Methodology

Pollster house effects, in plain language

Some polling firms consistently show a race a point or two more Democratic, or more Republican, than everyone else polling the same thing at the same time. That gap is called a house effect. It is not the same as being wrong — it is measured against other pollsters, not against an election result that hasn't happened — but leaving it in means the average partly reflects who happened to poll recently rather than what the electorate thinks.

1. Measure one poll against its contemporaries. For every poll, we build an average of what other firms found on the same question — the generic ballot, or that specific race — within 45 days either side of it, weighted the same way the model's own averages are. The gap between the poll and that average is the poll's residual. The firm's own polls are excluded from the comparison, so nobody is ever measured against themselves.
2. Only compare what is comparable. A residual needs at least 3 other firms in the window; against one or two, the "gap from the field" is mostly the other firm's own lean. Polls in races where the polling disagrees about who is actually running, and polls fought against a placeholder opponent, are left out entirely — the same rule that keeps them out of the averages.
3. Pool a firm's residuals into one number. All of a firm's residuals, across the generic ballot and every race, average into a single estimated lean. Older polls count for less, halving every 180 days, because a firm's methods change.
4. Shrink it toward zero, then cap it. A firm with three polls might look 6 points off the field purely by chance, so the raw number is pulled toward zero by how much evidence stands behind it: a firm keeps n / (n + 5) of its raw effect, where n is its number of usable polls. Whatever survives that is capped at 3.0 points, just above the largest house effect published research reports. Firms with fewer than 3 usable polls have their estimate shown but not applied.
5. Subtract it, and publish it. The surviving number comes off each of that firm's polls before they are averaged, and every firm's figures are on the pollsters page — raw, shrunk, and whether it was used. Race pages show the adjustment poll by poll.

One simplification, stated plainly. House effects and polling averages depend on each other: the "field" a poll is compared against is itself made of polls with their own leans. 538 handles that by running its aggregation three times, and Silver Bulletin loops the average and the house effects until the numbers stop moving. We compute house effects once, from unadjusted averages, and then use them. That is a simplification, not a refinement, and it means our effects are slightly biased toward zero for firms polling alongside other firms that lean the same way.

Known limitations

• One correlated error component is missing. 538's published House model decomposes error into five parts — national 3, regional 2, state 2, demographic-cluster 2, and 6 at the district level. We carry four of them at their values. The one we omit is the demographic-cluster term, which correlates districts that resemble each other regardless of where they are; building it needs a clustering of all 435 districts that we have no data for, and inventing one would be inventing the correlation it is supposed to measure. Our correlated total is therefore 4.12 against their 4.58. Measured on the board of 2026-08-12 — same polls, same seed, only the error model changed — adding a fifth component of 2 points takes between 0.4 and 3.4 points off the House control probability, depending on whether the missing term is assumed to behave like a state, a regional or a national one. The far end of that range assumes it correlates every district in the country at once, which a clustering can't do — it sorts the 435 districts into groups, so a shock hits some of them and the rest of the chamber absorbs it. The honest reading of the House figure is therefore a point or so softer than printed, and at the outside three, in whichever party's favour it leans. The Senate moves by under a point in either direction across the same three assumptions, which at 10,000 simulations is inside sampling noise — so we don't claim a direction for it. (For scale: the House control probability was 93.3% on model run 19, 2026-08-12. That is a figure from one run, not a live reading — the current one is on the front page.) All of this is a much smaller gap than this page reported before 2026-08-12, when the model had one correlated term instead of four; the arithmetic is in docs/BUILD_NOTES.md Phase 10 §C, and the run behind the scale figure is in §H.
• House effects are computed in a single pass. The correction described above is estimated once, from averages that have not themselves been corrected, where the two published models that do this iterate until the numbers settle. The effect of the shortcut is small and one-directional: where several firms polling the same thing lean the same way, each is measured partly against the others' lean, so all of their estimated effects come out a little too close to zero. The correction is therefore conservative rather than overcooked.
• 2024 House baselines are redistricting-stale. Several states redrew their House districts for 2026. A district's baseline is still its 2024 result on 2024 lines — for a redrawn district, that describes different territory than the one actually on the 2026 ballot.
• Imputed baselines. 37 of 435 House districts (8.5%) had no major-party opponent in 2024 at all, so their baseline is a fixed ±35.0-point assumption signed toward the winner, rather than a real two-party result. These carry a small "est." marker on the House table.
• Unsettled nominees. The Senate candidate list (db/seed_data/senate_2026.yml) is maintained by hand, not scraped — re-running the seed task does not pick up a primary result; only editing that file does. 10 of 35 Senate races (28.6%) are currently affected: where a party's nominee isn't settled yet, the model uses the declared front-runner where one has emerged, or a generic party prior where it hasn't — never a guess at a specific name.