Rating and adjusting American pollsters

Summary

American polling is large, fragmented and partly paid for by the people it measures. In the races we forecast this year, 402 of the 811 Senate polls, 214 of the 663 governor polls and 277 of the 358 House district polls were paid for by a campaign, party or allied group, or come from a firm that works mainly for one side. This paper describes the two separate things we do with pollsters.

The first is for readers: a rating out of 100 for every pollster with a final poll on record, on the same scale as our ratings in other countries, combining the accuracy of its final polls since 2000, its miss on the margin and how much of its working it publishes. On 15 October 2026, 502 US pollsters are graded. The rating does not weight any poll.

The second is for the model, and it has three parts, each measured from final polls since 2000: a house effect for each firm (how far its final polls sat from other firms' polls of the same races), a record weight (its typical miss against the typical pollster's), and the treatment of partisan and sponsored polls, which leaned 2.52 points to their side and are both moved back by that amount and counted at half weight. The paper ends with the split of Hart Research Associates and Public Opinion Strategies into three codes, made on 14 October 2026.

Two uses, kept apart

A rating answers a reader's question: which pollsters have a good record and can be checked. A model adjustment answers a different one: given this poll from this firm, the best estimate of the race. The two can disagree. A firm with a steady lean can have a modest rating and still be very useful to the model, because a steady lean can be taken off; a firm with a good record on a handful of races can be rated well but tell the model little. We therefore compute them separately, from the same final polls, and say which is which.

The ratings

The rating uses the method of every non-British country on the site (Rating the pollsters: accuracy, the Labour–Conservative lead and openness describes the British one), scored out of 100:

• Accuracy (50 points). For each final poll (the pollster's last poll of a race, ending in the month before polling day) of a Senate, governor or presidential race since 2000, the average miss on the two main parties' shares of the decided vote. Each race counts as one election. The pollster's average is ranked against every US final poll: the more of them it beats, the more points, with points falling with rank on a curve that gives the typical (median) final poll about 30 points and one in the best tenth about 46.
• Lead (20 points). Its miss on the margin between the two leaders, ranked the same way.
• Openness (30 points). What it publishes so its figures can be checked: raw or unweighted figures (8), full tables or crosstabs (8), a method note (4), the sample size (3), the undecided share (3), fieldwork dates (2) and membership of the AAPOR Transparency Initiative (2). An item not yet checked counts neither for nor against; a pollster none of whose items has been checked is rated on its record alone, scaled to 100, and marked "record only".

A pollster with a short record is pulled towards the typical final poll as if it had two more elections at the typical level, so that one lucky race cannot earn a top grade:

Grades: A from 75, B from 60, C from 45, D from 30, E below 30. Because a US Senate race misses by more than a Dutch or German election, each pollster is ranked against the US's own final polls, and US grades cannot be compared with other countries'.

Some of today's grades:

Table: Selected US pollster ratings, 15 October 2026. "–": openness not yet checked (rated on record alone). Races: final polls on record.

The grades of the largest pollsters cluster between 54 and 71. That is what a ranking against one's own country produces when the big firms have polled hundreds of races: their averages sit near the median, and the median scores about 30 of the 50 accuracy points. The top grades go to firms with good records on fewer races and full disclosure. The ratings page shows every pollster, its polls now, its lean at past elections and its latest poll against today's average.

What the model does with pollsters

House effects

A pollster's house effect is how far its final polls sat from the other final polls of the same races. For every race since 2000 with three or more final polls (Senate, governor and presidential), and every final poll \(j\) by firm \(f\) in it,

the poll's margin minus the mean of the race's other final polls. The firm's house effect is the mean of its gaps, shrunk towards zero as if it had ten more races with no lean:

A positive house effect means the firm has leaned to the Democrats against its peers. It is taken off each of its polls before averaging (State polls: the pollster-adjusted average and the blend). It measures a firm against other firms, not against results: if every firm overstates the Democrats, no house effect shows it, and that shared miss is what the statewide correction and the simulation's national error deal with.

Table: House effects (equation 3, points of margin, positive towards the Democrats), typical miss (equation 4) and record weight (equation 5), from final polls of races with three or more final polls, 2000–2025.

The spread of house effects among established firms (those with ten or more such races) is 1.83 points, measured as the root mean square of their unshrunk effects. That number is used as the uncertainty about the lean of a firm with no record: a new firm's poll is taken as it stands, with no house effect, but the average it enters is given a wider error (State polls: the pollster-adjusted average and the blend).

Record weights

Each firm's typical miss is the mean absolute miss of its final polls against the result, in the same races with three or more final polls, pulled towards the typical miss of all those polls (4.62 points) as if it had ten more races at that level:

Its record weight is the typical pollster's miss over its own, held between 0.6 and 1.4:

and a firm with no record is treated as missing by 1.1 times the typical, which gives it a weight of 0.91. The weight multiplies each poll's weight in the race averages and the generic ballot.

The record weight and the house effect measure different things. Trafalgar Group's polls have leaned 3.6 points to the Republicans against other firms', the largest house effect in the table, but its typical miss against results is below average: its final polls date from 2016 on, and in most of those cycles the field as a whole overstated the Democrats. The model takes off its lean and gives its polls slightly more than average weight. AtlasIntel, with the smallest typical miss of any firm with 30 or more races, reaches the 1.4 cap.

Partisan and sponsored polls

A poll is partisan in the model if one of three marks is on it:

• a party mark: a poll Wikipedia marks "(D)" or "(R)", or by a firm that works for one side;
• an internal mark: a poll paid for by a candidate's campaign, a party committee or an allied group;
• a sponsor: the named client, from which the side is read: Democratic, DCCC, DSCC, DGA or "progressive" groups, or a campaign written "(D, ...)", count for the Democrats; Republican, NRCC, NRSC, RGA, "conservative" or GOP groups, or "(R, ...)", for the Republicans; "(I, ...)" marks an independent candidate's own campaign.

