Hi,

I would like to implement a model where the response y~dnorm(mu,tau) is left-

censored but the censoring threshold (cutpoint) depends on both mu and tau.

Basically the censoring threshold is unknown, all I know is the probability of

censoring (pcens_) and I use that with mu and tau to compute the censoring

quantile (called upper_ in my model).

I'm unfamiliar with the syntax of dinterval in case of left censoring. I used

an indicator variable is.observed_ which takes value 0 if y_ is censored, and

value 1 if y_ is observed. This is the reverse situation to the right

censoring case mentionned in the JAGS manual.

My model is the following:

model{

for(i in 1:N){

is.observed_ ~ dinterval(y_, upper_)

y_ ~ dnorm(mu_,tau)

mu_ <- beta + beta*age_

upper_ <- qnorm(pcens_,mu_,tau)

}

tau <- pow(sigma,-2)

sigma ~ dunif(0,100)

beta ~ dnorm(0,0.001)

beta ~ dnorm(0,0.001)

}

When I run that model I get:

Error in node is.observed

Observed node inconsistent with unobserved parents at initialization

...so I guess the issue is that initial values for missing y's do not agree

with the censoring constraints. But then how may I solve that since the

censoring quantiles depends on mu and tau? As far as I know, this kind of

construct is prohibited in WinBUGS/OpenBUGS and the censoring thresholds must

not depend on mu or tau if the latter are unobserved as well. Is this actually

possible in JAGS? What am I missing here?

Thanks for any input