Dear Statalist,
I have an issue that I simply cannot figure out, despite a lot of googling and looking at papers. I'm running a panel data regression, with both time and id dimensions, but I have an endogeneity issue with one of my variables, which I regress with an interaction effect with an exogenous variable.
Fortunately, I seem to have found a instrument (potentially two) for that variable, but this instrument does <not> vary over the time dimension. My questions are as follows: the standard estimator that Stata implements when running xtivreg uses is the Balestra and Varadharajan-Krishnakuma estimator which, according to what I've read, implements the panel IV by first time de-trending your instruments... which obviously would mean that I'd be left with nothing! Can anyone confirm that this is the case?
Secondly, the ec2sls option / Baltagi estimator uses both the de-trended instrument and the time average instrument, so if my interpretation of the above is correct, would it be solve these issues and give consistent estimates?
Thirdly, I have used both estimators and the results for both are odd. The B&V-K estimator seems to work sometimes, and sometime not. The EC2SLS estimator seems to work, but if I ask for the first stage estimates, the de-trended instrument, denoted by _d, is not simply omitted, but used and sometimes it gives significant results, generally with opposite signs to the mean; on the flip side, the constant is then omitted, so is it the case that it's simply using the de-trended instrument as a constant? More generally, how could the B&V-K estimator even work in this situation?
Finally, if I want to the first stage F-statistics, can I use xtoverid, noisily? Even if, when I use EC2SLS, it omits some variables due to multicollinearity?
Many thanks.
I have an issue that I simply cannot figure out, despite a lot of googling and looking at papers. I'm running a panel data regression, with both time and id dimensions, but I have an endogeneity issue with one of my variables, which I regress with an interaction effect with an exogenous variable.
Fortunately, I seem to have found a instrument (potentially two) for that variable, but this instrument does <not> vary over the time dimension. My questions are as follows: the standard estimator that Stata implements when running xtivreg uses is the Balestra and Varadharajan-Krishnakuma estimator which, according to what I've read, implements the panel IV by first time de-trending your instruments... which obviously would mean that I'd be left with nothing! Can anyone confirm that this is the case?
Secondly, the ec2sls option / Baltagi estimator uses both the de-trended instrument and the time average instrument, so if my interpretation of the above is correct, would it be solve these issues and give consistent estimates?
Thirdly, I have used both estimators and the results for both are odd. The B&V-K estimator seems to work sometimes, and sometime not. The EC2SLS estimator seems to work, but if I ask for the first stage estimates, the de-trended instrument, denoted by _d, is not simply omitted, but used and sometimes it gives significant results, generally with opposite signs to the mean; on the flip side, the constant is then omitted, so is it the case that it's simply using the de-trended instrument as a constant? More generally, how could the B&V-K estimator even work in this situation?
Finally, if I want to the first stage F-statistics, can I use xtoverid, noisily? Even if, when I use EC2SLS, it omits some variables due to multicollinearity?
Many thanks.