This isn't a random factor, as you only have one response (the
proportion) per mother. You may well get extra-binomial variation if the
mothers differ in ways not explained by the `traits', so if you do, look
up how to handle `over-dispersion' in GLMs.
On Mon, 20 Dec 2004, Henrik Parn wrote:
Dear S-users,
I am working on a bird where a female either mate with her social partner, or
with other males in the surrounding territories. The offspring in a nest thus
all have the same mother, but can be sired by one or more males. I divide the
offspring into two groups: within-pair offspring (WPO) or extra-pair
offspring (EPO). My data set is arranged with one female (nest) per row and
with data on: number of WPO, number of EPO and different female traits.
My object is to relate the proportion of extra-pair offspring in a nest to
different traits off the mother. I have defined an EP-offspring as success,
and WP-offspring as failure. I first planned to use an ordinary glm:
y<-cbind(number of EPO, number of WPO)
glm(y~mother.trait, family=binomial)
However, now I am a unsure whether I need to include 'mother ID' as a random
factor. I.e whether to treat the respons variable (EPO vs WPO per nest) as a
single measure per female. Or should I instead 'imagine' the response as
several measures per female: one measure (EPO or WPO) per offspring in her
nest, and all offspring in her nest having the same random factor 'mother
ID'? If I need a random factor, which glmm do you suggest?
--
Brian D. Ripley, ripley@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
|