When the Final Polls Said One Thing and Voters Did Another
Every midterm, the final polls set the expectations for election night. Most of the time they land within a few points of the result. Sometimes they miss badly enough to pick the wrong winner.I went back through the Senate and governor races that missed by the most, took the final polling average for each one, compared it with the certified result, and measured the gap in points.
Since 2010, the final average has picked the wrong winner in at least 19 Senate and governor races. The misses do not lean one way, either. In 2014 the polls underestimated Republicans almost everywhere. In 2022 the averages underestimated Democrats in most of the closest races.
Here are the biggest misses on one chart, what went wrong in each, and the years when the polls were right and the experts were not.
The Biggest Polling Misses in Midterm Election History
The final polling average against the certified result, race by race. Some missed by more than ten points. Nineteen picked the wrong winner.
Final RealClearPolitics polling average, as archived, against the certified result, for Senate and governor races in midterm years. The miss is the gap between the two margins, in percentage points. Kansas Senate 2014 was Pat Roberts against independent Greg Orman, who sits on the left side.
For every race, I took the final RealClearPolitics polling average as it appears in RCP's archive and compared it with the certified result from the FEC or the state election office. The miss is the difference between the two margins. If the average had a candidate up 3 and he lost by 2, that is a 5-point miss.
I used RCP because it has the longest archive of final averages. FiveThirtyEight's averages, which weight polls by pollster quality, were noticeably closer in some 2022 races, and I note where that matters. Races from before averages existed are covered separately below, because a single poll isn't the same thing.
The races where the polls had the wrong winner
A miss of five points in a race decided by twenty is a footnote. A miss of five points in a close race changes who gets sworn in. Since 2010, there have been at least 19 midterm Senate and governor races where the candidate leading the final average lost:
Same measure. Every line in this chart crosses the centre: the grey dot and the coloured dot are on opposite sides of a tie.
Georgia's 2022 Senate race nearly made the list too. Herschel Walker led the final average by 1.4, Raphael Warnock finished first on election night, and Warnock won the runoff a month later.
Most of these were races the average had inside two points, which is about as close to a coin flip as a poll gets. The ones that stand out are the races where the eventual loser led by three points or more going into election day: Illinois governor in 2010 (Bill Brady up 4.7, Pat Quinn won), Florida governor in 2018 (Andrew Gillum up 3.6, Ron DeSantis won), Arizona governor in 2022 (Kari Lake up 3.5, Katie Hobbs won), Nevada Senate in 2022 (Adam Laxalt up 3.4, Catherine Cortez Masto won) and Colorado Senate in 2010 (Ken Buck up 3.0, Michael Bennet won).
Kansas 2014: an 11-point miss in a toss-up
The single biggest wrong-winner miss is Kansas in 2014. Senator Pat Roberts was in trouble against independent Greg Orman, and the final average had Orman ahead by 0.8 points. Roberts won by 10.6. That is an 11.4-point miss in a race the polls called a dead heat.
The governor's race on the same ballot missed the same way. The average had Democrat Paul Davis up 2.0 on Governor Sam Brownback. Brownback won by 3.7. Chapman Rackaway, a political science professor at Fort Hays State, summed it up after the election: Republicans turned out in greater numbers than the polls expected, and the samples did not reflect who actually voted.
2014: the year the polls leaned left
Kansas was not an isolated case. In 2014 the polls underestimated Republicans almost everywhere. Nate Silver calculated that the average Senate poll in the final three weeks overstated the Democrat's performance by about 4 points, and governor polls were off by about 3.4 points in the same direction.
Virginia is the one people remember. Warner was treated as safe all fall and nearly lost to Ed Gillespie. Nate Cohn of the New York Times traced it to turnout: the largest drops were in Democratic-leaning areas.
Maryland isn't on the chart because RCP had no governor average there that year, but it belongs in this story. The three final nonpartisan polls had Democrat Anthony Brown up 2, 7 and 13 points. Republican Larry Hogan won by 3.8. Silver's summary was blunt: literally every nonpartisan poll called the race wrong.
2022: the averages missed the other way
Eight years later, the misses ran in the opposite direction. In most of the closest races of 2022, the RCP average underestimated Democrats:
Not every 2022 miss went that way. In Florida, the average had Ron DeSantis up 12.2 for governor, and he won by 19.4, a 7.2-point miss in the Republican direction.
