Daniel Hoechle wrote:
Snip <<As Mark already said, you should not rely on
heteroscedasticity-consistent SEs but rather on panel robust
SEs as the former are typically inconsistent. Moreover, I am
quite sure that your panel will exhibit considerable
cross-sectional dependence. Therefore, without performing a
test for cross-sectional dependence with the -xtcsd- command,
you can be quite sure that -xtscc , fe- would produce more
appropriate standard error estimates. In fact, you can check
this by estimating both a model with cluster-robust standard
errors and one with Driscoll-Kraay standard errors. If the
t-stats from the latter model are significantly smaller than
those of the former model, then it is very likely that
cross-sectional dependence is present in your dataset. From my
experience with similar-sized microeconometric panel datasets,
I would expect that t-stats from Driscoll-Kraay standard
errors are less than half the size of cluster-robust t-stats.>>
Daniel,
First, thanks for contributing xtscc. I’ve been exploring its
capabilities and it appears to be a really valuable addition
to Stata’s panel analysis options. I have a question. In the
passage above, you note a situation when xtscc produces larger
standard errors than cluster-robust. I have seen the same in
a couple of my own applications. However, I have also
encountered the opposite. I have a situation where xtscc
produces notably smaller standard errors (about one-quarter
the size) compared to cluster-robust. I’m having a hard time
understanding why this would be. According to xtcsd
(pesaran), I have significant cross-sectional correlation of
0.43. Do you have any sense of what conditions would cause
the Driscoll-Kraay standard errors to be so much smaller than
cluster-robust? FYI, my data set has n=348, t=9. Thanks for
any suggestions.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
