[R] glmmPQL, log-likelihoods issue

[Prof Brian Ripley]

glmmPQL does not fit by maximum likelihood, and what is being quoted is 
not a likelihood for the original problem.

On Fri, 21 Nov 2003, Andrew Perrin wrote:

> Greetings-
> 
> a reviewer for a paper of mine noted an anomaly in some models I ran using
> glmmPQL (from the MASS package).  Specifically, the models are three-level
> hierarchical probit models estimated using PQL under R.  The anomaly is
> that the log-likelihoods decrease (or, alternatively -2logLik increases)
> as variables are added to the null model. This is unusual, and I'm trying
> to figure out how to interpret it.  I've found some indication (e.g., at
> http://www.ssicentral.com/hlm/hlm00150.htm) that PQL estimation doesn't
> produce meaningful log-likelihoods, but I'm suspicious of that claim. Any
> comments or advice would be helpful.

-- 
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