So this question is about the BUGS modeling language. So you either know it or have no clue. I'm a newbie to this so it's been driving me mad. I want to define a simple two-state hidden Markov model (HMM) where the emission of each state follows a Normal distribution. I have an array data of Nxl dimension where each row is a subject and column is a continuous variable Time associated with the subject. So in WinBUGS, I would define the model as:

 for(i in 1:N) { # for each subject
    # Sample the initial observation
    Time[i, 1] ~ dnorm(mu[State[1]], tau[State[1]])

    for(j in 2:l[i]) { # for each observation
        # Sample the observed variables
        Time[i, j] ~ dnorm(mu[State[j]], tau[State[j]])
        # Sample the hidden states
        State[j] ~ dcat(P[State[j-1], ])

# P is the transition matrix of the hidden Markov chain
P[1, 1:2] ~ ddirch(Pinit)
P[2, 1:2] ~ ddirch(Pinit)

# Set the initial states
State[1] ~ dcat(Pinit)

# Sample the initial params
mu[1] ~ dnorm(200, 1.0E-6)
mu[2] ~ dnorm(400, 1.0E-6)

# The precision params
tau[1] ~ dgamma(0.001, 0.001)
tau[2] ~ dgamma(0.001, 0.001)

Initial values have also been provided via R2WinBUGS. The model is syntactically correct. But when I run, I got this error: "multiple definitions of node State[2]"

Can you please tell me why and how to solve this? I've searched around on the error but each has their own specific case and there's not a generic solution or explanation as to why this arises.


I don't know about HMM, but I can see that in every loop of i you are defining State[j]. This is causing the error in WinBUGS as it can not sample from each node (such as State[2]) more than once in each iteration in the MCMC.

You need to either somehow define State[j] out of the i loop (as you did for State[1]) in its own loop or switch to a [i,j] index for State. (Without knowing HMM, I am not sure which of these is the correct solution).

  • $\begingroup$ What is a commonly used / decent / proper prior for the precision parameter $\tau = \frac{1}{\sigma^2}$ if you don't have much info about it? That is, it is non-informative, so to say. $\endgroup$ – Joe Jul 31 '13 at 14:09
  • 1
    $\begingroup$ i usually use an inverse gamma for the precision, as you did. can be problematic though in hierarchical models... andrewgelman.com/2006/02/06/prior_distribut $\endgroup$ – gjabel Aug 1 '13 at 11:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.