[R] Multilevel logistic regression using lmer vs glmmPQL vs. gllamm in Stata

[Prof Brian Ripley]

On Wed, 3 Aug 2005, Bernd Weiss wrote:

> I am trying to replicate some multilevel models with binary outcomes
> using R's "lmer" and "glmmPQL" and Stata's gllmm, respectively.

That's not going to happen as they are not using the same criteria.

> The data can be found at <http://www.uni-koeln.de/~ahf34/xerop.dta>.
>
> The relevant Stata output can be found at  <http://www.uni-
> koeln.de/~ahf34/stataoutput.txt>. First, you will find the
> unconditional model, i.e. no level1- or 2-predictor variables. The
> second model contains some level 1-predictor variables
>
> My R file can be found at <http://www.uni-koeln.de/~ahf34/xerop.R>.
>
> Beside the fact that there is a difference between the estimates of
> the intercept (unconditional model: R: -2.76459 and Stata: -2.698923)
> I am especially interested in the level 2 variance.
>
> In Stata the level 2 variance is about 1.03, while in R it is  4.68.
>
> Using glmmPQL from package MASS again gives different results for the
> level 2 variance component. What is meant by "Residual"? I thought
> the level 1 variance is fixed to (pi^2)/3.

Please read the book for which this is support software, as it definitely 
does not say that, and it does explain how such differences can occur.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595