Will do. Thanks, Patrick.
JP
On Thu, Apr 9, 2009 at 3:54 PM, Patrick Lam <plam at fas.harvard.edu> wrote:
I don't think the GEE models are quite right for
this framework. The
"correlation" that GEE models deal with is correlations between
observations, kind of like with clustered standard errors. For your example
where you have some type of two stage model with selection issues, something
like a bivariate probit would be more fitting. You might also look at
Heckman selection models, which deal with selection issues. Actually, you
could probably build your own model that is very similar to the
Zero-Inflated Poisson that we looked at for this week's HW assignment. You
just have to model the two stages in some way.
On Thu, Apr 9, 2009 at 11:21 AM, John-Paul Ferguson <jpferg at mit.edu>wrote:
Hello all,
Sometimes you have an event whose outcome is only observed if another
event happens. Say for example that you only observe whether someone votes
Democratic if they decide to vote at all. You could think of this as a pair
of probits, where the party variable is P(D=1 | V=1).
In such cases, we often think that the first stage is endogenous to the
second stage; that is, you're less likely to go out and vote at all if you
think that your party is going to lose. This is correlated data, and there
are a set of techniques for dealing with it, like bivariate probits (which I
think I understand pretty well) and generalized estimating equations (which
I am reading up on).
Christopher Zorn from Emory has a review of this stuff in the 2001 *APSR*.
He mentions that GEE models in particular "offer a number of advantages for
researchers interested in modeling correlated data, including applicability
to continuous, dichotomous, polychotomous, ordinal and event-count response
variables."
My question: I have data where the first "stage" is best modeled using an
event-history framework. The second "stage" is a dichotomous response
variable. I don't see duration variables listed in Zorn's lsit of
applications, and so I'm wondering: is there any work on techniques for
correlated data where one or more stages are duration models?
Best,
JP
John-Paul Ferguson
PhD Candidate, Economic Sociology
MIT Sloan School of Management
50 Memorial Drive, E52-533
Cambridge, MA 02142
617.253.3940 (w)
617.549.8482 (c)
_______________________________________________
gov2001-l mailing list
gov2001-l at
lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
Patrick Lam
Department of Government and Institute for Quantitative Social Science,
Harvard University
http://www.people.fas.harvard.edu/~plam<http://www.people.fas.harvard.ed…
_______________________________________________
gov2001-l mailing list
gov2001-l at
lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l