Node inconsistent with parents in JAGS model (R)

Question

I'm new to JAGS, and I'm trying to run a simple logistic regression. My data file is very simple: the response is binary and the one predictor I'm using has three levels. Like this:

col1: 1 2 2 2 1 1 1 2 1 2 ... 
col2: HLL, HLL, LHL, LLL, LHL, HLL ...

Dummy coding

The levels in col2 are HLL, LHL, LLL. I dummy coded it and created a data frame that looks like this:

(intercept) HLL LHL LLL
1           1   0   0   1
2           1   0   0   1
4           1   0   0   1
5           1   0   1   0
6           1   0   1   0
7           1   0   0   1

Data list

My data file (myList), then, looks like this:

List of 5
$ y  : num [1:107881] 2 2 2 2 2 2 2 2 2 2 ...
$ N  : num 500
$ HLL: num [1:107881] 0 0 0 0 0 0 0 0 0 0 ...
$ LHL: num [1:107881] 0 0 0 1 1 0 0 0 0 1 ...
$ LLL: num [1:107881] 1 1 1 0 0 1 1 1 1 0 ...

I'm using N=500 because the full data frame is huge and I just want to test it.

Model

cat(

    "model {
        for( i in 1 : N ){

            y[i] ~ dbern(mu[i])
            mu[i] <- 1/(1+exp(-(a + b*HLL[i] + c*LHL[i] + d*LLL[i])))
            }

            a ~ dnorm(0, 1.0e-12)
            b ~ dnorm(0, 1.0e-12)
            c ~ dnorm(0, 1.0e-12)
            d ~ dnorm(0, 1.0e-12)

            }", file = "model.txt"

)

Running model + error

model = jags.model(file = "model.txt", 
    data = myList,
    n.chains = 3, n.adapt = 500)

Error I get

Error in jags.model(file = "model.txt", data = antPenList, n.chains = 3,  : 
Error in node y[1]
Node inconsistent with parents

Answer 1

The dbern distribution expects response in {0,1} rather than {1,2} as it seems you have coded it, so you need to subtract 1 from your values of y.

It is a bit strange that you get this error, as dbern does not usually give an error for other response values (it basically makes <0 = 0 and >1 = 1). The error is probably stemming from the fact that the response is fitting all the same value, but if that doesn't fix it then you could try the following:

1) Try increasing the precision of your priors for a/b/c/d slightly - a variance of 10^12 is quite a lot

2) Instead of:

mu[i] <- 1/(1+exp(-(a + b*HLL[i] + c*LHL[i] + d*LLL[i])))

You could write:

logit(mu[i]) <- -(a + b*HLL[i] + c*LHL[i] + d*LLL[i])

This might also help JAGS to recognise this as a GLM and initiate the appropriate samplers - remember to load the glm module.

3) Set some initial values for a/b/c/d that are vaguely consistent with your data (perhaps obtained using a fit with glm() in R)

Answer 2

I solved it with

mu[i] <- 1/(1.000001+exp(-(a + b*HLL[i] + c*LHL[i] + d*LLL[i])))