Why the averages leaned Republican in the swing states is still argued about. Critics said RCP included too many Republican-leaning pollsters. RCP's Sean Trende disputed that. FiveThirtyEight's averages, which down-weight lower-rated pollsters, were closer: Michigan at Whitmer +4.8 instead of +1.0, and Washington at Murray +4.7 instead of +3.0. When the election analysis site Split Ticket went back over 2022 two years later, it found the overall effect of Republican-leaning polls on the averages was small, about a third of a point, with Pennsylvania the main exception.
Nevada keeps doing this
One state shows up on the wrong-winner list three times. Nevada polls have repeatedly underestimated Democrats:
The usual explanation goes back to Reid's race in 2010. The polling firm Latino Decisions found that the public polls in Nevada had only 50 to 100 Latino voters in their samples, with few or no interviews in Spanish. Those polls had Reid winning 55 to 70% of Latinos. Latino Decisions measured 90%. FiveThirtyEight's Clare Malone later added two more reasons Nevada is hard to poll: the population is highly transitory, and many voters work night shifts at hotels and casinos, so they are not home when pollsters call.
Before there were polling averages
Go back before the mid-2000s and there usually isn't an average to measure against, just a handful of final polls. A single poll can miss for reasons that have nothing to do with the electorate, so I kept these off the chart. But five of them are too big to leave out.
Georgia 2002 is the biggest gap I found. Barnes was an incumbent Democrat who led in every public poll and outspent Perdue six to one. Political scientists Danny Hayes and Seth McKee later traced the upset to rural white voters, who turned out in greater numbers and swung hard to Perdue, angry over Barnes's decision to shrink the Confederate emblem on the state flag and over his education changes. The polls did not pick up how strongly that played. On the same day, Saxby Chambliss beat Senator Max Cleland after trailing in most of the final polls.
Minnesota 1998 is the most famous. The final Mason-Dixon poll taken October 23 to 25 had Democrat Skip Humphrey at 34, Republican Norm Coleman at 33 and Ventura, the Reform Party candidate, at 23. Ventura won with 37.0%, Coleman had 34.3 and Humphrey finished third at 28.1. The polls did see him coming. Mason-Dixon had Ventura at 7% in June and 15% in mid-October, and the Star Tribune's final Minnesota Poll had him at 27. They still had him in third. Minnesota lets voters register on election day, and a record 332,540 people did. In the exit poll, 28% of Ventura's voters said they would not have voted at all if he hadn't been on the ballot. A likely-voter screen built on past turnout was never going to find them.
New York 1994 was a wrong winner in a wave year. The New York Times, Quinnipiac, the Daily News and Marist all had Cuomo ahead in their final polls. Only a New York Post poll had Pataki in front. Pataki won 48.8% to 45.5%.
California 1982 is still argued about. Bradley, the Black mayor of Los Angeles, led the final Field Poll 47 to 41 and lost by 1.2 points. The race gave its name to the Bradley effect, the idea that some voters tell pollsters they will support a Black candidate and then do not. Whether that is what happened is disputed. Deukmejian's own pollster had the race nearly tied by the final weekend, and a Republican absentee-ballot drive made the difference: Bradley actually won the votes cast on election day.
New Jersey 1990 is the reverse of an upset: the favourite won, barely. Bradley was seeking a third term and outspent Christine Todd Whitman about four to one. Whitman, then little known, came within 3 points during a voter revolt over Governor Jim Florio's $2.8 billion tax increase.
When the polls were right and the experts were wrong
Some of the most famous midterm surprises were not polling misses at all. The national polls landed close to the result. What missed was the expectation built on top of them.
2022. Going in, the talk was of a red wave. FiveThirtyEight's main forecast gave Republicans a 59% chance of taking the Senate and an 84% chance of the House. Prediction markets gave Democrats roughly a one-in-three chance of keeping the Senate. Republicans won the House by 222 seats to 213, a gain of nine, and Democrats picked up a Senate seat and kept control. The national polls were close: the final RCP generic ballot was Republicans +2.5, and Republicans won the national House vote by 2.8. Split Ticket's verdict on every race outside Pennsylvania: "it was a failure of punditry rather than a failure of polling."
