[R-sig-ME] zibinomial in MCMCglmm

[Adrian Jaeggi]

Dear list,

I have a data set where the successes part of a binomial response  
variable are zero-inflated:

str(mydata)
'data.frame':   30 obs. of  12 variables:
  $ approacher            : Factor w/ 6 levels "Dj","Ja","Lc",..: 2 3  
4 5 6 1 3 4 5 6 ...
  $ owner                 : Factor w/ 6 levels "Dj","Ja","Lc",..: 1 1  
1 1 1 2 2 2 2 2 ...
  $ Successes             : int  0 1 1 5 0 0 1 0 0 0 ...
  $ Failures              : int  25 16 14 22 6 3 19 2 4 1 ...
  $ owner.fgs             : num  0 0 0.04 0 0 ...
  $ davdif                : num  -1.123 -2.18 1.288 0.534 -2.663 ...
  $ matkin                : int  0 0 0 0 0 0 0 0 0 0 ...
  $ sexcom                : Factor w/ 4 levels "ff","fm","mf",..: 1 3  
1 3 3 1 3 1 3 3 ...
  $ relq                  : num  0.962 -0.878 -0.324 -0.65 -0.672 ...
  $ polq                  : num  2.6378 0.16 -0.0731 -0.3261 0.0379 ...

y<- cbind(mydata$Successes, mydata$Failures)
summary(y)
        V1               V2
  Min.   : 0.000   Min.   : 0.00
  1st Qu.: 0.000   1st Qu.: 4.00
  Median : 0.000   Median :15.00
  Mean   : 3.310   Mean   :20.66
  3rd Qu.: 1.000   3rd Qu.:24.00
  Max.   :72.000   Max.   :80.00

Running analyses in lme4() with family=binomial yields warnings:

glmm<-  
glmer(y~owner.fgs+relq+polq+matkin+davdif+sexcom+(1|approacher)+(1|owner),  
data=mydata, family=binomial)
  Error messages:
  1: glm.fit: fitted probabilities numerically 0 or 1 occurred
  2: In mer_finalize(ans) : singular convergence (7)

Furthermore, using glm() with family=quasibinomial indicates high  
overdispersion:
glm<-  
glm(y~Owner.vol.success+relq+polq+matkin+davdif+sexcom+approacher+owner,  
data=PLD2010bags, family=quasibinomial)
summary(glm)
> (Dispersion parameter for quasibinomial family taken to be 8.624607e+15)


Hence, I am trying to run a zero-inflated binomial model in MCMCglmm,  
but I am getting an error message:

priorzi <- list(R = list(V = diag(2), fix=2), G = list(G1 = list(V =  
diag(2), nu = 0.002), G2 = list(V = diag(2), nu= 0.002)))

glmm<-  
MCMCglmm(y~trait:owner.fgs+trait:relq+trait:polq+trait:matkin+trait:davdif+trait:sexcom,
  random=~idh(trait):approacher+idh(trait):owner, data=mydata,  
family="zibinomial", rcov=~idh(trait):units, prior=priorzi)

> Error in `[<-.data.frame`(`*tmp*`, , response.names, value = c(0L,  
> 1L,  :  missing values are not allowed in subscripted assignments of  
> data frames


There are no missing values in the data frame and I couldn't figure  
out what the error message refers to.
Thanks for any help!
Suggestions for more appropriate prior specifications are also  
welcome, I am never quite sure if I get them right..


Cheers,
Adrian



-- 
Adrian Jaeggi, PhD
Postdoctoral Researcher
Sage Center for the Study of Mind /
Department of Anthropology
University of California Santa Barbara
Santa Barbara, CA 93106-3210
Phone: 805-455-8587