Is it possible to have a model in Pyro where some of the parameters are just point-estimated as fixed effects with no prior distribution when using NUTS, while the rest of the parameters are normal Bayesian variables with priors and posteriors? Thank you
that doesn’t really work with MCMC (i.e. not pyro specific).
you can e.g. fit your model with some other inference technique like MAP inference, fix the fixed effects to the point estimates obtained from MAP, and then infer the remaining latent variables using NUTS.
but you can’t do it all in one go because MCMC doesn’t do point estimates.
Thank you for the response! I may also use ImproperUniform (as I understand it, the equivalent of not specifying a prior in Stan), although of course this is not a point estimation as in the original question. This route seems to be common in what I’ve read for “mixed effects models” with MCMC.