When the candidate pays with their own CPF (n_self=18), β = +19.12 pp (p=0.006) — roughly 2× the committee/party routes. Smallest cell but the sharpest signal; suggests strategic incentives concentrate where the sponsor has most personal stake.
Results
Table: Sponsor-route heterogeneity in β (error ~ sponsored_by × route)
| Route | β | SE | p | n_self |
|---|---|---|---|---|
| CPF (own-pocket) | +19.12 | 6.89 | 0.006 | 18 |
| Committee (CNPJ) | +8.73 | 1.56 | <0.001 | 429 |
| Party (CNPJ via despesa_partidaria) | +7.70 | 2.35 | 0.001 | 42 |
| Party-name parse | +6.33 | 2.59 | 0.015 | 152 |
(from build/table/heterogeneity.csv)
Interpretation
Route A (CPF — the candidate paying with their own CPF) shows β ≈ +19, more than twice the +6-9 in the CNPJ-mediated routes. Two readings sit on the table, and the 18-observation cell is too small to discriminate between them:
- Strategic-incentives reading. Candidates who pay personally (rather than through a committee or party) have the strongest direct stake and exert the most pressure on the pollster. They might also choose less-reputable firms (no intermediary to manage the relationship).
- Selection reading. Candidates who pay with their own CPF are systematically different — likely smaller-budget, possibly without a formal committee structure, possibly first-time candidates. They could be more desperate (or less reputationally constrained).
- Magnitude is large despite the small cell. The 18-observation cell precludes between-route SE comparison being tight, but the difference of ~10-12 pp (CPF vs the others) is economically large and statistically distinguishable from each other route's β at conventional levels.
- Route B is the workhorse. Committee sponsorship (n_self=429) is where most of the headline +7-8 average comes from. Routes C and D (party CNPJs and party-name parses) sit between, consistent with weaker direct candidate control over party-level commissioning.
- Named-committee vs party-name gap. β = +8.73 (Route B, named committee) vs +6.33 (Route D, party-name parse). The named-committee polls are more clearly tied to a specific candidate; the party-name polls may reflect broader party initiatives that affect multiple candidates. Lower direct stake → lower β is consistent.
Confidence rationale (yellow). The four-route ordering is consistent with a direct-stake gradient, but the sharpest signal (CPF, +19.12) rests on only 18 self-sponsored observations and cannot separate the strategic-incentives reading from a selection story. The committee/party/party-name contrasts are tighter but the headline CPF result needs corroboration before it can be load-bearing.
Follow-ups
- The CPF cell (n=18) needs more careful look: are these specific candidates with unusual profiles? Worth running a candidate-level list-out to see who they are and whether they share characteristics.
- This connects to [hyp:funding-source-gradient] which predicts a
similar gradient based on
DS_ORIGEM_RECURSO(Recursos Próprios vs Doações Eleitorais vs Fundo Partidário). The CPF/committee/party split here is sponsor-identity-based; the DS_ORIGEM_RECURSO split is funding-flow-based. Both should agree if the "direct stake → more slant" reading is right. Worth running. - The committee vs party-name split is interesting: β = +8.73 (Route B, named committee) vs +6.33 (Route D, party-name parse). The named-committee polls are more clearly tied to a specific candidate; the party-name polls may reflect broader party initiatives that affect multiple candidates. Lower direct stake → lower β is consistent.