RE: st: different standard errors with gllamm vs. xtmelogit

["de Vries, Robert" <r.de-vries08@imperial.ac.uk>]

The log likelihood's with 7 integration points are:

Xtmelogit: -15047.008
Gllamm:  -15056.655

I tried increasing the number of integration points to 15 as you suggested. This yielded yet more strange results.

The gllamm and xtmelogit likelihood's are more similar for these models, but with gllamm's still being slightly higher (15047.916 vs. 15059.682)

The standard errors of the coefficient I'm interested in are still wildly different between gllamm and xtmelogit, but now so is the coefficient itself:

Xtmelogit: -.034(.039663)
Gllamm: -.0045(.004751)

Most of the other coefficients and standard errors are almost identical between the two models so initially I thought it might just be a problem with the level 2 variable I've been discussing above. However the coefficients and SE's of the individual level binary gender variable are also very different between gllamm and xtmelogit.

I'm stumped...



-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Stas Kolenikov
Sent: 26 October 2010 16:56
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: different standard errors with gllamm vs. xtmelogit

Are the likelihoods the same? Sometimes I find small differences in
the likelihood, too, which is indicative of lack of accuracy (in at
least one program). The first thing I would do would be to increase
the number of integration points (in both models) to say 15 and see if
the results change. In most -gllamm- examples from Sophia, the number
of integration points is even, although I don't know if there any
particular reason for that.

On Tue, Oct 26, 2010 at 8:11 AM, de Vries, Robert
<r.de-vries08@imperial.ac.uk> wrote:
> Hello everyone. I'm having a weird problem with gllamm and xtmelogit when r= unning a fairly simple 2-level random intercepts model.
>
> The model is predicting a binary health outcome from 1 level 2 variable (me=
> angini) and several level 1 variables. It is a sample if 5,4410 people in 1=
> 6 countries, with 'country' as the level 2 cluster.
>
> The xtmelogit model looks like this:
>
> xi: xtmelogit poorhealth meangini age47 gndr i.education if poorhealth_samp=
> le=3D=3D1 || country1 :
>
> (note that the number of integration points is left at the default of 7)
>
> It converges fine on iteration 3 with a log likelihood if -15046.989. The r= esult I am interested in is for meangini and in this model it is -0.030 (SE=  =3D 0.035)
>
> The gllamm model is identical (as far as I can tell)
>
> xi: gllamm poorhealth meangini age47 gndr i.education if poorhealth_sample= =3D=3D1, i(country1) nip(7) link(logit) f(binomial)
>
> However the coefficient for the same variable is different (-0.041). And th= e Standard Error (0.0043) is over 8 times smaller.
>
> This is obviously extremely important in interpreting the statistical signi= ficance of the results so I'd appreciate any help anyone might be able to o= ffer as to what's going on.
>
> Cheers
> Rob
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>



-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/