System GMM on a smaller sample

[Alex Anohenko]

Dear Stata users, I am a newbie here but I have googled relevant information before coming here. First of all, I have very little econometric background but my policy research have forced me to register here in order to get some advice, considering how little I know by now.

I am doing a policy research on health issues. My topic is how tuberculosis has effect on economic growth of former communist countries. This issue is I believe very vital at the moment. My sample consists of 26 countries for years 1990 - 2015. I have not used Stata before, but I was able to teach myself how to run the basic fixed effects regression first.

xtreg gdp_growth tuberculosis l.log_gdp_percapita investment popul_growth trade inflation gov_spending life_expectancy, fe r

As you see below as expected TUBERCULOSIS is negative and significant. I have done a lot of literature review and noticed that I need to use system GMM with my panel data.

I have done reading and understood that I have to run xtdpdsys in Stata.
I have run the following command
xtdpdsys gdp_growth tuberculosis l.log_gdp_percapita investment popul_growth trade inflation gov_spending life_expectancy , lags(2) twostep vce(robust) artests(2)

But notice how the p-value has changed now. Moreover, everything is just insignificant now.

Can some one let me know what I have done wrong? I am very newbie to something outside of fixed effects regression so pardon me.
I have uploaded my dataset just in case.

I appreaciate any advice.

[Marcos Almeida]

It seems you have an average of 15 measures by country. Modeling the xtreg model (Yvar and xvar transformations, selecting the best model, etc.) could be helpful (help xtreg postestimation). I'm not sure a dynamic model is the best approach. If you have Stata 15.1, I recommend to take a look at the ERMs (extended regression models), mainly in terms of dealing with instrument variables and particularities of Yvar in both equations.

[Alex Anohenko]

buy why system GMM would not work in my case?

[Sebastian Kripfganz]

Your number of instruments for system GMM is far too large. You are using about 3 instruments for every 4 observations. Given the small cross-sectional dimension of your sample, the number of instruments would still be relatively large even if you reduce them using the maxldep() option of xtdpdsys. Hence, I do not recommend to use xtdpdsys in your context. Moreover, I do not recommend to use a GMM estimator for dynamic panel models at all without being familiar with the respective econometric literature. Otherwise, you might get easily caught in other traps that turn your results completely unreliable.