Vincenzo,
Your gllamm setup looks ok to me.
1) gllamm isn't better than xtlogit as such but it is much more flexible. If you want to do standard multilevel models with a normal random effect xtlogit is faster than gllamm. If you want e.g. random slopes or non-parametric random effects gllamm will do this (and xtlogit can't).
2) The level-1 variance is not estimated when you use the logit link for the outcome but is assumed to follow the logistic distribution, i.e., with variance phi2/3. You can use this estimate to calculate the intraclass correlation.
3) Depends on what you want. If you want to model the multilevel structure the multilevel model is the way ahead. If you're interested in the level 1 equation then clustered standard errors are an easy way to get around this problem.
Mads
-----Oprindelig meddelelse-----
Fra: [email protected] [mailto:[email protected]] På vegne af Vincenzo Carrieri
Sendt: 28. januar 2009 13:02
Til: [email protected]
Emne: st: some problems with gllamm
Dear statalisters,
I am trying to estimate a multilevel logit with this variables:
Dep var: hpoor (health conditions; 1 if bad)
Indep var: income quintiles (4 quintiles, highest quintile is the
reference category), age, age2, sex and finally a contextual variable
(named meanistr) measured at regional level (the id is reg) . My
purpose is to see if the socio-economic context has an influence on
the dep var. So I estimate the following, using gllamm command:
glamm hpoor quintpos1 quintpos2 quintpos3 quintpos4 age agesq age3
male meanistr, link (logit) fam (binom) nocons i (reg)
and I got the following:
number of level 1 units = 39777
number of level 2 units = 21
Condition Number = 1309484.3
gllamm model
log likelihood = -20815.479
------------------------------------------------------------------------------
hpoor | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
quintpos1 | .5610976 .0410717 13.66 0.000 .4805985 .6415968
quintpos2 | .4684868 .039587 11.83 0.000 .3908977 .5460759
quintpos3 | .4766715 .038078 12.52 0.000 .40204 .551303
quintpos4 | .2694107 .0372995 7.22 0.000 .196305 .3425163
age | .1018888 .0121446 8.39 0.000 .0780858 .1256917
agesq | -.000771 .0002509 -3.07 0.002 -.0012627 -.0002792
age3 | 6.21e-06 1.69e-06 3.67 0.000 2.90e-06 9.52e-06
male | -.0921198 .025821 -3.57 0.000 -.142728 -.0415115
meanistr | -1.977139 .08596 -23.00 0.000 -2.145617 -1.80866
------------------------------------------------------------------------------
Variances and covariances of random effects
------------------------------------------------------------------------------
***level 2 (reg)
var(1): .02703922 (.00502425)
------------------------------------------------------------------------------
My questions are:
1) is right gllamm command for my research purpose (asses the context
role)? or is it better an xtlogit?
2) why I get just 2-level variance? In this way I am not able to
mesaure the percentage of total variance explained by 2-level variance
(that is the role of context)
3) If I use already a variable mesaured at regional level to asses the
role of context, Would I need to use a multi-level model or I could
just use a cluster option to take off the correlation intra-region?
I appreciate any hint.
Many thanks in advance
Vincenzo
*
* 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/
