I think there is some confusion going on about censored and truncated because… the statistical definition of truncation sounds like it’s doing the common sense definition of censoring (hiding values beyond a range) whereas statistical censoring defines the common sense of truncation (clipping values above or below an edge to that edge). I am following STAN’s definition here: Stan User’s Guide
This discourse discusses an example of exponentially truncated data: Help with pm.Potential
And there is one example of censored data in this Notebook (resources/DataAnalysis.ipynb at master · pymc-devs/resources · GitHub ), section 5.5.
Also we might have an issue going on with both the TruncatedNormal distribution and pm.Bound which affect the gradient. If you think you have the right model and it’s not working try to use a non gradient based sampler such as Metropolis or Slice: pm.Bound and TruncatedNormal generate wrong gradients · Issue #4417 · pymc-devs/pymc3 · GitHub
A tutorial clarifying these things would be really awesome btw!