> ## Content Index
> Fetch the complete content index at: https://www.ourcivicdata.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# How to Read and Critique Pre-Election Polls
- URL: https://www.ourcivicdata.com/how-to-read-and-critique-pre-election-polls/
- Published: 2026-11-02T19:46:25.000Z
- Updated: 2026-11-02T20:50:14.000Z
- Description: Polls are not promises; they are snapshots. Find out what election polls can and cannot tell you, how to evaluate their accuracy, and why averages are still subject to error.
- Author: Alexandar McCulloch
- Tags: Elections

The United States mid-term elections are approaching on November 3, 2026\. With election season comes pre-election polls. Even those who try to avoid polling numbers will still run into armchair pundits and their reactions on social media. After such a high level of inaccuracy from pollsters during the 2016 presidential election, many are still distrustful of their results. But what does it mean for a poll to be accurate, and how can you gauge trust in them? 

# What a poll actually captures

A survey records respondents' opinions and attitudes at the moment the survey is administered. It is easy to jump ahead and predict election outcomes based on a single poll, but when and where it's conducted tells us more about those specific individuals' attitudes and intentions at that time, attitudes likely to change by election day. 

Polls help us understand the meaning of the election. They clarify who votes, issues important to voters, why some vote while others do not, and what voters and non-voters are saying with their actions. Take the 2024 Democratic Presidential primary as an example. Surveys found a large subgroup of Democratic Michigan voters, particularly Arab-Americans, were [unhappy with the Biden administration's](https://www.nbcnews.com/politics/2024-election/poll-shows-biden-support-slumping-michigan-muslims-rcna123732?ref=ourcivicdata.com) policy on the Israel-Hamas war. Indeed, more than 100,000 Michigan Democrats [voted "uncommitted"](https://bridgemi.com/michigan-politics/michigan-2024-results-13-democrats-uncommitted-will-it-matter-biden/?ref=ourcivicdata.com) in the 2024 primary, signaling disapproval of Biden's foreign policy. The pre-election polls contextualized these uncommitted voters in a key swing state that Harris would go on to lose. 

# What "accurate" means

There are really two different questions hiding inside "were the polls right?"

1. Did the poll measure public opinion well when it was taken?
2. Did the poll line up with the final result?

The latter is easy to check, and the first is not, because opinion at a past moment can't be re-measured. This is why 2016 is a useful case study. A task force of the [American Association for Public Opinion Research (AAPOR)](https://aapor.org/wp-content/uploads/2022/11/AAPOR-2016-Election-Polling-Report.pdf?ref=ourcivicdata.com) found that national polls were largely accurate that year: Hillary Clinton won the national popular vote by 2.1 points, which was close to where national polling had her. The larger miss came at the state level, especially in Wisconsin, Michigan, and Pennsylvania, where Clinton led narrowly in pre-election polls and then lost narrowly. Because the Electoral College runs through states, a modest error in a handful of them was enough to change the impression of who was ahead.

The task force identified a few contributing factors. Many state polls did not adjust their samples for education, at a time when education had become strongly tied to presidential vote choice. Furthermore, with the unpopularity of both Donald Trump and Hillary Clinton, a meaningful share of voters made up their minds late, after many polls had already stopped asking.

Overall, contrary to popular wisdom, polling in and outside the U.S.[ has a good track record.](https://www.nature.com/articles/s41562-018-0315-6.epdf?author%5Faccess%5Ftoken=Dv9fzB5mxgATeTZwsxLioNRgN0jAjWel9jnR3ZoTv0PalVG3evbr7AhAkXUDTrGPuxQLUa5my1W6hESsmpNFdYsKXZoBT0m-zNSyqIupcNM3PbrEMvayUOKxPYc%5FacJq1F0WQDhCRL87YXwQ3q4FRg%3D%3D&ref=ourcivicdata.com) According to [Scott Keeter of Pew Research](https://www.pewresearch.org/course/public-opinion-polling-basics/?ref=ourcivicdata.com#what-s-the-deal-with-election-polls), "A comprehensive study found that polls taken shortly before an election had an average error of less than 2 percentage points in estimates of support for a given party or candidate."

