Ordered Probit in Jags Using scaled inverse wishart - jags

Question

I am trying to use the following code (adapted from the code given in Gelman and Hill's Book) to estimate a varying coefficient/intercept ordered probit model in Jags. However, it is giving me a "Observed node inconsistent with unobserved parents at initialization.Try setting appropriate initial values". Where am I going wrong? Could somebody please help me? Thanks in advance !!
rm(list=ls(all=TRUE));
options(warn=-1)
library(mvtnorm)
library(arm)
library(foreign)
library("R2jags")
library(MCMCpack)
set.seed(1)
standardizeCols = function( dataMat ) {
zDataMat = dataMat
for ( colIdx in 1:NCOL( dataMat ) ) {
mCol = mean( dataMat[,colIdx] )
sdCol = sd( dataMat[,colIdx] )
zDataMat[,colIdx] = ( dataMat[,colIdx] - mCol ) / sdCol
}
return( zDataMat )
}
keep<-1
nobs = 150;
nis<-sample(1:40,nobs,replace=T) # number obs per subject
id<-rep(1:nobs,nis)
N<-length(id)
corr_beta = 0.6;
Sigma_beta = matrix(c(1, corr_beta, corr_beta, corr_beta,
corr_beta, 1, corr_beta, corr_beta,
corr_beta, corr_beta, 1, corr_beta,
corr_beta, corr_beta, corr_beta, 1), ncol=4);
betas <- rmvnorm(n=N, mean=c(-1.45, 0.90, 0.25, -2.3), sigma=Sigma_beta);
#Generate the data
x3 = matrix(0, nrow=N,ncol=3);
y3 = matrix(0, nrow=N,ncol=1);
for (i in 1:N) {
error_v = rnorm(1,0,1);
x3[i,1] = rnorm(1,0,1);
x3[i,2] = rnorm(1,0,1);
x3[i,3] = rnorm(1,0,1);
y3[i,1] = betas[id[i], 1] + betas[id[i], 2]*x3[i,1] + betas[id[i], 3]*x3[i,2] + betas[id[i], 4]*x3[i,3] + error_v;
}
cutoff=c(-100, 0, 1.5, 2.4, 100)
k=length(cutoff)-1;
Y3<-cut(y3, br = cutoff, right=TRUE, include.lowest = TRUE, labels = FALSE)
Y3=Y3
X3=x3
m1=max(Y3)
y = as.vector( Y3 )
n = length(y)
J<-length(unique(id))
X = cbind(1, standardizeCols( X3 ))
nPred = NCOL(X)
subjects<-as.vector(as.numeric(id))
K=nPred
W <- diag (K)
# MCMC settings
ni <- 5000; nb <- 2500; nt <- 6; nc <- 3
tau1u=c(0,1,2)
jags_data <- list ("n", "J", "K", "y", "subjects", "X", "W", "m1")
inits <- function (){
list (B.raw=array(rnorm(J*K),c(J,K)), mu.raw=rnorm(K), sigma.y=runif(1), Tau.B.raw=rwish(K+1,diag(K)), xi=runif(K))
}
params <- c ("B", "mu", "sigma.B", "rho.B", "tau1u")
cat("model {
for (i in 1:n){
y.hat[i] <- inprod(B[subjects[i],],X[i,])
y[i] ~ dcat(p[i,])
estar[i]~dnorm (y.hat[i], tau.y);
for (j in 1:(m1-1)) {
Q1[i,j]<-pnorm(tau1[j]-estar[i],0,1)
}
p[i,1] <- Q1[i,1]
for(j in 2:(m1-1)) {
p[i,j] <- Q1[i,j] - Q1[i,j-1]
}
p[i,m1] <- 1 - Q1[i,m1-1]
}
tau.y <- pow(sigma.y, -2)
sigma.y ~ dunif (0, 100)
# thresholds (unordered priors)
for(j in 1:(m1-1)){
tau1u[j] ~ dnorm(0,.01)
}
# ordered thresholds
tau1 <- sort(tau1u)
for (j in 1:J){
for (k in 1:K){
B[j,k] <- xi[k]*B.raw[j,k]
}
B.raw[j,1:K] ~ dmnorm (mu.raw[], Tau.B.raw[,])
}
for (k in 1:K){
mu[k] <- xi[k]*mu.raw[k]
mu.raw[k] ~ dnorm (0, .0001)
xi[k] ~ dunif (0, 100)
}
Tau.B.raw[1:K,1:K] ~ dwish (W[,], df)
df <- K+1
Sigma.B.raw[1:K,1:K] <- inverse(Tau.B.raw[,])
for (k in 1:K){
for (k.prime in 1:K){
rho.B[k,k.prime] <- Sigma.B.raw[k,k.prime]/sqrt(Sigma.B.raw[k,k]*Sigma.B.raw[k.prime,k.prime])
}
sigma.B[k] <- abs(xi[k])*sqrt(Sigma.B.raw[k,k])
}
}", fill=TRUE, file="wishart2.txt")
# Start Gibbs sampler
outj <- jags(jags_data, inits=inits, parameters.to.save=params, model.file="wishart2.txt", n.thin=nt, n.chains=nc, n.burnin=nb, n.iter=ni)

Answer 1

Your initial values function returns random numbers from normal and uniform distributions, which it appears are not close enough to sensible values to allow a non-0 posterior value to be calculated. I think you need to choose your initial values more carefully, and perhaps based on the values generated in the data, to ensure that the model compiles. Do Gelman and Hill give initial values for their model that you could start with?
Update: you could also try removing your 'inits=inits' argument to allow JAGS to select its own initial values, which works for most (although not all) models. I dont use R2JAGS though so I'm not sure if this is allowed for the jags function (but it is for rjags and runjags).