Posterior Predictive for hierarchical model without further hyper-parameter inference

[maxbiostat]

@Heiko_Schuett

Of course I can simulate new groups and observations, but that is not what I want to do.

What you want to do is not posterior prediction, then. If you really want to learn the group parameters only for the new data, you either fix the nuisance parameters to their posterior means/medians or approximate their posteriors using some parametric family and the plug those in as informative priors.