可以参考baltige 的书

实际上，只要数据量大就可以

XTABOND or XTABOND2 could be used in dynamic panel data GMM estimation (xtabond2 is better - it can do what xtabond can do, but not the reverse). For the instruments, by default, XTABOND2 uses all lagged values (you can certainly restrict the instrument set) of dependent and independent variables in the first differenced equation, and the lagged first-differerces of the series (dependent and independent variables) dated t-1 to be used in the level equations.

By the way, the IVREG2 mentioned by maoxinshu above is to be used for cross-section estimation for which a suitable instrument is needed - the instrument is always hard to get, not the lagged vales as I mentioned above.

Hope these help.

 ivgmm0 我看到过用这个命令做GMM

GMM并不像很多人想象的那么神奇，我觉得还是应该根据自身研究的需要酌情考虑是否要用GMM。

刚刚做了GMM，发现真如Roodman（2006）所说，必须是T小N很大的样本才行，如果样本中N的数量小于30，如很多人做的全国各省份面板数据的GMM研究，其本身就会极易出现工具变量过多的问题。所以，建议只在微观面板数据或世界经济研究中应用比较好。否则，GMM的估计效果不比固定效应好多少，甚至比OLS还差。

有道理，现在用GMM方法去估计省份面板数据都快成一种风气了，是该好好想想方法本身的适用性了！