Dear ASREML'ers
I have some experience in using ASREML GLMM for binomial data
sets. In one set of ASREML analysis on dairy cattle data on diseases
(>70000 records, incidence range 2-15%), I compared genetic
parameters of 6 binomial disease traits from univariate-threshold animal
model, threshold 'sire' model (both are GLMM with probit link) and linear
animal model (GLMM without any link functions). Results were different
for animal versus sire (threshold) models. I kept linear model as a base
to which threshold model estimates were compared (after some
transformation to 0/1 scale). Only results from sire model seemed
"plausible" and animal model estimates went wild especially at extreme
incidences.
We concluded that sire models are more appropriate to use
because
i) threshold models are based on Taylor series expansion and/or
ii) the animal model works at individual level where as sire model works
in progeny groups, where the later benefit from central limit theorem that
has much impact in terms of progeny group size and incidence.
iii) the fact that threshold models are still sensitive to incidences (as
Arthur also noted), suggests that threshold models are not independant
of the mean which is bit worrying.
Obviously these problems are global and so we would expect this
to happen with any type of software designed to do the job.
I think most of this is in line with those of Arthur and Luc.
Regards
Haja
--------------------------------------------------
Dr. Haja Kadarmideen
Geneticist
Animal Breeding and Genetics Department
Animal Biology Division
Scottish Agricultural College (Edinburgh)
Bush Estates
Penicuik, Midlothian EH26 0PH
Scotland, UK
Telephone: +44 0131 535 3246
Fax : +44 0131 535 3121
e-mail : h.kadarmideen@ed.sac.ac.uk
--------------------------------------------------
--
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml