Week 4 — Causal Inference II: DiD, RD, Synthetic Control, and Shift-Share

The modern panel/quasi-experimental toolkit — and the recent literature showing the “obvious” estimators are often wrong. See ../../TOC.md for the full plan.

Chapters

1. Ch 4.1 — Difference-in-Differences & Event Studies (parallel trends, leads/lags)
2. Ch 4.2 — The Staggered-Adoption Crisis (Goodman-Bacon, negative weights, Callaway–Sant’Anna, Sun–Abraham, BJS)
3. Ch 4.3 — Regression Discontinuity (sharp/fuzzy, local polynomial, IK/CCT bandwidth, McCrary)
4. Ch 4.4 — Synthetic Control & Synthetic DiD (Abadie, placebo inference, Arkhangelsky et al.)
5. Ch 4.5 — Bartik / Shift-Share Designs (GPSS shares vs. BHJ shifts; Rotemberg weights)

Notebooks (../../../notebooks/week-04/)

• nb4.1 Event-study & parallel-trends tests · nb4.2 Staggered DiD: TWFE vs. Callaway–Sant’Anna (differences) · nb4.3 RD with rdrobust (CCT) · nb4.4 Synthetic control & synthetic DiD (pysyncon) · nb4.5 Bartik decomposition
• All verified to run headless. nb4.2 reproduces the TWFE sign-flip (+0.40) vs. CS (−2.571); nb4.5’s Rotemberg weights reassemble the 2SLS estimate (−1.50).

Problem sets (solutions in Appendix E)

• PS 4.1 2×2 DiD & event study · PS 4.2 Goodman-Bacon decomposition · PS 4.3 RD & bandwidth · PS 4.4 Synthetic control & placebo · PS 4.5 Shift-share & identification critique

Lab, mentor, assessment

• Lab 4 — A Clean DiD on HMDA + a State Policy Shock (real CFPB HMDA path + seeded synthetic fallback; Callaway–Sant’Anna)
• Mentor Session 4 — “Detecting discrimination with a clean design” (Sun & Gao 2019, PNAS)
• Week 4 Assessment + Rubric