> Working with asreml using polynomial random regressions I have the same
problem than Dave Johnson: 0 covariance between the polynomial coefficients. If
I specify a US matrix with 0 covariances or a DIAG matrix the program works,
but trying to specify covariances different from 0 leads to non-convergence.
Does anyone know why this happen?
> Another related question: If I use just a DIAG structure and then check the
.sln file there is a problem with the identities of the solutions. Each
identity looks like a series of black squares followed by .001, .002, etc. So,
I know what solution corresponds to each term of the polynomial, but I can't
know to what individual.
> If it is of any help, I'm using asreml version April 16 1998 for win95.
> Thanks a lot,
> Luis &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Luis Apiolaza
Institute of Veterinary, Animal and Biomedical Sciences Massey University
Palmerston North New Zealand L.A.Apiolaza@massey.ac.nz
> "Quote me as saying I was misquoted " - Groucho Marx
This is not an uncommon problem. The AI algorithm can be unreliable for some
models. You need good starting values and the best way we have of doing this at
the moment is to build up the models. so for example, fit
1. diag mean,lin,quad say get the variance components. make sure all the
components are sequentially significant, as this can lead to problems with a
flat likelihood surface.
2. fit a model with only the mean,lin covariance as nonzero, this can be done
using !GZ option, ask arthur for details,
3. if ok, then include the final two covariance terms
i know this is tedious but.... We are in the process of writing up a set of
notes for the ASREML 'manual' on these issues for asreml and genstat
implementations of AI. these issues will be covered in detail, and lots of