Yes, the scaling is most likely to blame. Stata scales to the variance
of individual random effects, and LIMDEP may be scaling to the overall
variance of the residual term. What you can do is to see if the ratio
of the coefficients estimated in Stata to the overall variance of the
residual gives you the same estimates as the LIMDEP ones. Try
something like
nlcom [depvar]x/(1+exp([lnsig2u]_cons))
or whatever syntax modification of that would be required to make it
work :)). The "1" in denominator may be _pi^2/6, or whatever is the
variance of logistic distribution is.
I think Tom Mroz of UNC is so tired of all those scaling things that
he is writing a paper now where he compares a bunch of papers that
used different models with different software, got different
coefficients and engaged in a hot discussion as to who did this right,
and who's got wrong estimates. He says, "nobody did, it's just that
they are scaled differently". BTW, this is also an issue of the
difference between the logit and probit estimates.
On Tue, 21 Sep 2004 14:12:54 +0100, Alfonso Miranda Caso Luengo
<[email protected]> wrote:
> Dear All,
>
> Wiji has asked me to post the following message on her behalf (she's having some problems posting to statalist).
>
> Best wishes,
>
> Alfonso.
>
> ------------------------------------------------------
> Dear All,
> I hope someone can help me with this.
> I have tried running a simple random effects logit in stata, gllamm and limdep and I get different results from the first two compared to limdep. The maximised log like value is different and also the coefficients are different. Here is a summary:
>
> Nlogit Stata gllamm
> log L -23392.59 -23785.15 -23788.25
>
> intercept -0.9202 -1.2643 -1.25434
> (Std Error) (0.0203) (0.0285) (0.02812)
>
> beta 1 -0.35761 -0.52618 -0.5310
> (0.02872) (0.0402) (0.0403)
>
> rho 0.38615 0.38699
> (0.00503) (0.00844)
>
> This is a simple model with just one dummy variable with a coefficient beta 1.
>
> THE DIFFERENCE IN THE MAX LOG LIK
>
> I have been discussing this issue with Bill Greene and he has checked his programme and he thinks that one possibility for the difference in the log likelihood could be that stata is omitting a scaling factor . The difference in the log like between this model and another model with an extra variable is the same in both nlogit and stata. That is, the difference between nlogit and stata is a fixed factor. This factor could be the sqr(pi).
>
> the difference in the coeff
> Note, the estimated rho is very nearly the same. The coefficients and std errors are different.
>
> The command I use is
>
> xtlogit unidec male, i(anonsn) quad(16).
>
> The sample used is exactly the same and changing the number of quadrature points does not change the results significantly.
>
> Importantly, xtprobit results are consistent across the programmes. the problem is with the RE logit.
>
> Someone please help me. I have spent at least 5 days on this........ What is being reported in xtlogit?
>
> Many thanks
> best
> wiji
>
> =================================
> Professor Wiji Arulampalam,
> Department of Economics,
> University of Warwick,
> Coventry,
> CV4 7AL,
> UK.
> Tel: +44 (24) 7652 3471
> Sec. Tel: +44 (24) 7652 3202
> Fax: +44 (24) 7652 3032
> email: [email protected]
> http://www.warwick.ac.uk/go/arulampalam/
>
> ------------------------------------------------------
>
> =====================================
> Alfonso Miranda
> PhD Student
>
> Economics Department
> University of Warwick
> Coventry CV4 7AL
> E-mail:[email protected]
> WebPage: http://www.warwick.ac.uk/go/amiranda
> =====================================
>
> *
> * 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/
>
--
Stas Kolenikov
http://stas.kolenikov.name
*
* 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/
