Why Political Polls Sometimes Get Elections Wrong

A Brief History of Polling Errors in Major Elections

The history of election polling is punctuated by a small number of highly visible failures that have shaped how the field evaluates itself. The most dramatic early example came in 1948, when the major polling organizations — including Gallup — predicted that Republican Thomas Dewey would defeat incumbent President Harry Truman. As documented in academic reviews of polling methodology, that failure was attributed largely to non-random sampling and to the practice of stopping data collection weeks before Election Day, thereby missing a late surge toward Truman.

The discipline largely recovered from 1948 and performed with reasonable accuracy through many subsequent election cycles. However, the 2016 U.S. presidential contest reignited concern about election poll accuracy. Research published by the American Association for Public Opinion Research found that 2016 national polls, while directionally off, were closer to the final outcome than widely perceived; the more significant errors occurred in state-level polls, particularly in key Midwestern states that proved decisive in the Electoral College. In 2020, errors at both the state and national level were broader. The AAPOR post-election task force described the 2020 results as the polls’ worst performance in presidential elections in approximately four decades. The 2024 cycle offered a partial correction, with AAPOR concluding that state-level polls were the most accurate since 1944, though a consistent pattern of underestimating Republican vote share persisted for the third consecutive presidential election in which Trump was a candidate.

The Sampling Problem: Who Picks Up the Phone?

One of the most foundational challenges in election polling is obtaining a sample of respondents that accurately reflects the electorate. Telephone surveys — once the gold standard of political polling — have faced dramatically declining response rates as caller ID, call screening, and the shift away from landline phones have made it progressively harder to reach potential respondents. Research cited by Suffolk University political science professor Rachael Cobb found that where pollsters once might need to contact around 20 people to complete one interview, the figure has roughly doubled in recent years, making each completed poll more expensive and potentially more selective in who it captures.

When those who respond to polls differ systematically from those who do not, the resulting data contains what researchers call nonresponse bias. A report commissioned by AAPOR found that in both the 2016 and 2020 elections, Republicans and independents were less likely to respond to pollsters than Democrats, and that among Republican respondents, those who did answer were less likely to support Trump than those who refused. This created a structural tilt in the data that demographic weighting alone could not fully correct, because the political preferences of non-respondents within demographic groups differed from those who agreed to participate.

Erin Caldwell, a survey research principal with more than three decades of experience conducting civic surveys, has noted that political polls face a compounding version of this problem. Pollsters rely on voter registration files to identify potential respondents, but within that pool only a fraction will actually vote. The challenge is not simply reaching registered voters but identifying which among them will turn out — a question that requires predicting behavior rather than merely recording stated intent.

Likely Voter Screens and the Limits of Turnout Prediction

Because not every registered voter participates in each election, pollsters use what are known as likely voter models — screening criteria designed to filter respondents down to those deemed most probable to cast a ballot. These models typically incorporate factors such as past voting history, stated intention to vote, and interest in the current election. However, the models involve significant judgment, and different methodological choices can yield starkly different results from the same underlying data.

A 2016 exercise published by The New York Times illustrated this vividly: four separate teams of analysts were each given identical raw polling data from a Florida survey and asked to process it using their own weighting and likely voter methodologies. The four resulting estimates ranged from Clinton leading by four percentage points to Trump leading by one — a spread of five points from the same raw numbers. The exercise demonstrated, as the Rasmussen Reports commentary on the AAPOR task force findings noted, that likely voter modeling done by rational, expert analysts can result in significant divergence, and that in close elections these differences can determine which direction a poll points.

Turnout can also be affected by events outside pollsters’ control. The COVID-19 pandemic significantly altered voting behavior in 2020, with an unprecedented expansion of mail-in voting in many states and record-setting overall turnout. AAPOR noted that the states with higher pandemic case rates at the time of polling showed larger polling errors, suggesting that the extraordinary circumstances complicated the task of modeling who would actually vote.

Late-Deciding Voters and the Problem of Election Day Uncertainty

Even a methodologically sound poll conducted weeks before an election cannot account for shifts in voter sentiment that occur in the final days of a campaign. In 2016, AAPOR’s post-election analysis found that an unusually high proportion of voters — as many as 13 percent, according to some estimates at the time — were still undecided or considering third-party options on Election Day, and that among those who decided in the final week, a majority ultimately voted for Trump. Polls completed several days earlier had no way to capture these last-minute movements.

The late-deciding phenomenon was less pronounced in 2020, with only around 4 percent of poll respondents in final state-level surveys giving an answer other than Biden or Trump when asked their preference in the closing two weeks, according to the AAPOR task force report on that election. This suggested that a different set of factors — primarily nonresponse bias — drove the 2020 errors rather than a repeat of the 2016 undecided-voter dynamic.

The relationship between stated intention and actual behavior adds another layer of complexity. A poll can accurately record what a respondent says they plan to do without that intention translating into a vote. Party registration files show that Democrats historically report higher participation intentions in polls but, according to survey research specialist Erin Caldwell, are in some contexts less likely to follow through on voting day. Republicans have shown the reverse pattern: lower reported poll participation but higher rates of actual ballot-casting. These behavioral asymmetries mean that even a well-sampled poll may not translate directly into an accurate electoral forecast.

