Module 5 — Modern Causal Analysis
Use flexible prediction, heterogeneity, sensitivity and transportability without weakening identification discipline
Module 5 — Modern Causal Analysis
Modern causal methods are most useful when they separate two tasks:
Machine learning helps with the second task. Cross-fitting and orthogonal scores reduce sensitivity to nuisance-model error; they do not remove unmeasured confounding, interference or poor overlap.
Prepare
Bring one causal design from Modules 0–3. Mark:
- the target parameter;
- the treatment-assignment assumption;
- nuisance functions such as outcome regression or propensity score;
- variables available before treatment;
- intended target population.
Route
- Double Machine Learning: orthogonalisation and cross-fitting.
- Heterogeneity and Policy Learning: CATEs, honest discovery and constrained decisions.
- Sensitivity and Partial Identification: replace binary robustness claims with calibrated ranges.
- External Validity: move an effect across populations, treatments and contexts.
Workshop: model card for a causal estimator
Teams construct a model card with six fields:
| Field | Required statement |
|---|---|
| estimand | ATE, ATT, CATE, policy value or transported effect |
| identification | experiment, conditional exchangeability, IV, DID or other design |
| learners | nuisance functions and why flexibility is needed |
| sample splitting | training, tuning and estimation roles |
| stress tests | overlap, split stability, sensitivity and subgroup support |
| decision boundary | cost, capacity, fairness and target population |
Follow-up deliverable
Submit either a cross-fitted estimate with a split-stability audit or a calibrated sensitivity analysis. Postgraduate work should state the orthogonal moment or identified set and explain the rate/regularity conditions in words.