[Limdep Nlogit List] panel random effects with cluster variance adjustment

[William Greene]

Peter.
The random effects logit model does not provide for an additional layer
of "clustering," so I'm afraid the answer is no.  On the other hand, this
is one of those cases, where the validity of the second level correction
is a little dubious.  If there is a connection across the observations at
this higher level of aggregation, it is operating through a common omitted
"random effect." (If it is a "fixed effect" all bets are off, and your
estimator is inconsistent to begin with.)  Given that it is a random effect,
in a logit model, the force of not including it in the model to begin with
means that your estimates are inconsistent, by a scale factor that equals
1/sqr(1 - rho**2), where rho is the correlation across observations at the
higher level.  You'll find this discussion in, e.g., Wooldridge's panel
data book.  This is precisely the case that is not getting enough 
attention in the literature, as analysts almost reflexively apply the 
"clustering" correction whenever data seem to be aggregated at a higher
level than is accounted for in the estimator used.  Now, the better news.
You will find the estimator that accounts for random effects at more than 
one level of aggregation in the NLOGIT/LIMDEP manual in the chapter on
panel data, 17, Section 17.9, "Multilevel and Multiple Effects Random 
Parameter Models."  The binary logit model is one of the models supported
by that procedure, and yours is a template application.  Your random effects 
model is an RPM in which only the constant term is random.
Regards,
Bill Greene



----- Original Message -----
From: "Peter Glick" <pjg4 at cornell.edu>
To: limdep at limdep.itls.usyd.edu.au
Sent: Thursday, April 3, 2008 9:18:55 AM (GMT-0500) America/Bogota
Subject: [Limdep Nlogit List] panel random effects with cluster variance adjustment

Bill or subscribers,

Will Nlogit allow me to estimate a panel random effects logit and also do 
the variance correction for clustering?  The panel consists of multiple 
observations of individuals over time, with the random effects used to 
model the within-person error correlation over time.  The clustering 
adjustment would deal with the survey structure, i.e., the fact that people 
are grouped in sample clusters, i.e., villages (I have a number of 
village-level covariates).

I have not tried to estimate this model yet, but given the formidable data 
setup involved, I would just like to make sure ultimately I will be able to 
do what I would like.

Many thanks,

Peter Glick
Cornell University

   

_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au
http://limdep.itls.usyd.edu.au


-- 
Professor William Greene
Department of Economics
Stern School of Business
New York University
44 West 4th St., Rm. 7-78
New York, NY   10012
http://www.stern.nyu.edu/~wgreene
212.998.0876