Plot_ppc() with discrete data - How to control bins?

Questions
1

So I have some count data that is overdispersed and I want to visually show how a negative binomial model is giving a better fit than a poisson model. The issue is that plot_ppc() is using varying bin sizes in an inconsistent manner and I can’t seem to find a way to pass some sort of kw_arg to fix this.

Here are some example images so you can see what I mean:
Poisson plot:

image
image731×491 23.7 KB

Negative Binomial plot:

image
image731×491 24.7 KB

You can see how most of the Poisson distributions are using more bins than the observed data and the Negative binomial distribution. If you look closely though you can see some of the light-blue/alpha Negative binomial fits use more bins, and some of the Poisson ones use fewer bins.

Does anyone know of some way that I can make the number of bins fixed/consistent? This doesn’t seem to be an issue when I plot the cumulative distributions, which is fine from a technical point of view, but I really like these pdf views as they make you appreciate how the Poisson fit finds the mean of the distribution pretty well but it struggles with the spread of the data since for Poisson mean=var.

Here is some example code of how I am creating the above plots right now:

with pois_model:
    ppc = pm.sample_posterior_predictive(pois_trace)
    ax=az.plot_ppc(az.from_pymc3(posterior_predictive=ppc, model=pois_model),
                   kind='kde', var_names=['home_points'], num_pp_samples=2000)
    ax[0].set_xlim(-5, 20)
1 Like
2

It is not possible right now, one (or maybe a couple) kwarg arguments should be added to make that possible.

If you are interested in working on this I can guide you through the process to submit a PR to ArviZ.