Re: st: ivregress with2sls and clustered standard errors

["Averett, Susan L" <averetts@lafayette.edu>]

I didn't have any luck with the wmatrix option. In that case, I actually am told that I have no observations. Could this be because I am using school fixed effects as well as trying to cluster on schools?

Susan Averett
----- Original Message -----
From: "Stas Kolenikov" <skolenik@gmail.com>
To: statalist@hsphsun2.harvard.edu
Sent: Saturday, March 26, 2011 4:35:21 PM
Subject: Re: st: ivregress with2sls and clustered standard errors

Susan Averett is working with education data estimating an
instrumental variable model with standard errors clustered at the
school level, and is looking for ways to conduct overidentification
tests.

On Sat, Mar 26, 2011 at 3:08 PM, Averett, Susan L
<averetts@lafayette.edu> wrote:
> I am trying to run a model using ivregress with 2sls as my estimator.
> I have a data set of individuals who are clustered in various schools.
> I am using school fixed effects in my model but I am trying to cluster
> my standard errors on the school variable

If you used fixed effect dummies for schools, then you would have more
regressors than you have clusters, which should've triggered an error
message somewhere.

> AND also obtain the overidentification tests from
>
> estat overid
>
> However, when I use the cluster command, I get an error:
>
> robust tests of overidentifying restrictions after 2SLS
>  estimation not available with cluster-robust standard errors
>
> I do not get this same error if I use robust standard errors.

Apparently, the way -overid- is currently coded, it does not even try
to compute the overidentification test with clustered standard errors.
Either it was too difficult to code, or, more likely, there was (and
in all likelihood, still no) solid theory regarding overid tests with
clustered data.

> What is the option if this error is received? Is is possible I do not
> need to cluster my standard errors?

I see three ways to get out of this.

1. At a descriptive level, you can compare the standard errors with
and without clustering (-robust- only). If they are not drastically
different (indicating that the schools are rather alike, and student
homogeneity within schools does not affect your results much), then
you can probably use the test without clustering, although of course
you would need to describe everything in detail to convince the
readers of your paper.

2. You can reverse engineer -overid-, figure how this test is
obtained, and compute it by hand.

3. You will get the overidentification test from -gmm- (assuming you
are using Stata 11 which you did not state). See examples in the
manual as to how run the analogue of -ivregress-; you would need to
specify the -wmatrix(cluster school) to obtain both estimates that are
efficient under clustering, and the overid test (using -estat
overid-).

-- 
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/

*
*   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/