On 14 October 2026 script 73 added party marks that the import had missed: polls marked "(D)" or "(R)" on Wikipedia, Lake Research Partners' South Dakota Senate poll among them, and the past final polls of firms that work for one side (on the Democratic side Lake, GQR, Global Strategy Group, GBAO, Impact, Normington Petts, Garin-Hart-Yang, Benenson, ALG, Mellman, Hickman, Hamilton, Harstad and Democracy Corps; on the Republican side Public Opinion Strategies, Tarrance, OnMessage, Harper, Fabrizio, McLaughlin, Peak, Moore, Wilson and WPA). Media polls and joint polls by a Democratic and a Republican firm are left unmarked.

To measure how far partisan polls lean, in every race since 2000 with both kinds of final poll each partisan final poll is compared with the mean of the race's non-partisan final polls, signed towards its side:

Across 556 polls, \(\pi\) = 2.52 points (median 2.33). The model then does two things to every partisan poll: it moves the poll back by 2.52 points against its side, and it counts it at half weight. Both are applied in the same way to both parties.

Table: How far partisan final polls sat from the independent final polls of the same race, 2000–2025. Points of margin.

Republican-side polls have leaned further, and we use one figure for both sides. That is a choice for symmetry: the 155 Democratic-side polls are fewer, and a correction that treats the two parties differently would need a stronger test than this record can give. A site article found the same order this year among 216 pairs of 2026 polls taken within 30 days of each other: Democratic-side polls ran 3.4 points more Democratic than non-partisan polls of the same race, and Republican-side polls 2.1 points more Republican.

Independent challengers. Where an independent is the main challenger to a Republican (Nebraska, Idaho, South Dakota and Montana this year), a Democratic sponsor and the independent's own campaign both count as working for the challenger, and a Republican sponsor for the Republican. Before 14 October 2026 Todd Achilles's own polls in Idaho had no side and were not moved; they now are.

How much partisan polls carry today. In the races we forecast, 402 of 811 Senate polls (195 Democratic-side, 200 Republican-side), 214 of 663 governor polls (78 and 130) and 277 of 358 House district polls (212 and 59) are partisan. In the House, where Democratic-side releases outnumber Republican ones by more than three to one, halving their weight and moving them back keeps them from setting district averages on their own; where a district has only partisan polls, the fundamentals carry more of the forecast (From the generic ballot to 435 districts).

The people sampled

Midterm turnout is low and uneven, so likely voters decide the race. In race averages a poll of likely voters counts in full, registered voters 0.7 and all adults 0.5. A registered-voter poll of a race whose main challenger is an independent counts 0.5, because such candidates draw most on the voters least sure to turn out and no past race lets us measure that gap. These factors are set, not fitted, and are shown on the method page.

Firm splits: Hart Research and Public Opinion Strategies

Hart Research Associates is a Democratic firm and Public Opinion Strategies a Republican one. Together they conduct the NBC News and CNBC polls. Until 14 October 2026 all three were one code, named "Public Opinion Strategies", so a poll Hart conducted for a Democratic candidate (Jamie Davis's Louisiana Senate poll was the case that exposed it) appeared as a Republican firm's poll, and the joint media polls shared a record with both firms' partisan work.

They are now three codes. Each poll goes to the firm named by Wikipedia and in the cited release: Democratic-sponsored polls to Hart Research (31 polls, all from this cycle and all marked Democratic), Republican-sponsored polls to Public Opinion Strategies (32 polls, 15 of them past final polls, all marked Republican), and the joint NBC News and CNBC polls to Hart Research Associates / Public Opinion Strategies (9 polls, among them a final generic ballot of 2024). "Illinois Public Opinion Strategies", which polled Illinois in 2016, is a different firm with its own code. Each code now has its own house effect, record and rating.

The same principle applies to other firms with several roles: a pollster's code is the firm that did the fieldwork, and its client goes into the sponsor mark.

Checks on the record

Since 13 October 2026 a past final poll whose figures, undecided included, add to more than 103 is left out of house effects, records and ratings. A check of our ratings against an independent compilation found 31 such polls, most counting the same voters as both "other" and "undecided" (up to 127 in one 2016 Oregon poll), and one Siena poll of New York's 2010 governor race that counted its undecided twice.

This cycle's house effects

A firm's lean can also be measured from this cycle's polls: each race poll against the mean of other firms' polls of the same race within 21 days, and each generic-ballot poll against the others within 10 days, partisan lean taken off first. An owner's setting blends that with the past-final-polls effect, as if the past effect were worth ten of this cycle's polls. It is off. Run today, the tests give:

Table: Average miss on the margin, points. The 2026 rows compare the model without this setting and with it (the generic-ballot row with the weighted average in both).

The gain on past races is a few hundredths of a point, the 2026 race polls are predicted worse and the generic ballot better. Without a clear gain the setting stays off.

Limits

House effects are relative. They measure a firm against its peers. A firm that is right when all others are wrong shows a house effect towards the result; a lean shared by every firm is invisible to them.

The record is uneven. Most firms have fewer than ten races of three or more final polls, and their house effects are shrunk more than half way to zero. Firms change methods and owners: a record describes what a firm did then.

One partisan lean for both sides. The measured lean is larger for Republican-side polls; the model uses one figure.

Marks come from secondary sources. Whether a poll is sponsored is read from Wikipedia and the release. A sponsored poll published without its sponsor is treated as independent.

References

American Association for Public Opinion Research. Transparency Initiative member list.

Figures as of 15 October 2026. The pages they come from are rebuilt with each new poll.