[R-sig-eco] correlated parameters, absolute model fit, and MCMC
Nicholas Lewin-Koh
nikko at hailmail.net
Sat Dec 20 05:07:51 CET 2008
Hi Dan,
Since you are fitting a fully Bayesian model, the output from your MCMC
sample
is a sample from the joint posterior distribution. So given that your
chain is mixing
properly and converged (a whole nother kettle of fish) I don't see why
the correlation
is a problem. You can assess the model fit using posterior predictive
checks,
ie from the joint posterior distribution you can generate the posterior
predictive
distribution. There are may papers on this a good place to start is
Gelman et al. (2004) "Bayesian Data Analysis"
Hope this helps
Nicholas
----------------------------------------------------------------------------
Hello -
I'm investigating model adequacy for several stochastic models; I've
obtained posterior distributions for all parameters using approximate
MCMC. I'd now like to know, in an absolute sense, how good these
parameters actually are. In other words, given my data D, and a set
of summary statistics on those data S(D), I would like to simulate
date D' under the model parameters and look at the distribution of S
(D') and see how well it matches up with S(D). This should be a
straightforward evaluation of absolute model fit via parametric
simulation.
However, because some of the parameters are correlated, I cannot
simply use the parameter set with the maximum posterior probability.
Nor can I use maximum likelihood estimates for the parameters (I am
using approximate MCMC because no likelihood function can be
specified, but the simulations are easy, and the data are well-
characterized by several summary statistics).
Does anyone have any suggestions on how to deal with correlated
parameters like this, or alternative approaches for evaluating model
fit?
One solution would be simply to run simulations with parameters
sampled at random from the complete converged chain (or set of
chains), with the constraint that I would have to sample all
parameters from a given generation simultaneously. Parameter sets
would thus be sampled roughly in proportion to their joint posterior
probability. But I wonder whether there may be other approaches out
there that I should investigate.
This may be more stats than R, but it is an ecological analysis I'm
running in R!
Thanks-
~Dan Rabosky
More information about the R-sig-ecology
mailing list