In the nutshell: I need to build a Bayesian Logistic Regression, setting strictly negative prior distributions for some coefficients and defining them outside the model. How is it possible to do without pm.Bound ?
Description: I need to create a Bayesian logistic regression using pymc3, where some of the coefficients are sure to be below 0. As pymc3 does not provide options to reflect distributions (e.g. use pm.Lognormal which is surely above zero and reflect it by multiplying by -1), I need to make a regression where features with these negative coefficients will be negative (because the coefficients will be positive), so I cannot use pm.glm.GLM.from_formula . However, if I use likelihood = pm.Bernoulli() , I cannot use my prior distributions, as I define them outside the model and thus have to set them like pm.Lognormal.dist(...) . Using such coefficients in the regression gives AssertionError: <class 'pymc3.distributions.continuous.Lognormal'> . What are the possible solutions? Thanks in advance!