If I recall correctly, the tobit model itself (even without xt) is very
sensitive to non-normality. I thought it was in the manual but a quick
search doesn't show this. Amemiya wrote on this and I think this was
the source of the issue.
At any rate, bootstrapping this estimate would seem to be a good plan.
Tony
Peter A. Lachenbruch
Department of Public Health
Oregon State University
Corvallis, OR 97330
Phone: 541-737-3832
FAX: 541-737-4001
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Even Bergseng
Sent: Monday, April 14, 2008 1:32 PM
To: [email protected]
Subject: SV: st: RE: xttobit initial value not feasible
The RE Tobit is sensitive to extreme values in continuous independent
variables. One suggestion would be to check for extreme values and in
their presence try to estimate the model without these observations
where the extreme values occur or even winsorise these variables for
example to the 99th percentile. Also, RE Tobit is sensitive to the
number of quadrature points in the integral, especially as the group
size increases. Do a - quadchk - and possibly increase the number of
points (be aware that the computation time increases as it is
proportional to the number of points).
Regards,
Even
________________________________________
Fra: [email protected]
[[email protected]] på vegne av David Jacobs
[[email protected]]
Sendt: 14. april 2008 21:22
Til: [email protected]
Emne: Re: st: RE: xttobit initial value not feasible
I like Nick's (?) last advice. In my experience the random-effects
pooled time-series estimators are much more sensitive to data problems
than a purely cross-sectional estimator such as Tobit probably because
the default random-effects estimation method is based on a mvaghermite
approach.
If you get reasonable Tobit (rather than xttobit results, try clustering
on the panel identifier. Perhaps (?? a bit of a long shot) that
correction may even be sufficient for referees in your discipline.
The count estimators suggested below, of course, involve different
assumptions about the distribution of your dependent variable, but
xtpoisson or xt negative binomial can be estimated with an xtgee
approach, which again I find to be far more robust than the
random-effects count model alternatives to data problems. And if you
really have too many zeros (which I also doubt), you can estimate such
count models with a two equation zip or zinb approach and again correct
the panel standard errors by clustering on your panel identifier (Stata
as yet doesn't offer pooled time-series zip or zinb estimators, but
LImdep does).
But if you are an economist, do check with an econometrician first.
Dave Jacobs
At 02:45 PM 4/14/2008, you wrote:
>I doubt the large number of zeros is the ulitimate cause of your error
>message. There is probably something odd about your data. Can you
>run a pooled model with -tobit- instead of -xttobit-? You may also
>want to consider -poisson- and others in that family.
>
>On Mon, Apr 14, 2008 at 2:24 PM, sarah young
><[email protected]> wrote:
> > Nick,
> >
> > Thanks so much for your reply again. you are right that 2/3 firms
do
> > not export anything in my data, i.e. their export intensity is
zero.
> > export intensity is measued as the ratio of exports and total
sales.
> > Many existing literature on determinants of export intensity use
> > Tobit, I understand a reason is that they treat export intensity as
a
> > censored variable. I read somewhere that for a variable like this,
> > i.e. with over half of the observation as zero, Tobit model is more
> > appropriate than xtreg or xtabond2. Is that right? Thanks a lot.
> >
> > Sarah
> >
> >
> >
> > On Mon, Apr 14, 2008 at 6:48 PM, Nick Cox <[email protected]>
wrote:
> > > My point had I think nothing to do with censoring, but with the
geometry
> > > of what you are
> > > asking Stata to do.
> > >
> > > Please appreciate that I don't know anything about your data, nor
> > > I am an economist.
> > >
> > > If I see that 2/3 of firms have zero "export intensity", then
unless
> > > intensity
> > > has some special meaning, I infer from that that 2/3 firms do not
export
> > > anything -- at least
> > > in the data period.
> > >
> > > sarah young
> > >
> > >
> > > Thanks for the response. I am just starting to learn
econometrics. I
> > > thought that for a variable with over half of the observations
are
> > > zero, like export intensity in my model, the variable could be
treated
> > > as a censored variable. Or probably I am wrong?
> > >
> > > I have used xtreg and xtabond2 to estimate the model and they
performed
> > > well.
> > >
> > >
> > >
> > > On Mon, Apr 14, 2008 at 5:31 PM, Nick Cox <[email protected]>
wrote:
> > > > Consider a standard regression of your response (call it y) on
any two
> > > > predictors.
> > > > Then a simplified analogue of your problem is that two-thirds
of your
> > > > data lie on
> > > > the plane y = 0 and the other third lies on one side of that
plane.
> > > > Geometrically that sounds
> > > > a pretty awkward situation to model. Having more predictors and
using
> > > > tobit don't I think
> > > > make it any easier.
> > > >
> > > > There is no positive suggestion embedded here, just a thought
that the
> > > > message doesn't sound that
> > > > surprising in such a light.
> > > >
> > > > Nick
> > > > [email protected]
> > > >
> > > > sarah young
> > > >
> > > > I have been trying to run xttobit without success. The problem
is
> > > > "initial value not feasible".
> > > >
> > > > I have a panel of more than 4000 firms for 15 years, which
altoghther
> > > > is about 62,000 firm-year observations. The dependent variable
is
> > > > export intensity, two thirds of which are zeros. So I chose to
use
> > > > Tobit model. The model is on the determinants of export
intensity,
> > > > theoretically the choosing of the independent variables is
fine, which
> > > > include TFP, age, size, R&D intensity etc, as well as year and
> > > > industry dummies. I have tried by using different sets of
independent
> > > > variables, for example, including only one variable on the
right side,
> > > > excluding year and industry dummies, all of these not working
and the
> > > > message was the same, "initial value not feasible."
> > > >
> > > > I have searched on Statalist archive and found this thread:
> > > > http://www.stata.com/statalist/archive/2007-05/msg01036.html
> > > >
> > > > I tried to solve the problem according to the suggestions in
the
> > > > thread but still it is not working. Could anyone give some
advice?
> > > > What are the possible problems behind this? How to solve them?
>*
>* 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/
*
* 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/
