The current version of the Sontag model seems to be having convergence problems. The total log-likelihood is pretty variable, and parameters do not state at the generative parameters.
In particular, B seems to be sampling from 0 to 1 randomly and B0 similarly behaves as if the only force it feels is from the monotonicity constraint. Z looks to all get too low while the leak terms L get too high.
It is not 100% clear whether this is caused by a bug in the samplers/distributions, just problematic behavior from a shallower probability landscape, or maybe even issues interfacing with or directly from pymc3 (we love the dev team, but we've run into some bugs already from using multidimensional transformed distributions).
The current version of the Sontag model seems to be having convergence problems. The total log-likelihood is pretty variable, and parameters do not state at the generative parameters.
In particular, B seems to be sampling from 0 to 1 randomly and B0 similarly behaves as if the only force it feels is from the monotonicity constraint. Z looks to all get too low while the leak terms L get too high.
It is not 100% clear whether this is caused by a bug in the samplers/distributions, just problematic behavior from a shallower probability landscape, or maybe even issues interfacing with or directly from pymc3 (we love the dev team, but we've run into some bugs already from using multidimensional transformed distributions).