Re: st: Robust Standard Errors in Paneldatasets

[Stas Kolenikov <skolenik@gmail.com>]

On Tue, Oct 26, 2010 at 10:13 AM, Léon Bueckins
<l.bueckins@zeppelin-university.net> wrote:
> Thank you for clearing this up to me. So using Petersens code or ivreg2 will be unbiased as they account both for 2 dimensional clustering, right?

Cluster standard errors are NEVER unbiased. Just forget about it. With
small number of clusters, they will be very notably biased down.
Because of this bias, you should expect 10-20% rejection probabilities
in the tails instead of the nominal 5%.

The best you can hope for is that the standard errors are consistent
if you are lucky, i.e., if the technical assumptions under which the
proofs were derived are satisfied. I am not happy with the proofs in
Cameron, Gelbach and Miller as they only deal with asymptotics in one
dimension, not in both dimensions simultaneously. Angrist & Pischke's
"Mostly Harmless Econometrics" suggest the number 42 as the practical
definition of "infinity" in terms of the number of clusters, but this
is inevitably a judgment call.

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.

*
*   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/