GMM problem - Sargan Arellano-Bond test

[Petko Bachvarov]

Dear Statalist members,

I have been trying to implement xtdpdsys in my paper. However, based on my data set the Sargan test is always zero, where as the Arellano-Bond test is showing second autocorrelation. When I use three lags on the dependent variable I have only first autocorrelation. Having said that, could you please let me know whether is correct to use more lags in order to get the right result or it is wrong because my model is overloaded with instruments.I am adding L(3/3) to the dependent variable. Although I tried to read what is the purpose of the latter, I still cannot understand how to interpret it? In addition, can you please help me and explain how can I use Hansen test with my data set?

The first table is showing the results based on the real model (i.e. only one lag on the dependent variable and on the explanatory variable).

Below you can see the results from the table with lags:

I would really appreciate your help if you have any recommendations or comments.

Apologies in advance if the format is not correct (will fix it immediately).

[Sebastian Kripfganz]

In the second specification, you are using the 3rd lag of llrgl as your dependent variable. That is probably not what you want. You would like to have llrgl as your dependent variable but then use lags 1 to 3 as right-hand side variables, i.e. you would specify something like xtdpdsys llrgl ... , lags(3).

xtdpdsys does not compute the Hansen test. You might want to consider the community-contributed xtdpdgmm command as an alternative:

• XTDPDGMM: new Stata command for GMM estimation of linear (dynamic) panel data models

My recent London Stata Conference presentation explains how to use the xtdpdgmm command and it might answer some more of your questions:

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

Regarding formatting, please see point 12.3 of the Statalist FAQ on how to use CODE delimiters.

[Petko Bachvarov]

Dear Sebastian,

I do appreciate your answer.

Thanks for the recommendations. I wanted to ask you just for clarification. Is it okay to add more lags in the model overall? As some of the comments in the forum state that it is fine but others are saying it is not.

Thanks in advance.

[Sebastian Kripfganz]

As long as your sample size is large enough, adding lags should not be a problem.

I just realize that your sample has a very small cross-sectional and a very large time dimension. That is not a typical setup for these dynamic panel model GMM estimators that are designed for situations with short time horizon and large cross-sectional dimension. Given that you are not treating any of the variables as endogenous, you can just use the conventional fixed-effects estimator with xtreg, fe. The bias due to the lagged dependent variable should be absolutely negligible given your more than 200 time periods.

Alternatively, a (pooled) mean-group estimation might be reasonable.

[Petko Bachvarov]

Dear Sebastian,

Thank you very much for your help!

All the best!