1998. With Bill Clinton facing impeachment, Republicans were expected to gain House seats. Democrats gained five. It was the first time since 1934 that the president's party picked up House seats in a midterm. This one was partly a polling miss too: FiveThirtyEight's database shows the race-level polls overstated Republicans by 4.3 points on average that year, the largest bias of any midterm it measured.
1994. Gallup's final poll had Republicans winning 53.5% of the two-party House vote. They got 53.5%. The poll nailed the vote. Few people saw what that meant for seats: Republicans gained 54 seats and took the House for the first time in 40 years. House Speaker Tom Foley lost his own seat in Washington state.
2010 went the other way. Gallup's final likely-voter poll had Republicans up 15 on the national ballot. They won it by 6.8, so Gallup overstated them by about 8 points, and Republicans still gained 63 seats.
The national polls: closer than the races
The generic ballot, which asks people which party they will vote for in their House race, has been more reliable than individual races in recent midterms:
| Midterm | Final RCP average | National House vote | Miss |
|---|---|---|---|
| 2010 | R +9.4 | R +6.8 | 2.6 |
| 2014 | R +2.4 | R +5.7 | 3.3 |
| 2018 | D +7.3 | D +8.4 | 1.1 |
| 2022 | R +2.5 | R +2.8 | 0.3 |
National House vote margins as reported by RCP. The Clerk of the House figures put 2018 at about D +8.6.
Which way do the polls lean?
If you are looking for a pattern, there isn't one that holds up. FiveThirtyEight measured the average bias of race-level polls in every midterm since 1998. The direction flips:
Average statistical bias of polls in the final three weeks, all races. Source: FiveThirtyEight (1998 to 2014) and Nate Silver's Silver Bulletin (2018 and 2022).
The size of the miss is fairly steady. In FiveThirtyEight's numbers, the average midterm poll missed the final margin by about 5 to 6 points from 2002 to 2014, and by 7.5 in 1998. What changes from year to year is the direction. After 2014, Silver described the bias as largely unpredictable from one election to the next, and the record since then has borne that out.
Why midterm polls miss
Turnout. Midterm turnout is much lower than in presidential years, so pollsters have to guess who will actually vote. Pew Research tested this with validated voter records and found that more than a third of the registered voters in its sample did not vote, even though many of them said they had. That makes the likely-voter screen one of the biggest judgment calls in a midterm poll. Kansas and Virginia in 2014 both came down to turnout.
Who answers. If one group of voters is less willing to take a poll, weighting by age, race and education doesn't always fix it. The American Association for Public Opinion Research's review of the 2020 election concluded that at least some of the error came from people who did not respond, and that weighting to party and past vote did not fully correct it.
Who gets reached. Nevada's Latino voters in 2010 and Minnesota's first-time voters in 1998 are the same story: voters the polls were not reaching.
What goes into the average. An average is only as good as the polls in it. 2022 is still argued about for exactly that reason.
The final polling average usually lands within a few points. But in a midterm, a few points decides a lot of races, and since 2010 the average has picked the wrong winner at least 19 times in Senate and governor races alone.
The misses do not reliably favour either party. The 2014 polls underestimated Republicans, and the 2022 averages underestimated Democrats in most of the close races. A two-point lead in a final average is not a safe lead, whichever side it is on.
Sources: RealClearPolitics archived final polling averages (realclearpolling.com); Federal Election Commission, Federal Elections compilations of certified results, 1990 to 2022; state election offices for governor results (Kansas Secretary of State, Minnesota State Canvassing Board, Illinois State Board of Elections, Colorado Secretary of State and others); FiveThirtyEight polling averages, forecasts and pollster accuracy analyses (Nate Silver, 2014, 2018, 2023); Silver Bulletin (2026); Gallup final pre-election estimates; Pew Research Center (2006, 2016); American Association for Public Opinion Research, 2020 Task Force report; Split Ticket (2022, 2024); Latino Decisions (2010); Minnesota Public Radio, Mason-Dixon poll (October 1998); Hayes and McKee, Booting Barnes, Politics and Policy (2004); Eagleton Institute, Rutgers (1990); Field Poll (1982); Minnesota Secretary of State election statistics; Clare Malone, FiveThirtyEight (2016); New York Times; CBS News (1998).
Filed under: General Knowledge
