Add dCDH estimator LinkedIn carousel

[igerber]

Summary

• Add LinkedIn carousel PDF announcing the dCDH (de Chaisemartin-D'Haultfoeuille) estimator
• 9-slide carousel with slate+teal palette: hook, problem statement with treatment timeline visual, real-world scenarios, DID_M equation, feature grid, dynamic event study plot, code example, R validation parity, and CTA
• Generator script produces the PDF via fpdf2 + matplotlib

Methodology references (required if estimator / math changes)

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

Validation

• Tests added/updated: No test changes
• Backtest / simulation / notebook evidence (if applicable): N/A

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall Assessment

⚠️ Needs changes

Executive Summary

• No estimator/runtime code changed in this PR; the review findings are about the new carousel’s public methodology/API claims.
• Slide 8 overstates validation evidence: it claims 245 dCDH tests and blanket ATT/SE parity with R, but the cited files currently define 173 tests by source count (158 + 15), and the parity suite intentionally does not assert mixed-direction SE parity.
• Slide 7’s code sample is not accurate as shown: with L_max=3, dCDH suppresses the joiner/leaver decomposition, and the repo’s plotting entrypoint is plot_event_study(results), not results.plot_event_study().
• Slide 5/6’s placebo wording (“validate/validated”) is stronger than the project’s own documented guidance on parallel-trends diagnostics.
• I did not find performance or security issues in the generator/PDF beyond a normal PDF open action.

Methodology

• Severity: P1. Impact: The validation slide in carousel/generate_dcdh_carousel.py:L837-L880 says “ATT and SE parity with R’s DIDmultiplegtDYN” and hardcodes 245 tests, but the committed parity suite only checks mixed-direction point estimates (not SEs) in tests/test_chaisemartin_dhaultfoeuille_parity.py:L91-L180, only checks multi-horizon SE parity on pure-direction scenarios in tests/test_chaisemartin_dhaultfoeuille_parity.py:L221-L344, and only covers covariate/trend parity on joiners-only scenarios in tests/test_chaisemartin_dhaultfoeuille_parity.py:L365-L466. The underlying mixed-direction SE limitation is already tracked in TODO.md:L59-L60, so the implementation limitation itself is mitigated; the problem is the new slide copy overclaiming past that documented boundary. Concrete fix: rewrite slide 8 to the tested scope and correct the count (173 in the two cited files), or remove the numeric/blanket parity claim entirely.
• Severity: P2. Impact: The carousel says dynamic placebos “validate” parallel trends / “pre-trends validated” in carousel/generate_dcdh_carousel.py:L683-L699 and carousel/generate_dcdh_carousel.py:L757-L763. That is stronger than the repo’s own guidance: docs/llms.txt:L18-L23 says insignificant pre-trends do not prove PT, docs/methodology/REGISTRY.md:L197-L198 says low power does not validate PT, and dCDH’s own results text only states the one-way implication in diff_diff/chaisemartin_dhaultfoeuille_results.py:L724-L731. Concrete fix: rephrase to “placebos near zero support/are consistent with the design” rather than “validate/validated.”

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• Severity: P3. Impact: The mixed-direction SE-parity limitation is already tracked in TODO.md:L59-L60, so I am not treating the estimator implementation difference itself as a blocker. Concrete fix: none required in estimator code for this PR; just align the new slide copy with the tracked limitation.

Security

• No findings. I did not see PDF markers for JavaScript, launch actions, URIs, or embedded files; only a standard /OpenAction entry was present.

Documentation/Tests

• Severity: P2. Impact: The slide 7 example in carousel/generate_dcdh_carousel.py:L767-L833 is inaccurate in two ways. First, with L_max >= 1, dCDH suppresses the joiner/leaver decomposition by design in diff_diff/chaisemartin_dhaultfoeuille.py:L2293-L2333, and that behavior is explicitly tested in tests/test_chaisemartin_dhaultfoeuille.py:L1844-L1865, so “Joiners, leavers, placebos, event study - all from one fit() call” is false for the shown L_max=3 example. Second, I could not find a plot_event_study() method on ChaisemartinDHaultfoeuilleResults; the documented plotting API is the top-level helper shown in diff_diff/visualization/_event_study.py:L138-L157. Concrete fix: either switch the example to Phase 1 (L_max=None) if you want joiners/leavers/placebo in one result object, or keep L_max=3 and limit the caption to event-study/placebo outputs; in both cases, replace results.plot_event_study() with plot_event_study(results).

