[R] residual variance in lmer

[Prof Brian Ripley]

On Fri, 4 Nov 2005, Douglas Bates wrote:

> On 11/4/05, Tom Van Dooren <vdooren at rulsfb.leidenuniv.nl> wrote:
>
>> when running lmer with family set to quasibinomial,
>> is the "residual" variance reported among the random effects, the
>> dispersion parameter as in glm, or something else?
>
> It should be the dispersion parameter.  I say "should be" because I am
> not that familiar with the quasibinomial family and I don't know
> exactly how the dispersion parameter is defined.  I can tell you that
> the 'residual' variance is evaluated as the ML estimate of the
> variance from the last weighted penalized least squares problem used
> in the iterative estimation.  Does that help?

I would say that is an estimate of the dispersion parameter, different 
from those used in glm() as indeed it should be as this is a GLM.

Estimation of the dispersion in a quasibinomial family is pretty tricky, 
even for a glm.  Be prepared for estimates to be biased.

-- 
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