JAGS Error when fitting Gamma GLM: Invalid parent values

Question

I am trying to fit a Gamma GLM to my data.Here's my code:

Data <- read.csv("ODcost.csv")
###Define data matrix
X <- cbind(1,Data[,c(1:16)])
C <- Data$C
E <- Data$E
N <- length(C)

Gamma.data <- list("C","N","X","E")

Gamma2 <- function() {
# Priors:
for(j in 1:17){beta[j] ~ dnorm(0.0, 0.001)}
alpha ~ dunif(0,100)

# Likelihood data model:
for (i in 1:N) {
log(mu[i]) <- log(E[i]) + inprod(X[i,],beta[])
# dgamma(shape, rate):
C[i] ~ dgamma(alpha, mu[i]/alpha)
 }
}

Gamma.inits <- function(){list("beta"=rep(0.001,17))}

Gamma.params <- c(paste("beta[",i=1:17,"]",sep=""))

Gammafit <- jags(data = Gamma.data,inits=Gamma.inits,
            parameters.to.save = Gamma.params, n.chains=2, n.iter=10000,
            n.burnin=5000, n.thin=2, model.file = Gamma2)

Here's what Data looks like:

head(Data,n=10)

   x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15 x16     C     E
1   0  0  0  0  0  0  0  0  0   0   0   0   0   0   0   0  4462    34
2   0  0  0  0  0  0  0  0  0   0   0   0   1   0   0   0 11405    11
3   0  0  0  0  0  0  0  0  0   0   0   0   0   1   0   0  4720   184
4   0  0  0  0  0  0  0  0  0   0   0   0   0   0   1   0  4631    14
5   0  0  0  0  0  0  0  0  1   0   0   0   0   0   0   0  7128 17881
6   0  0  0  0  0  0  0  0  1   0   0   0   1   0   0   0  7043  7581
7   0  0  0  0  0  0  0  0  1   0   0   0   0   1   0   0  5688 19699
8   0  0  0  0  0  0  0  0  1   0   0   0   0   0   1   0  5819  8268
9   0  0  0  0  0  0  0  0  1   0   0   0   0   0   0   1  5527 12229
10  0  0  0  0  0  0  0  0  0   1   0   0   0   0   0   0  9853   241

I get the error:

Error in jags.model(model.file, data = data, inits = init.values, n.chains = 
n.chains, :
Error in node C[497]
Invalid parent values

And here's what's inside Data[497,]:

    x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15 x16    C E
497  0  0  0  1  0  0  0  1  1   0   0   0   0   0   0   1 1140 0

I do not know why the code is stuck at that particular row. Thanks in advance for any help.

Answer 1

I have several rather random ideas:

• Your j loop goes from 1 to 17, but you seem to only have 16 predictors.

• I would try to experiment with somewhat more informative priors for beta[j], say dnorm(0, 0.01). I vaguely remember that this used to be an issue once when I was also using log link (but with Poisson distribution).

• I would try to experiment with initializing your alpha. Also, maybe try using some continuous and positive alternative to dunif() (gamma?). Or maybe try to make the prior a bit narrower.

• Does the model run with just a single predictor?