I have been using MCMCglmm in R to fit bivariate (two response traits) mixed models in R, but now I need to move to lme to account for temporal autocorrelation of the residuals. In MCMCglmm I can fit the following model (data in wide format):
mod.mcmc=MCMCglmm(cbind(dvm,sdl)~trait-1+poly(month,3),random=~us(trait):id,rcov=~us(trait):units,data=wide21,family=c("gaussian","gaussian"),verbose=FALSE,prior=prior1,pr=TRUE)
And from it extract the variance-covariance matrices (posterior modes) at the random and residual level, with this code:
cor.matrix.bet=matrix(posterior.mode(posterior.cor(mcmc2$VCV[,1:4])),2,2,dimnames=list(c("dvm", "sdl"), c("dvm", "sdl")))
cor.matrix.res=matrix(posterior.mode(posterior.cor(mcmc2$VCV[,5:8])),2,2,dimnames=list(c("dvm", "sdl"), c("dvm", "sdl")))
In lme, my models looks like this right now (data in long format):
mod.lme=lme(value~trait*poly(month,3)-1,data=long21,random=~trait-1|id,weights=varIdent(form=~1|trait),correlation=corAR1(form=~tim|id/trait))
This model differs from the one in MCMCglmm in that I specified a temporal correlation structure in the residuals, something that cannot be done in MCMCglmm. Specifically, "tim" is a timer variable, and the temporal correlation is assessed within each individual ("id") and within each of the two traits ("trait").
I can extract the random variance-covariance matrix of mod.lme wit the getVarCov() function. But how can I get the residual variance-covariance matrix?
And also, the residual variance-covariance matrix is unstructured ("us") in mod.MCMC, which is exactly what I want because then, all the covariances are estimated. In mod.lme I have somehow contrained this (I guess) by using the correlation function, but can I still incorporate somehow in the model that I want a "us" structure for the residuals, or is it completely non-sense since I already have the corAR1 in the model?
Thank you
getVarCov(mod.lme, type = "conditional")
has atype
argument in which you can specify that you want the "conditional" covariance matrix, which correspond to the residual covariance. $\endgroup$