RE: st: ivreg2 missing SE's when using two cluster variables

["Schaffer, Mark E" <M.E.Schaffer@hw.ac.uk>]

Max,

Some other ideas:

1. If you have some exogenous regressors that aren't of interest, you
can use the partial() option to partial them out and see if that makes a
difference.

2. -ivreg2- has two undocumented options, psd0 and psda, either of which
can be used to make the S matrix (the estimated matrix of orthogonality
conditions, used to construct the VCV) positive definite.  psd0 uses a
Stock-Watson 2008 Econometrica suggestion and replaces negative
eigenvalues with the absolute value of the eigenvalues.  psda uses a
Politis 2007 suggestion to replace negative eigenvalues with zeros.
(Apologies, I don't have more specific references to hand.)  Personally,
for me this is a little too much like torturing the data until it
confesses ("We have ways of making you invertible..."), and this is
partly why the options aren't documented.  But caveat emptor, and feel
free to give them a try.

--Mark

> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-
> statalist@hsphsun2.harvard.edu] On Behalf Of Austin Nichols
> Sent: 17 September 2012 20:09
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: ivreg2 missing SE's when using two cluster variables
> 
> The 2-way cluster-robust SE is constructed from V1+V2-V12 and that
matrix is
> not guaranteed to be PSD.  I would use the nearest PSD matrix to the
> defective matrix instead.
> Or use the max over SE for clustering in either dimension and both
> dimensions, where a max over two numbers and missing is the max over
two
> numbers.
> 
> On Mon, Sep 17, 2012 at 3:02 PM,  <Max.Livingston@ny.frb.org> wrote:
> > Hello,
> >
> > I'm performing some analyses using -ivreg2- to perform OLS with two
> > cluster variables. I also have a relatively small sample (around 200
> > obs) so I'm using the small option. I found that in some of my
models,
> > there is
> >
> > no std error. While this is not incredibly surprising given the
small
> > n and multiple clusters, I'm a little confused as to why I
experience
> > this problem in some models and not others, and what potential
> > solutions may exist. The help file explains that if the small option
> > is used with clustering, ivreg2 makes the following adjustment: qc =
> > (N-1)/(N-K)*M/(M-1), where M is the minimum # of clusters from the
two
> > cluster variables. Yet, when I run the same regression with either
one
> > of the clusters, ivreg2 is able to calculate the std. error. Further
details:
> >
> > N=251, cluster1 has 119 clusters, cluster2 has 60. This is not a
case
> > of singleton dummies, as there is only one independent variable.
> >
> > Any advice you may have would be greatly appreciated.
> >
> > Thanks,
> > Max
> *
> *   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/


-- 
Heriot-Watt University is the Sunday Times
Scottish University of the Year 2011-2012

We invite research leaders and ambitious early career researchers to 
join us in leading and driving research in key inter-disciplinary themes. 
Please see www.hw.ac.uk/researchleaders for further information and how
to apply.

Heriot-Watt University is a Scottish charity
registered under charity number SC000278.


*
*   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/