Dear Ramon,

to answer you question i need more information:

first, i guess the RE and FE specifications are without the lagged dependent variable, in this case they are long run impacts whereas the coefficients of the DPD are shor run impacts (long run imapcts being b/(1-rho)) this is a first reason for divergence in the results and you should not be worried about it. If you try dynamic RE and FE (although biased and inconsistent) the coefficient should get closer to those of the GMM.

Secondly, you should check the validity of your instrumentation procedure: how many instruments are you using? As a rule oh thumb you should not exceed the number of panels.

Finally, Is there any other endogenous regressor?

Once you checked all these steps, in my opinion you should not be worried about the difference between estimators. If a DPD is the best specification then the results of GMM of AH should be considered more reliable.

I hope this helped,

best,

Piero

Dear Ramon,

Usually, if OLS and FE estimates are close it indicates that the pooled model (based on OLS) is enough. But since the estimates are insignificant I would guess there is smth wrong with the model in respect to variable choice. This could be the reason why OLS and FE are "close" and that only the individual effects are (jointly) significant in FE. For GMM you are probably using some instrumental variables, which handle slightly better.

Hope this help,

Kamil

Dear Ramon,

First, most (if not all) panel data models have the implicit assumption that T

Thank you all for contributing your thoughts. To clarify my question, the FE and OLS models contain the lagged dependent variable and their estimates are therefore necessarily biased.

Based on different robustness checks (limiting the instrument count) and the pertinent statistics (Hansen), the instrumentation procedure should not be an issue.

Apart from the lagged dependent variables there are other endogenous regressors, which is accounted for.

Having clarified these points, it seems to me that the problem is not in the sGMM specification, but the OLS model. Nickel suggests that these estimates are upward biased, which is not the case with my data. The question is why I don't observe this upward bias.

I would suggest to use the Hahn Kuhersteiner which is designed for N and T medium to large.

https://onlinelibrary.wiley.com/doi/abs/10.1111/1468-0262.00344

It is implemented here extended for exogenous variables.

https://www.researchgate.net/publication/312165764_Panel_Vector_Autoregression_in_R_with_the_Package_panelvar

We are sorry but instead of Hahn we wrote Hank in the help file of the code. There will be a package update soon.

Type install.packages(“panelvar“) in R to install all functions.

Now i have same problem.

Do you think system gmm estimate should be between ols and fe estimate.

Of course, it seems to be specified well.

I have a similar condition, but the GMM is lower than OLS and FE. I also use lag dependen var in my model. Is there any reson for this?

I have the firm characteristics as my IV and Zscore as my DV, is it possible if i am using GMM as if i proceed with this, i need to look into system gmm and difference gmm. after run the system and difference, the result is not good only one iv is signficant.