The problem MAY be that the data is Poisson and not overdispersed. I ran a
test of Poisson simulated data, showing the fact that there is no extra
dispersion (that is why I used GLM rather than POISSON, which does not give you
many diagnostics). When I ran the same Poisson data using NBREG the same
problems that you mentioned appeared. The run can be found below.
Note that you cannot put 0 in as the negative binomial ancillary parameter
using GLM. To get close you can do something like fam(nb 0.001).
To check if your data is Poisson, run the data using the GLM command and
check the dispersion. If it is 0, or very close, it is Poisson and that is
likely the source of the problem. If it is greater than 0, then there is another
problem.
Joe Hilbe
. set obs 5000
obs was 0, now 5000
. gen x1=invnorm(uniform())
. gen x2=invnorm(uniform())
. gen xb= 1 + 0.25*x1 - 0.75*x2
. gen mu=exp(mu)
mu not found
r(111);
. gen mu=exp(xb)
. rndpoix mu
( Generating
.................................................................. )
Variable xp created.
. glm mu x1 x2, fam(poi) nolog
Generalized linear models No. of obs =
5000
Optimization : ML: Newton-Raphson Residual df =
4997
Scale parameter = 1
Deviance = 4.97544e-11 (1/df) Deviance =
9.96e-15
Pearson = 4.97669e-11 (1/df) Pearson =
9.96e-15
Variance function: V(u) = u [Poisson]
Link function : g(u) = ln(u) [Log]
Standard errors : OIM
Log likelihood = -7317.213353 AIC =
2.928085
BIC = -42560.41438
------------------------------------------------------------------------------
mu | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
x1 | .25 .0073479 34.02 0.000 .2355984 .2644016
x2 | -.75 .0071323 -105.16 0.000 -.763979 -.736021
_cons | 1 .0093382 107.09 0.000 .9816975 1.018302
------------------------------------------------------------------------------
. nbreg mu x1 x2, nolog
Negative binomial regression Number of obs =
5000
LR chi2(1) = 8688.40
Prob > chi2 = 0.0000
Log likelihood = -7317.2134 Pseudo R2 =
0.3725
------------------------------------------------------------------------------
mu | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
x1 | .25 .0073479 34.02 0.000 .2355984 .2644016
x2 | -.75 .0071323 -105.16 0.000 -.763979 -.736021
_cons | 1 .0093382 107.09 0.000 .9816975 1.018302
-------------+----------------------------------------------------------------
/lnalpha | -49.44545 . . .
-------------+----------------------------------------------------------------
alpha | 3.36e-22 . . .
------------------------------------------------------------------------------
Likelihood-ratio test of alpha=0: chibar2(01) = 0.00 Prob>=chibar2 =
1.000
=============================================
In a message dated 3/10/2005 1:40:51 PM US Mountain Standard Time,
[email protected] writes:
Hi, all,
Usually there is a log likelikhood ratio test statistics reported after a
negative binominal regression (I ran possion regression first, and then
used 'nbreg' regression), and sometimes the Z score for alpha is reported
too. But in my case, I didn't see the ratio reported after the
regression, and there is no z score for alpha. Then I tried to use
'lrtest' after the model, but it says
"lrtest not valid after robust
specify force option to perform test anyway".
So I put 'force' in option, but it still didn't give me the result.
Can anybody tell me what's going on?
Ying
*
* 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/
