I am fiting a model of about 10 parameters and I obtain good ‘Intra’ convergence of the chains but not ‘inter’ convergences for all parameters.
I try different things but I never reach \hat{R}<1.1 for all the parameters.
My understanding is that for some of the model parameters, there are different spaces that give me good model fits (not a unique space solution). In this case, I feel that my problem is more related to modelling rather than MCMC.
I would like to have your insigths on how do you deal with a model with some “sloppy” parameters when you want to do bayesian inference?
Hi,
I don’t think your question can be answered in such a general form beyond “we think hard and try stuff out” :-) Everything beyond this is model specific, so sharing your model is preferable (possibly in a new topic with more descriptive name). Divergent transitions - a primer has a list of hints for degenerate models (the causes of divergences have substantial overlap with causes of high Rhat).
That’s very likely correct.
That’s a likely possibility (the keyword would probably be “multimodal” or “non-identified” or “label switching” for mixture models) but a lot of other stuff can be happening, especially if you have increased adapt_delta or max_treedepth above their defaults.