test: anchor LPDiD regression-adjustment SE against Stata teffects (first Stata parity arm)

[igerber]

Summary

• Adds the repo's first Stata parity arm to externally anchor the LP-DiD (Dube, Girardi, Jorda & Taylor 2025) regression-adjustment (RA) standard error, which had no runnable R reference and was previously pinned against itself.
• benchmarks/stata/generate_lpdid_ra_golden.do independently reconstructs each horizon's clean sample (porting the R prep/clean_h recipe to tsset + L./F.; tsfill mirrors R fill_gaps() across the deliberate interior gap) and runs teffects ra (Dy x i.time) (tdiff), atet vce(cluster unit).
• Result: the library RA IF SE matches teffects to ~1e-16 at all 7 event-study horizons; the SE ratio is a flat 1.0 (no G/(G-1), no (n-1)/(n-k)) — confirming the no-finite-sample-factor convention.
• teffects is native to Stata (no SSC dependency), so version 19 pins numerical behavior. Stata is node-locked and cannot run in CI, so — exactly like the R arm — the golden is committed and the test skips when absent; the CI-runnable guards (RA_SE_PIN + coverage_lpdid_ra.py) are unchanged.
• No diff_diff/ source changed.

Methodology references

• Method name(s): LPDiD regression-adjustment event-study SE (covariate path, reweight=True).
• Paper / source link(s): Dube, Girardi, Jordà & Taylor (2025), Local Projections Difference-in-Differences — footnote-9 Stata recipe teffects ra (Dhy i.time) (dtreat) if D.treat==1 | Fh.treat==0, atet vce(cluster unit). Reference implementation: Stata teffects ra (native).
• Intentional deviations from the source: none introduced. REGISTRY LPDiD Deviation 2 is upgraded from inferred (read off the authors' .do comments) to measured: the no-finite-sample-factor RA convention is now confirmed against a runnable teffects to machine precision.

Validation

• Tests added: tests/test_lpdid_ra_stata_parity.py (22 cases) — three gates per horizon, each a purely absolute bound: point vs the R-anchored ra_cov[h][0] (1e-10), Stata (e(N), e(N_clust)) vs the library (n_obs, n_clusters) (exact), and library SE vs teffects SE (1e-7 cross-platform; ~1e-16 local). Strong corroboration of the same realized clean sample (not a formal membership proof).
• Golden regenerates byte-identically; existing LPDiD R-parity suite (RA_SE_PIN) and doc-IA / doc-deps / refresh gates all pass.
• Notebook evidence: N/A.

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01X9m4hRxwqRnEJsv5ZhHY3i

[github-actions]

Overall assessment: ✅ Looks good

Executive summary

• No estimator or default behavior changed.
• Stata sample construction matches the paper, Registry, and existing R recipe.
• Golden ATETs, sample sizes, clusters, and IF SEs independently reproduce within (10^{-15}).
• No P0/P1 findings.
• Full pytest execution was unavailable because the review image lacks dependencies.

Methodology

• P3 — Documented implementation choice
  • Impact: The no-finite-sample-factor RA variance is explicitly documented and externally anchored against Stata at REGISTRY.md:L2130. It is not a defect.
  • Concrete fix: None required.

The post/pre clean-control filters and teffects ra ... atet vce(cluster unit) specification at generate_lpdid_ra_golden.do:L85 conform to the cited methodology.

Code Quality

No findings.

Performance

No findings. The library fit is module-scoped and reused across parameterized tests.

Maintainability

No findings.

Tech Debt

• P3 — Tracked
  • Impact: The optional R reference remains explicitly tracked in DEFERRED.md:L60; it does not block approval.
  • Concrete fix: None required.

Security

No secrets, unsafe external commands, or sensitive data found.

Documentation/Tests

• P3 — Minor test hardening
  • Impact: The new point gate compares Stata against R, while library_fit retains only SE/counts at test_lpdid_ra_stata_parity.py:L88. Direct Python–Stata point agreement relies on the existing, looser R-parity test.
  • Concrete fix: Optionally retain each library coefficient and compare it directly with the Stata ATET at POINT_ATOL.

Static syntax, JSON validity, diff hygiene, and documentation dependency paths passed.

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: cf1cd76cc1a098306a6788c3968395f5bb821c42

---

Overall assessment: ✅ Looks good

Executive summary

• No estimator or default behavior changed.
• Stata sample construction matches the paper, Registry, and R recipe.
• Golden points, sample counts, clusters, and SEs are internally consistent.
• The prior direct Python–Stata point-comparison gap is resolved.
• No P0/P1 findings.

Methodology

• P3 — Documented implementation choice
  • Impact: The no-finite-sample-factor RA variance is explicitly documented at docs/methodology/REGISTRY.md:L2130 and externally anchored across seven horizons. The post/pre filters at benchmarks/stata/generate_lpdid_ra_golden.do:L105-L130 match the cited clean-sample recipe.
  • Concrete fix: None required.

Code Quality

No findings.

Performance

No findings. The library fit is module-scoped and reused.

Maintainability

No findings.

Tech Debt

• P3 — Tracked
  • Impact: The optional R implementation remains tracked at DEFERRED.md:L60.
  • Concrete fix: None required.

Security

No secrets, unsafe external commands, or sensitive data found.

Documentation/Tests

• P3 — Prior informational item resolved
  • Impact: tests/test_lpdid_ra_stata_parity.py:L159-L171 now directly compares recomputed library coefficients with Stata ATETs.
  • Concrete fix: None required.

Static Python syntax, JSON validity, diff hygiene, sample counts, and committed numerical anchors passed. Full pytest execution was unavailable because the review environment lacks NumPy, pandas, SciPy, and pytest.