Understanding reparameterization of nonlinear hierarchical models with brms

Interfaces brms
,
4

Hi Paul,

Thanks so much for the response! Sorry to be a bit slow to get back to you - I’ve been away this week.

To follow up:

  1. I understand that the z_* parameters are not directly utilised for interpretation of the model, but, from my understanding, this is an important place to check for pathological posterior geometry when making use of the non-centred parameterisation in hierarchical models. From the case study I mentioned earlier, the z_* parameters represent eta[*], which is shown here and here. Or should we be including z_* parameters in our posterior with additions to the generated parameters code block if we really want to look into them?

  2. Is there a way to apply a centred parameterisation with brms in which not only the location, but the scale is also centred (i.e. no z_* parameters ), either using lf() or stanvar(), or some other way? Or ought I to start look into modifying the generated STAN code and running it with cmdstanr directly? Of course brms makes everything so much easier and so much more convenient, so I’d love to stay with it if at all possible, but I also understand that supporting any kind of model configuration or parameterisation is highly burdensome for you as the maintainer of the package. Or is something planned or in the works? If so, I could probably help to contribute if there were to be something planned (especially allowing centring of some random effects and not of others).

Thanks so much for the help!

2 Likes