Error-tolerant adaption

Michael
2012-08-29
2012-09-01
  • Michael
    Michael
    2012-08-29

    Dear all,

    right at the start: can't thank Martyn enough for the good work that goes into
    JAGS; makes MCMC worthwile!!

    But still I get problems (that's life):

    I'm currently dealing with a slow-mixing chain.. that is, mixing does appear
    to improve the longer I run it.. and I've already tried e.g. parameter
    expansion to improve mixing..

    There are a large number of parameters to be estimated: four parameters each
    describe the time series measured in every individual.. a four-dimensional
    multivariate normal random effect is used - so not all those parameters are
    estimated freely..

    Still, I would now just like to brute-force this thing into convergence, adapt
    for half a million iterations, if necessary.. Computationally, it's feasible.
    Problem: very rarely I get an error that the log-likelihood cannot be
    calculated, always for a different node.

    Is there a way to run an error-tolerant version of "adapt" in rjags.. a way
    where errors are just ignored, rather than having the function exit .. after
    all, the chain is not used for inference during adaptation.. ?

    thanks

    Michael

     
  • Martyn Plummer
    Martyn Plummer
    2012-08-31

    It may be inconvenient, but I think it is a bad idea to gloss over numerical
    problems. Failure to calculate the likelihood indicates a fairly serious
    numerical problem. It may be a problem in your model (Gelman's folk theorem)
    or a bug in the software that should be fixed.