I have just done a series of bivariate analyses for 171 trait combinations.
Incidentally, I was most impressed with how little 'fiddling' I had to do
in ASREML to achieve convergence for the majority of these trait
combinations - so I am getting more impressed by the day with ASREML :-).
Anyway, it does leave me with a general query as to what one should do when
a correlation (for example) goes just outside the parameter space? I prefer
to do all my analyses in the land of no constraint - and even many highly
correlated traits fell quite nicely into place within the parameter space,
so this was generally not an issue. However, there are just a couple of
trait combinations which would seem to benefit from using !GP or similar to
constrain the estimates, but I don't see much point. For example, each time
I have obtained an estimate >1 (eg. 1.003 or 1.01 - you can see they're not
that wild, and could even be mistaken for a rounding error at the lowest
level), the associated standard error would still indicate that the
correlation was no different to unity anyway. I think I should just present
them as is (against the common trend of not presenting estimates outside
the parameter space?). I was wondering of the prevailing attitudes of
others to this dilemma, or do you all tend to use constraints?? I hesitate
to use constraints myself because of the potential to bias other estimates.
So, I am looking forward to your ideas!
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