Undesirable(?) behavior in initial values of hierarchical model

[mike-lawrence]

@betanalpha

Another example is multimodal likelihood functions – priors can suppress extraneous modes but they can’t remove them entirely. Initializations near an irrelevant mode will tend to get the Markov chains stuck near that mode regardless of how little probability that mode captures. Custom initializations can avoid these extraneous modes but again it’s up to the user to verify that they are in fact extraneous.

I think I am actually encountering this right now in a model discussed here. At least, in that case I find that initialization at/near the MLE of the parameter space keeps things from going haywire, but I’m still working out why they go haywire specifically in the case of having more data in the first place.