Hey, thank you for taking your time to have a look at this. In what way does the Metropolis sampler behave inefficient in this case?
Even the posterior sample for ‘pm_event_prob’ is somewhat contradictory to what I expect. For once, despite a high number events and experts, the shape is very much dependent on the prior I pass for it. If I give a uniform prior, the histogram is always triangular, while for a different beta prior it is smoother. Additionally, independent of the event and the consensus per event, the posterior of ‘pm_event_prob’ is always taking on one of two “shapes”, with most probability mass at 0 or 1, but there are no gradual differences which would indicate higher or lower uncertainty.
Is my model misspecified or am I missing something about the internal mechanics of pymc3?