Bounded variable transformation when not using HMC?

Which sampler are you using? Setting tune=0 might mean the initial proposal size is not too bad (and it never had the opportunity to explore something worse… or better). If your step sampler is not too complicated (Metropolis or Slice) you can hand-tune the proposals yourself. For NUTS that would be quite an enterprise :smiley:

And yes, it may be fine to disable the transforms. The default are there mostly for use with NUTS which cannot cope at all with sharp boundaries in the logp. Other samplers are much more okay with that and it may justify not using a default transform. There is nothing “sacred” about them.

I think you are not veering anywhere too dangerous. Just be cautious to run many chains at different initial points to confirm they converge well. You loose a lot of “warning diagnositics” by not using NUTS so it’s easier to be blindsided.

2 Likes