Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Clustering Problem


From   Austin Nichols <austinnichols@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Clustering Problem
Date   Fri, 22 May 2009 12:03:39 -0400

chris <cc.basten@googlemail.com> :

Kish (1965) gives a correction for a mean; see
http://www.stata.com/statalist/archive/2006-12/msg00738.html for a
reference.  See also
http://www.stata.com/statalist/archive/2008-07/msg00671.html

See page 15 of http://repec.org/usug2007/crse.pdf referencing
http://www.nber.org/papers/t0327 for an approx correction for
regression.  I'm not aware that anyone has adapted that to IV
regression, but you can probably derive a similar approximation.  Note
that in your case, the intracluster correlation of the excluded
instrument is 1.

There is no guarantee that the cluster-robust SE estimator will give
you good answers in any finite sample, and my guess is that the
approximations will also be wildly off in many real applications, but
it may be better than nothing. You should probably run some
simulations where you know the DGP but you impose various levels of
clustering on residuals, and check your SEs and rejection rates,
uncorrected/het-robust/cluster-corrected.

On Fri, May 22, 2009 at 3:47 AM, statachris <cc.basten@googlemail.com> wrote:
> Good morning,
>
> to exploit a (quasi-) experimental setting, I would like to run an IV
> regression in a setting in which there is clustering with only 2
> clusters, and furthermore the two clusters correspond to the two values
> of the instrument dummy.
>
> It doesn't seem to make much sense here to use cluster-adjusted standard
> errors. But since the t-stats without cluster-adjustment are very large
> indeed, I was wondering whether I could make an argument that they would
> remain significant even if I could and did adjust for the clustering. To
> do so, I would need to have an idea by what factor the standard errors
> would have to be multiplied if I did adjust them. So I was wondering if
> there exists any formula or rule-of-thumb that would give me an idea of
> that?
>
> Or is there another, better way to deal with this?
>
> Many thanks!
>
> Chris
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index