iframe-proxy

igerber · 2026-05-25T14:21:22Z

Summary

Adds carousel/generate_spillover_carousel.py (an 11-slide portrait LinkedIn carousel generator for the SpilloverDiD launch) plus the generated carousel/diff-diff-spillover-carousel.pdf artifact.
Mirrors the well-received HAD carousel architecture but introduces a fresh earth-tone palette (slate-blue / warm terra / muted gold / cream gradient) signaling the spatial / geographic identity of the estimator.
Marketing artifact only — no library code, tests, or methodology files touched. SpilloverDiD itself shipped through v3.4.1 in Waves A / C / D / E.1 / E.2 / E.3.
Two AI review rounds (local Codex) caught and fixed: (1) "implemented as written" → "documented synthesis" reframing on slide 10 card 1; (2) "Monte Carlo coverage" → "Monte Carlo recovery" on slide 10 card 4 (the actual tests are recovery tests, not CI-coverage tests); (3) outer ring label [100, 200) → [100, 200] to match the public SpilloverDiDResults.spillover_effects schema; (4) slide-6 bias-annotation arrow now points specifically at tau_spill(0) (previously ambiguous under the equation center).

Methodology references (required if estimator / math changes)

Method name(s): N/A — no methodology changes. Carousel references but does not modify the existing SpilloverDiD methodology.
Paper / source link(s): N/A
Any intentional deviations from the source (and why): N/A

Validation

Tests added/updated: No test changes. Carousel is a one-shot marketing artifact with no CI gate.
Backtest / simulation / notebook evidence (if applicable): Visually verified all 11 slides render correctly (sidebar tick advances slide-by-slide, palette consistent, no clipped text). PDF size 372KB.

Security / privacy

Confirm no secrets/PII in this PR: Yes

Generated with Claude Code

Marketing artifact (carousel PDF + generator script). NOT a library code change — SpilloverDiD shipped through v3.4.1 in Waves A/C/D/E.1/E.2/E.3. Methodology claims made in the deck — please validate against the source of truth on review: 1. Slide 1 hook: "There's a clean line between treated and control. / Until there isn't." Hook framing, not a methodological claim — but verify it doesn't overstate (we're saying SOME contamination, not wholesale corruption). 2. Slide 4 footnote: "Built on two-stage Gardner (2022) DiD." — verify Gardner year matches docs/methodology/REGISTRY.md SpilloverDiD section. 3. Slide 6 equation: "beta_DiD ~ tau_total - tau_spill(0)" attributed to Butts 2021 Proposition 2.1. Verify sign convention + citation against docs/methodology/papers/butts-2021-review.md L101-113 and diff_diff/spillover.py module docstring. 4. Slide 8 code: SpilloverDiD constructor (rings, conley_coords) + fit(data, outcome, unit, time, treatment) + result.att + result.spillover_effects DataFrame schema. Verify against diff_diff/spillover.py and diff_diff/results.py SpilloverDiDResults. 5. Slide 9 production-ready cards — verify each is actually shipped: - Spatial-HAC SEs (Conley 1999 panel-block, kdtree fast path) - Cluster-Robust (HC1 / CR1 with Gardner GMM correction) - Survey Design (pweights / strata / PSU / FPC via Binder TSL) - Subpopulation Domains (full-design retention via zero-pad scores) - Event-Study Mode (per-event-time x ring decomposition) - Staggered Timing (Gardner two-stage; non-staggered too) 6. Slide 10 validation cards: - Butts Eq 5/6 + Proposition 2.1 — implemented as written - Gardner Two-Stage GMM Section 4 — stage-1 FE + first-stage uncertainty correction - Conley (1999) Spatial-HAC — matched to R conleyreg at atol=1e-6 on parity fixtures (verified against tests/test_conley_vcov.py PARITY_TOL = 1e-6) - Monte Carlo Coverage — recovers known tau_total + delta_j; coverage near nominal level 7. Slide 3 real-world examples (place-based econ + OOH + geo-targeted digital + retail footprint) — fair characterizations of spillover dynamics, or overstated? Out of scope for the deck (intentional): - Self-citation of Gerber (2026) Binder TSL — kept out of the PDF; surfaced only in the LinkedIn caption copy. - vcov_type="classical" restriction; replicate-weight NotImplementedError gates; HC2/HC2_BM not yet routed for stage-2; data-driven d_bar selection — all queued follow-ups documented in spillover.rst. Earth-tone palette (slate-blue / warm terra / muted gold / cream) is a fresh family variation; HAD's indigo palette unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Addresses /ai-review-local findings (review at .claude/reviews/local-review-latest.md): P1 #1 — Slide 10 card 1 "implemented as written" overclaimed paper fidelity. SpilloverDiD is a documented synthesis (REGISTRY.md L3274-L3288 + butts-2021-review.md L123-L139): the library uses the time-varying ring-exposure form, not Butts's literal unit-static Eq. 5 (which is rank-deficient under TWFE). Rewrote to "Time-varying ring-exposure synthesis (Butts Sec 5 + Gardner + Conley)." P1 #2 — Slide 10 card 4 "Monte Carlo coverage / coverage near nominal" overclaimed inference validation. Current tests at tests/test_spillover.py:L741-L823 + L3625-L3702 are recovery tests, not CI coverage tests. Renamed card "Monte Carlo Coverage" → "Monte Carlo Recovery" with desc "Recovers known tau_total + delta_j on synthetic spillover DGPs." Subtitle "Paper equations + Monte Carlo coverage" → "Documented synthesis + Monte Carlo recovery". P3 — Outer ring label "[100, 200)" → "[100, 200]" on slide 7 chart and slide 8 code sample. Matches the public SpilloverDiDResults.spillover_effects schema at diff_diff/results.py:L368-L372 (last interval is closed-right). PDF regenerated. No methodology code touched. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address Codex re-review P3: internal docstring still said "paper-equation parity + Monte Carlo coverage" after the prior commit corrected the user-facing slide copy to "documented synthesis + Monte Carlo recovery". Stale local description; doesn't render in the PDF but could reintroduce the over-claim on future edits. PDF unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The "the hidden bias term" caption was centered under the full equation, which visually landed closer to tau_total than to tau_spill(0). Readers could plausibly interpret tau_total as the hidden bias, which is the opposite of the truth: tau_total is the estimand we want to recover, and tau_spill(0) is the contaminating bias term. Rewrote slide 6 to render the equation + an actual annotation arrow as a single matplotlib figure via a new _render_bias_equation() helper. The arrow originates near the base of tau_spill(0) and a "hidden bias term" label sits below it in TERRA — unambiguously naming the correct term. The standalone TERRA italic caption beneath the equation is removed (superseded by the inline arrow + label). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Arrow x-position 0.715 → 0.668 in the matplotlib figure coordinate space. Previous position landed on the right edge of "(0)"; intermediate 0.645 overcorrected to the left edge of the tau glyph. 0.668 sits between the "spill" subscript and the "(0)" — visual center of the compound symbol. PDF regenerated. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