Social Desirability Bias and the Contested “Shy Voter” Hypothesis

One of the most discussed explanations for polling errors in recent elections is the possibility that some voters conceal their true preferences from pollsters. This concept — related to what social scientists call social desirability bias, or the tendency to give answers perceived as socially acceptable rather than honest ones — has been applied to Trump-era elections under the label of the “shy Trump voter” hypothesis. The argument holds that some individuals who intended to vote for Trump were reluctant to say so in a survey context, leading to an artificial undercount of his support.

The empirical evidence on this hypothesis is contested. A nationally representative list experiment conducted by researcher Alexander Coppock in 2016, involving more than 5,000 American adults, found no statistically significant evidence that social desirability was causing respondents to understate their Trump support; the list experiment estimate of Trump support was slightly lower, not higher, than the direct question estimate. However, research published in the Journal of Behavioral and Experimental Economics by Andy Brownback and Aaron Novotny found marginally significant evidence that social desirability responses caused some respondents to understate agreement with Trump and overstate agreement with Clinton. The AAPOR task forces on both the 2016 and 2020 elections ultimately concluded that nonresponse bias and weighting failures were the more significant contributors to polling error, while characterizing the shy-voter question as warranting continued research rather than settled.

Education Weighting, Sample Composition, and the Representation Gap

Following the 2016 election, one of the most broadly accepted findings of the AAPOR task force was that state-level polls had significantly over-represented college-educated voters, who lean more Democratic, without adequately correcting for this imbalance through weighting. Non-college-educated white voters — a group that swung heavily toward Trump — were systematically underrepresented in samples. The national polls had applied education-based weighting more consistently, which is one reason national-level polling was somewhat more accurate than state polls in 2016 even if both underestimated Trump’s support.

In response, a large majority of pollsters adopted education weighting ahead of the 2020 election. The AAPOR task force on 2020 found that 92 percent of state-level presidential polls conducted in the final two weeks included education in their statistical adjustments — an improvement from 2016. Yet the errors in 2020 were in some respects larger than in 2016, indicating that education weighting alone did not resolve the underlying problem. The task force found that when poll data was reweighted to match certified 2020 outcomes, it was necessary to increase the percentage of Republicans and 2016 Trump voters in the sample, suggesting that the composition problem went beyond education level to something more difficult to measure and correct: the political characteristics of non-respondents within demographic groups.

Over-representation of college-educated voters can also interact with geographic sampling challenges. Rural communities and geographically dispersed populations can be harder to reach and may be underrepresented relative to densely populated urban areas. Because education levels, urbanicity, and partisan preference are correlated, errors in geographic or demographic representation can compound one another in ways that are difficult to untangle through post-hoc weighting.

Technological Change and the Declining Poll Response Rate

The structural difficulties facing election polling have been amplified by rapid changes in communication technology. Telephone polls, which became the dominant method following the lessons learned from the 1936 Literary Digest debacle, depended on reaching a broad cross-section of the public by phone. As landline telephone ownership has declined and caller ID has made it easy to screen unfamiliar numbers, the share of people who answer polling calls has dropped substantially. Rachael Cobb of Suffolk University noted in reporting by CNBC that what once required calling around 20 people to complete a single interview now requires roughly 40, doubling the cost and time of each poll while simultaneously raising questions about whether those who do answer are representative of those who don’t.

Online polling has emerged as an alternative, but it carries its own set of methodological challenges. Opt-in online panels — where respondents volunteer to participate — are not randomly sampled and can introduce self-selection biases. Address-based sampling, where invitations are mailed to a random selection of physical addresses, offers a closer approximation to true random sampling but is expensive and slow. Research published in academic proceedings, including analyses of international polling failures reviewed in an arXiv paper by researchers at the City College of New York, has argued that these structural limitations of traditional polling methods are contributing to predictive failures not only in the United States but across multiple democracies, including Brexit-related polling in the United Kingdom and presidential elections in Argentina.

Industry Efforts to Improve Polling Accuracy After High-Profile Misses

The polling industry has not been passive in the face of these challenges. The American Association for Public Opinion Research has convened dedicated task forces after each major election cycle to identify sources of error and recommend methodological improvements. After 1948, pollsters committed to continuing data collection through Election Day rather than stopping weeks in advance. After 2016, education weighting was widely adopted. Between 2020 and 2024, a greater number of firms experimented with new respondent recruitment strategies, harder-to-reach sampling techniques, and revised weighting approaches in an attempt to capture the voters most likely to be missed by conventional methods.

The 2024 AAPOR task force analysis of 116 general election polls found evidence that these reforms produced measurable improvements. State-level polls in 2024 were described as the most accurate since 1944, and the overall picture of a close race between Kamala Harris and Donald Trump proved broadly correct. Methodological differences among pollsters mattered: the task force found that polls from higher-volume firms, those weighting on partisan self-identification, those using detailed likely-voter models, and Republican-affiliated pollsters were on balance slightly more accurate than others. Still, the persistent underestimation of Republican vote share across three consecutive Trump-era elections suggests that the structural challenges — particularly around reaching and accurately representing voters skeptical of institutions — have not been fully resolved.

Frequently Asked Questions About Election Polling Accuracy

Sources Referenced