H7: β peaks just below the front-runner (bandwagon)

When a candidate sits just below the apparent leader of a race, slanting their poll number up by a few points can flip the perceived ordering — pushing them from "close 2nd" to "apparent 1st" — and so trigger the conformity, expressive-utility, and mobilization-asymmetry premia attached to looking like the winner. The bandwagon prediction is that sponsor bias β should be largest precisely for these candidates, and the discriminating contrast against the coordination-peak prediction (H6) is the cross-muni pattern: bandwagon predicts the same peak rank in small and runoff-eligible munis, whereas coordination predicts the peak shifts with the M+1 cutoff.

Evidence strength: Mixed by AN-045, AN-050 (2026-06-02). The descriptive cell-mean cut by final_rank × tight_race finds the biggest tight-race sponsor-bias amplification at rank 1, not rank 2 (+4.80 pp non-tight → +12.20 pp tight, a 2.5× jump). The rank-at-commission sharpening attenuates the pattern by half (rac1-in-tight β = +6.59 pp, n_sp = 20) but the leader-in-tight signal survives. Within-candidate FE specs are all underpowered for the 3-way; yellow confidence pending thin-cell triage.

Theory

The framework is Polls as bandwagon triggers (theory.md §"Polls as bandwagon triggers"). Voters do not explicitly compute pivotal probabilities but are nonetheless pulled toward whoever appears to be leading, through three non-exclusive channels: conformity / social desirability, expressive utility from voting for a winner, and mobilization asymmetry (leader supporters exert more effort, trailer supporters stay home). The attractor in this framework is the apparent front-runner, which is what makes the prediction distinct from coordination: the relevant cutoff is rank 1, not M+1.

Prediction

β-heterogeneity by final-rank position peaks for candidates just below the front-runner — rank 2 — because these are the candidates with the most marginal product from slant: a +7 pp shift can move them across the perceived-leader threshold and capture the bandwagon premium. Sponsors of a clear 4th-place candidate gain less (they cannot plausibly look like leaders); sponsors of clear 1st-place candidates also gain less (they already do). Crucially, the peak position is the same in small and runoff-eligible munis — the bandwagon attractor does not respect Cox's M+1 cutoff.

Competing predictions

Coordination at M+1. H6 predicts the peak position shifts across the 200k-voter institutional threshold: rank 2 in small munis (M+1 = 2 under plurality) and rank 3 in runoff-eligible munis (M+1 = 3 once the runoff round is folded in). The cross-muni shift is the cleanest discriminating contrast against the bandwagon prediction; the same regression cube identifies both.

Underdog effect. The counter-prediction within the bandwagon literature is that voters sympathize with trailing candidates, pulling votes toward them. If underdog dominated bandwagon in Brazilian mayoral elections, sponsors would slant their candidates down to look like sympathetic trailers, not up. The observed sign of the headline sponsor effect is itself weak evidence underdog is not the dominant force here.

Ex-post-rank aggregation. A pattern centered at final rank 1 may not reflect leader-at-commission behaviour: candidates who ended rank 1 were polled at varied positions throughout the campaign, so a final-rank cut mixes leaders, runners-up, and trailers within the same candidate's polling history. The rank-at-commission swap (AN-050) is the test that strips out this Simpson's-style aggregation.

Prior research

The Brazilian 2018 natural experiment showed voters who cast ballots after seeing partial results exhibited a bandwagon (not underdog) pull Araujo & Gatto (2021) — the cleanest Brazilian evidence for the mechanism. Cross-national bandwagon and underdog evidence comes from McAllister & Studlar's classic on British polls McAllister & Studlar (1991), Chatterjee et al's analysis of the Indian exit-poll ban Chatterjee & Kamal (2020), and Farjam's field experiment with real political organizations Farjam (2020). The fact that observed sponsor effects in this project are positive (slant up, not down) is itself a weak piece of evidence against underdog dominance — sponsors would slant down if it were the governing force.

Evidence

Analysis	Bearing	Key takeaway
AN-045	Supports (descriptive)	Cell-mean decomposition of the headline by final_rank × tight_race: biggest tight-race amplification at rank 1 (+4.80 pp non-tight → +12.20 pp tight, a 2.5× jump), consistent with leaders in close races over-stating to manufacture inevitability. Pooled OLS gives sp × tight = +9.26 pp (p<0.001). Within-candidate FE attenuates to n.s. given thin sponsored cells (n=117 for rank-1-tight-sponsored). Yellow confidence; rank-at-commission swap queued as the cleanest sharpening of "leader".
AN-050	Partial support (rank-at-commission sharpening)	Replaces AN-045's final_rank with rank-at-commission. The rac1-in-tight bandwagon pattern survives but attenuates: sponsor effect at rac1-tight is +6.59 pp (n_sp = 20) vs +12.20 pp under final_rank; non-tight rac1 baseline is +2.71 pp. So the bandwagon signal is real but ~half the AN-045 magnitude — the AN-045 estimate was inflated by Simpson's-style aggregation across heterogeneous within-candidate trajectories.

Open tests

Cross-muni peak-position contrast against H6

The discriminating prediction against H6 is not the existence of a peak but its position across small vs runoff-eligible munis. Bandwagon predicts the peak stays at rank 2 in both; coordination predicts it shifts from rank 2 (small munis) to rank 3 (runoff-eligible). The same final-rank-quintile × small/large cube as for H6 answers both jointly. Queued; the descriptive 2×K plot is the first cut.

Rac3+ thin-cell viability-grab puzzle

AN-050 surfaced a thin-cell surprise: rac3+ in tight races shows a sponsor effect of +18.64 pp (n_sp = 14), suggesting a complementary viability-grab supply-side mechanism — trailers in close races slanting hard to manufacture viability where the M+1 cutoff puts them at risk of strategic abandonment. If this survives thin-cell triage and a larger-n replication, it deserves its own hypothesis page parallel to this one. Currently underpowered.

Within-candidate FE confirmation

All within-candidate FE specs in AN-045 and AN-050 attenuate to n.s. because sponsored cells per (rank × tight) cell are thin (n=12–39 in AN-050; n=51–208 in AN-045) and few candidates have within-candidate variation across multiple cells. The leader-in-tight pattern is so far a cross-candidate descriptive finding; a panel design with more sponsored cell mass — either pooling across cycles or extending the sample to the full 14.9k registered universe — is the route to elevating confidence from yellow.

Supporting analyses