I sent this to Bart, but perhaps should have sent it to the list (depends
if the info is useful or not I guess :-)). Any results I mention below
occur even where the incidence is 0.5 (which is the most informative you
can get for 0/1 data) and for very large data sets (>20000 records).
>Date: Mon, 17 Jul 2000 20:53:21 +1000
>From: Kim Bunter <firstname.lastname@example.org>
>Subject: Re: non-normality
>in my experience the animal model does not perform well for binomial
traits using a GLMM, whereas sire and/or dam models perform well. If you
look back through the discussion list, Arthur explained why this was the
case. This result is most unfortunate given that you may wish to try and
estimate additional random effects which are usually estimated in an
expansion of the animal model (eg. additive+maternal). In fact, I have
performed a simulation study which shows that not only does the animal
model not perform well under a binomial GLMM, but it can also lead to
spurious estimates for additional random effects (such as maternal
effects). So, beware! I doubt very much if this is just a problem with
ASREML, although I have not had the time to test my theory with other
software (presumably also implementations of Schall's method). The problems
may be less dramatic if you have more than two categories in your data, but
you cannot analyse categorical traits (other than binary responses) in
ASREML under a threshold model.
>However, the animal model does seem to perform well if you ignore the fact
that your data is not normal, and the incidence(s) is(are) not too extreme.
In this case, estimates under an animal model seem pretty robust even under
the incorrect random effects models (also from binary trait simulation
studies). Once you get towards extreme incidences, your estimate and
transformation to an underlying scale may be a bit wild. Regardless, you
will always wonder what you would have got had you used the technically
correct approach (ie one that is supposed to deal with categorical traits).
And it seems to me, that options are still somewhat limited when it comes
to finding software available to do the job.
>It is tempting to believe results from an animal model, especially when
the estimates seem 'about right', and we have probably spent half our lives
being told how superior this model is. It just does not yet seem to hold
under (this should have read: implementations of) threshold models? Have we
got anyone with experience using Gibbs sampling etc etc (or are there only
frequentists on this list :-)). It seems to me that the only way to be
reasonably sure that you've got it right though is to see how much your
answers change under different models and approaches.
>So, have fun!
>At 09:51 17/07/00 +0200, you wrote:
>>Continuing on Hermann's question about non-normal data, what is your
>>on/ experience with using an animal model in analysing non-normal data?
>>I have applied it succesfully (i.e. convergence to reasonable estimates)
>>data. I tried to get a confirmation of the results from applying a sire +
>>model on the same data, but the iteration did not converge. This is very
>>likely caused by small family sizes in this particular case.
>>I know that other users have bad experience with animal models, and they
>>sire+dam models instead.
>>But first of all, it is not quite clear to me, if animal models applied to
>>discrete data lead to proper estimates?
>>Animal Breeding & Genetics group
>>Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml
Kim Bunter (M.Rur.Sc)
Animal Genetics and Breeding Unit
University of New England
Armidale, NSW, 2351
Ph (ISD): -61-2-67733788
Fax (ISD): -61-2-67733266
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml