Re: AW: weighted bivariate analysis[Scanned.]

Dear Dorte

The answer to the original question is that for
Cov(e1,e2)=( sigma_e1^2*diag(w1)^{-1},
rho*sigma_e1*sigma_e2*diag(sqrt(w1*w2))^{-1) )
           ( rho*sigma_e1*sigma_e2*diag(sqrt(w1*w2))^{-1),
sigma_e2^2*diag(w2)^{-1} )

For example, if E was
  1 1
  1 4

and we have 2 pairs and a weight vector 1 4 16 9

the resulting variance structure is

1 1/2 0 0
1/2 4/4 0 0
0 0 1/16 1/12
0 0 1/12 4/9

However, It appears from a small test I'm running that it is not
producing the correct variance components. For example,
putting in a constant weight of 2 gave components that were 12/13 of
the original values when my expectation was that the components would
be doubled. I am still looking
at the issue but thought it best to advise of progress so far.
 
In your example, doubling the weights should result in doubling the
error variance structure to compensate. It seems I'm not accounting for
the weights
properly when calculating the SCORE.

    <><><><><><><><><><><><><><><><><><><><><><><><><>

Wittenburg, Dörte <wittenburg_at_FBN-DUMMERSTORF.DE>
Sent by: ASReml users discussion group <ASREML-L_at_AGRIC.NSW.GOV.AU>
22/08/2007 05:26 PM
Please respond to ASReml users discussion group

 
        To: ASREML-L_at_AGRIC.NSW.GOV.AU
        cc:
        Subject: AW: weighted bivariate analysis[Scanned.]

Hi list members, dear Arthur,

it seems that ASReml does not ignore the weights. There are essential
differences in (co-)variance components between a weighted analysis and a
non-weighted. You find attached the different results of ASReml. File
"..._w1.asr" with weights and "..._w2.asr" without weights. The run with
doubled weights also yielded other results.

Many thanks,
Dörte

-----Ursprüngliche Nachricht-----
Von: ASReml users discussion group [mailto:ASREML-L_at_AGRIC.NSW.GOV.AU] Im
Auftrag von arthur.gilmour_at_DPI.NSW.GOV.AU
Gesendet: Mittwoch, 22. August 2007 07:54
An: ASREML-L_at_AGRIC.NSW.GOV.AU
Betreff: Re: weighted bivariate analysis[Scanned.]

Dear Dorte,

I suspect ASReml is ignoring the weights. The first thing to check
is whether you get the same results if you either drop the !wt qualifier,
or rescale it (say double it).

ASReml could handle weights in direct product with the variance matrix
(i.e. same weight for both traits within a unit)
but as you indicate, it is not obvious how to incorprate individual
weights which differ between traits.

    <><><><><><><><><><><><><><><><><><><><><><><><><>

Wittenburg, Dörte <wittenburg_at_FBN-DUMMERSTORF.DE>
Sent by: ASReml users discussion group <ASREML-L_at_AGRIC.NSW.GOV.AU>
21/08/2007 06:52 PM
Please respond to ASReml users discussion group

        To: ASREML-L_at_AGRIC.NSW.GOV.AU
        cc:
        Subject: weighted bivariate analysis

Hi,

I am trying a weighted bivariate analysis with data that are prepared in a
univariate way. For the two traits I have generated different weights, w1
and w2. ASReml 2.0 works fine. Now my question is, how does the residual
covariance structure looks like when I define the R structure as "I times
US" ? If I use the notation units within traits (US times I), then I think
it is

Cov(e1,e2)=( sigma_e1^2*diag(w1){-1}, rho*sigma_e1*sigma_e2*I )
           ( rho*sigma_e1*sigma_e2*I, sigma_e2^2*diag(w2){-1} )

Am I right?

My .asr file:

Bivariate analysis
 mmutter !A !P
 dam !A !P
 gewicht
 log_var
 sau * !I
 vater !A
 paritaet 4
 saison * !I
 merkmal 2

sort_pedigree.ped !SKIP 1 !ALPHA !MAKE
data_univariat_w.asd !SKIP 1 !MAXIT 200 !DOPART $1 !ASMV 2
!TYPEIIISS

!PART 1
log_var !WT gewicht ~ merkmal merkmal.paritaet at(merkmal,1).saison,
 !r ![ merkmal.mmutter at(merkmal,2).dam !] at(merkmal,1).vater
 at(merkmal,1).sau !f mv

 1 2 1

 0 0 IDV
 merkmal 0 US
 1 -0.5 18

 merkmal.mmutter 2
 3 0 US 1 0.1 1 0.1 0.2 1 !GP
 0 0 AINV

Many thanks for response,
Dörte Wittenburg

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
FBN Dummerstorf
Genetik & Biometrie
Wilhelm-Stahl-Allee 2
18196 Dummerstorf, Germany
email wittenburg_at_fbn-dummerstorf.de
phone +4938208 68-931
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This message is intended for the addressee named and may contain
confidential information. If you are not the intended recipient or
received it in error, please delete the message and notify sender. Views
expressed are those of the individual sender and are not necessarily the
views of their organisation.

This message is intended for the addressee named and may contain
confidential information. If you are not the intended recipient or
received it in error, please delete the message and notify sender. Views
expressed are those of the individual sender and are not necessarily the
views of their organisation.