Finally A Way to Model Discrete Parameters in Stan

Modeling
18

I’d suggest something like

  • simple mixtures (one discrete parameter per data point)

  • change point (one discrete parameter with 200 possible values)

  • HMM decoding (lots of linked discrete parameters); that is, set up training data (x, y), use it to estimate parameters p(theta | x, y), then give it some new x’ and let it predict new y’.

  • Cormack-Jolly-Seber mark-recapture : the animals live/death state is discrete and marginalized out in the Stan example

  • Dawid-Skene model – discrete parameter per item (not quit per data point), marginalized out in the Stan model

2 Likes