I need help to understand what methods I can use to evaluate my fitted model if I don’t know how to sample from the custom distribution I implemented. In a bit more detail,
A lot of times, I have to implement a custom likelihood function because the distribution is not implemented in PyMC, e.g., even though Wald distribution is implemented in PyMC, a recently proposed censored shifted Wald distribution is not. So for such custom distributions, I don’t always know how to sample from the distribution. Hence, I am not able to use functions such as pm.sample_posterior_predictive for model fit evaluation because it throws a NotImplementedError, which is correct because I have not implemented the random function.
However, this has become a big roadblock for me to try new models/likelihood functions using Bayesian inference methods. Hence, please advise if there is a way in PyMC to do a model fit evaluation, e.g., posterior predictive analysis or some other methods, without knowing how to generate samples from the custom distribution/likelihood function.