Add Wooldridge ETWFE LinkedIn carousel

[igerber]

Summary

• Add 9-slide LinkedIn carousel PDF announcing the Wooldridge ETWFE estimator (v2.9)
• Light theme with emerald green accent, targeting applied researchers
• Narrative arc: nonlinear DiD gap → Wooldridge (2023, 2025) ETWFE → three likelihoods (OLS/logit/Poisson) → ASF + delta-method SEs → event study → code example → milestone + community credit → CTA
• All claims validated against source code, methodology docs, and external academic literature
• Softened initial claims after validation: "When Linear DiD Isn't Enough" (not "breaks"), "Standard parallel trends may not hold on the level scale" (not "biased ATT")

Methodology references (required if estimator / math changes)

• Method name(s): N/A - no methodology code changes (carousel/marketing only)
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): N/A

Validation

• Tests added/updated: No test changes (PDF generator script only)
• Backtest / simulation / notebook evidence (if applicable): N/A
• Manual verification: PDF renders 9 pages, all text readable, equations render cleanly, no footer overlap

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall Assessment

⚠️ Needs changes — one unmitigated P1 methodology documentation mismatch.

No estimator code or inference code changed in this PR. The review risk is in newly added docs/marketing copy drifting from the Methodology Registry and current API.

Executive Summary

• The estimator implementation itself looks unchanged; I did not find a new numerical regression in ETWFE code paths.
• P1: new Wooldridge/ETWFE copy repeatedly says OLS uses ASF-based ATT, but the registry and source limit ASF to the nonlinear logit/Poisson paths; OLS uses direct β_{g,t} extraction.
• P2: new survey-support copy says all estimators accept survey_design / survey weights, but WooldridgeDiD.fit() has no survey_design parameter and the same PR documents that gap.
• P2: the new SurveyDesign.subpopulation() example in llms-full shows the wrong return contract.
• I did not find a security or performance blocker in the new carousel generator.

Methodology

• P1 WooldridgeDiD / ETWFE is described incorrectly in the new public copy.
  Impact: the changed docs/marketing now describe ETWFE as “ASF-based ATT” for OLS/logit/Poisson, which contradicts both the Methodology Registry and the implementation. The registry limits ASF to the nonlinear extensions, while the OLS path takes the estimated treatment-cell coefficients directly. Affected changed files: ROADMAP.md, docs/choosing_estimator.rst, docs/llms.txt, docs/llms-full.txt, carousel/generate_wooldridge_carousel.py. Source of truth: docs/methodology/REGISTRY.md, docs/methodology/REGISTRY.md, diff_diff/wooldridge.py, diff_diff/wooldridge.py.
  Concrete fix: rewrite the changed copy to distinguish paths explicitly, e.g. “OLS estimates ATT(g,t) directly from the saturated regression; logit and Poisson use ASF-based ATT with delta-method SEs.”

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings. A basic strings scan of the committed PDF did not show obvious secrets or local-path leakage. Full PDF metadata inspection was limited because pdfinfo is not available in this sandbox.

Documentation/Tests

• P2 The new survey-support copy overstates current support.
  Impact: multiple changed files now say all estimators accept survey weights or survey_design, but WooldridgeDiD.fit() does not accept that parameter today. This will cause users/agents to generate invalid code and overstate supported weighting/inference paths. Problematic changed copy: ROADMAP.md, docs/llms.txt, docs/llms-full.txt. Contradictory source in the same PR: ROADMAP.md, docs/survey-roadmap.md, diff_diff/wooldridge.py.
  Concrete fix: change the new copy to say WooldridgeDiD is the current exception, or state the supported count explicitly instead of “all estimators.”

• P2 The new subpopulation() example is wrong.
  Impact: docs/llms-full.txt shows sd_female = sd.subpopulation(...), but SurveyDesign.subpopulation() returns a (SurveyDesign, DataFrame) pair, and the returned DataFrame must be used with the returned design. This will mislead users and LLM consumers of this file. Source of truth: diff_diff/survey.py, docs/methodology/REGISTRY.md.
  Concrete fix: replace the example with sd_sub, data_sub = sd.subpopulation(...) and show fit(..., survey_design=sd_sub) on data_sub.

