Re: RE SE's for boundary fixed parameters
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: RE SE's for boundary fixed parameters

Thank you Arthur, Julius and Bruce for your replies!

With regards to Arthur's and Bruce's comments: there is indeed a large (
around -.8) sampling correlation between additive and pe effects and the
log-likelihoods barely change through adding the animal effect for the
'boundary' traits. I had already checked this and had no doubt exactly what
was causing the additive to move towards the boundary. My query was just
really can you state that boundary=0 estimate and whether SE's are at least
as good as any other we may have got had the estimate been close to the
boundary (rather than fixed on it)?

Unfortunately, Bruce's comments that this model (a+pe) is not appropriate
is also correct in the sense that we can properly test whether it is better
than fitting a or pe alone (which I had already done). However, ostrich
data will never be well structured (at least in the short term) and an
estimate of additive effects with a large SE is better than none at all.
Think of all those years gone by where parameters were published with large
SEs (and probably worse approximations of SEs than that provided by
ASREML). With a new industry needing information (and a thesis partially
riding on this type of data) I can not afford to be so picky with my
models. It is simply the interpretation of results which must be made clear.

I also know there are large pe effects from the results so far and it is
not possible to ignore these by fitting animal as the only random term (ie
BIG bias problem) as Bruce suggested. There are too few animals to dump all
the repeated records (this would result in using only 28% of the available
records). Ostrich data is also unsuitable for models often used for poultry
because the mating structure is not hierachical (ie PAIRS). A moderate
proportion of sires and dams are fully confounded so both estimates will be
inflated by maternal effects. The confounding I have already tried to deal
with by systematically fitting a number of different models to highlight
which ones are the most appropriate.

With regards to Arthurs comment 'It could be that there is another major
source of variation in the data which has been ignored, or that there is an
interaction of the genetic effect with time.' I have tested all factors I
have information on, but you may well be right that there is still some
source of variation which I do not have info on. However, could you give me
an example of what you mean by genetic*time interaction? I am not really
clear on this one.

Anyway, thanks for all the input so far. The moral of this story is 'never
do a PhD with data from a new industy' OR 'blood WILL come from a stone if
you beat it hard enough, but it won't have the properties of real blood
when you finally get it' :-)



At 08:20  24/02/00 -0600, Bruce Southey wrote:
>Following up Arthur's comments, what is the log likelihood with and
>without the additive genetic variance?  The difference is probably very
>small.  Note that the difference is not distributed Chi-squared but
>probably a mixture of Chi-squared.
>With a variance component on the boundary condition, I doubt that SE are
>valid.  But, I don't remember my math stats about this and relates to
>the conditions required.
>Given your comments, it does not sound that you should be fitting this
>model.  Basically, you don't have sufficient information to make any
>conclusion about additive genetic variance.  What would be informative
>is to profile the likelihood with respect to the additive and permanent
>variance components.  I think that you would see that this is rather
>flat with respect to the additive genetic variance.  If so, then you are
>best to just to fit animal as a random effect.  This causes a bias and,
>if I recall correctly, John James gave a paper about ignoring repeated
>records in an animal model at one of the AAABG meeting around 1990.
>If you have a suitable pedigree structure, you could try a
>parent-offspring regression or sire-dam model etc.  
>Best of luck,
>Bruce Southey
>Asreml mailinglist archive:
Kim Bunter (M.Rur.Sc)
PhD Student
Animal Genetics and Breeding Unit
University of New England
Armidale, NSW, 2351

Ph:  (02) 6773 3788
Fax: (02) 6773 3266
Asreml mailinglist archive: