Problem to reproduce Figure 1.5 about bookcode_edition1 HOT 4 OPEN

If your postererior_predictive is InferenceData type, then try something like: az.plot_dist(posterior_pred_values.sum('y_obs_dim_0').isel(chain=0).y_obs). Each observation is a Bernoulli, but there is a coordinate y_obs_dim_0 (or something similar) whose length is the number of trials. If you sum over this axes, then you get Binomial samples with n=20 (or however many samples were in your observed data used for the likelihood function).

from bookcode_edition1.

Thank-you, @brendan-m-murphy that was very helpful.

The following with var_names appears to work in PyMC 5:

pred_dists = (
    pm.sample_prior_predictive(1000, model).prior_predictive["y_obs"],
    pm.sample_posterior_predictive(idata, model, var_names = ["y_obs"])
)

Likewise,

pm.sample_prior_predictive(1000, model, var_names=["y_obs"]).prior_predictive

with the var_names arg seems to work as well.

Cheers,
Avram

from bookcode_edition1.

Is there erratum for the book using the pymc library?

The book code on page 12 (Code 1.7) below gives an error of: KeyError: 'y_obs'

pred_dists = (                                                                                                                                       
    pm.sample_prior_predictive( 1000, model)["y_obs"],                                                                                                    
    pm.sample_posterior_predictive(idata, 1000, model)["y_obs"]                                                                                           
)                                                                                                                                                        

from bookcode_edition1.

@aaelony The "sample_ ... _predictive" functions return arviz inference data, so you need to use e.g. pm.sample_prior_predictive(1000, model).prior_predictive["y_obs"] (I think.)

Also, for PyMC 5, it seems like you need to use pm.sample_posterior_predictive(idata, model). There is no argument to specify the number of samples (I think it just gives you the same number of posterior predictive samples as samples in the idata, so 1000 samples per chain by default).

from bookcode_edition1.