[R-sig-ME] MCMCglmm covariance matrix specification

[Walid Crampton-Mawass]

Hi Jarrod,

Actually, I do have a shared pedigree and the cross-env covariances were
estimable. The goal of this analysis is to see if precision improves in any
way if I can set cross-env covariances to zero using the initial way my
matrices were constructed which is simply: ~us(env:trait):animal - which
includes cross-env covariances

I will try your proposal and get back to you. Though, I am thinking that a
model that does include cross-env covariances in my covariance matrix would
be a better choice so as to analyse cross-env genetic correlations (unless
I can still do that with your approach)

Thanks!
-- 
Walid Crampton-Mawass
Ph.D. candidate in Evolutionary Biology
Population Genetics Laboratory
University of Québec at Trois-Rivières
3351, boul. des Forges, C.P. 500
Trois-Rivières (Québec) G9A 5H7
Telephone: 819-376-5011 poste 3384


On Wed, Feb 24, 2021 at 4:05 PM Jarrod Hadfield <j.hadfield using ed.ac.uk> wrote:

> Hi,
>
> As I understand it you have 3 2x2 covariance matrices to be estimated,
> one for each environment?
>
> ~us(at.level(env, 1):trait):animal+us(at.level(env,
> 2):trait):animal+us(at.level(env, 3):trait):animal
>
> should work. I presume you have no shared pedigree between the
> envrionments hence the cross-env covariances are not estimable? In the
> computation time is long, get back to me; there are ways to
> reparameterise it to make it faster but it's a bit fiddly.
>
> For harder problems (where the covariance matrix can't be permuted such
> that the estimable bits fall in blocks along the diagonal, as here) then
> fixing elements to zero is probably not a good idea even if you could do
> it (for example in asreml). The zero elements will force patterns in the
> estimable elements to ensure positive-defitness. The antedependence
> solution I posted earlier gets round this issue I believe.
>
> Cheers,
>
> Jarrod
>
>
> On 24/02/2021 18:29, Walid Crampton-Mawass wrote:
> > This email was sent to you by someone outside the University.
> > You should only click on links or attachments if you are certain that
> the email is genuine and the content is safe.
> >
> > Hey all,
> >
> > Hope you are doing well during this time!
> >
> > I have been racking my brain for weeks on how to do model this issue but
> I
> > have found nothing other than one old answer by Jarrod Hadfield (
> > https://stat.ethz.ch/pipermail/r-sig-mixed-models/2015q4/024036.html)
> which
> > recommends using an antedepedence model. Here is the issue:
> >
> > I have constructed a bivariate animal model (trait1, trait2) with a
> random
> > interaction with the additive genetic random effect and the residual
> > variance,i.e. (trait:env):animal. The interaction variable is a
> categorical
> > environmental variable of 3 levels (Low, Mid, High). So my
> > variance-covariance matrix has a 6x6 shape (2traitsx3env). Hence, the
> > matrix would include both among-trait covariances within the same env and
> > between env, and cross-env covariances for the same trait:
> >
> > trait1:low trait1:mid trait1:high trait2:low trait2:mid trait2:high
> > 1 0 0 0 0 0
> > 0 1 0 0 0 0
> > 0 0 1 0 0 0
> > 0 0 0 1 0 0
> > 0 0 0 0 1 0
> > 0 0 0 0 0 1
> > (1 represent variances, 0 represent covariances)
> >
> > I have already run the model with both the idh() and us() specification.
> In
> > the first case, no covariances are calculated at all, only variances are
> > calculated. In the second case, all types of covariances are calculated.
> >
> > I need help figuring out how to specify the variance-covariance matrix in
> > MCMCglmm (and prior) in a way to tell the model not to estimate the
> > cross-env covariances, only the among-trait covariances should be
> > estimated:
> > trait1:low trait1:mid trait1:high trait2:low trait2:mid trait2:high
> > 1 x x 0 x x
> > x 1 x x 0 x
> > x x 1 x x 0
> > 0 x x 1 x x
> > x 0 x x 1 x
> > x x 0 x x 1
> > (1 represent variances, 0 represent covariances to be estimated, x
> > represent covariances fixed at 0, i.e. not estimated)
> >
> > any help would be appreciated!
> > --
> > Walid Crampton-Mawass
> > Ph.D. candidate in Evolutionary Biology
> > Population Genetics Laboratory
> > University of Québec at Trois-Rivières
> > 3351, boul. des Forges, C.P. 500
> > Trois-Rivières (Québec) G9A 5H7
> > Telephone: 819-376-5011 poste 3384
> >
> >          [[alternative HTML version deleted]]
> >
> > _______________________________________________
> > R-sig-mixed-models using r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> The University of Edinburgh is a charitable body, registered in Scotland,
> with registration number SC005336. Is e buidheann carthannais a th’ ann an
> Oilthigh Dhùn Èideann, clàraichte an Alba, àireamh clàraidh SC005336.
>

	[[alternative HTML version deleted]]