Michael's Wiki

Notes on model diagnostics

What we really want to see is the posterior probability of the observed data after each iteration.

  • After each step, what is the likelihood of the observed data as explained by the model
  • Should show convergence behavior

How do we get that out of PyMC3….?

  • We don't. PyMC3 is too high-level for this analysis.
  • PyMC3 is performing automatic MAP and sample method selection
  • PyMC3’s step_methods submodule contains the following samplers: NUTS, Metropolis, Slice, HamiltonianMC, and BinaryMetropolis