EM/DA/Bayesian IMP/Full Bayesian Estimation

Hi all,

May I ask that what’s the difference and connection between the above inference techniques, especially how the idea of EM algorithm connects/differs from the typical full Bayesian inference approach we did in Stan. Is it possible to implement EM algorithm in Stan?


There aren’t any tools for doing EM with general Stan models that I know of.

I like this EM description by @Bob_Carpenter which is kinda relevant to the question I suppose: https://github.com/bob-carpenter/case-studies/tree/master/monte-carlo-em

I don’t know what DA or IMP is, but I assume they’re different approximations to doing Bayesian inference. They’ll have different assumptions/limitations. The general way to evaluate some sort of Bayesian inference is SBC: https://arxiv.org/pdf/1804.06788.pdf (it’s in the manual too: https://mc-stan.org/docs/2_25/stan-users-guide/simulation-based-calibration.html)

Thanks so much for your help!