I sent a similar question about a week ago. There was one response. Both
are copied below; my question is interspersed with Scott Merryman's
answer.
Parenthetically, I am still puzzled why stata would change the
widely-understood meaning of fixed effect; is there some statistical
reason why fixed effects as normally understood would be a problem for
nbreg with a varying dispersion parameter? If anyone can answer that, I'd
appreciate it. However, anyway, hope the info below helps.
Sam
Date: Tue, 7 Jun 2005 18:16:34 -0500
From: Scott Merryman <[email protected]>
Reply-To: [email protected]
To: [email protected]
Subject: st: RE: "Regular" fixed effects and nbreg, in stata8.2?
> -----Original Message-----
> From: [email protected] [mailto:owner-
> [email protected]] On Behalf Of SamL
> Sent: Tuesday, June 07, 2005 9:24 AM
> To: Stata Listserve
> Cc: [email protected]
> Subject: st: "Regular" fixed effects and nbreg, in stata8.2?
>
> The manual states that for the command -xtnbreg- with the -fe- option, the
> term "fixed effects" applies "to the distribution of the dispersion
> parameter, and not to the xB term in the model." I am understanding this
> to mean that there is no fixed effect--in the usual sense of a beta
> coefficient for the unit (i) to which the observation belongs. I have two
> questions.
>
> 1)Is my understanding correct?
Yes
> 2)If my understanding is correct, may one obtain a fixed effects nbreg
> model that *does* have a fixed effect, in the usual xB sense, for the unit
> to which each observation belongs and, if so, how? (I don't need to see
> the fixed effect--if it drops out of the conditional model, that's fine--I
> just need for the usual fixed effect to drop out, not the dispersion
> parameter. For my analysis an equal dispersion parameter across units
> would be fine.).
You could use -poisson- with dummy variables. Unlike other nonlinear
estimators, it does not suffer from an incidental parameters problem. See
Cameron and Trivedi 1998, "Regression Analysis of Count Data" p. 280-282.
If you a large number of observations per cross section (enough that you
would be willing to estimate each cross section separately) you may be able
to get away with the use dummy variables (the inconsistency disappears as
T-> infinity). On a similar topic, you may find the post by Bill Gould,
Vince Wiggins, and David Drukker helpful:
http://www.stata.com/statalist/archive/2003-09/msg00103.html
On Thu, 16 Jun 2005, Khan, Nasreen wrote:
> Hi everyone,
>
> I need some expertise.
>
>
> I have an unbalanced panel data with around 68,000 observation and 23,000
> individuals (or groups). The dependent variable is in counts. And i am
> controlling for time invarient individual heterogenity by introducing
> fixed effect.
>
> Ideally, i would like to use poisson with robust standard errors (there is
> overdispersion in the DV) as i find use of dispersion parameter in neg
> binomial somewhat arbitary. However, stata 8 SE does not allow the robust
> option with xtpoisson. So i am restricted to negative binomial. However
> in the nbreg model variables (like race and sex) do not get absorbed in
> the fixed effect (they do so in xtpoisson). My questions are
>
> 1. Why doesnt the time invarient variables gets absorb in fixed effect in
> xtnbreg?
> 2. If the fe is for dispersion parameter, is there a way to introduce the
> individual level fe in nbreg ?
> 3. Is there any program available to adjust the SE in fixed effect xtpoisson?
>
>
> Thank you
>
> Nancy, PhD. Candidate
> University of Illinois at Chicago
>
>
>
>
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
