the fixed values by the command

 

predict var,u

 

and then I have done an OLS using:

 

reg var timeinvariant. Is this procedure correct?

 

2)To be sure of my procedure I have estimated my model by an LSDV. Why are the coefficients of the time-invariant variables roughly similar (but not equal) to those one worked out with the Fixed Effect model? And Why are the standard errors completely different?

 

3) I suppose to have heteroskedasticity in my dataset. How Can I test this one? Is "xttest3" sufficient? I also tried the White’s test for heteroskedasticity, not using "whitetst", but following the theoretical procedure. Is it correct for Panel Analysis?

 

4) All these test showed the presence of heteroskedasticity. So I decided to use a GLS model (for STATA xtgls varlist, p(h)). Why do I receive different results, using the iterated methodology (igls) instead of the basic one? And what is the best?

 

5) I've tried to test the autocorrelation by the command xttest2 but I have alwaiys the same error: the correlation matrix is singular. What does it really mean?

 

6) If the theory suggests me that the Fixed Effect model is better than the others, How can I do in presence of heteroskedasticity or autocorrelation? Could I estimate a Fixed Effect GLS?

 

Thank you

 

Giorgio