US Election Polling Error History: A Complete Timeline

A polling error is not the gap between a poll’s number for one candidate and that candidate’s final vote share. It is the gap between the margin the poll implied and the margin the election actually produced. A final poll showing the Democratic candidate at 49% and the Republican at 46% implies a Democratic margin of +3. If the real result is Democratic 48%, Republican 50% — a Republican margin of +2 — the poll’s margin error is not 2 points. It is the distance between +3 (D) and -2 (D), which is 5 points. This distinction is why AAPOR and most professional post-election reviews always report error on the two-party margin, not on any single candidate’s number in isolation.

Professional reviews also separate signed error (which direction the polls leaned, and by how much) from absolute error (how far off the polls were, regardless of direction). A pollster average that overstates Democratic margins by 4 points nationally in ten different states has a signed error of +4 toward Democrats even if the absolute size of the miss varies state to state. Averaging many biased polls together reduces random noise, but it does not reduce a bias every pollster shares — which is exactly what happened in 2020.

It is equally important to separate a poll from a forecast. A poll reports what a sample of respondents said when asked who they support. A forecast — the kind that produces “70% chance to win” — combines polls with modelling assumptions to estimate a probability. National polls showing Hillary Clinton roughly 3 points ahead in 2016 were reasonably close to the actual 2-point popular-vote result; it was largely the win-probability forecasts built on top of those polls, several of which implied Clinton was overwhelmingly likely to win, that created the sharper sense of shock on election night. A poll that gives one candidate a modest lead is not the same claim as a forecast that calls the race decided.

Before scientific polling existed, magazines and newspapers ran “straw polls” by mailing ballots to as many readers, subscribers or listed households as they could reach. In the 1936 presidential race between incumbent Franklin D. Roosevelt and Republican challenger Alf Landon, the magazine Literary Digest mailed roughly 10 million ballots and received about 2.4 million responses — a sample size far larger than any professional poll uses today. The Digest predicted a comfortable Landon victory. Roosevelt won in a landslide, carrying 46 of 48 states.

The failure traced to two compounding problems. First, the Digest built much of its mailing list from telephone directories and automobile-registration records — in Depression-era America, phone and car ownership skewed sharply toward wealthier households, who leaned more Republican than the electorate as a whole. Second, the roughly 76% of recipients who did not respond were not a random subset of who did — a classic case of nonresponse bias. The lesson the industry drew from 1936 became foundational: 2.4 million biased respondents can produce a worse estimate than a few thousand carefully, randomly selected ones. George Gallup’s newer probability-sampling methods, which correctly anticipated Roosevelt’s win using a dramatically smaller sample, gained credibility from the Digest’s public humiliation and helped establish scientific sampling as the industry standard.

Incumbent Harry Truman faced Republican challenger Thomas Dewey in 1948. Major pollsters showed Dewey comfortably ahead for most of the fall, and some organizations grew confident enough in the result that they stopped interviewing weeks before Election Day. Truman won. The Chicago Tribune, working against an early print deadline and relying on that polling consensus, ran the now-famous incorrect front page: “DEWEY DEFEATS TRUMAN,” a copy of which Truman held up for photographers after his victory.

Two methodological problems were identified afterward. Pollsters had stopped surveying too early, missing a late shift of undecided and wavering voters toward Truman in the campaign’s final weeks. And several major pollsters used quota sampling, in which interviewers were assigned demographic quotas to fill but retained personal discretion over exactly whom within those quotas to approach — a design that can introduce selection bias in ways random sampling avoids. The two lessons the industry took from 1948 reshaped polling practice for decades: keep interviewing all the way to Election Day, and reduce interviewer discretion in who gets selected.

The decades between 1948 and 2012 were not free of surprises, but most of them were not polling failures in the way 1936 or 1948 were. John F. Kennedy’s razor-thin 1960 win over Richard Nixon was broadly anticipated as a close race by pre-election polling. Ronald Reagan’s roughly 10-point 1980 win over Jimmy Carter came after many late polls showed a tighter contest — a case where a meaningful share of voters appear to have moved late, rather than pollsters having measured the electorate wrongly at the time they surveyed it. Bill Clinton’s 1992 three-way race against George H. W. Bush and Ross Perot tested pollsters’ ability to model turnout with a serious third-party candidate in the mix. And 2000’s Bush-Gore race demonstrated, starkly, that an accurate national popular-vote estimate and an accurate Electoral College outcome are not the same test: Gore narrowly led the national popular vote, while Bush won the Electoral College after Florida’s disputed recount.

2012 is worth flagging on its own: several high-profile polls and averages that cycle underestimated Barack Obama’s margin over Mitt Romney — a reminder that polling error has, historically, run in both partisan directions rather than consistently favoring one side. That fact matters directly when interpreting the pattern that emerges starting in 2016.

Final national polling ahead of the 2016 Trump-Clinton race implied roughly a Clinton +3 popular-vote margin. The actual national result was close to Clinton +2 — a modest, unremarkable miss by historical standards, and not the “historic polling disaster” the immediate post-election narrative described. AAPOR’s own analysis found the 2016 national presidential polls performed reasonably well relative to prior cycles, with a 2.2-point average absolute error.

