I sample my model with two chains and 600 draws each. Then I sample the
posterior predictive, using
pm.sample_posterior_predictive() (or its speedy cousin
pm.fast_sample_posterior_predictive()). The posterior predictive dataset has one chain and 1200 draws.
Why does sampling the posterior predictive use only one chain instead of the two in the posterior? Maybe I shouldn’t care about this: 1x1200 is maybe equivalent to 2x600. But a subsequent call to
pm.summary() on the posterior predictive group makes numerous warnings about the change in shape:
Should these warnings be ignored?