[R-sig-ME] priors for univariate model with very low expected variance contributions (mcmcglmm)

[Robert Griffin]

Dear list,

I have sampled ~35 copies of a chromosome from a population because I want
to estimate how much contribution that part of the genome makes to the
variance in the traits. Therefore I want to estimate the additive genetic
variance. I will do this by using making a univariate response model in
MCMCglmm. The data for the trait was collected from 300-500 offspring per
sampled chromosome, and measured in males. This was done across 4
experimental *blocks *and within each *line *and *block *there were 4 *vials
*of many individuals, each sourced from one of two *sets *of parents.

Within the model there are *block*, parental set (*set*), *vial*, and *line
*effects to model. I have done this in the following way:

chain1 = MCMCglmm(trait ~1 + block,
random = ~line + set + vial,
rcov = ~units,
nitt = nitt,
thin = thin,
burnin = burnin,
prior = prior,
family = "gaussian",
start = list(QUASI = FALSE),
data = df1)


However, the phenotypic variance in this trait is large [var(trait) =
~150], and I am expecting an extremely large part of the variance to be
environmental & measurement error (residual), and the variables of line,
set, block, and vial to contribute very little (probably <5% of total
variation each) - visual examination of the data suggests that there is
almost no variance among lines, blocks, vials, or parental sets. Which
leads me to my call for help.

I am mainly concerned about how to choose priors for variances which are
expected to be near zero (when the aim is to test if line variance is not
0) - can this affect the outcome of the model? How should I define my
priors in such a case? Currently my best estimate from reading the
literature is to use the following:

prior = list(G = list( G1 = list(V = var(trait)/4, nu=0.002),
G2 = list(V = var(trait)/4, nu=0.002),
G3 = list(V = var(trait)/4, nu=0.002)),
R  = list(V = var(trait)/4, nu=0.002))


Advice about the priors (and the model in general if you happen spot
anything- e.g. should the family be Gaussian?) would be greatly appreciated,

Rob


-----------------------------
Robert Griffin
PhD candidate, Uppsala University
griffinevo.wordpress.com

	[[alternative HTML version deleted]]