[R-sig-ME] blocked diagonal R structures model convergence

[Robert Griffin]

Hi list members,

I am currently running a model to estimate G matrices for 4 traits
(size/weight, and number offspring, in males and females). I am using
blocked diagonal R structures because resdiual cross-sex covariance and
females is inestimable.

Looking at the traces from the chains, I see that there are some not well
converged cases - there's a lot of autocorrelation going on - but these are
only in cases where the traits being estimated are present in different
sexes (e.g. t3:at.level(sex,"1"):t2:at.level(sex,"1").units).

Would it be right to not worry about this? I think these are irrelevant
values which will be ignored later (e.g. when calculating phenotypic
variation) and I can be satisfied with the model because these are the only
problematic traces, however, could they still perhaps be influencing the
model in ways I haven't considered?

Script is below, thanks (and tips on priors or other aspects of the model
are of course appreciated if you spot anything odd!)
Rob

#### Prior
p1 = list(
G = list(
G1 = list(V = diag(4), nu = 1.002),
G2 = list(V = diag(4), nu = 1.002)),
R = list(
R1 = list(V = diag(4), nu = 1.002),
R2 = list(V = diag(4), nu = 1.002))
)

## Model:
# t1-t4 = traits (t1 & t2 in males, t3 & t4 are the same traits but in
females)
# r1-r2 = categorical random effects
# animal and dam effects to allow partitioning of additive genetic and
maternal effects
# ped is the pedigree
# using blocked diagonal R structures because cross-sex residual
covariances are indentifiable
# data consists of ~7500 individuals, ~50:50 sex ratio and both traits
measured in all individuals
# t1 & t3 are continuous variables
# t2 & t4 are count data, ranging from 1-20 (with lots of 1's, few 20's),
so using poisson family

mod1 <- MCMCglmm(cbind(t1, t2, t3, t4) ~ trait-1 + trait:r1 + trait:r2,
random   = ~us(trait):animal + us(trait):dam,
rcov = ~us(trait:at.level(sex, "1")):units+us(trait:at.level(sex,
"2")):units,
family = c("gaussian", "poisson", "gaussian", "poisson"),
pedigree = ped,
nitt = 6500000,
burnin = 1500000,
thin =    5000,
data = df1,
prior = p1)


----------------------------------------------------------

Robert M. Griffin
Postdoctoral Researcher, University of Turku
www.griffinevo.com

	[[alternative HTML version deleted]]