[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: Large differences between standard errors with and without clustering

From   "Austin Nichols" <>
Subject   Re: st: Large differences between standard errors with and without clustering
Date   Tue, 30 Oct 2007 14:34:33 -0400

See also

Unlike Maarten, I don't think -xtreg, re- will offer you much.  The
result with the -cluster- option is telling you (essentially) you
really have 13 observations, not 13*T, so the SEs are big.  The size
of increase in SEs you observe is large, but not unexpectedly large.
I've often seen SEs triple and quadruple with the -cluster- option.

But note that with the -cluster- option, you are getting
asymptotically correct inference, meaning when the number of clusters
is arbitrarily large, and 13 is not close enough to infinity for my
taste.  That said, the cluster-robust estimator is biased *downward*
in small samples!

Short answer: get more countries and more years of data.

You could also try different SE corrections from -xtivreg2- (ssc
install xtivreg2) which does not require endogenous variables.

On 10/30/07, Maarten buis <> wrote:
> --- "M.C.D. van Damme" <> wrote:
> > I am estimating an OLS regression with variables on the individual
> > and country level, correcting for intragroup correlation (,cl
> > (country)). I have 13 countries.
> > After correction, not only the standard error of the country level
> > variable is corrected, but also the standard errors of some of the
> > individual level variables are much higher. Normally, I do not find
> > such big differences.
> > What can be the reason for these large adjustments in standard errors
> > on the individual level?
> > In other words, what is the cluster option exactly doing?
> Normally standard errors are based on the assumption that you have
> drawn a simple random sample. The cluster option corrects the standard
> errors (but not the coefficients) when your sample isn't random but
> contains clusters. Robust standard errors (including -cluster-) are
> discussed in the User's guide (for Stata 10) chapter 20.15. It was also
> discussed in the User's guide of previous version but my Stata 9
> manuals now reside in the office of my advisor, who is gone now, so I
> can't give you the chapter.
> If you are worried by this try estimating your model with -xtreg- or
> -xtmixed- with the -re- option. 13 countries is not a lot, so you can't
> enter a lot of country level variables, but that is as it should be: on
> the country level you have only 13 observations, so more than 1 or 2
> country level variables doesn't make sense in any model even if Stata
> gives you results.
> Hope this helps,
> Maarten
*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index