[Limdep Nlogit List] Model opinion

[Wells, Aaron]

Professor Greene,

Thank you sincerely for your assistance.  I computed the LRT for the
Tobit vs. Cragg model specification, with the result showing the latter
to provide a better fit to the data.

Best regards,

Aaron

-----Original Message-----
From: limdep-bounces at limdep.itls.usyd.edu.au
[mailto:limdep-bounces at limdep.itls.usyd.edu.au] On Behalf Of William
Greene
Sent: Wednesday, February 20, 2008 10:52 AM
To: Limdep and Nlogit Mailing List
Subject: Re: [Limdep Nlogit List] Model opinion

Aaron. I think you are jumping too quickly to the notion of 'selection'
bias.  If there is an explanation for why the observations are zero (or
not), you can model that separately using a binary choice model. If so,
then the other 94% would presumably be handled in a truncated regression
model. You will find this discussed in the early literature as Cragg's
model.  If the coefficients in the zero/not zero equation are the same
as in the truncated regression, then the familiar tobit model emerges.
This is a testable hypothesis, using a likelihood ratio test.  (It's
documented in my text.) /Bill Greene

----- Original Message -----
From: "Aaron Wells" <Aaron.Wells at healthways.com>
To: "Limdep and Nlogit Mailing List" <limdep at limdep.itls.usyd.edu.au>,
limdep-bounces at limdep.itls.usyd.edu.au
Sent: Wednesday, February 20, 2008 9:05:57 AM (GMT-0500)
America/New_York
Subject: [Limdep Nlogit List] Model opinion

Good morning all,

If possible, I would like to assess the scholastic opinion of those on
the listserv regarding the following issue:

I am modeling the change in medical expenditures (logged) between two
periods for 7296 individuals.  Of the 14592 observations represented by
these individuals, 6% are equal to zero; the remaining observations are
normally distributed.  To model medical expenditures over the two
periods, I have considered a random effects linear regression model, a
tobit model (limit=0), a truncated model (all expenditures > 0) and a
Heckman model.  Ideally, I would like a linear regression counterpart to
the hurdle or ZIP model for count data, though I am not sure what that
model is (the Heckman model seems to imply, from a theoretical
perspective, that there is a selection bias to be modeled, which I do
not believe is the case for whether an individual is observed as having
y = 0 or y > 0).  Does anyone have a thought on whether these options
sound legitimate, or, whether one is most appropriate?

Thank you sincerely,

Aaron


************************************************************************
******
This email contains confidential and proprietary information and is not
to be used or disclosed to anyone other than the named recipient of this
email, and is to be used only for the intended purpose of this
communication.
************************************************************************
******
_______________________________________________
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

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


******************************************************************************
This email contains confidential and proprietary information and is not to be used 
or disclosed to anyone other than the named recipient of this email, 
and is to be used only for the intended purpose of this communication.
******************************************************************************