id: an-097 hypothesis: shell-contratante headline: |error| degradation on sponsored rows is concentrated in the firm tails. T3 (highest pro-sponsor β, 7 firms, 310 polls): +6.82 pp on |error| under race + cand FE (p<0.05). T1 (low or anti-sponsor β, includes CENSUS and EVA FRANCIELI): +3.54 pp (p<0.10) — slant in either direction inflates |error|. T2 (moderate β): null. Untested firms (no AN-016 β, 19,337 cand-poll rows, 80% of universe): −2.14 pp (p<0.10) on |error| but +5.73 pp (p<0.001) on SIGNED error — the bias is real and large but doesn't translate into measurable accuracy degradation. This is the cleanest demonstration that the market for poll accuracy cannot discipline pollsters: 90% of the universe of polls shows null-or-negative |error| degradation on sponsored rows even when the +7 pp slant is precisely measured. type: heterogeneity question: "Does |error| degradation on sponsored rows scale with firm slant intensity (within-firm β from AN-016)?" tags: ["hyp:shell-contratante", "within-firm", "heterogeneity", "noise-floor", "all-brazil"] status: interpreted status_date: 2026-06-17 confidence: green created: 2026-06-17 script: source/analysis/an-097-sponsored-row-by-beta-tercile.py target: build/table/an-097-sponsored-row-by-beta-tercile.csv

AN-097: |error| degradation by firm β tercile

Extends AN-096 by stratifying firms into terciles based on the AN-016 within-firm β. Direct test of the user's intuition that "for very high β firms, |error| should be very bad."

Tercile definitions (signed β from AN-016)

Tercile β range n firms n polls (cand × poll) Mean β
T1_low β ∈ [−8.4, +2.3] 8 2,207 −1.96
T2_mid β ∈ (+2.3, +10.0] 7 811 +5.90
T3_high β ∈ (+10.0, +23.4] 7 310 +16.86
Untested (no β estimable) 402 19,337

T1 includes the anti-sponsor firms CENSUS (β=−8.2, n=213) and EVA FRANCIELI (β=−8.4, n=75). T3 includes the most slant-y firms METHODUS (β=+23.4) and CAMARGO E MEDINA (β=+22.4). "Untested" = ~80% of the universe (402 firms with no within-firm β estimable from AN-016 due to too-few self- sponsored polls).

Headline: sponsored_by coefficient on |error|

Race + Candidate FE; cluster SE at race. Outcome = |error|.

Tercile S0 No FE S1 Race FE S2 Race + Cand FE S3 Race + Cand + Firm FE
T1_low +6.37*** (1.28) +4.20*** (1.18) +3.54* (1.84) +3.24* (1.90)
T2_mid −0.34 (0.75) −1.67 (1.44) −0.74 (1.16) +0.66 (1.51)
T3_high +1.76 (1.67) +8.38*** (0.78) +6.82** (3.09) +6.71** (2.97)
Untested +1.84** (0.83) −0.19 (1.10) −2.14* (1.11) −1.96* (1.14)

Raw cell means show the same pattern:

Tercile Mean |err| unsponsored Mean |err| sponsored Raw diff
T1_low 6.4 13.8 +7.4
T2_mid 6.3 6.3 +0.0
T3_high 8.6 12.3 +3.7
Untested 7.4 9.7 +2.3

Sanity check: signed-error coefficient (the slant) by tercile

Race + Candidate FE.

Tercile S2 signed error S2 |error|
T1_low +3.56* +3.54
T2_mid −0.78 −0.74
T3_high +5.11 +6.82
Untested +5.73*** −2.14*

Compare the Untested row: the slant is +5.73 pp (p<0.001) — real and large — but |error| coefficient is −2.14 pp (p<0.10, negative). For the 80% of the universe in untested firms, the +5.73 pp slant translates into LOWER |error| on sponsored rows.

The likely mechanism: in firms that rarely take candidate work, sponsored polls go to front-running candidates whose true vote share is easier to predict; even with a +5.73 pp boost on the poll share, the cand's actual final share is in the right ballpark.

Interpretation: the market doesn't discipline pollsters

Three results combine into a substantive market-structure claim:

  1. The signed slant is precise and broad. +6.7 to +7.0 pp across every FE spec (AN-096); +5.73 pp even in untested firms (this AN). The bias is real, large, and consistent.

  2. The slant does NOT translate into measurable |error| degradation for the typical firm. Only T3 (the top |β| firms, 7 firms, ~5% of the polls in AN-016's testable universe) shows positive |error| coefficients robustly. T2 is null. The 80% of polls in untested firms shows negative |error| coefficient — the slant is fully absorbed by sample selection on race difficulty.

  3. Without sponsor identity, the bias signal is statistically invisible. A typical sponsored poll's deviation from the typical unsponsored poll for the same cand in the same race is small (<1σ of natural variance — see AN-098). An outside auditor with access to poll accuracy data but not sponsor identity could not flag a sponsored poll as anomalous.

The economic implication. Standard reputation models for sponsored-content markets assume the sponsor's bias degrades the content quality enough that outside observers can detect it, creating a reputational cost to the producer. That mechanism doesn't operate here:

The disclosure-regime implication. The only piece of information that reveals the slant — who actually paid for the poll — is precisely the piece the regulatory regime requires to be disclosed. Shell sponsoring (AN-082, AN-083, AN-093) makes this disclosure mechanically possible to evade while staying within formal compliance. Enforcement gains from accuracy auditing are mechanically small; the lever that has bite is substantive sponsor-identity disclosure (i.e., requiring the actual paymaster, not the formal contratante).

Caveats

Follow-ups for the paper

  1. Move §Discussion to lead with the no-discipline argument. The empirical chain (AN-016AN-095AN-096AN-097AN-098) supports a substantive theoretical claim that goes beyond "Brazilian polls are biased." It's "the polling market structure permits bias because the accuracy signal is too noisy to discipline producers."

  2. Quantify the firm-revenue mechanism. If the high-β firms (T3) ARE detectable but continue producing slanted polls, are they covering their reputational costs from candidate- market revenue? Cross-check T3 firms' poll volume mix with AN-007 / AN-017 customer-mix data: do they have any media work, or are they purely candidate-revenue firms?

  3. Extend the β table beyond 22 firms. AN-016's restrictive sample is the load-bearing constraint here. Lowering the n_self threshold from ≥5 to ≥3 would roughly double the firms with usable β estimates.

Artifacts