Statalist The Stata Listserver

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

st: RE: RE: ivreg vs ivreg2, both with cluster

From   <>
To   <>
Subject   st: RE: RE: ivreg vs ivreg2, both with cluster
Date   Fri, 23 Jun 2006 21:17:44 +0100

Hi Mark

Yes I am very much interested in your beta-Version. If you could sen it
to me....

Based on your comment: does it imply that the stata command ivreg uses a
somewhat different estimator from ivreg2, obviously not a GMM estimator
(looks like an OLS ?) ?

Best and thanks


-----Original Message-----
[] On Behalf Of Schaffer,
Mark E
Sent: 21 June 2006 22:58
Subject: st: RE: ivreg vs ivreg2, both with cluster


> -----Original Message-----
> From:
> [] On Behalf Of 
> Sent: 21 June 2006 13:39
> To:
> Subject: st: ivreg vs ivreg2, both with cluster
> Hi
> I am estimating a pooled cross section with 26 states, 10 time 
> variables, and, many additional explanatory variables, of which one is

> endogenous. Clustering at the state level is supposed to account for 
> serial autocorrelation of any kind.
> Using ivreg2, which I prefer for its test statistics, I receive the 
> error message:
> "number of clusters must be greater than instruments" (in principle 
> fine with me)
> Using ivreg and the same model specification, however, carries out the

> estimation without problems.
> Does anybody know why this is the case ? How can I make
> ivreg2 work for me without having to reduce the number of explanatory 
> variables ?

This restriction in the current version of ivreg2 arises because when
the number of clusters is smaller than the number of instruments, the
variance-covariance matrix of orthogonality conditions (central to GMM)
is not of full rank, and this can cause problems.

It's also a problem when the number of clusters is small in an absolute
sense.  Put another way, in your application you are relying on only 26
observations when estimating the standard errors in your regression.

That said, we have a version of ivreg2 on the way that has a new option,
-fwl- (for Frisch-Waugh-Lovell), which provides a way of addressing the
issue.  The variables in the varlist provided to -fwl- are exogenous
regressors that are "partialled out" of the remaining variables.  This
solves the problem because the var-cov matrix of orthog conditions then
becomes full rank.  ivreg2 will report the coeffs for the remaining
variables only.  All you need to do is specify enough exogenous
regressors in -fwl- so that the number of remaining instruments
(exogenous regressors + excluded instruments) is no longer greater than
the number of clusters.

We're somewhere between programming and testing this option, so I'm not
sure when it will be generally available, but if the need is urgent on
your part, I can send you a beta version.


Prof. Mark Schaffer
Director, CERT
Department of Economics
School of Management & Languages
Heriot-Watt University, Edinburgh EH14 4AS tel +44-131-451-3494 / fax

> Thanks
> Justina
> *
> *   For searches and help try:
> *
> *
> *

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2023 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index