xtabond2 postestimation

[Gaurav Bhattacharya]

Dear Stata Users,

Could someone brief me about the post estimation tests in case of GMM estimation?
My results on Sargan Hansen test are as follows:

Sargan test of overid. restrictions: chi2(51) = 93.19 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(51) = 48.91 Prob > chi2 = 0.557
(Robust, but weakened by many instruments.)

Aren't they contradictory?

[Sebastian Kripfganz]

Unless you are using only instruments for the first-differenced equation and are assuming homoskedasticity of the errors, the Sargan test is based on a weighting matrix that is not optimal. You can generally just ignore it. The large number of the degrees of freedom (51) indicates that you are probably using too many instruments in your estimation. It is advisable to reduce this number, for example with the collapse option.

[Gaurav Bhattacharya]

Thank you so much! The number of cross-sectional units in the estimation is 66 (>51). In that case, is it necessary to use "collapse"?
And, I was keeping a tab on other query threads where you have mentioned that it is essential to include eq(level) with iv(.). Why is it necessary?
Lastly, could you please suggest any reading material other that Roodman which provides a lucid explanation of the theory of system GMM? I think I am still not thorough with the underlying theory and intuition.

Thanks in advance.

[Sebastian Kripfganz]

With only 66 cross-sectional units, I would always recommend to collapse the instruments to obtain robust estimates.

Regarding the iv() option, please see my comment in another topic:
https://www.statalist.org/forums/for...74#post1361774

You should be able to find a chapter on system GMM estimation in any modern econometrics textbook that covers (advanced) panel data topics.

[Gaurav Bhattacharya]

Thank you so much, Sebastian!