## Margin of error and absolute error

The margin of error you see reported captures one thing: random sampling error, the natural variation you'd get from interviewing a sample instead of everyone. For a poll of about 1,000 respondents, it works out to roughly plus or minus 3 percentage points at the usual 95% confidence level.

In plain terms, the margin of error tells you how much a poll's number could vary just because the pollster asked a sample of people instead of everyone. If a candidate is at 48% with a margin of error of plus or minus 3 points, read that as "somewhere around 45% to 51%," not exactly 48%.

**Two cautions follow:**

- The margin applies to each candidate's share separately. The uncertainty on the gap between two candidates is roughly double, so a candidate "leading by 4" in a poll with a 3-point margin is not clearly leading.
- The margin says nothing about everything else that can go wrong: who agreed to respond, how the sample was weighted, how the questions were worded, or how the pollster decided who counts as a likely voter.

An error rate can be calculated for pollsters to assess the accuracy of not only the winner of the election, but the margin of victory. Using [FiveThirtyEight's polls dataset](https://github.com/fivethirtyeight/data/tree/master/pollster-ratings?ref=ourcivicdata.com), election outcomes and margin errors have already been calculated. The margin of error is calculated by subtracting the Republican vote share from the Democrat vote share. Meaning a negative margin indicates a Democrat victory and a positive margin represents a Republican victory. The average of all polls' absolute error– Mean Absolute Error (MAE)–informs us of the number of points off the pollster is on average. 

For example, a poll shows voters' preferences in the local election are 48% Democrat and 52% Republican (a margin of -4, or R+4). However, the true outcome is the inverse: 52% Democrat and 48% Republican (a margin of 4, or D+4). This is a difference of -8 (-4-4), signaling the poll understated Democratic support by 8 points. Averaging these differences tells us which party a pollster over/underestimates support. By taking the absolute difference and dropping the sign, we have the overall error rate. 

The following visuals show the MAE for each poll conducted between 2016 and 2023 by several major pollsters. If a point falls within the blue region, that poll overstates Democratic support, and the red region indicates overestimated Republican support. Ideally, points will fall on the diagonal line, which indicates 0 difference between predicted and observed margins. Thus, points further from the line have larger error rates. The top right and bottom left quadrants represent correct outcome predictions. 

These graphs further contextualize each pollster's error rates. For example, Marquette University Law School–located in Milwaukee, Wisconsin–almost exclusively polls swing states, specifically Wisconsin. Therefore, their error rate represents their ability to measure voter intent within Wisconsin specifically. Others have a healthy mix of swing state, non-swing state, and national polls. 

![](https://storage.ghost.io/c/0d/3f/0d3fdaf8-2127-46b6-a597-11fe18718a02/content/images/2026/11/pollsters_lowest_error-1.png)

This is important to consider, since more experience generalizing results to specific regions will lead to greater expertise in measuring said region. If Marquette University polls non-swing states and has a large error, that is not an indication that they are a poor pollster, just that their expertise falls in measuring attitudes of Wisconsin voters. 

Also notice how some points cluster around the center. Those points represent competitive races, where the margin of victory is slim. Some of these pollsters appear to favor polling locations with competitive races, such as The New York Times/Siena College. Keep this in mind when assessing a poll. The nature of competitive races makes it more challenging for a poll to reflect the eventual outcome. Thus, a slightly larger error for The New York Times/Siena is reasonable. 

![](https://storage.ghost.io/c/0d/3f/0d3fdaf8-2127-46b6-a597-11fe18718a02/content/images/2026/11/pollsters_highest_error-2.png)

The margin of error is only one tool to use to assess the quality of a pollster, and a relatively high margin is not immediately suggestive of a poorly designed or executed survey. Indeed, many pollsters in the chart above are well respected in the industry. Ipsos, for example, has an average absolute error of 5.5, meaning Ipsos polls' predicted vote margin is off by 5.5 percentage points on average. The Ipsos graph shows a clear trend for non-competitive races, particularly races where the Democrat is expected to win (see the high number of points in the bottom right quadrant in the blue region). 