github-actions · 2026-05-25T14:26:02Z

Overall Assessment

✅ Looks good

Executive Summary

This PR is marketing-only in code terms: it adds a standalone carousel generator and a generated PDF, with no estimator, weighting, inference, or default-behavior changes.
Affected methods referenced, but not modified, are SpilloverDiD and its Gardner/Conley variance surface.
Cross-checking the deck against the Methodology Registry and the shipped SpilloverDiD docs/code found no undocumented methodology mismatch.
The deck’s API/schema claims match the current public surface, including result.att, result.spillover_effects, and the closed outer-ring label [100, 200].
Residual risk is low and operational: the PDF is a binary artifact with no CI smoke test.

Methodology

No findings. The deck’s substantive claims align with the committed SpilloverDiD contract: documented-synthesis framing, time-varying ring regressor, Gardner two-stage/GMM variance, event-study × ring output, and survey/subpopulation support are all consistent with docs/methodology/REGISTRY.md:L3272-L3485, diff_diff/spillover.py:L1651-L1739, and diff_diff/spillover.py:L2150-L2189. The slide-8 output example also matches SpilloverDiDResults.spillover_effects, including the closed outermost ring label [100, 200] in diff_diff/results.py:L368-L445 and tests/test_spillover.py:L498-L504.

Code Quality

No findings. The generator is self-contained, uses fixed RNG seeds for illustrative plots, and cleans temporary raster files in a finally path (carousel/generate_spillover_carousel.py:L86-L97, carousel/generate_spillover_carousel.py:L1235-L1254).

Performance

No findings. For a one-shot asset generator, the matplotlib → PNG → FPDF pipeline is acceptable.

Maintainability

Severity: P3. Impact: carousel/diff-diff-spillover-carousel.pdf is a binary artifact with run-specific metadata (/CreationDate), so identical-content regenerations still create opaque diffs and review noise. Concrete fix: normalize PDF metadata after generation or treat the PDF as a generated release artifact rather than committed source. Location: carousel/diff-diff-spillover-carousel.pdf, carousel/generate_spillover_carousel.py:L1250-L1252.

Tech Debt

No findings beyond the low-severity binary-artifact reviewability note above.

Security

No findings. I did not see any new network surface, secret-handling path, or obvious embedded secret material in the added script/artifact.

Documentation/Tests

Severity: P3. Impact: there is no smoke test or content-regression check for the new generator, so future claim/layout drift would be caught only by manual inspection of a binary PDF. Concrete fix: add a lightweight test that runs the generator to a temp output and asserts successful generation plus a few stable invariants such as page count or extracted key strings. Location: carousel/generate_spillover_carousel.py:L1235-L1258.

igerber and others added 5 commits May 25, 2026 10:20

igerber added the ready-for-ci Triggers CI test workflows label May 25, 2026

igerber merged commit c5416ae into main May 25, 2026
5 of 6 checks passed

igerber deleted the spillover-launch branch May 25, 2026 14:39

Sunbelt Computer Software

PL/B Language Development and Support

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

LinkedIn launch carousel for SpilloverDiD (Butts 2021)#490

LinkedIn launch carousel for SpilloverDiD (Butts 2021)#490
igerber merged 5 commits into
mainfrom
spillover-launch

igerber commented May 25, 2026

Uh oh!

github-actions Bot commented May 25, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Sunbelt Computer Software

PL/B Language Development and Support

Conversation

igerber commented May 25, 2026

Summary

Methodology references (required if estimator / math changes)

Validation

Security / privacy

Uh oh!

github-actions Bot commented May 25, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant