Hi,
I realize that some of the things have been said or implied.
>
> For example, it is my understanding that when your variance heterogenity
> factor (deviance/DF) is approximately 1 (as occurs under and animal model)
> there is no need to allow for overdispersion because it doesn't exist when
> cluster size=1 (ie don't re-run using !disp).
Under the standard binomial model, this should be 1 or very close - no
extra-dispersion present. As a result, using the !disp option in Asreml should
not change anything (if anything, it should be numerical error and adjustment
for dispersion parameter). If not then you need to make appropriate
adjustments.
>Further, this unaccounted for additive variance
> can be picked up by additional random effects which have inherently high
> sampling correlations with the additive effects (eg maternal). ie resulting
> in spurious maternal effects for example.
What are the standard errors?
>It is also not possible to
> compare the Log-likelihoods of additive vs additive-maternal effect models
> (my simulations often show the incorrect direction of change under more
> parameters). Is this something to do with not being able to use deviance
> as an appropriate measure of fit when cluster size=1 (although cluster size
> presumably varies for the different effects: ie., = 1 for animal effects
> but is large for sire effects)?
Do you have sufficient information to estimate the parameters you require?
One of the manual examples shows this with maternal effects.
It would appear that the likelihood is rather flat and/or local maxima. But
without any results, one can only guess. There is potentially more
information for the sire model but also a potentially incorrect model.
The sire and animal models are not equivalent models - you have to account for
the correlations between individuals and the sire model ignores the
information from the dams.
> What if the implied dispersion is incorrect?
Generally there is more variation than expected or over-dispersion. The option
available in Asreml is to use the !disp option that changes the variance
structure. Hence standard errors will change, but the magnitude depends on
the degree of overdispersion. Failing that, you would have to directly model
the extra-dispersion such as via a beta-binomial (I think there was a genetics
paper on this a few years ago).
If it is under-dispersion, probably the binomial distribution is an incorrect
assumption.
>
> Also, regarding models where we want to estimate maternal effects (for
> example), what are the best options for 0/1 data?
There are many options but really depend on the dataset as it is the the
extremes where any differences will occur. It is best to check a variety of
options and ensure that the model is adequate. Commonly normal
distribution is used and is okay within reason - frequency between ~0.2 to
~0.8 and large number of observations -> normal approximation theory. Of course
you need to live with the fact that estimates and standard errors are often
outside the parameter space. Using the logistic or probit models should not
lead to any differences (e.g. Lopez-Villalobos and Garrick,1999 Proc. NZ Soc.
Anim. Prod. 59:121) - these link functions are very similar except for the
extremes frequencies.
Asreml uses Schall's method as noted in the manual and this was explained on
this list.
If you have any questions, just ask,
Regards Bruce
--
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml