I think u are following right procedure.

Hi,

To check how such your data are poolable, see this simple discussion:

to check how such your data are poolable

Similar questions that might help:

https://www.researchgate.net/post/I_am_running_a_model_with_panel_data-can_anyone_help

https://www.researchgate.net/post/Should_I_include_Pooled_OLS_Random_Effects_and_Fixed_Effects_in_my_results_table_when_the_GMM_estimator_is_superior

Thank you both for your time. I have checked the data and I think it can be pooled. I could make use of GMM as well but the problem is that my data is a long panel. I have 11 countries and 15 years. Besides, I was wondering if running Pooled OLS with Robust Standard Errors would help to improve the validity of results. Because POLS is usually criticized due to the underestimation of standard errors which could be biased. Therefore, I was also thinking if I could average my data into say periods, such as 2002-2004, 2005-2008 and so on. Do you think I could use GMM to analyze my data? Please share your thoughts on this...

Averaging across sub-periods will hide some of the information in your data. I would therefore say: Only do it if you have a very strong reason to do so and if you do it then think about how you might use the full data for robustnes checks regrading the conclusions you get from the averaged data.

Hi,

I would suggest to run additional tests to see whether you need a GMM framework. First of all, you could run arellano-bond test for first and second order correlation since the DW statistics can be critcized when using a panel framework. In case of singificant correlation you should try a dynamic specification like GMM-DIF of GMM-SYS.

Second, are all the regressors exogenous or is there a problem of endogeneity? Why the papers you have read use the IVV-GMM approach?

In general, journals ask for robustenss checks, so you might, in any case, add other estimators alongside the pooled OLS. Given the importance of time invariant variables, I would suggest the random effect and the hausman taylor estimator (or a standard IV-GMM), HAC estimates (they can be implemented in STATA with the xtivreg2 command) and, according to the previous discussion, a dynamic panel specification.

Last suggestion, since your time span is quite long you might wanna to test for stationarity of the series and eventually use a cointegration analysis (for this sample size the CCEMG estmator by Chudik and Pesaran 2015 could be good, especially because it can impemente a dynamic specification or control for endogeneity).

Two final suggestions: first, add time dummies ot the pooled old (and all other estimators except for the CCEMG); second, try to cluster standard errors by country.

Maybe you don't need to show all the different attempts I suggested but in my opinion these are steps worth trying.

I hope this was helpful,

best,

Piero

Dear Vince Daly and Piero Esposito,

Thank you very much for your time and efforts.

I am clustering by Country in Pooled OLS.

I decided not to go for data averaging and the use of GMM.

In order to show the Robustness of my analysis, I have been recommended to use Pooled OLS along with 2SLS and Random Effects models. Since I don't have valid instruments to be used in 2SLS, I will be using Lewbel's 2SLS with special regressor estimators.

Many researchers would be happy when they receive such test results. And if you feel more as an economist than an econometrician then simply finish the study, submit it and wait for the reviewers remarks. I agree with Vince Daly: one should have good (economic) reasons to use other methods. Of yourse, one can always try to improve the efficiency, but one can also overstress the issue. If you want to do something additional, try simply to apply heterogeneity tests, for example the tests on variance equality. If you use Eviews, they are easily to handle.

Dear,

You are not on good truck. Because the you are going to compare two regions, i.e, in this case pooled data is not appropriate model.A pooled model is one where the data on different individuals are simply pooled together with no provision for individual differences that might lead to different coefficients. That mean it doesnot all for possible individual heterogeneity and it assumes them as homogeneous (It assume all regions as they are the same). Thus, I highly recommend you to use either Fixed Effect model or Random effect model based on Hausman test if you are not a such familiar with panel analysis. Otherwise you can search for Advanced panel analysis.

Dear Hubert Gabrisch and Tesfaye Etensa

First of all, I would like to thank you for your time and effort and respond to my question.

I have run my analysis and arrived at good results. However, as you have mentioned, should I keep things simple or use complex methods to arrive at pretty much similar results. As of now, I ran three different analysis using, Pooled OLS, Random Effects and FGLS models to check the robustness of my results across similar methods. Among them, both POLS and RE present similar results, but FGLS provides a little different outcome. For instance, two of the variables that were insignificant became statistically significant in FGLS. Therefore, I was wondering whether I should keep one of the models or keep all three together ?