Why does the error occur in the prior prediction but not in the posterior prediction?

Questions v5
5

A self contained example of this is:

import pymc as pm

with pm.Model() as m:
  sigma = pm.Normal("sigma", 0.001, 1)  # bad prior for sigma
  y = pm.Normal("y", mu=0, sigma=sigma, observed=[-1, 0, 1])

with m:
  try:
      pm.sample_prior_predictive()
  except Exception as exc:
     print(f"{type(exc).__name__}: {exc}")  # ValueError: scale < 0

  idata = pm.sample()
  pm.sample_posterior_predictive(idata)  # Fine

Posterior predictive works because there are only positive draws of sigma, whereas in the unconstrained prior predictive, negative draws are suggested (roughly half the time)

2 Likes