Path to Approval

1. Fix the Wooldridge/ETWFE methodology wording in the changed docs/carousel so ASF is attached only to the nonlinear logit/Poisson paths, not OLS.
2. Fix the new survey-support copy in the changed files to explicitly exclude WooldridgeDiD until survey support is implemented.
3. Correct the subpopulation() example in docs/llms-full.txt to unpack the returned (SurveyDesign, DataFrame) pair and use the returned DataFrame in fit().

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: bc1beea3d5b7551af49723165d272deb07539fa7

---

Overall Assessment

⚠️ Needs changes — the prior ASF/OLS blocker appears resolved, but one unmitigated P1 methodology mismatch remains in the new carousel copy.

Executive Summary

• The previous P1 about ASF wording looks fixed: the current carousel now separates OLS direct-coefficient ATT from nonlinear ASF ATT.
• The earlier survey-doc findings from the last review are no longer in the current diff; this re-review is limited to the new carousel script/PDF, and no estimator or inference code changed here.
• P1 [Newly identified] Slide 3 presents an OLS-only unit-FE ETWFE equation as the general method while adjacent copy frames the estimator as one framework for OLS, logit, and Poisson.
• P2 The slide 7 code example uses results.aggregate('event').summary(), but summary() defaults to the simple overall ATT table, not the event-study table.
• P2 The generator script depends on PIL and fpdf, but those packages are not declared in pyproject.toml, so the committed PDF is not reproducible from the repo’s declared environment.
• No security blocker found. The committed PDF is structurally 9 pages, and a basic strings scan did not reveal obvious secrets or local-path leakage.

Methodology

• P1 [Newly identified] carousel/generate_wooldridge_carousel.py:L393-L413 shows Y_it = α_i + γ_t + ... as “Wooldridge ETWFE” while the same slide positions the estimator as covering OLS, logit, and Poisson.
  Impact: that equation matches the OLS ETWFE path in the registry (docs/methodology/REGISTRY.md:L1094-L1103), but the documented nonlinear implementation uses additive cohort and time dummies rather than unit fixed effects (docs/methodology/REGISTRY.md:L1152-L1154, diff_diff/wooldridge.py:L735-L739, diff_diff/wooldridge.py:L923-L929). That distinction is methodological, not just cosmetic notation.
  Concrete fix: label the displayed equation as the OLS specification only, or replace it with wording/notation that matches the nonlinear paths as implemented: additive cohort/time effects plus ASF ATT for logit/Poisson.

Code Quality

No findings.

Performance

No findings.

Maintainability

• P2 carousel/generate_wooldridge_carousel.py:L8-L14 imports matplotlib, PIL, and fpdf, but pyproject.toml:L46-L75 only declares matplotlib as an optional dependency; Pillow and the PDF package are undeclared.
  Impact: contributors cannot reliably regenerate carousel/diff-diff-wooldridge-carousel.pdf from the repo’s declared environment.
  Concrete fix: add the exact image/PDF packages to a suitable extra (dev, docs, or a new carousel extra), or add a short carousel/README with the required install command and regeneration step.

Tech Debt

No findings.

Security

No findings. A basic scan of carousel/diff-diff-wooldridge-carousel.pdf showed a 9-page page tree and no obvious embedded secrets. Full metadata inspection was limited because pdfinfo/exiftool are not available in this environment.

Documentation/Tests

• P2 carousel/generate_wooldridge_carousel.py:L680-L682 uses results.aggregate('event').summary(), but diff_diff/wooldridge_results.py:L165-L171 and diff_diff/wooldridge_results.py:L203-L214 default summary() to aggregation="simple".
  Impact: a user who copies the slide code will see the simple overall ATT table, not the event-study output the slide is advertising.
  Concrete fix: change the example to print(results.aggregate('event').summary('event')) or split it into results.aggregate('event') followed by print(results.summary('event')).

