[R-sig-ME] Trivariate MCMCglmm

[Matthew Wolak]

Dear Gemma,

I can offer a couple of quick things to check. First, do you have any inbreeding occurring in your study design? This could cause the additive effects to change depending on the presence of the dominance term in the model.

You might also want to consider a larger thinning interval. It is hard to say anything particularly useful without knowing more information or having quantitative descriptors of the model [e.g., what is the output from: autocorr(model3.2) or the effective sample sizes for each random effect?]. However, I suspect the autocorrelation between thinned samples is probably pretty high. Depending on what autocorr(model3.2) tells you, you might want to consider changing the following three arguments like so:

   nitt = 5005000, burnin = 5000, thin = 5000

It might also be hard for the model to separate dominance and maternal effects from one another, but it is hard to say without knowing more about the breeding design. In general, you could be pushing the model too hard with 3 traits, 4 random terms, and only 1063 animals. If you don't have much missing phenotypic data, univariate models might be a lot easier to start with so that you can see if dominance and/or maternal effects are even necessary to include in the trivariate model.

I was a little uncertain about your statement:

"On the other hand, characters with high dominance effects have a change in additive effects in the model without dominance. Additive effects are highly reduced when dominance is added to the model. Is this an overestimation of the additive variance in the model without dominance caused by its lack?"

To be clear, are you being careful to use precise language when talking about the individual animal effects for the additive or dominance "effects" (i.e., the model BLUPs) versus the model estimate of the variance in these effects when talking about the "additive variance"?

Sincerely,
Matthew

....................................................
Dr. Matthew E. Wolak
School of Biological Sciences
Zoology Building
University of Aberdeen
Tillydrone Avenue
Aberdeen AB24 2TZ
office phone: +44 (0)1224 273255

On 19/08/14 13:51, Gemma Palomar García wrote:


Hello,

We are using MCMCglmm package with a dataset of 1063 toads. We have three different characters: infection rate, fresh body weight and development rate and we would like to infer additive variance, dominance and maternal effect which it is possible with our breeding design. We want to study heritability and genetic correlations between the characters. The three were normalized with a logarithmic transformation and block was used as random factor to control the variation caused by the position of the containers.

Inferring dominance effect, we used nadiv package to obtain dominance matrix:

library(nadiv)
pedDom=read.table("Ped_Dom.txt",header=T)
Dom<-makeD(pedDom)
Dinv=Dom$Dinv
data$dom=data$animal

The construction of the trivariate model was the following:

model3.2=MCMCglmm(cbind(Infection_rate,Weight,Developm)~trait-1,
                  random = ~us(trait):animal+us(trait):Mother
                  +us(trait):Block+us(trait):dom,
                  rcov=~us(trait):units,
                  ginverse=list(dom=Dinv),
                  family = c("gaussian", "gaussian","gaussian"),
                  pedigree = pedDom,data = data, nitt = 5000000,
                  thin = 100, burnin = 5000, prior = prior3.2)

And we used this prior:

prior3.2=list(G = list(G1 = list(V = diag(3), n = 2.002),
                       G2 = list(V = diag(3), n = 2.002),
                       G3 = list(V = diag(3), n = 2.002),
                       G4 = list(V = diag(3), n = 2.002)),
              R = list(V = diag(3), n = 2.002))

We obtained very bad mixing but DIC is lower than DIC of the model without dominance. Scaling the variables doesn�t seem to improve the model. Any advice of the model or the prior will be welcome to improve the mixing.

On the other hand, characters with high dominance effects have a change in additive effects in the model without dominance. Additive effects are highly reduced when dominance is added to the model. Is this an overestimation of the additive variance in the model without dominance caused by its lack?

Thank you,

Gemma Palomar
PhD student of University of Oviedo, Spain






        [[alternative HTML version deleted]]





_______________________________________________
R-sig-mixed-models at r-project.org<mailto:R-sig-mixed-models at r-project.org> mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models




The University of Aberdeen is a charity registered in Scotland, No SC013683.
Tha Oilthigh Obar Dheathain na charthannas clàraichte ann an Alba, Àir. SC013683.

	[[alternative HTML version deleted]]