Heterogeneous Effects

This notebook fits a CATE model on simulated data with a known per-unit effect. On a second cohort with no heterogeneity, the in-sample top group overstates the truth, while held-out re-estimation recovers it and targeting_rule refuses to target.

Open the notebook in a full browser tab

References and assumptions

The heterogeneous-treatment-effect discussion follows Chernozhukov et al., “Generic Machine Learning Inference on Heterogeneous Treatment Effects in Randomized Experiments”. Segment effects and targeting decisions require separate validation on held-out or otherwise independent data; a credible overall average effect does not validate a subgroup ranking or targeting rule.