save deviance or log probability for the entire model

[Quresh Latif]

Hi all. Is there a simple way of monitoring the log probability or deviance for the entire model at every saved posterior sample? I have a large model for which chains are not readily converging. I'd like to look at and compare the log probabilities or deviances of each of my chains to understand if the model is basically not identifiable or if there are local maxima that a chain is getting stuck in, and so if I can possibly nudge the stubborn chain to where it needs to go (e.g., with better initial values). In any case, I'd like to save some overall measure of relative fit at each posterior sample. In JAGS, we can save "deviance". Can we do that in NIMBLE as well? I am having trouble finding that in the manual.

[Daniel Turek]

Quresh, great question.  I think there's a vignette on exactly that.

Cheers,

Daniel

On Tue, Apr 9, 2024 at 10:39 AM Quresh Latif <quresh...@birdconservancy.org> wrote:

Hi all. Is there a simple way of monitoring the log probability or deviance for the entire model at every saved posterior sample? I have a large model for which chains are not readily converging. I'd like to look at and compare the log probabilities or deviances of each of my chains to understand if the model is basically not identifiable or if there are local maxima that a chain is getting stuck in, and so if I can possibly nudge the stubborn chain to where it needs to go (e.g., with better initial values). In any case, I'd like to save some overall measure of relative fit at each posterior sample. In JAGS, we can save "deviance". Can we do that in NIMBLE as well? I am having trouble finding that in the manual.

--
You received this message because you are subscribed to the Google Groups "nimble-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nimble-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nimble-users/0ad80403-62a4-431d-a051-efd30fc9ba52n%40googlegroups.com.