System-GMM: autoregressive dependent & lag choices/number of instruments (xtabond2)

[Moritz Wolf]

Dear all,
I have two main concerns regarding a system-GMM regression with xtabond2, and hope some of you can help me with these questions.
I added the code and highlighted the critical terms.

xtabond2 L(0/1).$dependent $regressors $dummies $controlls, gmm(L1.$dependent $regressors, laglimits(2 .) collapse) iv($dummies $contro lls) svm twostep robust noconst artests(4)

Is it always necessary to include the autoregressive term of the dependent variable in the gmm(..) part of xtabond2?
And if so, could some of you explain to me why this is the case and what this means for the model? My results change a lot depending on whether I include the lagged dependent here or not.

xtabond2 L(0/1).$dependent $regressors $dummies $controlls, gmm(L1.$dependent $regressors, laglimits(2 .) collapse) iv($dummies $controlls) svm twostep robust noconst artests(4)

How many lags and, therefore, how many instruments are appropriate? I am aware of the potential problems with too many instruments, but how many are too many?
I have 950 observations, is it thus better to go for 54 or 15 instruments?
Moreover, I am curious how the Hansen-test statistic relates to this. As far as I understand, a p-value close to 1 implies that I have too many instruments. A p-value of lower than 0.1 indicates that the IV's are not exogenous, right? But is it feasible to use the test statistic to decide how many lags/Nr of IV's I should include in my model?

Thank you in advance for your help!

Best,
Moritz

[Sebastian Kripfganz]

You do not necessarily have to use lags of the dependent variable as instruments, as long as lags of your other regressors (used as instruments) are sufficiently correlated with the lagged dependent variable. If your results change a lot, this could either mean that your other instruments are insufficiently correlated with the lagged dependent variable (when not using lags of the dependent variable as instruments) or that the lags of the dependent variable are invalid instruments (when using them). The latter could happen if there is remaining serial error correlation.

There is no general answer to your question of how many instruments are appropriate. If in doubt, keep the number low. You cannot generally use the Hansen test statistic to decide about the number of instruments.

More on dynamic panel data GMM estimation:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.

[Moritz Wolf]

Thanks a lot for your very helpful response!
Is there any possibility to test either the relation between my instruments & the lagged dependent or the validity of the lagged dependent IVs? Or is it sufficient to assume that my insturments are valid IVs for the lagged dependent, because they are valid to the dependent and my time series follows a dynamic process?
Thanks again for you help!