AN-089: Within-firm sponsor effect on accuracy

id: an-089 hypothesis: shell-contratante headline: Within less-trusted firms (n=5,097, 419 firms) holding firm + race fixed: candidate-sponsored polls reduce margin_error by 1.94 pp vs media-sponsored (p=0.07); other_firm-sponsored reduce by 1.04 pp (p=0.11). Directionally consistent with the cross-sectional finding at ~2/3 magnitude. Same SIGN, smaller magnitude — most of the cross-sectional shell signal is CROSS-FIRM composition (different sponsors hire different firms), not within-firm slant. AN-088 already showed sponsor contributes only 0.1-0.2 % marginal R² beyond firm + race; this AN-089 result is the within-firm correlate of that. NOTE: in trusted firms (n=430, 5 firms) sponsor effects are statistically weak overall but point estimates on margin_error are LARGER (−9.4 / −8.8 pp at the firm-only spec, p≈0.07), suggesting trust × sponsor interaction worth flagging but with very thin n. type: causal question: "Within the same pollster, does accuracy differ by sponsor type? Especially for less-trusted firms?" tags: ["hyp:shell-contratante", "within-firm", "robustness", "all-brazil"] status: interpreted status_date: 2026-06-17 confidence: yellow created: 2026-06-17 script: source/analysis/an-089-within-firm-sponsor-accuracy.py target: build/table/an-089-within-firm-sponsor-accuracy.csv

AN-089: Within-firm sponsor effect on accuracy

User question (2026-06-17): "Can we argue that the same pollster might produce different accuracy polls depending on type of sponsor? Focusing on the less trusted pollsters."

Companion to AN-088 (variance decomposition showed sponsor = 0.1-0.2 % marginal R² beyond firm + race). This script isolates the within-firm sponsor effect on accuracy.

Sample partition

Within-firm + race FE results

Reference category = media-sponsored polls within the same firm × race.

Sample Outcome n is_candidate is_pollster_self is_other_firm
LESS_TRUSTED mean_abs_error 5,097 +0.17 (p=0.73) +0.22 (p=0.60) +0.02 (p=0.94)
LESS_TRUSTED poll_calls_winner_first 5,097 +0.047 (p=0.18) +0.055 (p=0.089) +0.017 (p=0.50)
LESS_TRUSTED margin_error 4,879 −1.94 (p=0.070) +0.64 (p=0.46) −1.04 (p=0.108)
ALL mean_abs_error 5,655 +0.12 (p=0.79) −0.01 (p=0.98) +0.13 (p=0.66)
ALL poll_calls_winner_first 5,655 +0.057 (p=0.10) +0.045 (p=0.12) +0.010 (p=0.69)
ALL margin_error 5,424 −1.75 (p=0.072) +0.22 (p=0.78) −0.91 (p=0.14)
TRUSTED mean_abs_error 430 −0.90 (p=0.56) −0.87 (p=0.47) +0.71 (p=0.62)
TRUSTED poll_calls_winner_first 430 −0.20 (p=0.28) −0.07 (p=0.52) −0.23 (p=0.040)
TRUSTED margin_error 420 −3.66 (p=0.18) −2.00 (p=0.29) −0.81 (p=0.69)

Reading

LESS_TRUSTED firms (the user's focus)

Direct answer to the user's question: yes, the same firm produces different accuracy by sponsor type, but the effect is directional and not statistically clean.

Within a less-trusted firm × race, candidate-sponsored polls reduce margin_error by 1.94 pp (p=0.07); other_firm-sponsored polls reduce by 1.04 pp (p=0.11). The within-firm coefficient on candidate is similar in sign and ~2/3 the cross-sectional magnitude (AN-085 had β=−2.53 cross-sectional vs −1.94 within- firm here for the LESS_TRUSTED partition). The within-firm effect on other_firm is about half the cross-sectional magnitude (cross-sectional was −1.95; within-firm is −1.04).

Interpretation: most of the cross-sectional sponsor effect is CROSS-FIRM composition, not within-firm slant. A LESS_TRUSTED firm doesn't dramatically change its accuracy depending on who pays — what differs cross-sectionally is which firms tend to get hired by different sponsor types. The within-firm slant is present but smaller and harder to identify.

Why the AN-082 / AN-085 / AN-086 effects shrink within firm

AN-088's variance decomposition gave the structural answer: sponsor's marginal R² over firm + race is 0.1-0.2 %. The within-firm regression is the regression-coefficient version of that decomposition. Both are honest readings of the data.

Two ways to read it:

The two readings co-exist. The cross-firm composition channel likely dominates the cross-sectional finding, but the within- firm slant channel is present.

LESS_TRUSTED on poll_calls_winner_first

is_pollster_self β = +0.055 (p=0.089). Within less-trusted firms, pollster_self polls call the winner 5.5 pp more often than media polls — borderline significant. Other coefficients are null. Less-trusted firms doing self-contracted work appear to do them more carefully (consistent with the AN-085 selection- into-easy-races story for pollster_self).

TRUSTED firms (n=430, 5 firms) — comparison sample

Most coefficients null or directional. One sig finding: is_other_firm × poll_calls_winner_first = −0.229 (p=0.040). Within the 5 trusted firms × race, polls sponsored by other_firm contratantes call the eventual winner 22.9 pp LESS often than the same firm's media-sponsored polls in the same race. This is dramatic but the sample is thin (n=430, 5 firms, trusted-firm × other_firm cell n=70 from AN-086).

The directional pattern on margin_error in trusted firms is striking even without statistical significance: cand β=−3.66 pp; other_firm β=−0.81 pp; pollster_self β=−2.00 pp. The candidate and pollster_self margins shrink substantially within trusted firms, just under-powered. If real, the trusted- firm × sponsor interaction means trusted firms DO slant their sponsored polls in the direction of the broader pattern (smaller margins), but the n=42 candidate-sponsored and n=168 pollster_self cells aren't large enough to call it.

What this means for the headline story

  1. Sponsor effects are real but small relative to firm quality and race difficulty (AN-088: 0.1-0.2 % marginal R²; this AN-089: −1 to −2 pp within firm + race vs −2 to −3 pp cross-sectional).

  2. The shell-bucket margin-error finding (AN-082 / AN-085 / AN-086) is partly cross-firm sorting: shell-suspect sponsors disproportionately hire firms that have lower accuracy in general, in addition to the within-firm slant. AN-089's within-firm coefficient on other_firm margin_error is −1.04 pp (p=0.11) — about half the cross-sectional magnitude.

  3. The candidate-bucket margin-error finding is more consistent within-firm. The within-firm coefficient (−1.94 pp, p=0.07) is ~75 % of the cross-sectional coefficient (−2.53). Less cross-firm sorting; more direct sponsor slant.

  4. No within-firm-FE evidence that less-trusted firms are slanting their sponsored polls dramatically more than their media polls. The signal is there but small. If the user's intuition was that less-trusted firms produce dramatically different accuracy across sponsor types, the answer is "no, not dramatically." If the intuition was "trace amounts of slant within firm," the answer is "yes, around 1-2 pp on margin_error."

Caveats

Follow-ups

  1. Re-spec with firm × race interactions. The current within-firm + race FE assumes firm and race effects are additive. A firm × race interaction would test whether specific firm-race combinations show selectivity.
  2. Split LESS_TRUSTED further by AN-016 |β| (the most trustworthy non-D1 firms vs the most-slanting non-D1 firms). The within-firm sponsor effect might be concentrated in the most-slanting tail.
  3. Restrict to late polls (≤14 days to election). The timing-clean window with within-firm + race FE is the sharpest test possible; small n but cleanest interpretation.
  4. Wire to paper. The within-firm finding tempers the AN-085 / AN-086 cross-sectional claims. Worth one paragraph in §Identification distinguishing cross-firm sorting from within-firm slant, and an appendix table.

Artifacts