-----Original Message-----
From: Political Methodology Society [mailto:[log in to unmask]] On
Behalf Of Achim Kemmerling
Sent: Thursday, February 08, 2007 11:32 AM
To: [log in to unmask]
Subject: [POLMETH] Heckman
Dear Polmeth-members:
I have a problem with Heckman models. I have found several applications
of this model for comparative political economy purposes. In some of
these rho is close to or equal to 1. In the STATA reference manual I
have found that such a rho should lead you to reject the Heckman
approach. Since I am also writing on a paper using a Heckman model and I
am facing similar problems I would like to get some opinion on this. I
use some 120 European regions of which only 30-60 are observed on the
(final) dependent variable. My guess is that a rho equal to 1 - implying
that the selection and the regression process are equal - reveal two
problems: either it shows that sample size is not large enough for an
iterative estimation procedure, or that specification of the selection
process is poor.
Can anyone help?
Best
Achim
--
Dr. Achim Kemmerling
Research Fellow
Social Science Research Centre Berlin (WZB)
Reichpietschufer 50
10785 Berlin/ Germany
Tel.: +49-(0)30-25491-150
Fax.: +49-(0)30-25491-222
Email: [log in to unmask]
________________________________________________________
Here is a copy of an exchange on a similar matter a few years back
From H-PolMeth <[log in to unmask]>
Subject H-POLMETH: RE: Heckprob (stata) problems
Date posted 2003-07-30 07:46:18
Hi, all-
We've received several detailed responses to the recent posting on Heckman
probit models. They follow, with the original posting, for your
reference.
KL
---------- Forwarded message ----------
Date: Fri, 25 Jul 2003 19:35:57 -0400
From: Thomas Schmeling <[log in to unmask]>
Subject: Heckprob (stata) problems
Hi all,
I'm runing the heckprob procedure in stata. The selection equation
predicts the occurance of a court case and Equation 2 predicts the outcome
of the case (both binary dependent variables). There are 1542 observations
for 50 state courts over 42 years. The results look just fine, except for
the reported RHO.
The value of RHO given is 1, with std. err. of 3.16e-08. the 95%
confidence bounds are -1 and 1.
This is clearly a problem, but I'm having a hard time figuring out where it
comes from or what do do about it. I've queried the STATA list, but
haven't gotten a response that helps.
Any pointers?
Tom
--------------------------
Dept. of Political Science
Rhode Island College
600 Mount Pleasant Ave. Providence, RI 02908
Phone: (401) 456-8722
From: Patrick T. Brandt <[log in to unmask]>
Date: 28 Jul 2003 13:47:19 -0500
Subject: [Fwd: H-POLMETH: Heckprob (stata) problems]
Tom--
You may want to look at
Butler, J.S. 1996 "Estimating the Correlation in Censored Probit Models."
Review of Economics and Statistics 78(2): 356-358.
The problem you observe generally happens when there is not much
variation in the regressors across the 2 equations -- typical when most
of the regressors are dummy variables. The Butler piece presents a nice
discussion of the problem and solutions.
There is also a recent Political Analysis piece by Anne Sartori as well
that addresses the case of assuming the same regressors in each equation,
which is equivalent to rho=1.
Patrick Brandt
Tom--
--
Patrick T. Brandt
Assistant Professor
Department of Political Science
University of North Texas
[log in to unmask]
<a
href="http://www.psci.unt.edu/~brandt">http://www.psci.unt.edu/~brandt</a>
From: "Anne E. Sartori" <[log in to unmask]>
Date: Mon, 28 Jul 2003 14:56:32 -0400
Subject: Re: H-POLMETH: Heckprob (stata) problems
My guess is that you don't have enough information to use heckprob. You
might want to think about whether you can find a better exclusion
restriction or whether heckprob is the best choice of estimators.
On the difference between the SE of rho and the confidence interval: as
you probably know, STATA does not estimate rho directly; rather, it
estimates a function of rho, atanh rho (Atanh
rho=(1/2)ln(((1+rho)/(1-rho))).
STATA calculates the standard error of rho from the standard error of
atanh rho using an approximation (the delta method). However, it
calculates the confidence interval around rho without an approximation; it
calculates a confidence interval around atanh rho and transforms the
endpoints into values of rho using the formula for atanh rho. In the
limit, these methods give the same answer. However, in a finite sample,
the approximation may be far from the truth.
You might want to take a look at my Political Analysis article (11:2,
Spring of 2003). The article proposes an estimator for the special case in
which identical explanatory factors influence selection and the subsequent
outcome of interest. I also present some simulation results about what
happens with heckprob if one has a valid but weak exclusion restriction. I
use the estimator in a paper on enduring rivalries that's on my web site:
<a
href="http://www.princeton.edu/~asartori/">http://www.princeton.edu/~asartor
i/</a> . A STATA program that implements the
estimator is also there.
Hope this helps.
Anne
From: "Adam J. Berinsky" <[log in to unmask]>
Date: Mon, 28 Jul 2003 15:52:49 -0400
Subject: Re: H-POLMETH: Heckprob (stata) problems
Tom (and others).
Having run many heckprob runs in stata and LIMDEP what you have is a
problem that occurs more often than we would like. I don't have my manual
here, but if you look at the command for heckprob (and heckman), you'll
see that you don't actually estimate rho, but a non-linear transformation
of rho -- arctan of rho. stata estimates the SE of rho indirectly -- it
doesn't have a SE of its own. Usually the transformation gives you nice
results. But sometimes (as in your case) it doesn't. You should look at
the SE on the arctan of rho. From my experience my guess is that you have
a flat likelihood -- you might see this directly if you have very large
value on the arctan with an even larger SE -- STATA just can't pin down
any value for rho, so it's giving you nonsenical results -- a boundary
estiamte with an impossibly small SE. Also, take a look the LR test of
independent equations at the bottom of the output (assuming you didn't use
robust SEs). You will probably get a non-signific! ant result. So, don't
trust the SE on rho -- look at the other statisitics and the confidence
interval around rho.
Usually in my own runs I get results like those when the model has had
problems converging -- look for multiple iterations with a "nonconcave"
warning. When I do get results like yours I try to go back and see what
went wrong -- maybe your system of equations isn't identified.
So the bottom line is that I would look carefully at your output for other
warning signs (SE on the arctan of rho; LR test on independent equations
where you can't reject the null; convergence problems) and try another
specification. I have found these to be tricky models to actually
implement.
hope that helps
ab
Often w
Date: Mon, 28 Jul 2003 14:00:52 -0600
From: [log in to unmask]
To: H-NET/APSA List for Political Methodology <[log in to unmask]>,
H-PolMeth <[log in to unmask]>
Subject: Re: H-POLMETH: Heckprob (stata) problems
You probably have an identification problem. Make sure that you have some
exclusionary condition satisfied in those equations.
ab
Adam J. Berinsky
Associate Professor
Department of Political Science
Massachusetts Institute of Technology
77 Massachusetts Avenue
Cambridge, MA 02139 E53-459
Tel: (617) 253-8190
Fax: (617) 258-6164
E-mail: [log in to unmask]
Web Page: http://web.mit.edu/berinsky/www/
|
