β × days_to_election slope = +0.037 pp/day (p=0.054). Polls 6 months out are slanted by ~13 pp; polls in the final week by ~7 pp. Slant shrinks toward the election as verification looms — direction matches the verifiable-disclosure prediction.
Design
Results
Table: Days-to-election interaction coefficients
| Coefficient | Estimate | SE | p |
|---|---|---|---|
sponsored_by (β at days = 0) |
+6.74 | 1.36 | <0.001 |
sponsored_by × days_to_election |
+0.037 | 0.019 | 0.054 |
(from source/analysis/heterogeneity.py)
Table: Implied β at representative time-to-election
| Days before election | Implied β |
|---|---|
| 7 | +7.00 |
| 30 | +7.85 |
| 90 | +10.07 |
| 180 | +13.41 |
(from build/table/heterogeneity.csv)
Slant magnitude almost doubles moving from the final week to ~6 months out.
Interpretation
- Direction matches the verifiable-disclosure prediction ([hyp:verifiability-decay-b]). Slant carries a future cost because the election eventually verifies the poll. A poll fielded 6 months out has very weak verification pressure (the campaign hasn't crystallized; the verification step is half a year of news cycles away). A poll fielded in the final week has very strong verification pressure (the result is known days later, and the gap between poll and result is precisely what political actors, journalists, and the lawsuit channel scrutinize).
- Marginally significant slope. p = 0.054. It survives the sample restriction to sponsored-only rows (not shown — informally verified). Magnitude is economically large: half the headline β appears to decay over the 6 months before the election.
- Consistent with Channel B but not discriminating. Tracks the
Channel B (residual / fabrication) story in the project's
docs/theory.md§ Bayesian persuasion — Channel B should be most active where verification is weakest. But it does NOT discriminate Channel A from B per se; design-driven slant could also decay if methodology choices have to look more defensible closer to the verifiable event.
Confidence rationale (yellow). Slope is borderline significant (p = 0.054) and the direction is consistent with theory, but the test does not separate Channel A from Channel B and rests on a single specification. A 2022-cycle extension and a sample-split estimator would tighten the inference.
Follow-ups
- p = 0.054 is borderline. The 2022 cycle extension would help — more time-to-election variation, more candidates contributing.
- The decay slope's interaction with sponsor route (CPF, committee, party) would tell us which sponsor types let slant decay most. Probably Route B (committees) — they have professional staff who understand verification dynamics.
- Sharpest version of this test: a sample-split estimator (last-month polls vs early polls) rather than a continuous interaction. Larger gains in interpretability.