plotting piecewise ITS results

[glow-27]

Hi, does the .plot() function for piecewise ITS also plot the 94% highest density intervals for the observed vs. fitted vs. counterfactual plot?

Example screenshot attached. Thanks!

[drbenvincent]

Thanks for this question. This should be a simple matter of looking it up in the api docs but it isn't there... so I put this quick issue together #886

[drbenvincent]

Thanks for the question @glow-27. I've now closed #886. Your question actually made me discover some minor plot related errors and shortcomings in the API docs. So right now you'll be able to see much richer information in the API docs for the plot method. We've got another PR which we'll fix soon bringing the same transparency to the other public facing methods.

[glow-27]

Thank you, Benjamin!

[glow-27]

I'd like to ask a follow-up question if you don't mind! Are the HDIs shown on the plot posterior mean credible intervals or posterior predictive intervals? I'd like to get a sense of the uncertainty in the counterfactual predictions, which I believe refers to the posterior predictive distribution.

I don't have the strongest background in Bayesian methods, so forgive me if my interpretations are incorrect. Thank you so much!

[drbenvincent]

I'd like to ask a follow-up question if you don't mind! Are the HDIs shown on the plot posterior mean credible intervals or posterior predictive intervals? I'd like to get a sense of the uncertainty in the counterfactual predictions, which I believe refers to the posterior predictive distribution.

They are the posterior mean credible intervals by default, but there's an option to change. I think I need to write an issue to make that much clearer.

Though this should also be more visible when the api docs pr gets merged.

[Abdeltoto]

Yes — for the Bayesian PiecewiseITS plot, HDIs are intended on the top panel

In current CausalPy, the Bayesian plot builds the top axes by calling plot_xY(...) twice:

• once for fitted mu (plot_hdi_kwargs={"color": "C0"})
• once for counterfactual mu (plot_hdi_kwargs={"color": "C1"})

Those plot_xY calls are what add the mean line + HDI band (default mass is hdi_prob=0.94, from HDI_PROB, i.e. a 94% HDI — not 95%).

So if you’re looking for 94% bands on fitted vs counterfactual in the first subplot: they should already be there as the shaded patches paired with each line in the legend (Fitted / Counterfactual).

If you don’t see obvious bands

• Very tight posteriors → the band can look almost identical to the line at the plot scale.
• Check the legend — the fitted/counterfactual entries should be (line, patch) pairs when HDI shading is present.
• OLS vs PyMC — the OLS path is different; HDI-style credible bands are a Bayesian thing. Make sure you’re using the PyMC/Bayesian workflow when you expect those ribbons.

Inspect the numbers directly

PiecewiseITS.get_plot_data_bayesian(hdi_prob=0.94) returns columns like fitted_hdi_lower_94 / fitted_hdi_upper_94 and the counterfactual counterparts — useful to confirm the interval width even when the ribbon is hard to see.

Docs pointer: get_plot_data_bayesian.

[glow-27]

Thank you, Abdel! Good to know that the params for get_plot_data_bayesian directly connects to plot()

[Abdeltoto]

@glow-27 — quick follow-up on credible intervals vs posterior predictive:

In the current Bayesian plot, the HDI bands for Fitted and Counterfactual are built from mu only (posterior_predictive["mu"]), and plot_xY computes HDIs over those draws. The same objects feed get_plot_data_bayesian (HDI via az.hdi on those mu series).

So in practice these ribbons summarize posterior uncertainty about the mean trajectory (μ) for the fitted vs counterfactual structural predictions — credible bands on μ, not full posterior predictive intervals for observations y that would also fold in observation noise from the likelihood. If you need uncertainty for counterfactual observations as they’d be realized, you’d want predictive draws for y (if your model exposes them), not only mu.

If that doesn’t match your model or how you read the API, @drbenvincent can confirm.

If this answers your question completely, please go ahead and mark this reply as the accepted answer — thanks!

[glow-27]

Thanks for your response! This is very helpful. I will be extracting the posterior predictive draws from my model.