It seems both the book and the Numpyro example are using the Laplace Approximation, whereas in PyMC3 we are using full MCMC + NUTS. I would expect the PyMC3 posterior to be more accurate than either of those two.
It seems both the book and the Numpyro example are using the Laplace Approximation, whereas in PyMC3 we are using full MCMC + NUTS. I would expect the PyMC3 posterior to be more accurate than either of those two.