H9: Slant is largest in polls financed by partisan funds

Among self-sponsored polls, the size of the sponsor-bias coefficient should vary with the source of the money that paid for the poll. The prediction is a gradient: β is largest for polls financed through Fundo Partidário or Doações Eleitorais (party apparatus / public money), smaller for Recursos Próprios (the candidate's personal money), smaller still for Outros (third-party donors). The quantity of interest is the heterogeneity of β across DS_ORIGEM_RECURSO categories, holding sponsor identity fixed.

Evidence strength: Partial by AN-006, AN-031 (2026-06-02). No analysis yet uses the DS_ORIGEM_RECURSO field directly. The available proxy is the sponsor-identity route split: polls paid from the candidate's own CPF give β = +19.12 pp (SE 6.89, p = 0.006, n_self = 18), versus +8.73 (committee CNPJ), +7.70 (party CNPJ), +6.33 (party-name parse). Wild-cluster bootstrap on the CPF cell rejects H₀: β = 0 at p = 0.009. The direction of the gradient on the identity axis is the opposite of what the party-apparatus-laxer-discipline reading predicts, which sharpens the case for running the funding-flow split directly.

Theory

The framework is Polls as Bayesian persuasion (theory.md §"Polls as Bayesian persuasion (supply-side / Channel A)"). H1 fixes the sender as "the sponsor"; H9 partitions senders by who actually puts up the money. Different senders have different cost functions: the candidate paying with personal money internalizes the full reputational and electoral cost of slanting; a party apparatus spending Fundo Partidário (public money, routed through the party treasury) is a step removed from the candidate and has weaker discipline against pushing slant up. The canonical Kamenica & Gentzkow (2011) setup is one sender with one objective; the funding-source split is the natural empirical proxy for sender heterogeneity in the cost of slanting.

Prediction

A triple interaction sponsored_by × DS_ORIGEM_RECURSO on the spec-3 regression should produce an ordered β gradient:

β(Fundo Partidário) ≥ β(Doações Eleitorais) > β(Recursos Próprios)

β(Outros).

The mechanism is principal–agent: as the spender becomes more removed from the candidate's personal stake, and as the money becomes more public (Fundo Partidário is statutory party money), the marginal cost of slanting falls. Sample-size limit is sharp: 712 mayoral protocols (~4.7%) have explicitly political funding on the recurso side, spread across treatment + control cells.

Competing predictions

Direct-stake reading (opposite gradient). If the cost of slanting is reputational to the candidate personally rather than disciplinary from the party apparatus, the gradient should flip: the candidate paying out of pocket has the most to lose if the poll is exposed as fabricated and so should be more careful, not less. Under this reading, β(Recursos Próprios) < β(Fundo Partidário) is the wrong sign. The route split (AN-006) already produces the opposite of this reading on the identity axis (own-CPF β is the largest, not the smallest), which is why both readings need the direct funding-flow test before being adjudicated.

Selection on candidate type. Candidates who self-finance polls (Recursos Próprios) may be systematically different from those whose party pays — smaller budget, possibly first-time, possibly without a formal committee structure. Any β gradient across DS_ORIGEM_RECURSO then conflates funding-source mechanics with candidate-type heterogeneity. The within-candidate FE absorbs this on the headline axis but does not absorb it for the funding-source interaction, because the funding source varies across polls of the same candidate only rarely (most candidates use one funding type throughout the cycle).

Prior research

The Paraná Pesquisas case is the load-bearing anecdote: Folha (2022) documented the PL paying R$ 2.7M to Paraná Pesquisas via Fundo Partidário in the 2022 pre-campaign while not appearing as TSE contratante on any of the firm's 63 presidential polls, which consistently reported a Lula–Bolsonaro technical tie while Datafolha, Ipec, and Quaest showed Lula leading [stories.csv 2022-19]. The institutional anchor is statutory: DS_ORIGEM_RECURSO is disclosed per LE.33.II (institutions.md §"Statutory mapping"), so the funding-source field is on every protocol by design rather than selectively reported. There is no academic literature pricing sponsor-bias gradient by funding source — the literature on sponsor effects in online surveys [cite:leeper2019sponsorship; cite:crabtree2020sponsorship] does not distinguish among types of political sponsors.

Evidence

Analysis	Bearing	Key takeaway
AN-006	Supports (proxy axis)	Sponsor-route split: CPF (own-pocket) β = +19.12 (SE 6.89, p = 0.006, n_self = 18) vs committee CNPJ +8.73, party CNPJ +7.70, party-name +6.33. Splits sponsor identity, not funding flow — proxy for the H9 gradient.
AN-031	Confirms (inference)	Wild-cluster bootstrap on the AN-006 route split: CPF WCR p = 0.009, percentile CI [+8.68, +29.47]; bootstrap lower bound sits at or above the other routes' point estimates. The route gradient survives finite-sample-valid inference.
DS_ORIGEM_RECURSO direct test	Not tested	Triple interaction sponsored_by × DS_ORIGEM_RECURSO on spec-3 is queued. Field is in hand (pipelines/politica/build/clean/poll_sponsor_2024.parquet); 712 politically-flagged protocols spread across treated + control cells.

Open tests

Direct funding-flow gradient

Run sponsored_by × DS_ORIGEM_RECURSO on spec-3 with the four canonical categories (Fundo Partidário / Doações Eleitorais / Recursos Próprios / Outros) plus a #NULO# bin. The 712-protocol politically-flagged cell determines the precision ceiling. This is the load-bearing test that adjudicates between the principal-agent reading (gradient pointing toward party/public money) and the direct-stake reading (gradient pointing toward own pocket). Without it, the H9 page is on the AN-006 proxy axis only.

CPF cell heterogeneity

AN-031 follow-up: with n = 18 in the CPF cell, the bootstrap takes the cell's marginal structure as given. A descriptive look at those 18 candidates (race competitiveness, final outcome, party, candidate experience) could surface whether the +19 pp comes from a homogeneous "personal-stake" pattern or is driven by a subset of unusual cases. Either reading would sharpen the interpretation of the funding-source gradient when it lands.

Cross-reference with the partisan-fund anecdote

The Paraná Pesquisas / PL case is a single firm-year datapoint where Fundo Partidário money flowed outside the TSE contratante field — the firm appears as the contratante on its own polls while the PL is the upstream funder. This is closer to H11 (pollster-self-cover) than to the H9 funding-source split, since the DS_ORIGEM_RECURSO field on those polls would not flag the PL money. Worth flagging in the funding-source paper section as the limit case the registry field misses.

Supporting analyses