Re: st: xtlogit vs gllamm

["Bryan W. Griffin" <bwgriffin@GeorgiaSouthern.edu>]

There are three reasons results from xtlogit and gllamm do not agree 
here for you.

First, your data apparently have a large random effect variance which 
creates problems, often, with random effect estimation for logistic 
regression. Or it may be simply that you have spare data (e.g., 
little information, small sample size, or few 0 or 1's for the 
outcome). The condition code from gllamm you include suggest and the 
large differences in parameter estimates suggest something odd with 
your data.  Most likey the issue is the large random effect -- note 
rho is about .78 (that is very large for logistic regression).

Second, xtlogit and gllamm use different implementations of adaptive 
quadrature. I have run simulations on this and reported to Stata that 
xtlogit's adaptive quadrature provides large biased estimates when 
large random effects are present and that gllamm estimates and 
estimates using adaptive quadrature from SAS are more accurate (and 
agree with each other). I hope that with a future update the 
implementation in xtlogit changes to agree more with gllamm and SAS. 
Note, however, that xtlogit with non-adaptive quadrature works very 
well, so try estimating your model with that approach.

Third, despite belief among some that only a few integration points 
are needed, when large variances are present in random effects, many 
integration points are required. You used only 12; up the number to 
150 for gllamm and xtlogit then check estimates. Rerun both with 
number of points at 165 or 170 for comparisons to see if number of 
points still produces differing results. Run xtlogit with both 
adaptive and nonadaptive, and do the same with gllamm and compare results.




> Date: Mon, 15 May 2006 14:49:33 -0400
> From: =?iso-8859-1?Q?B=E9gin_Karine?= <k.begin@umontreal.ca>
> Subject: st: xtlogit vs gllamm
>
> Hi,
>
> I have a dichotomous outcome variable taken at various time 
> points.  I performed a regression on this variable using both the 
> "gllamm" command with a logit link, a binomial distribution and the 
> adaptative quadrature (ado file) and the "xtlogit" command (running 
> with the same number of adaptative quadrature points as with 
> gllamm).  Both commands produce quiet different results and I am 
> wondering why since both gllamm and xtlogit seem to be conceived for 
> similar purpose.  An example of the results is listed below.
>
> Thanks,
> Karine
>
>
> . gllamm emploi sexe FAC3_, i(no_seque) link(logit) family(binom) 
> nip(12) adapt
> Condition Number = 8.8565296
> - 
> ------------------------------------------------------------------------------
>       emploi |      Coef.   Std. Err.      z    P>|z|     [95% 
> Conf. Interval]
> - 
> -------------+----------------------------------------------------------------
>         sexe 
> |  -2.772054   .3028192    -9.15   0.000    -3.365569    -2.17854
>        FAC3_ 
> |   1.382007   .1186971    11.64   0.000     1.149365    1.614649
>        _cons 
> |  -9.172404   .4245712   -21.60   0.000    -10.00455    -8.34026
> - 
> ------------------------------------------------------------------------------
>
> Variances and covariances of random effects
> - 
> ------------------------------------------------------------------------------
> ***level 2 (no_seque)
>     var(1): 227.29026 (21.232383)
> - 
> ------------------------------------------------------------------------------
>
>
> xtlogit
> . xtlogit emploi sexe FAC3_, i(no_seque)
> - 
> ------------------------------------------------------------------------------
>       emploi |      Coef.   Std. Err.      z    P>|z|     [95% 
> Conf. Interval]
> - 
> -------------+----------------------------------------------------------------
>         sexe 
> |  -.1323433   .1503605    -0.88   0.379    -.4270445    .1623579
>        FAC3_ 
> |   .2813436    .065331     4.31   0.000     .1532971    .4093901
>        _cons 
> |  -2.290372    .085298   -26.85   0.000    -2.457553   -2.123191
> - 
> -------------+----------------------------------------------------------------
>     /lnsig2u 
> |   2.481279   .0429478                      2.397103    2.565455
> - 
> -------------+----------------------------------------------------------------
>      sigma_u 
> |   3.457824    .074253                       3.31531    3.606463
>          rho 
> |   .7842202   .0072676                      .7696356    .7981235
> - 
> ------------------------------------------------------------------------------
> Likelihood-ratio test of rho=0: chibar2(01) =  7631.32 Prob >= 
> chibar2 = 0.000
___________________________________________________________________
Bryan W. Griffin
Curriculum, Foundations, & Reading
P.O. Box 8144
Georgia Southern University
Statesboro, GA 30460-8144

Phone: 912-681-0488
E-Mail: bwgriffin@GeorgiaSouthern.edu
WWW: http://coe.georgiasouthern.edu/foundations/bwgriffin/

*
*   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/