[R-sig-ME] MCMCglmm R-structure problem in heteroskedastic model

[Paul Johnson]

Hello,

I'm using MCMCglmm to estimate heritability of lung function in a data set with 1346 nuclear families, each family having 2 parents and at least 1 offspring (total n=4777). I'm having trouble with a particular model:

########################
>     prior.model3<-list(R=list(nu=0.02,V=diag(2)),G=list(G1=list(nu=0.02,V=1)))
>     model3<-
+       MCMCglmm(fixed.form,
+         random=~animal,
+         pedigree=pedigree,
+         rcov=~idh(asthma.family):units,
+         prior=prior.model3,
+         data=mfs.fam,nitt=nitt,thin=thin,burnin=burnin)
Warning in MCMCglmm(fixed.form, random = ~animal, pedigree = pedigree, rcov = ~idh(asthma.family):units,  :
  some combinations in animal do not exist and 2692 missing records have been generated
Error in MCMCglmm(fixed.form, random = ~animal, pedigree = pedigree, rcov = ~idh(asthma.family):units,  :
  R-structure does not define unique residual for each data point
##########################

What I'm trying to do is estimate two different residual variances, one in families with at least 1 asthmatic (n=722 individuals) and one in the rest (n=4055 individuals), but as you can see I get an error. Can anyone see what the problem is?

Similar models work fine, including
1. a model with two variances [random=~animal, rcov=~units]
2. a model with 4 variances [random=~idh(asthma.family):animal, rcov=~idh(asthma.family):units].
3. a 3-variance model with a single residual variance and two animal variances [random=~idh(asthma.family):animal, rcov=~units].

(An odd thing about 2 and 3 above is that MCMCglmm warns me that "6730 missing records have been generated" which is exactly 2.5 times the 2692 that I'd expect).

Thanks for any help,
Paul Johnson

PS other details:

> R.Version()$version.string
[1] "R version 2.12.0 (2010-10-15)"

> Sys.info()
     sysname      release      version     nodename      machine        login         user
   "Windows"      "7 x64" "build 7600"    "PAUL-PC"        "x86"       "Paul"       "Paul"

> cbind(installed.packages()["MCMCglmm",])
          [,1]
Package   "MCMCglmm"
LibPath   "C:\\Users\\Paul\\Documents/R/win-library/2.12"
Version   "2.10"
Priority  NA
Depends   "tensorA, Matrix, coda, ape, corpcor"
Imports   NA
LinkingTo NA
Suggests  "rgl, combinat, mvtnorm, orthopolynom"
Enhances  NA
OS_type   NA
License   "GPL (>= 2)"
Archs     "i386, x64"
Built     "2.12.1"


The University of Glasgow, charity number SC004401