The bias is sender-specific, not a generic pollster house effect
🟡 The +7 pp average is concentrated in candidate-touched commissioning, not pollster-firm-level house effects.
Draws on:
- opponent-sponsored-mirror — β_opp = −1.93 means opponent-sponsored polls understate the candidate by ~2 pp, with the same FE specification. β_self − β_opp ≈ +10 pp.
- pollster-heterogeneity — several mid-volume firms (Verita, IIP, Census) show β near zero. If the +7 pp were a generic pollster house effect, every firm's β should cluster around +7.
Neither pattern is consistent with a generic house effect; both match sender-specific bias.
Strength: 🟡 because the per-pollster β estimates are noisy at the firm level — we can rule out "generic house effect" as the dominant mechanism but can't fully characterize the firm-specific distribution.
Sources.