Re: fitting random regressions
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: fitting random regressions




>I'm interested into analysing repeated measurements of a progeny test
>using a random regression model, with a econd order polynomial. The data
>structure in the data file is:
<snip>

It may not be appropriate to suggest this
on an ASREML discussion group, but you may
find it easier to approach this analysis using
the new version of DFREML, which includes
a program to do estimation of Covariance
Functions. Sorry Arthur.

While I am here, I have encountered a
couple of problems which someone else
may have suggestions about:-

(a) a multivariate run which will not
    remain PD, even with !GP specified.
    The traits are highly correlated, and
    eventually the program reports that
    it cannot form R inverse and gives up.

    If I use the cholesky decomposition,
    the problem persists. I suspect the
    decomposition fails.
 
    Does anybody have any experience of
    forcing these types of problems through?
    I could try fixing some parameters, but would
    like a (fairly) robust suggestion to
    work with as each run takes a long time
    even if it eventually crashes.

(b) I have data on two traits, one of which
    has repeated records. I can do a bivariate
    run estimating fixing the residual covariance
    at zero and fitting a code common to all records
    on the animal to estimate the residual covariance
    and part of the permanent environmental effect
    (thanks, Arthur). This run works fine.

    In univariate analyses of these traits, I can
    also fit a second random effect (Service Sire,
    the traits are number born alive in sows - parity
    1 and later parities) with either an identity
    or using the NRM. If I try to fit this effect
    in a bivariate model as above it causes a Fortran error
    (Floating invalid) after estimating the first
    likelihood. This happens both with an Identity
    or NRM associated with the service sire effect,
    and when I fix the between trait service sire
    covariance to zero (that being the parameter
    I haven't previously managed to estimate from
    either the uni or simple bivariate analyses). Again,
    does anyone have any suggestions?

    I am running the programs on an Alpha based
    machine running unix, with the ASREML programs
    compiled at fixed sizes (ie an executable
    corresponding to each of S1 to S9).

Ron Crump.