Dear all,
This is a small "update" to the problem I posted before.
Using a different ordered categorical dependent variable, Stata 8.2 doesn't
tell me that the "outcome does not vary" this time. However, the output is
as mysterious as before. Besides, -search r(2000)- is not particularly
instructive.
. rfprobit y2 x1 x2, i(country)
Constant-only model
rho = 0.0 Log Likelihood = -115.33963
rho = 0.1 Log Likelihood ~ -71.47475
rho = 0.2 Log Likelihood ~ -53.679395
rho = 0.3 Log Likelihood ~ -43.642044
rho = 0.4 Log Likelihood ~ -37.060202
rho = 0.5 Log Likelihood ~ -31.396172
rho = 0.6 Log Likelihood ~ -25.898749
rho = 0.7 Log Likelihood ~ -21.31849
rho = 0.8 Log Likelihood ~ -18.051591
rho = 0.9 Log Likelihood ~ -15.209674
Iteration 0: Log Likelihood = -15.209674
(nonconcave function encountered)
rho >= 1, set to rho = 0.99
rho >= 1, set to rho = 0.99
rho >= 1, set to rho = 0.99
Iteration 1: Log Likelihood = -10.397214
(nonconcave function encountered)
no observations
r(2000);
Does anyone have an idea?
Yours,
Tobias
PS: If this question (or the last one) is utterly stupid, please tell me!
;-]
-----Original Message-----
From: Tobias Hofmann
Sent: Monday, September 06, 2004 4:53 PM
To: '[email protected]'
Subject: RE: st: Fixed-effects and the ordered logit/probit model
Dear Statalisters,
Having read Marc's quick response to Eric's question, I just wanted to give
those two commands (-reoprob- and -gllamm-) a try. However, the
(non-existent) results were rather disappointing. Running -reoprob- Stata
8.2 told me "outcome does not vary" and running -gllamm- Stata has not even
bother to tell me anything within the last 25 minutes. What could be the
reasons for these strange results?
Thanks for your help,
Tobias
PS: The model consists of one dependent variable with 3 different values (1,
2 and 3), two continuous independent variables, 15 countries and 20 years.
-----Original Message-----
From: [email protected]
Sent: Monday, September 06, 2004 3:33 PM
To: [email protected]; Eric Neumayer
Subject: Re: st: Fixed-effects and the ordered logit/probit model
Eric,
Quoting Eric Neumayer <[email protected]>:
> Hi,
>
> the statistic for computing an ordered logit/probit model with fixed
> effects is extremely complex, which is why Stata has no routine such
> as xtoprobit or xtologit, right? So, how would one interpret the
> results if someone were to introduce country dummies in an oprobit
> model to estimate the following?
>
> xi: oprobit y x i.country
>
This came up on Statalist about a year ago, and generated a thread which
included a very informative response from Bill Gould, Vince Wiggins and
David Drukker:
http://www.stata.com/statalist/archive/2003-09/msg00103.html
The bottom line is that the estimator isn't consistent, and if you use it
you'll probably get "awful" results. As others in the thread pointed out,
the most appealing alternative is probably random effects ordered probit
using -reoprob- or -gllamm-.
Cheers,
Mark
*
* 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/