The real failure was concentrated in state-level polling in a handful of critical battlegrounds — Wisconsin, Michigan and Pennsylvania chief among them — where polls understated Donald Trump’s support by enough to flip the eventual Electoral College outcome away from what national coverage had led audiences to expect. AAPOR identified two specific, well-supported contributors: many state polls did not adequately weight their samples by education, a variable that became strongly correlated with vote choice in 2016 as college-educated respondents (who participate in surveys at higher rates) leaned more Democratic than non-college respondents; and a meaningful share of late-deciding voters in key states broke disproportionately toward Trump, after many final polls had already stopped fielding. The popular “shy Trump voter” theory — that Trump supporters were embarrassed to admit their preference to pollsters — was widely discussed, but post-election research did not establish it as a primary driver of the error.

Joe Biden defeated Donald Trump in 2020, and polling correctly identified Biden as the winner both nationally and in the Electoral College. But the margins were substantially overstated. AAPOR’s post-election evaluation of national presidential polling in the campaign’s final two weeks found an average signed error of 3.9 points overstating Biden’s margin, and an average absolute error of 4.5 points — more than double the 2.2-point national absolute error in 2016. State-level presidential polling showed a 4.3-point average signed error toward Biden and a 5.1-point average absolute error — identical, in absolute terms, to the 2016 state-level error, even though the underlying causes were not identical. The overstatement was not limited to the presidential race: many Senate and gubernatorial polls that cycle also overstated Democratic performance.

Unlike 2016, AAPOR did not identify one clean explanation. Its investigation examined likely-voter modelling, survey mode, sampling and weighting methodology, and pandemic-era effects on who was reachable and willing to respond. The concern that gained the most traction was partisan nonresponse: the possibility that people who supported Trump were systematically less likely to participate in surveys than demographically similar Biden supporters, which ordinary demographic weighting (age, sex, race, education) cannot fully correct, because it does not know the political leaning of the people it never reached in the first place.

Ahead of the 2022 midterms, political commentary increasingly floated a possible large Republican wave. The actual outcome was more mixed — Democrats retained the Senate, Republicans narrowly took the House — and post-election reviews found 2022 polling performed among the most accurately of any recent cycle by historical measures. The gap between the “red wave” narrative and the polling record is itself a useful lesson: a handful of high-profile polls, forecasts or pundit predictions circulating in coverage are not the same thing as the aggregate performance of the polling industry, and conflating the two produces a misleading verdict on “the polls” either way.

After President Biden exited the race, Vice President Kamala Harris became the Democratic nominee against Donald Trump. Both national and swing-state polling showed an extremely close race through Election Day, with many surveys landing differences inside the margin of sampling error. Trump won the national popular vote and the Electoral College. AAPOR’s post-election review, covering 611 polls of presidential, Senate, gubernatorial and House races conducted in the campaign’s final two weeks, found an average absolute two-party-margin error of 3.3 points — down sharply from 5.2 points in 2016 and 5.3 points in 2020. National presidential polling averaged a 2.6-point absolute error; state presidential polling averaged 3.0 points, which AAPOR describes as the most accurate state-level presidential polling on this measure since 1944.

The pattern from 2016 and 2020 did not fully disappear, though. Averaged across offices, 2024 polls still overstated Democratic margins by 2.7 points, compared with 3.1 points in 2016 and 4.6 points in 2020 — smaller each time, but the same direction for a third consecutive presidential cycle. AAPOR’s task force also investigated whether many 2024 swing-state polls looked suspiciously similar to one another because pollsters were “herding” — adjusting results toward a consensus rather than publishing genuine outliers. It found no evidence supporting the herding explanation; the more likely driver of that similarity was broader, independent adoption across many pollsters of political weighting variables — party registration where available, recalled past vote, or partisan identification — layered on top of standard demographic weighting.

Turnout modelling remained an unresolved source of error even in an improved year. AAPOR found that many 2024 polls did not fully anticipate within-state turnout shifts: counties that had favored Trump in 2020 saw turnout grow more than counties that had favored Biden, and polls built around an electorate that resembled 2020’s turned out to understate that shift. Preference measurement and turnout measurement are two separate problems, and 2024 shows a year where the first improved substantially while the second remained genuinely hard.

No single cause explains every polling miss in this history. Each era’s failure traces to a different combination of these underlying error sources, and small errors from several sources at once can add up to a result that looks, from the outside, like one dramatic failure.

A single poll can be an outlier by chance; that is expected and is exactly what a margin of sampling error describes. Averaging several polls together reduces that kind of random noise, which is why poll averages are generally more stable and more reliable than any one poll in isolation. But averaging cannot correct a systematic error that most or all pollsters share — if ten different pollsters are all affected by the same nonresponse pattern, as AAPOR’s 2020 analysis suggests happened, averaging their results together does not cancel that bias out. It can even make the biased number look more precise than it is, because the average’s narrower range creates a false sense of certainty. This is also why the traditional published margin of sampling error should not be read as covering every kind of uncertainty in a poll: it describes sampling variability under standard statistical assumptions, not nonresponse bias, weighting error, turnout modelling error, or late movement in the electorate.