Panel data analysis - GMM-Model (xtabond2) & Hansen test

[Lorenz Franz]

Hello together,

for a thesis for the university, I am trying to analyze the effect of entrepreneurial education on economic growth.

For this, I use ln(GDPperCapita) as my dependent variable, basicschoolentrepreneurialedu and postschoolentrepreneurialedu as my key explanatory variables, and different control variables related to economic growth (R&D_expenditure, GrossCapitalFormation, Educ_Attainment, Unemployment, Governm_Expenditure, PopulationGrowth).

I'm pretty new to the analysis of panel data, but I was reading a lot that to avoid endogeneity problems in models related to economic growth a good regression model is the system GMM-Model.

To calculate it for my example I used stata17 and the xtabond2 command, but I face different problems regarding my results, and I'm not sure if I used the command, especially with the iv and lagged variables in the right way.

 xtabond2 lnGDPperCapita  Basicschoolentrepreneurialedu Postschoolentrepreneurialeduc RandD_expenditure GrossCapitalFormation Educ_Attainment Unemploymenttotaloftotal Governm_Expenditure Populationgrowthannual, gmm(L.lnGDPperCapita, lag(. 4)) ivstyle(Governm_Expenditure RandD_expenditure GrossCapitalFormation Educ_Attainment Unemploymenttotaloftotal ) robust twostep

My results look like this:

My problems:

1. My Hansen test shows a high p-value of 0.959. After roodman it should be between 0,1 and 0,3. what did I do wrong in the model?
2. I read that the ar(2) is important for interpreting. What is it resulting in my example?

I really appreciate every answer and your help. As a beginner in panel data analysis, I face a lot of difficulties and your help would bring me forward a lot.
Many thanks

[David Roodman]

33 instruments seems like a lot for only 25 individuals and 109 observations. The 0.959 on the Hansen test could be exactly what it appears to be--a reassuring result--or it could indeed be misleadingly elevated by having so many instruments. I would reduce the number of instruments by inserting the "collapse" option in the gmm() option.

The high p value on the AR(2) test in first differences suggests that there is no first-order serial correlation in the undifferenced errors, which is good.

[Lorenz Franz]

Thank you for your advice Mr. Roodman. That helped a lot. The only problem that I am facing right now is that the p-values of my independent variables are still insignificant. Is there a way to influence the p-value?

[Sougata Mondal]

Respected @ David Roodman, in my study, the AR(1), AR(2) and Sargan test are showing a significant results though the AR(3) and Hansen test are showing insignificant results, therefore should I consider my model to be valid or not? (I have no of instruments< No of groups)
Your suggestions are highly needed.