H10: Reputable national pollsters show smaller β than small regional firms

Pollster firms are repeat players. Each commissioned poll trades short-run sponsor revenue against long-run accuracy reputation among future clients. If career concerns discipline firm behaviour, the within-candidate sponsor effect β should be smaller (closer to zero) at firms with more reputational capital — the major national houses (Datafolha, Quaest, Ipec, AtlasIntel) — than at small regional shops. The empirical test is a triple interaction sponsored_by × pollster_volume_tercile on the spec-3 regression, with volume tercile as a coarse proxy for the reputation stock.

Evidence strength: Mixed by AN-007, AN-023 (2026-06-02). AN-007 finds the predicted positive direction at n=11 pollsters (slope +13.6 OLS / +6.3 WLS) but neither slope clears significance (p=0.40 / 0.47). AN-023 looks for the equilibrium-style fingerprint at n=216 pollsters and finds corr(candidate_share, substantive_share) = +0.058 — the reputation-disciplined disclosure pattern does not appear at the firm level.

Theory

The framework is Polls as career-concerns games (theory.md §"Polls as career-concerns games"). The pollster firm has reputational capital R_f accumulated from past poll-vs-result accuracy and chooses a slant level for each commission given the sponsor's offer and the depreciation function δ(τ, detection probability). Holmstr\"om (1999) is the canonical career-concerns model; Dewatripont et al. (1999) extends it to the precision of the public output signal — sharper verification yields sharper reputational discipline; Gentzkow & Shapiro (2006) is the structurally identical reputation-bias model in media markets, mapping cleanly onto polling firms with media clients in place of news consumers.

The mechanism is firm-level, not sponsor-level: this hypothesis is about which firms agree to slant and how much, conditional on a sponsorship offer arriving — distinct from H1 (why the sponsor demands slant in the first place).

Prediction

Per-firm β within the within-candidate FE design should decrease in firm reputational capital. The headline operationalisation is the triple interaction sponsored_by × pollster_volume_tercile, with volume serving as a coarse proxy for accumulated reputation. Pollster volume is computed from registry protocol counts (NR_CNPJ_EMPRESA) and is already in hand. A sharper variant uses years-since-first-TSE-registration from the multi-cycle registry.

Competing predictions

Capacity, not reputation. Pollster volume mixes reputation stock and field capacity: a major firm has many polls because clients trust it AND because it has the field operations to deliver them. Both predict the same heterogeneity sign through different channels. A sharper test would need a pollster-level shock that changes reputation without changing capacity (loss of a major media contract, statistician-level audit finding); the 2024 data do not support that yet.

Vertical integration with media outlets. Brazilian polling firms are often vertically linked to media outlets (Datafolha ↔ Folha; Globo's polling partnerships — [institutions.md §"Major firms"]). The media outlet's reputation may be the binding constraint rather than the polling firm's own. The model's predictions hold but the unit of accumulation is the outlet-pollster pair, not the pollster CNPJ.

Customer-mix sorting, not size. AN-007 originally framed the heterogeneity through candidate-share of customer mix — firms serving candidates can sustain a slant-friendly reputation, firms serving media cannot. AN-023's null on the methodology fingerprint pushes against the strong-form sorting story; the volume-discipline story remains intact but the customer-mix component is empirically subordinate.

Prior research

The "encomendada" poll is a recognised phenomenon in Brazilian political discourse but firm-level slant heterogeneity has not been quantified registry-wide. Two anecdotes are directly informative:

• Major firms regularly defend their methodology publicly under litigation pressure. Datafolha's response to the Russomanno censorship attempt explicitly recited its 35-year-old quota methodology [stories.csv #131]. Reputational stake is observable in such defenses; the career-concerns model predicts exactly this defensive behaviour at high-R_f firms.
• The cross-firm consistency check between Paraná Pesquisas (which showed empate in the 2022 presidential race) and Datafolha / Ipec / Quaest (which showed Lula leading) in the same cycle [stories.csv #077] is consistent with the reputation gradient — the firms with the largest stock of accumulated accuracy reputation produced the call that ex-post matched the outcome.

The closest direct supply-side predecessor is the formal pollster reputation game of Meirowitz (2005) — pre-election polls as strategic games between pollster firms and clients. The [institutions.md §"Brazilian polling industry"] reference catalogues the 448 pollsters in the all-Brazil sample, with the top firms accounting for the bulk of media-sponsored polls.

Evidence

Analysis	Bearing	Key takeaway
AN-007	Mixed (underpowered)	n=11 pollsters with usable β: slope of β on candidate_share is +13.6 (p=0.40, OLS) / +6.3 (p=0.47, WLS). Direction matches theory but underpowered. The two highest-volume firms (IIP, Census) sit near β=0 despite high candidate-share — consistent with the secondary volume → discipline prediction but against the strict monotone customer-mix version.
AN-023	Against	n=216 pollsters with ≥20 polls: corr(candidate_share, substantive_share) = +0.058 — essentially zero. Methodology disclosure style is uncorrelated with customer mix at the firm level. The reputation-equilibrium prediction that candidate-serving firms develop a recognisable boilerplate-heavy style does NOT hold in the cov_bucket fingerprint.

Open tests

Triple-interaction on the headline spec

The clean directly-on-prediction test is sponsored_by × pollster_volume_tercile on spec 3. Pollster volume from registry protocol counts is in hand; the interaction is queued. The current evidence base is firm-level cross-sections of β rather than the interaction itself, so this would convert a cross-firm scatter into a within-design heterogeneity coefficient.

Years-since-first-TSE-registration

A sharper proxy for R_f than 2024 volume alone is the firm's multi-cycle registration history. Newcomer firms (first electoral cycle) have no reputation to protect and are predicted to show the largest β. Computable from the multi-cycle TSE poll registry once the 2022 cycle is folded in (queued in docs/todo.md).

Customer-mix vs volume decomposition

AN-007 and AN-023 disagree on whether customer mix carries any independent signal once volume is conditioned out. A joint regression of per-firm β on log(n_total) + share_candidate on the larger 31-firm sample (already in hand from AN-016) would settle which axis is load-bearing; see [theory.md §"Pollster reputation: volume vs customer mix"] for the formal split.

Recommendation

Extend theory.md with a career-concerns / reputation section (run /theory --extend) before the triple-interaction test, so the within-firm volume-discipline result has a formal home rather than sitting under "Bayesian persuasion." The drafted section in theory.md §"Polls as career-concerns games" (2026-06-02) already covers most of this.

Supporting analyses