Which pollsters lean Democratic or Republican?

The short version

• Most of what looks like a pollster's lean is the year. In 2016 and 2020 nearly every pollster overstated Democrats; in 6 of the 14 election years since 1998 polls overstated Republicans.
• Measured against other pollsters in the same elections, 26 of 77 lean Democratic and 18 lean Republican by more than chance. The rest can't be told apart from even.
• The strongest leans belong to campaign pollsters, firms whose polls are mostly paid for by one party: 9 of the 10 most Democratic-leaning and 7 of the 10 most Republican-leaning. Among independent pollsters the leans are mostly a point or two.
• Leaning is not the same as being wrong. Which way a pollster leans has almost nothing to do with how accurate it is (correlation -0.03); how far it leans matters somewhat (0.31).

First, take out the year

A pollster's raw record mixes two things: how it polls, and which years it happened to poll. The year matters a lot. The chart shows how far the average final poll missed in each election year, counting each pollster once per election.

So a firm that only polled in 2016 and 2020 looks Democratic-leaning even if it polled exactly like everyone else. SurveyMonkey, which polled only from 2016 to 2020, overstated Democrats by 5.3 points on average, but by only 1.0 more than the other pollsters in the same elections. RMG Research looks D+4.7 on its raw record and is R+0.4 against its peers. It works the other way too: Trafalgar's raw record is R+2.2, but in the years it polled, everyone else was overstating Democrats, so against its peers it leans R+3.8.

That is why everything below compares each pollster with the other pollsters in the same elections: its miss minus their average miss. The year's shared error, and each election's own surprise, cancel out. It is the same idea as the “house effect” that poll averages, including ours, subtract.

Every pollster's lean

Each row is a pollster with at least 25 elections shared with other pollsters. The dot is its average lean against the others; the bar is the range that lean could plausibly be in (a 95% interval). A short bar means a long, consistent record. A bar that crosses zero means we can't say the pollster leans at all.

Most of the firms at the two ends are campaign pollsters, the hollow dots. Firms like Mellman, GQR and Lake Research on the Democratic side, and McLaughlin, Public Opinion Strategies and OnMessage on the Republican side, mostly poll for one party's candidates. Campaigns release the polls that help them, so the published ones lean their way even if the firm's methods are sound. Our model corrects for that: a party-sponsored poll is shifted back toward the other side before it enters the average.

Among the 59 independent pollsters, 15 lean Democratic and 11 lean Republican by more than chance, and 49 of the 59 are within 2 points of their peers. Well-known names like NYT/Siena, YouGov, Quinnipiac and Ipsos are within half a point of their peers.

Does leaning make a pollster less accurate?

Not much. The chart puts each pollster's lean against its accuracy, both measured against the other pollsters in the same elections. Higher is more accurate. If leaning made a pollster worse, the dots would sag toward the bottom on both sides; instead they scatter.

A pollster can lean one way every year and still be accurate, because the lean is small next to the ordinary error of a poll. AtlasIntel leaned R+2.6 against its peers and was still among the most accurate, because in the years it polled the others were overstating Democrats. What makes a pollster inaccurate is mostly noise, not direction.

How we measured it

• Polls: every Democrat-vs-Republican poll for president, Senate, governor or House taken in the final three weeks of a general election, 1998–2024, from FiveThirtyEight's archive and its 2024 poll feed. Elections where the two major parties got under 80% of the vote are left out.
• Each pollster counts once per election: its polls in that election are averaged, so a firm that polled a race ten times doesn't count ten times.
• Lean is the pollster's average error (its Democratic margin minus the result's) minus the average error of the other pollsters in the same election. Elections no other pollster covered are left out.
• The 95% range treats each election as independent. Elections in the same year share a polling miss, so for a firm whose record is mostly one or two years the true range is wider than shown.
• Campaign pollster: at least half of its polls in our data were paid for by a party or a partisan group, as FiveThirtyEight flags them.
• Pollsters are grouped by FiveThirtyEight's pollster ID and shown under their latest name; Rasmussen appears twice because its two eras have separate IDs.

For pollster accuracy ratings, see the pollster leaderboard. For how far any single poll can be trusted, see what “margin of error” really means. The same question for forecasters is in which forecasters lean Democratic or Republican.