Short version: Would someone like to join a sort-of-adversarial collaboration with me on a PyMCon talk? I have built a simple version of a model using a grid algorithm. If someone wants to build a better model with PyMC, we could compare results.
I am thinking of proposing a talk for PyMCon based on an example I used in Probably Overthinking It. The context is missing value imputation using item response theory. My implementation uses a grid algorithm, which is possible because I treat the “difficulty” parameters as fixed and estimate the “efficacy” parameters individually for each survey respondent.
I would be curious to see what difference it makes if we build a fully Bayesian version of the model that estimates difficulty and efficacy simultaneously. I think the results of the comparison would make a nice talk!
I can provide my implementation as a starting point and we can brainstorm options for the better model, but I’d need someone comfortable with implementing the better model in PyMC.