[R-sig-ME] observation level random effects/kinship model

["Pär K. Ingvarsson"]

> Date: Wed, 14 Mar 2012 14:05:14 +0400
> From: Yves Rousselle <yvesrousselle at gmail.com>
> To: r-sig-mixed-models at r-project.org
> Subject: Re: [R-sig-ME] observation level random effects/kinship model
> Message-ID:
>        <CAA8=r0DAnaByMy23C8tPuZn4W-PeccjTHviG7UtOwT8k40MoXw at mail.gmail.com>
> Content-Type: text/plain
> 
> Hi,
> 
> If I understand well, if you want to take into account the kinship between
> individuals in genetics models, you have to specify a random effect that is
> the individuals levels and specify that this effect follows a distribution
> with a variance/covariance matrix equal to 2*K*Vg (basically). 2*Vg is just
> a number and K is the kinship matrix.
> I am currently using R to do such genetics models to do association
> mapping. I ask to other people that have done that before me and if I
> understand well, no packages allows to specify such a variance/covariance
> matrix for a random effect except ASREML. But you have to pay a license to
> use it. I am using this package for my study.

There are several packages that can handle kinship matrices for doing association studies in R. Of the top of my head I can come up with:

kinship (http://cran.r-project.org/web/packages/kinship/index.html)
EMMA (http://mouse.cs.ucla.edu/emma/news.html)
GenABEL (http://www.genabel.org/)


> 
> Concerning the question of putting an observation-level effect, I begin to
> understand it but you have to check with others perhaps. I will take the
> association mapping case as an example (I hope you know it a bit). You have
> a sample of individuals that are evaluated within a repeated block design
> for example. So each individuals is repeated end therefore, the
> observations level is not **yet** the individual level. The classical first
> step is to estimate (predict is the good word I guess) BLUP for the
> individual level with a model that takes into account the experimental
> design parameters. After this step, you obtain a dataset in which the
> observation level is the individual level. The second step consists in
> testing the association between the BLUP and some markers. In this model,
> you specify a random effect which is the individual level for with you
> specify the variance/covariance matrix 2KVg. So, at this step, you use a
> random effect at the observation level. I talked about that with
> biostatiticians (I hope this traduction is good) because I was surprised
> that an effect could be at the same level as the observations level because
> there won't be enough degree of freedom in the model. They explained me
> that, the correlation between individuals, specified in the kinship matrix,
> acts like repeating each individuals in their common part in other
> individuals. Well, I am sorry that I'm not able to put this idea in words
> in a better way, I hope it will help.

You can do the analyses directly without going through the BLUPs estimation step, but that generally leads to similar results as with the two-step method. See for instance http://www.genetics.org/content/178/3/1745.abstract for more details


-Pelle

--
Pär K. Ingvarsson
Professor, Evolutionary Genetics
Umeå Plant Science Centre
Department of Ecology and Environmental Science
Linneaus väg 6
Umeå University, SE-901 87 Umeå, Sweden
tel. +46-(0)90-786-7414, fax. +46-(0)90-786-6705