Path to Approval

1. Update slide 8 copy and regenerate the PDF so the validation claims match the committed evidence: correct the hardcoded test count and remove blanket SE-parity wording.
2. Fix slide 7’s code sample/caption to match the current dCDH API and L_max behavior.
3. Reword the placebo/pre-trend language to diagnostic phrasing rather than proof-style phrasing.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 881f50b0d6c0566b2e01b5dd42a5eebc7f7ca868

---

Overall Assessment

⚠️ Needs changes

Executive Summary

• Previous re-review items on placebo wording and the plotting API look addressed: the carousel now uses “consistent with parallel trends,” and the code sample uses plot_event_study(results).
• Slide 8 still overstates R validation. It now scopes SE parity to pure-direction cases, but it still presents point-estimate parity as an unqualified “match,” which is broader than the registry and parity suite support.
• Slide 7’s sample is still not runnable as shown: L_max is passed to the estimator constructor instead of fit(), and the caption promises HonestDiD from a call that does not enable honest_did=True.
• No estimator/runtime code changed in this PR. I did not find performance or security issues in the generator or checked-in PDF artifact.

Methodology

• Severity: P1. Impact: The validation slide in carousel/generate_dcdh_carousel.py:L854-L873 still says “Point estimates match R’s DIDmultiplegtDYN” and “ATT matches R across joiners, leavers, and controls.” That is broader than the documented contract. The registry explicitly notes O(1%) mixed-direction point-estimate differences from Python’s period-based stable-control sets vs R’s cohort-based sets, and separately notes that parity holds on the committed one-observation-per-cell fixtures under the equal-cell weighting contract docs/methodology/REGISTRY.md:L536-L538 docs/methodology/REGISTRY.md:L608-L610. The parity tests reflect that boundary by allowing 2.5% tolerance on mixed-direction point estimates rather than exact equality tests/test_chaisemartin_dhaultfoeuille_parity.py:L97-L137 tests/test_chaisemartin_dhaultfoeuille_parity.py:L173-L183. Concrete fix: change the slide to a scoped claim such as “validated against committed R parity fixtures,” or explicitly qualify pure-direction / one-observation-per-cell conditions instead of saying point estimates “match R” unconditionally.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• Severity: P3. Impact: The estimator-side mixed-direction SE-comparison limitation is already tracked in TODO.md:L60-L60, so that limitation itself is not a blocker under the project’s deferred-work rules. Concrete fix: none in estimator code for this PR; keep the carousel copy inside the documented validation scope.

Security

• No findings. The checked-in PDF appears to contain only a standard page-fit open action and no JavaScript, launch actions, URIs, or embedded files.

Documentation/Tests

• Severity: P2. Impact: The slide 7 example in carousel/generate_dcdh_carousel.py:L801-L830 will not run as shown. L_max is not a constructor parameter on ChaisemartinDHaultfoeuille diff_diff/chaisemartin_dhaultfoeuille.py:L359-L370; it is a fit() parameter diff_diff/chaisemartin_dhaultfoeuille.py:L472-L491. The caption also promises “HonestDiD” from that one call, but honest_did defaults to False and must be enabled in fit() diff_diff/chaisemartin_dhaultfoeuille.py:L540-L549. Concrete fix: move L_max=3 into est.fit(..., L_max=3) and either add honest_did=True to the example or remove HonestDiD from the caption.
• Static review only: this sandbox does not have pytest or core runtime deps like numpy, so I could not execute the generator or test suite here.

Path to Approval

1. Reword slide 8 so the R-validation claim matches the documented scope: either say “validated against committed R parity fixtures” or qualify the claim to the scenarios where exact parity is documented.
2. Fix slide 7’s sample to use fit(..., L_max=3) and align the caption with the actual call; if HonestDiD stays in the caption, add honest_did=True.
3. Regenerate carousel/diff-diff-dcdh-carousel.pdf after updating the slide source so the checked-in PDF matches the corrected copy.