-----Original Message-----
From: Ian White
To: res.guest23@bbsrc.ac.uk
Sent: 5/2/00 4:17 PM
Subject: bivariate analysis (asreml)
Somebody just asked me the following question.
"I would like to set up an ASREML to do a bivariate genetic analysis
which has a simple trait on the one hand (eg progesterone measured
once) and a trait like condition score on the other.
Condition score is measured several times on some cows and only once on
others. However each sire probably has information on the rate of
decrease
of condition score.
Can you show us how to set up the as file to do the anaysis to get, say,
a
genetic correlation between the rate of decline and progesterone?"
I haven't the faintest idea how to do this. Any suggestions?
************************************************
* I.White *
* ICAPB, University of Edinburgh *
* Ashworth Laboratories, West Mains Road *
* Edinburgh EH9 3JT *
* Fax: 0131 667 3210 Tel: 0131 650 5490 *
* E-mail: i.m.s.white@ed.ac.uk *
************************************************
Dear Ian,
The following might give you some ideas.
The random model for CS probably include components for SIRE animal an
measure=residual
The random model for PG only has SIRE and animal=residual
So, to set up this model in ASREML, we cannot directly use the bivariate
feature
[we probably could use multivariate if there was a fixed number of
repeat observations but not in the general case]
So we must manually expand our data file so we can use the sections feature.
Say there are 100 sires, 1000 animals and 5000 CS measures
The PG data file might look like
ANIMAL SIRE TREAT Trait Measure PG
where ANIMAL SIRE TREAT and PG have obvious meanings, Trait is fixed at 1
and Measure at Zero.
The CS data file is
ANIMAL SIRE TREAT Trait Measure CS
where Trait is 2 and Measure codes for the repeat observation.
The data files are appended
The ASREML job might then be
Repeat obs bivariate
ANIMAL 1000
SIRE 100
TREAT !A
Trait 2
Measure
response
PGCS.dat
response ~ Trait Tr.TREAT !r Tr.SIRE ANIMAL at(Tr,2).ANIMAL !GU
2 1 1
1000 !S2=10
5000 !s2=1
Tr.SIRE 2
Tr 0 US 1 .1 .1
SIRE
Notice that I have !GU declared for the CS animal component.
If the CS variances are smaller than thye PG variances (as shown here)
this componet is likely to be negative (if the ANIMAL componrnt is +ve)
If the reverse is true, the PG error variance might be negative which will
be a problem.
You would then need to increase the scale of PG to get around this problem.
This analysis gives us SIRE variances and covariance for the two traits.
The error=animal variance for PG is given by PG residual + ANIMAL variance
The animal variance for CS is ANIMAL variance + at(Tr,2).ANIMAL variance.
The ANIMAL level covariance is given by the ANIMAL variance
The within animal variance for CS is the CS residual.
Obviously you can expand this for other aspects of the model.
But this should get you started.
Arthur
The error variance
There may be a difficulty fitting this because
The
--
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml