H11: Pollster-self-sponsored polls operate as a hidden-sponsor channel

H11: Pollster-self-sponsored polls operate as a hidden-sponsor channel

When a registered pollster files a poll with ST_PESQUISA_PROPRIA = S (no external contratante), the firm is declaring that no one paid for the poll — it is self-initiated, financed out of the institute's own balance sheet. ABEP's standing industry critique is that this field is sometimes used as cover for undisclosed candidate or party financing (auto-financiamento as a caixa-dois channel). If that is right, ST_PESQUISA_PROPRIA = S polls should look noisier and more right-tail-skewed than genuine media-sponsored polls of the same races, because they bundle two underlying populations: honest no-sponsor polls and disguised candidate/party commissions whose true sender is concealed.

Evidence strength: Not tested. Queued behind the main mechanism work. The hypothesis rests on two pieces of pre-existing evidence (one industry critique, one specific case with quantified flow) but has no in-project test result yet.

Theory

The framework is Polls as Bayesian persuasion (theory.md §"Polls as Bayesian persuasion (supply-side / Channel A)") with sender-identity concealment added to the standard setup. In the canonical Kamenica & Gentzkow (2011) sender-receiver problem the sender's identity is common knowledge — receivers know who is paying, and discount accordingly. Channel A operates despite known sender identity because the methodology commitment binds. When sender identity is itself concealed, receivers cannot apply the sponsor-aware discount at all: a ST_PESQUISA_PROPRIA = S poll is read as a disinterested industry product even if its true sender is a candidate or party. The predicted effect on observed numbers is therefore larger than for declared self-sponsored polls — concealment removes a free discount that receivers would otherwise apply.

The mechanism does not require every pollster-self poll to be a hidden sponsorship. It only requires that some fraction is, and that the fraction is large enough to leave a footprint in the variance and tail shape of the pollster-self distribution relative to true media-sponsored polls.

Prediction

Conditional on race and week, polls with ST_PESQUISA_PROPRIA = S should show (i) larger variance of error (deviation from the race-week consensus or final result) than media-sponsored polls, (ii) a heavier right tail in the direction of whichever candidate the firm is plausibly aligned with, and (iii) within-firm differences between the same pollster's media-sponsored output and its pollster-self output — the firm's own polls should not be noisier than its commissioned polls under the null of honest self-financing.

Competing predictions

Honest self-financing. Pollster-self polls are genuinely unsponsored — fielded as marketing or reputational investments by the firm. Under this null, the variance of error should be no larger than for media-sponsored polls of the same race, and the within-firm placebo (the firm's own polls vs its media-sponsored polls) should show no systematic shift. Any apparent excess variance would then trace to firm-level selection (small institutes self-finance more, small institutes are noisier) rather than concealed sponsorship.

Firm-quality selection. Self-financed polls are concentrated in smaller, less-resourced firms that produce noisier estimates for mechanical reasons (smaller samples, weaker field operations). This generates excess variance in the pollster-self distribution without any concealed sender. The within-firm placebo discriminates: if the same firm's pollster-self polls are noisier than its media-sponsored polls, firm quality is not the explanation.

Prior research

The auto-financiamento concern is a documented industry-level critique, not yet a quantified empirical result. The two anchors:

The closest direct academic comparators are sponsor-disclosure experiments [cite:leeper2019sponsorship; cite:crabtree2020sponsorship] where the manipulation is visible sender identity. The sender-identity-concealment case studied here has, to our knowledge, no published quantitative treatment.

Evidence

Analysis Bearing Key takeaway
(none) Pending No in-project AN page targets H11 yet. Planned tests: variance and tail-quantile tests of error on ST_PESQUISA_PROPRIA = S vs media-sponsored, a Benford / digit-frequency check on the upper tail (mirroring the AN-013 design), and a within-firm placebo comparing each firm's media-sponsored polls to its pollster-self polls.

Open tests

Variance and tail-quantile diagnostic

The headline test is a side-by-side comparison of the distribution of error conditional on sponsor type, holding race and week fixed. The diagnostic is non-parametric: Levene / Brown-Forsythe for variance, quantile-by-quantile comparison for the right tail. Data are in hand (sponsor table + cleaned poll panel); the test is laptop-resolvable once a candidate or party alignment proxy for pollster-self polls is defined (firm's modal partisan customer over the cycle is the obvious starting proxy).

Within-firm placebo

The cleanest design — and the one that rules out firm-quality selection — compares the same firm's media-sponsored polls to its pollster-self polls. Firms with both modes in the sample are the identifying set. The architecture is the same as AN-016 (per-firm β with PDF-style held fixed) but on the variance/skew of error rather than the mean.

Benford / digit-frequency on the upper tail

If the concealed-sender share is non-trivial, the right tail of pollster-self polls should accumulate evidence of post-fielding adjustment — the same digit signatures AN-013 looked for, but pooled across polls rather than within-poll across candidates. AN-013 found no within-poll digit anomaly on the headline sponsor-bias sample; running the same machinery on the pollster-self tail is a natural extension. See H4 and the channel-decomposition program for the broader Channel A vs Channel B framing.