Hello everyone,
I'm new with GMM, I'm using this estimator because I found it's the most suitable for my case. I have a panel dataset relative to the 20 Italian regions (n=20) for the period 2013-2017 (t=5). I want to use a dynamic model and regress GDP per capita on lagged GDP, other control variables (Investments, R&D, PublicExpenditure, Demography) and two variable of interest that are Corruption and an index of CriminalOrganization. I want to use a sys GMM since my data present kind of persistence. I use the command collapse to avoid the problem of "too many instruments" (I don't know if it is better to also limit the lags). I wanna treat PublicExpenditure and R&D as exogeneous. Data are in logs. I copied my command to ask if someone of you, experts, finds some contrast with the idea I presented:
xtabond2 logGDPpercapita L.logGDPpercapita logInvestments logPublicExpenditure logR&D logCorruption logCrimOrg logDemography, gmmstyle(L2.logGDPpercapita L.logInvestments L.logCorruption L.logCrimOrg L.logDemography, collapse) ivstyle(logR&D logPublicExpenditure) robust
The first question is: what are the main robustness checks to do? Are Sargan-Hansen, AR, Wald enough?
The problems are: 1)I obtain not significant estimation (for some coefficients) and mainly 2) the signs of coefficients are unexpected (to be clear, for example, the coefficient relative to Criminal Organization is positive). How it is possible considering that the correlation between GDP and CrimOrg is largely negative. Maybe Collinearity among regressors? Singularity? Too many instruments? Please, give me suggestions, it's real important for my research. PS: Results change dramatically if I exclude, add, "change" some variables or modify "something" in the command.
Thanks in advance
Regards