Truncated Dirichlet Process

Help
2011-02-08
2012-09-01
  • Jacopo Soriano

    Jacopo Soriano - 2011-02-08

    Hi,

    I am fitting a semiparametric generalized linear mixed effect model with a DP
    prior on the random effects and I use a truncated approximation of the DP. If
    I add a prior on the mass parameter of the DP, after several iterations I
    obtain this error message:

    Errore: RUNTIME ERROR:

    Singularity in likelihood found by Slicer

    What does it mean? Where is the problem?

    Thank you for your help,

    Jacopo

     
  • Will Bullock

    Will Bullock - 2011-02-17

    Hi Jacopo,

    I've received the same error with one of my more complex models when running
    JAGS 2.2.0 64-bit via R on Windows. While I can't explain the error, I do know
    that my particular model worked fined with JAGS 2.1.0.

    Good luck,

    --Will

     
  • Martyn Plummer

    Martyn Plummer - 2011-02-17

    The problem is that some distributions have unbounded probability density
    functions ( e.g. Dirichlet, Beta, Gamma when the relevant shape parameter is
    less than 1). If the slice sampler finds itself at a point where the current
    log-density is infinite then it cannot move away from it. But it is necessary
    to have a finite probability of moving away from the current state at all
    times in order to satisfy the regularity conditions on which the whole MCMC
    theory depends.

    Previously the slice sampler did not issue a warning but simply stayed where
    it was. Now it stops with an error message. Am I being too cautious? I don't
    know. In practice it seems that the sampler was able to move away from the
    singularity eventually, but I'm a lot more worried about JAGS appearing to
    work and giving the wrong answer than I am about some models not being able to
    run. You can usually work around the problem by strengthening the prior or
    bounding the distribution away from the boundary where the singularity lies.

     
  • Jacopo Soriano

    Jacopo Soriano - 2011-02-18

    Thank you!

    The model compiles, initializes correctly, an error occurs during updating,
    but the current state of the model doesn't dump to a file named jags.dump1.R.

    I would like to know which unbounded prior cause this problem.

    How can I obtain the current state of the chain?

     
  • annalix

    annalix - 2011-06-30

    Hi.

    I have the same problem.

    I am using a Dependent Dirichlet Process and I put an inverse-gamma prior on
    the mass parameter.

    After some iterations JAGS stops with the error:

    Singularity in likelihood found by Slicer.

    I don't understand where is exactly the problem. How can I obtain the state of
    the chain when JAGS stops?

    Thank you.

    Annalisa

     

Get latest updates about Open Source Projects, Conferences and News.

Sign up for the SourceForge newsletter:





No, thanks