# Re: Model selection in ASReml

From: Arthur <asremlforum_at_VSNI.CO.UK>
Date: Thu, 12 Mar 2009 05:11:30 +0000

Dear Donobi,

Your first model is the standard multivariate model but for clarity I recommend changing the last line so it reads
wt1 wt2~ Trait Trait.group !r Trait.animal
1 2 1
0 0 ID
Trait 0 US 0.1 0.3 0.5 !GP
Trait.animal 2
Trait 0 US 0.01 0.03 0.06 !GP
animal 0 AINV

if animal has a pedigree associated with it.

Now, the second model is not properly specified.
For 2 traits, I presume you want to test that the error variances are equal but the genetic variances are different. I cannot think of a situation where that would be a sensible model, but the easiest way to specify it would be
1 2 1
0 0 ID
Trait 0 US 0.1 0.2 0.5 !GP !=101

Such a model could easily fail with non positive definite R structure if there is actually a fivefold difference in variance.
Note that the initial values specify a matrix with correlation 3/sqrt(5) = 1.34 !!

If you were thinking that you wanted autoregressive correlattion model with different variances, the specification would need to be

!ASUV
wt1 wt2~ Trait Trait.group mv !r Trait.animal
1 2 1
0 0 ID
Trait 0 AR1H 0.9 0.1 0.5 !GP
Trait.animal 2
Trait 0 US 0.01 0.02 0.06 !GP
animal 0 AINV

but this is an equivalent model.

That is, with just two times, there is just a single correlation parameter.

However, if there were more times, you could consider AR1H,
but I would be inclined to try ANTE1 , if you didn't want US,
and probably put the restriction on the genetic, not the residual part.

i.e.
wt1 wt2 wt3 ~ Trait Trait.group !r Trait.animal
1 2 1
0 0 ID
Trait 0 US !GP
0 0 0 0 0 0 # Let ASReml work out the stating values
Trait.animal 2
Trait 0 ANTE1 !GP
0 0 0 0 0 0 # If this doesn't work, copy in a tenth of the values printed in the .asr file
animal 0 AINV

I trust that helps

------------------------
Arthur Gilmour

Principal Research Scientist (Biometrics)
NSW Dept Primary Industries

-------------------- m2f --------------------

Sent using Mail2Forum (http://www.mail2forum.com).