search

Genentech / jmpost

https://genentech.github.io/jmpost/
17 stars 4 forks source link

Enable "constant" priors #266

Open gowerc opened 8 months ago

gowerc commented 8 months ago

Enhancement idea raised during the PR (#265 ) for the SF model. Idea being that the SF model is just a simplification of the GSF model which potentially could be better implemented via "constant priors" rather than re-writing out the whole model.

gowerc commented 8 months ago

@danielinteractive , I guess its worth pointing out that as far as I can see Stan has no direct support for this. That is it would still require is to implement our own dynamic code. But alas potentially yes I see that if we did implement this it would allow people to specify constants rather than distributions for specific parameters. That being said can the same not already be achieved via very narrow priors or does this lead to model stability issues?

e.g.

par ~ U(2.999999, 3.00000001)
danielinteractive commented 8 months ago

Correct, Stan does not support this directly. We could learn from brms the dynamic code path to enable this in practice in jmpost. I don't think these "spike" priors will work well for the computations/ derivatives inside Stan etc.