Re: st: test for clustering in instrumental variables settings

[Austin Nichols <austinnichols@gmail.com>]

Stas, Sergio,
Indeed, Mark and I have a working version of Kezdi's test for
clustering, which Mark extended so one can consider clustering
relative to one variable or many.  But all of that is for linear
models; given that the recursive biprobit models you are estimating
work by estimating the correlation across equations in a SUR sense
(also a clustering problem), I think your case is harder.  With
unbalanced clusters, there are good reasons to cluster and good
reasons not to.  In that case, you probably have to report both
results and discuss the potential problems with each estimate.

But what do you mean by:
treatment variable is whether
>> treatment at certain
>> > type of hospital
?

On Thu, Feb 25, 2010 at 11:55 AM, Schaffer, Mark E
<M.E.Schaffer@hw.ac.uk> wrote:
> Stas, Sergio,
>
>> -----Original Message-----
>> From: owner-statalist@hsphsun2.harvard.edu
>> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of
>> Stas Kolenikov
>> Sent: Thursday, February 25, 2010 3:52 PM
>> To: statalist@hsphsun2.harvard.edu
>> Subject: Re: st: test for clustering in instrumental
>> variables settings
>>
>> With a binary endogenous variables, you need to think first
>> whether the
>> effect in the main dependent variable is due to the 0/1 value of the
>> endogenous variable, or due to the propensity (the linear
>> predictor part)
>> associated with that variable.
>>
>> I don't think there are formal tests for whether you do or do
>> not need the
>> clustered standard errors. But the folk wisdom is, if you
>> have clusters then
>> you have to use the clustered standard errors (which will
>> likely dilute the
>> significance of your results compared to the assumption of the i.i.d.
>> data). In a somewhat related problem of testing for
>> heteroskedasticity in
>> linear regression, econometricians use White's information matrix test
>> (-estat imtest- after -regress-). In all likelihood, it can
>> be generalized
>> to the clustered data situation, but I am not aware of
>> whether that was done
>> or not.
>
> Gabor Kezdi has done it.  Here's one version of his paper:
>
> http://papers.ssrn.com/sol3/papers.cfm?abstract_id=596988
>
> I had a go at implementing it (with Austin Nichols) in Stata.  We have a
> working beta version (somewhere...).
>
> --Mark
>
>>
>> On Wed, Feb 24, 2010 at 4:39 PM, Sergio I Prada
>> <sprada1@umbc.edu> wrote:
>>
>> > Dear users:
>> >
>> > I am trying to come up with a good way to test whether I
>> need to use SEs
>> > clustered in my estimation. But I could not.
>> > I have a binary outcome and a binary treatment variable. My
>> treatment
>> > variable is endogenous and I have two good instruments. The
>> model includes
>> > covariates.
>> > The problem is that my treatment variable is whether
>> treatment at certain
>> > type of hospital, and my clusters are hospitals. So with no
>> variation at
>> > the cluster level on the endogenous variable I cannot use
>> tricks like
>> > adding averages or deviations of the endogenous variable
>> (as recommended
>> > in the multilevel literature).
>> > I am using instead recursive biprobit models, and of course
>> the problem is
>> > that the significance of my results change with and without
>> SEs clustered
>> > at hospital level.
>> > I have 69 clusters, and they vary a lot by size (from 2
>> patients in one
>> > hospital to 122 in other)
>> > Are any of you aware of a way to test whether I have to
>> adjust SEs at the
>> > cluster
>> > level.
>> >
>> > --
>> > Sergio

*
*   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/