Hello,
I am using an unbalanced panel data set with N=31, T between 18 and 43. Because of evidence of heteroskedasticity, serial correlation and cross-section dependence, I am using Driscoll-Kraay standard errors. I always specify country and time fixed effects.
I investigate the presence of endogeneity using IV methods and IVREG2 specifically, estimating specifications like
xtivreg2 depvar exogvar yeardummies (endogvar = L.endogvar L2.endogvar), fe dkraay(10)
Now it appears that xtivreg2 implements Driscoll-Kraay standard errors by clustering on the time dimension (http://fmwww.bc.edu/RePEc/bocode/i/ivreg2.ado) - but based on the documentation, I am not actually sure that that is equivalent to how Driscoll-Kraay standard errors are implemented in xtscc, say (http://fmwww.bc.edu/repec/bocode/x/xtscc_paper).
I am asking because there is a huge difference in the reported underidentification test statistic (Kleibergen-Paap LM statistic) when I use the Driscoll-Kraay option (dkraay) vs. when I use the Newey-West option (bw) in xtivreg2, so my choice of (consistent and appropriately implemented) standard error really matters for the conclusions I draw about the IV results overall. (I'm aware that Newey-West standard errors don't address the presence of cross-section dependence.)
I would be extremely grateful for any thoughts on the matter! (My apologies if I'm not meeting all how-to-post-on-statalist requirements; this is my first one.)
Best,
Lea
I am using an unbalanced panel data set with N=31, T between 18 and 43. Because of evidence of heteroskedasticity, serial correlation and cross-section dependence, I am using Driscoll-Kraay standard errors. I always specify country and time fixed effects.
I investigate the presence of endogeneity using IV methods and IVREG2 specifically, estimating specifications like
xtivreg2 depvar exogvar yeardummies (endogvar = L.endogvar L2.endogvar), fe dkraay(10)
Now it appears that xtivreg2 implements Driscoll-Kraay standard errors by clustering on the time dimension (http://fmwww.bc.edu/RePEc/bocode/i/ivreg2.ado) - but based on the documentation, I am not actually sure that that is equivalent to how Driscoll-Kraay standard errors are implemented in xtscc, say (http://fmwww.bc.edu/repec/bocode/x/xtscc_paper).
I am asking because there is a huge difference in the reported underidentification test statistic (Kleibergen-Paap LM statistic) when I use the Driscoll-Kraay option (dkraay) vs. when I use the Newey-West option (bw) in xtivreg2, so my choice of (consistent and appropriately implemented) standard error really matters for the conclusions I draw about the IV results overall. (I'm aware that Newey-West standard errors don't address the presence of cross-section dependence.)
I would be extremely grateful for any thoughts on the matter! (My apologies if I'm not meeting all how-to-post-on-statalist requirements; this is my first one.)
Best,
Lea