Stroke case-study notebook (reproduces arXiv:2606.12623 on synthetic data)

[oduerr]

Adds notebooks/stroke_case_study.py — reproduces the case-study paper Estimating Individualized Treatment Effects in Acute Ischemic Stroke with TRAM-DAG (arXiv:2606.12623) on the public synthetic magic-mrclean cohort (the clinical MAGIC/MR CLEAN data is private).

Mirrors the paper's pipeline: fit one flexible TRAM-DAG on the confounded observational arm (Age/NIHSSa nonlinear, mRS_pre/T linear, ordinal outcome), then predict per-patient ITEs on the randomized trial arm and validate. Reproduces:

• Fig. 5 — ITE distribution, all positive, ATE = mean ITE +0.094;
• Table 3 — ATE consistency: TRAM-DAG +0.094 vs known true +0.104 vs synthetic trial +0.094; naive confounded contrast +0.301;
• Fig. 6 — model-simulated trial outcome shift under do(T=0/1);
• Fig. 7 — treated-arm discrimination (observed good-outcome frequency rises with predicted probability).

Plus the validation the real study explicitly could not do (the fundamental problem of causal inference): with a known SCM, predicted vs true per-patient ITE — rank-corr 0.83, and the saturation of predicted ITE vs the SCM's true age-heterogeneity cleanly exposes the homogeneous-T modeling choice the paper flags as future work.

Self-contained (pip install tramdag, uses the packaged generator); validated end-to-end headless, all figures render. jupytext .py (the .ipynb stays git-ignored).

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com