ROLLDID — Lee & Wooldridge rolling-transformation DiD

Path A (empirical), both applications of Lee & Wooldridge (2026), Simple Approaches to Inference with Difference-in-Differences Estimators with Small Cross-Sectional Sample Sizes. ROLLDID is a clean-room MIT implementation written from the paper’s equations; the AGPL lwdid package was used during development only as a black-box oracle (its point estimates agreed to ~4 decimals) and shares no code with mlsynth.

California Prop 99 (common timing, Table 3)

Data: basedata/smoking_data.csv — the Abadie, Diamond & Hainmueller (2010) panel, 39 states × 1970–2000, California treated from 1989; outcome = log per-capita cigarette sales, 38 never-treated controls.

Quantity	ROLLDID	Paper (Table 3)
Demeaning (Proc 2.1) avg ATT	−0.422 (se 0.121)	−0.422 (0.121)
Detrending (Proc 3.1) avg ATT	−0.227 (se 0.094)	−0.227 (0.094)
Detrend exact \(p\)	0.021	0.021
\(\tau_{1989}\) (demean)	−0.168	−0.168
\(\tau_{2000}\) (demean)	−0.667	−0.667
\(\tau_{2000}\) (detrend)	−0.403	−0.403

Castle laws (staggered rollout, §7.2)

Data: basedata/castle.csv — Cunningham (2021), 50 states × 2000–2010, 21 states adopting “castle” laws across 2005–2009 and 29 never-treated; outcome = log homicides per 100k. The cohort-share-weighted aggregate (eq. 7.18–7.19) uses never-treated states as the comparison.

Quantity	ROLLDID	Paper (§7.2)
Demeaning aggregate ATT	0.092 (OLS se 0.057)	0.092 (0.057), t ≈ 1.61
Demeaning HC3 \(t\)	1.50	1.50
Detrending aggregate ATT	0.067 (HC3 se 0.055)	0.067 (0.055), t ≈ 1.21

Reproduce

python benchmarks/run_benchmarks.py rolldid_lw

The durable case is benchmarks/cases/rolldid_lw.py; unit-level reproduction is pinned in mlsynth/tests/test_rolldid.py. Both run on the in-repo data, so no network or external reference is required.