Antwort: Re: Wilmik and random regression

Dear Nicola,

if I understood your email correctly, you would like to try various
mathematical functions for modelling random genetic effects. One of
the functions is Wilmink function. Based on my experience with
functions on modelling (co)variances of test day data in German
Holstein population, I would strongly discourage you to use Wilmink
function or functions containing exponential or log terms to model
random effects in ROUTINE APPLICATIONS. Of course, it is important
to do research on such topic to see how different functions behave
in modelling (co)variances.

Back in 2001 I estimated parameters of a test day model using
(co)variance function approach. First, geentic and permanent
environmental (co)variances of lactation stages were estimated, and
then various mathematic functions were fitted to the (co)variance
matrices in the second step. I compared derived genetic correlations
from Legendre polynomials in various orders, Wilmink function, and
mixed log function to the original genetic corrections estimated
from the first step. Polynomials with 3 or more parameters could
restore well the original (co)variance structure, but Wilmink
function and mixed log function gave very strange genetic
correlation estimates, particularly between extreme days in milk.
The correlations even went to negative range. Also the heritability
curves derived from Wilmink and mixed log functions just did not
make sense, i.e. U-shaped curve.

I think that the exponential or log terms in the functions were
responsible for the strange estimates. I would be very interested in
learning what you will find in your study with Wilmink function.

Best regards,

Zengting Liu
=================================================
 Dr. Zengting Liu, Geneticist
 VIT, Heideweg 1, D-27283 Verden/Aller, Germany
 Phone: 0049-4231-955178, Fax: 0049-4231-955166
 Email: Zengting.Liu@vit.de, http://www.vit.de
=================================================

                    Nicola Macciotta
                    <macciott_at_UNISS.I An: ASREML-L_at_AGRIC.NSW.GOV.AU
                    T> Kopie:
                    Gesendet von: Thema: Re: Wilmik and random regression
                    ASReml users
                    discussion group
                    <ASREML-L_at_AGRIC.N
                    SW.GOV.AU>

                    10.10.2005 17:13
                    Bitte antworten
                    an ASReml users
                    discussion group

Bruce,
thank you for your answer but probably I have not been very much
clear in
my question. The RR model I am trying to fit is

Milk (daily yield) = HTD AGE DIM ANIMAL

where

HTD = effect of a specific combination of herd-test date
AGE = covariable represented by the age at calving
DIM = fixed effect of the stage of lactation (10 intervals of 30
days each)
to account for the average lactation curve

is the fixed part of the model

for the random part ANIMAL, I would like to model individual
deviations
using different functions in order to test the effects of the
function used
on the estimated patterns of variances along the lactation. Among
the
function, apart from those that are currently used as Legendre
Polynomials,
I would like to use some common linear (or linearisable) functions
of
lactation curve such as the Wilmink model (that was used in one of
the
earlier version of the Canadian RRM). I will use it in the linear
form
because I am setting the k parameter to a fixed value (that I have
previously estimated by fitting the non-linear form). I am
previously
interested in estimating the (co)variance matrix of random
coefficients.
But my problem is how to enter the Wilmink function as random term
in ASREML.

regards

Nicolò

probablyAt 15.03 10/10/2005, you wrote:
>Hi,
>Wilmik's curve is not random regression but rather a nonlinear
function like
>other functions such as Wood's. Apart from the nonlinear
generalized linear
>models, these models can not be fitted as such in ASREML. In the
case of
>Wilmik
>if the c parameter is treated as known then it becomes a simple
linear
>model and
>can be fitted. However, this is usually rather stupid because
biases the
>results. Also if you knew this parameter then there is no need to
fit the
>model.
>
>Regards
>Bruce
>
>---- Original message ----
> >Date: Thu, 6 Oct 2005 12:54:23 +0200
> >From: Nicola Macciotta <macciott_at_UNISS.IT>
> >To: ASREML-L_at_AGRIC.NSW.GOV.AU
> >
> >Dear ASREML users,
> >I am trying to run a random regression model for milk yield where
I would
> >like to test for modeling individual random deviations wioh
different
> >functions, some of them are not implemented in the ASREML
software. For
> >example, the function proposed by Wilmink (1987, Liv. Prod. Sci.)
> >
> >Ydim= a + be-kdim + cdim
> >
> >where dim are the days in milk. The function can be linearised by
setting k
> >to a fixed value (0.1 for example).
> >
> >So the model I want to use has three terms
> >
> >a the intercept
> >
> >b that actually is (be-0.1dim )
> >
> >
> >c x dim
> >
> >So my question is: how can I put this function as random within
animal?
> >Could be this the correct syntax of the random part of the model?
> >
> >r! a.animal b.animal dim.animal
> >
> >
> >and how can I get the (co)variance matrix of coefficients?
> >
> >
> >Thank you very much for your help
> >
> >Nicolò
> >
> >Nicolò P.P. Macciotta
> >Dipartimento di Scienze Zootecniche
> >Università degli Studi di Sassari
> >Via De Nicola, 9
> > 07100 Sassari, Italia
> >tel 39.079.229298 fax 39.079.229302
> >e-mail macciott_at_uniss.it

Nicolò P.P. Macciotta
Dipartimento di Scienze Zootecniche
Università degli Studi di Sassari
Via De Nicola, 9
  07100 Sassari, Italia
tel 39.079.229298 fax 39.079.229302
e-mail macciott_at_uniss.it
Received on Mon Oct 11 2005 - 07:52:45 EST