PropensityScore.calculate_impact#

PropensityScore.calculate_impact(y_true, y_pred, response_type='expectation')#

Calculate the causal impact as the difference between observed and predicted values.

By default, the impact is calculated using the posterior expectation (mu) rather than the posterior predictive (y_hat). This means the causal impact represents the difference from the expected value of the model, excluding observation noise. This approach provides a cleaner measure of the causal effect by focusing on the systematic difference rather than including sampling variability from the observation noise term.

Parameters:

  • y_true (xr.DataArray) – The observed outcome values with dimensions [“obs_ind”, “treated_units”].

  • y_pred (az.InferenceData) – The posterior predictive samples containing the predicted values.

  • response_type ({"expectation", "prediction"}, default="expectation") –

    The response type to use for impact calculation:

    • "expectation": Uses the model expectation (μ). Excludes observation noise, focusing on the systematic causal effect.

    • "prediction": Uses the full posterior predictive (ŷ). Includes observation noise, showing the full predictive uncertainty.

Returns:

The causal impact with dimensions ending in “obs_ind”.

Return type:

xr.DataArray

Notes

When using response_type="expectation" (the posterior expectation), the uncertainty in the impact reflects:

  • Parameter uncertainty in the fitted model

  • Uncertainty in the counterfactual prediction

But excludes:

  • Observation-level noise (sigma)

When using response_type="prediction", the uncertainty also includes observation-level noise, resulting in wider intervals.

Note

REFACTOR TARGET: Currently, experiment classes store impacts using the default response_type="expectation" at fit time. When users call plot(response_type="prediction"), the prediction-based impact is calculated on-the-fly in the _bayesian_plot() methods. This approach works but duplicates the decision logic. A future refactor could unify this by either storing both variants or always calculating on-demand.