Key Findings — Poll Sponsor Bias

Curated index of headline empirical findings and interpretations from the 2024 Brazilian mayoral poll analysis. Each entry is a standalone page; this index is the scannable directory.

Use this index to scan what we believe; use docs/analyses/an-*.md for the raw numbers and methodology behind each finding.

How to read the confidence tags

A traffic-light convention runs across both empirical and interpretive claims: 🟢 = strongest confidence, 🟡 = middle, 🔴 = weakest.

Empirical findings — color reflects the source of confidence, not the size of the effect:

🟢 Replicated — appears across multiple independent identifying designs that agree in direction and magnitude.
🟡 Single source — one solid analysis, no independent replication yet.
🔴 Provisional — parser-dependent, sample-sensitive, or carries a known caveat.

Interpretations — parallel scheme:

🟢 Strong — multiple converging lines of evidence; alternatives considered and rejected.
🟡 Plausible — consistent with the evidence but other readings remain open.
🔴 Speculative — suggested by the data but unverified.

Findings overview

Empirical findings

🟢 The headline: sponsor-paid polls overstate the candidate by ~7 pp — within-candidate FE on 568 self-sponsored rows
🟢 Within-candidate jump from preceding independent poll is +6.7 pp (t=5.2) — placebo that kills the "self-sponsor when leading" alternative
🟢 The race × week FE design gives β = +6.95 pp on 60 cells — tightest timing-controlled spec, same magnitude as the other two
🟡 Opponent-sponsored polls understate the candidate by ~2 pp — sender-specific bias, not generic house effect
🟡 Peak β shifts from rank 2 (small munis) to rank 3 (runoff munis) — the M+1 cutoff moves at the 200k voters threshold
🟡 β is monotone in race competitiveness: +10 (tight) vs +4 (wide) — slant largest where margins matter
🟡 Slant decays as the election approaches: +0.037 pp / day — consistent with verifiable-disclosure pressure
🟡 Own-CPF route gives β ≈ +19 vs +6-9 for committee/party — n=18 cell but stable
🟡 Pollster β ranges from −10 (Paraná) to +30 (Intenção) across 33 firms — the "Brazilian pollster" average disguises wide firm variation
🟡 Sponsored polls flex the population frame and the bairro list — paired qualitative diff on 39 classical-bias pairs reveals population_reference=mixed and coverage_class=urban_only/specific_neighborhoods as the dominant Channel A signatures
🟢 The +7 pp is per-poll visible to a tight benchmark but buried in cross-firm noise; calibrated blind audit gives 2.6× lift — reputation + policy implications

Interpretations

🟢 Voter coordination dominates bandwagon as the consequential channel — the rank × runoff shift
🟡 The bias is concentrated in candidate-touched commissioning, not pollster house effects — symmetry test + per-pollster heterogeneity
🟡 Channel A vs Channel B — pilot landed, documented levers under-explain the +7 pp — 244-pair LLM pilot + 7+ universe-scale lever tests; size-mismatch finding published in paper §5

Open items for this page

2022 cycle extension (low priority in docs/todo.md) would add a within-pollster-over-time identification strategy that cleanly tests the customer-mix reputation theory.
Channel A vs B decomposition — pilot work landed (poll-methodology-paired + AN-024/041/042/043/107) and the size-mismatch finding is published in paper/paper.tex §5. The universe-scale poll_methodology LLM extractor would tighten SEs on small same- direction contrasts (population frame, ponderação) but is no longer load-bearing for the paper's conclusion. Open item is now interpretation of the unattributed 2–6 pp residual, not getting more data.
News-anchor citations: the project will eventually accumulate Brazilian press references on specific cases (e.g., Paraná Pesquisas / PL via Fundo Partidário). Should be folded into Sources footers as references/news/texts/... once collected.
Aggregate-report citations: no CNJ / IBGE / academic-survey reports cited yet. Add when relevant for institutional framing.