paulgstf
commented
4 years ago First pass at responses.
-
prior predictive distribution of a datapoint versus prior distribution of parameters.
-
Ditto. posterior predictive for y_next | y_collected, versus posterior for parameters given y_collected.
-
Yes, the posterior predictive checks advocated by Gelman et al. do receive some criticism for using the same data twice.
-
Hmmmm, that error message is a bit cryptic.
-
Not sure, think because we are "tracing" the MCMC output across iterations.
prior predictive distribution vs prior distribution:
distribution of y values vs distribution of parameters before data given
posterior predictive distribution vs posterior distribution
distribution of y values vs distribution of parameters after data given. PPD has wider range of samples since it accounts for uncertainty of parameters ?
posterior predictive check compare the given data, and the new samples from PPD. are we using the data twice, once for obtaining the posterior distribution, twice for getting PPD?
Mike's variational Bayes gives us trace and ppc while MCMC giving me an error: Not enough samples to build a trace. Is it an advantage of VB? being able to get samples in some not good enough situations? what does 'Not enough samples to build a trace.' mean? The 'samples' referring to the original data set? (I doubled the sample size to produce but got the same error).
why do we call trace as trace (trivial)
sometimes the process of sampling (Variational Bayes) stopped in the middle and lagged. Or sometimes it works.... I am confused. Is it a a. converging issue, not convergence -> infinite loop? b. any other reason? c. is it common thing to handle when we do Bayesian analysis ?
Adding time variation, overdose probability p is increasing from Mike's plot while mine is not. I am wondering it's just my coding issue (maybe something I did wrong) or a difference can be made from MCMC and VB.
(link for questions problem)