[R-sig-ME] R-sig-mixed-models Digest, Vol 9, Issue 9

[Douglas Bates]

On 9/9/07, Iasonas Lamprianou <lamprianou at yahoo.com> wrote:
> Dear friends,
> first of all I would like to thank everybody, and especially Dr Bates, for all the support on using lmer. I now have another question. When using lmer on binary responses using family=binomial, we do not get the residual variance. In the default option (normal model) we do get the variance for every rndom effect, as well as the residual variance, so that we can compute the % of variance accounted by each effect. How can we find the residual variance in the binomial case, so that we compute the% of variance explained by each random effect? Thank you for your help once again

The binomial distribution is determined by the mean (i.e. the
probability of success on each trial) and the number of trials.
Unlike the normal distribution there is no scale parameter in addition
to the parameters determining the mean.  If the distribution of the
response in a mixed model, conditional on the value of the linear
predictor, is binomial then there is no separate calculation of the
residual variance and you cannot meaningfully talk about % of variance
accounted for by each effect.