[R-sig-ME] How to assess significance of variance components(Please discard previous e mail... but read this one)

[Stephane Chantepie]

Sorry once again, I have problems with my e-mail software which send draft 
without my approbation...It is very annoying

First thank David for your response

So, I have tried to used "pedigreemm" but I am unable to run random regression 
models whith this package. I succeed in fitting "VA" and "permanent 
environment" model without interaction. 

m1<-pedigreemm(oeuf~as.factor(st_age)+(1|animal)+(1|
ID),data=oeuf,family="poisson",REML=TRUE,pedigree=list(animal=ped),verbose=TRUE)

But  when I try to add an interaction like 

m2<-pedigreemm(oeuf~as.factor(st_age)+(st_age|animal)+(st_age|
ID),data=oeuf,family="poisson",REML=TRUE,pedigree=list(animal=ped),verbose=TRUE)

it does not work. I do not find examples on the internet to help me...

if someone has already done this kind of models?

> or maybe you have no power to estimate these kinds of
> models, or you have to centre age. 

I use centered age for each model?

For the problematic trait (number of  spz), I have more than 100000 data 
(repeated measurment) on more than 1600 individuals with an exhaustive 
pedigree. The number of individuals decrease with age so I decided to use data 
until 15 years old (older 23 years old). 
To be more precise on the results I have for the spz trait, the univariate 
models tend to show a deacrease but when taking into account the 95% 
confidence intervals, it could be really acceptable for having no interaction. 
When I plot the results from random regression, the curve increase but the 
confidence interval is quite big and additive genetic variance appear bigger 
than phenotypic variance. So it looks like VAxAge is not significant without 
beeing able to test it (other than graphically).  

With another trait (number of eggs); Va is increasing and there is a perfect 
match between univariate and random regression models from  MCMCglmm. So, I 
think  that the power is not the problem. The problem is that random 
regression with MCMCglmm failed to catch the constant variance across age. 



> Does phenotypic variance fall off with
> age? 

Not at all


> Have you done diagnostic plots for poissonness? 

Yes, I have. The distribution is poisson without any doubt. I  have tried to 
log the data to have a gaussian distribution but it does not give perfect 
gaussian distribution.

> What do simple
> correlations between different classes of relatives look like?

I am not sure to understand your question! What do tou mean by "different 
classes of relatives"?


I will try to use the last version of asreml which seems (not sure) to perform 
animal model for poisson distribution and test the VaxAge interaction...

Stephane