Path to Approval

1. Update slide 3 in carousel/generate_wooldridge_carousel.py:L393-L413 so the displayed model matches the documented estimator variants: either explicitly label it as the OLS ETWFE equation, or replace it with copy that correctly states logit/Poisson use additive cohort/time effects plus ASF ATT. Then regenerate carousel/diff-diff-wooldridge-carousel.pdf.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 8e58af655975b26b3562ed3a4d08d4571d94153d

---

Overall Assessment

✅ Looks good — this re-review clears the prior P1 methodology issue and the prior slide-7 API-example issue. The only unmitigated item I found is P2 reproducibility for the generator environment.

Executive Summary

• The prior P1 is resolved: slide 3 now explicitly labels the displayed equation as the OLS ETWFE specification and separately states the nonlinear logit/Poisson path as cohort+time dummies plus ASF ATT (carousel/generate_wooldridge_carousel.py:L395-L414; docs/methodology/REGISTRY.md:L1094-L1103, docs/methodology/REGISTRY.md:L1152-L1154).
• The prior slide-7 example issue is also resolved: the carousel now uses .summary('event'), which matches the actual API behavior (carousel/generate_wooldridge_carousel.py:L693-L697; diff_diff/wooldridge_results.py:L165-L258).
• No estimator, weighting, variance, or identification code changed in this PR; methodology risk is confined to the claims in the new carousel assets.
• One prior P2 remains unresolved: the new generator imports PIL and fpdf, but those packages are still not declared in pyproject.toml, so the PDF is not reproducible from the repo’s declared environment (carousel/generate_wooldridge_carousel.py:L8-L14; pyproject.toml:L52-L75).
• P3 informational: slide 5’s “Matches Stata jwdid_estat output” wording is slightly broader than the registry-supported claim, because the registry documents a small nonlinear QMLE SE deviation from Stata (carousel/generate_wooldridge_carousel.py:L615-L619; docs/methodology/REGISTRY.md:L1115-L1121, docs/methodology/REGISTRY.md:L1123-L1130).

Methodology

No unmitigated P0/P1 findings. The previous slide-3 methodology mismatch is fixed in both the source and the committed PDF.

• Severity P3 — carousel/generate_wooldridge_carousel.py:L615-L619
  Impact: “Matches Stata jwdid_estat output” reads like exact parity, but docs/methodology/REGISTRY.md:L1115-L1121 documents a small nonlinear QMLE SE difference from Stata, while docs/methodology/REGISTRY.md:L1123-L1130 supports parity on aggregation structure/weighting. This is informational only because the deviation is explicitly documented in the registry.
  Concrete fix: tighten the wording to “Matches Stata jwdid_estat aggregation types/weighting” or add a small footnote pointing readers to the registry note on SE differences.

Code Quality

No findings.

Performance

No findings.

Maintainability

• Severity P2 — unresolved from the prior review — carousel/generate_wooldridge_carousel.py:L8-L14; pyproject.toml:L52-L75
  Impact: the new generator depends on matplotlib, PIL, and fpdf, but the repo only declares matplotlib in optional extras and does not declare the image/PDF packages at all. In the current environment, those imports are absent, so contributors cannot regenerate carousel/diff-diff-wooldridge-carousel.pdf from the project’s declared setup.
  Concrete fix: add the missing packages backing PIL and fpdf to an explicit extra such as carousel or dev, and document the install command; if the team wants to defer that, track it in TODO.md so the limitation is explicit.

Tech Debt

No findings.

Security

No findings. I decompressed the PDF streams and verified the binary has a 9-page page tree, contains the updated slide text, and does not expose obvious local paths or secret-like strings.

Documentation/Tests

• Severity P3 — carousel/generate_wooldridge_carousel.py:L805-L826; tests/test_imputation.py:L1248-L1255
  Impact: this PR adds a new scripted artifact plus committed PDF, but no smoke test covers importability/regeneration or even a basic page-count check. The repo already has precedent for validating marketing claims with tests, so this remains manual-only.
  Concrete fix: add a lightweight optional-deps smoke test that imports the generator and verifies a 9-page export/page tree when the relevant extra is installed, or at minimum add a test that the script’s dependencies are declared.