[Subscribe](#/portal/signup/free)

Ipsos correctly predicts the winner of these races, but as the points indicate, overestimates the Democrat's performance. When we remove these non-competitive races and consider only the competitive ones, Ipsos' error is smaller in elections with narrow vote margins. Disaggregating the data further, Ipsos' error for swing states is also less than their 5.5 absolute error rate: 4.31\. 

## Polling Accuracy by Pollster and State Type

Mean absolute error (MAE) across national, non-swing state, and swing state polls. Lower values indicate greater polling accuracy.

| Pollster                         | National | Non-Swing State | Swing State |
| -------------------------------- | -------- | --------------- | ----------- |
| ABC News/The Washington Post     | 2.28     | 3.40            | 5.83        |
| Data for Progress                | 4.52     | 5.87            | 4.24        |
| Ipsos                            | 4.05     | 6.29            | 4.31        |
| Marquette University Law School  | 7.12     | —               | 3.94        |
| Quinnipiac University            | 3.52     | 5.73            | 7.72        |
| Rasmussen Reports                | 3.45     | 5.02            | 3.87        |
| Siena College                    | —        | 5.93            | 1.50        |
| SurveyMonkey                     | 3.07     | 6.90            | 5.39        |
| The New York Times/Siena College | 3.06     | 4.99            | 4.65        |
| Trafalgar Group                  | 2.99     | 5.10            | 3.72        |

**How to read this table** 

Each number is the average gap, in percentage points, between a pollster's final margin and the actual result. A dash means the pollster had no polls in that category.

The table above indicates that different pollsters have greater success based on the location they are extrapolating to. ABC News/The Washington Post are quite accurate with national polling, while Siena College does exceptionally well in swing states. When assessing a pollster, learn who they are generalizing (national, specific states, etc.) and learn their track record doing so. 

Lastly, these averages are affected by when each survey was administered. If some pollsters conducted multiple surveys 6-8 months before election day and then stopped, it would make sense for them to have a fairly high error rate, since polls further out from election day are naturally less predictive. 

# Who answered, and who is counted

Most people contacted by pollsters don't respond, and the people who do are not a perfect miniature of the electorate. Pollsters correct for this by weighting, which means adjusting the sample so that its makeup matches known population benchmarks such as age, race, and, as 2016 made clear, education. Weighting is only as good as the variables chosen. If a group that votes differently is underrepresented and the weighting doesn't fix it, the poll will be off in a consistent direction.

Then there is the likely-voter problem. Every poll of an election has to decide whom to count, and pollsters do this differently. In a midterm, where turnout is lower and less predictable than in a presidential year, those choices matter more. Two polls fielded the same week can disagree simply because one assumes a different electorate.

## How people are contacted

It has become increasingly more difficult to contact and convince a person to take a survey. Response rates have [declined steadily](https://www.bostonreview.net/articles/claude-fischer-survey-poll/?ref=ourcivicdata.com) since the 1950s, down from 80%. Those who maintain a "high" response rate today (for example, 30%) must offer greater incentives to participate. 

The mode of contact impacts these numbers as well. Fewer have landlines today and mostly rely on cell phones. Those who do still use landlines tend to be older, and those with only cell phones tend to be younger and less likely to answer. Online and SMS surveys have relatively higher response rates, but respondents tend to be younger, more educated, and Democratic. 

![](https://storage.ghost.io/c/0d/3f/0d3fdaf8-2127-46b6-a597-11fe18718a02/content/images/2026/11/Survey-Methods-Count.png)

As a result, pollsters must identify who is likely to vote and weight their numbers accordingly. If they predict older Americans are more likely to vote, but used an online survey platform, they must adjust so their respondents reflect the proportion of younger to older voters. 

# Why state and district polls deserve extra skepticism

In the AAPOR review of 2016, state polls were less likely than national polls to use live interviewers and less likely to adjust for education, and they tended to be done more quickly with fewer resources. Pew's survey research director, who chaired the task force, also noted that state-level polls tend to carry more error over time than national ones.

That is worth keeping in mind in a midterm. No single national race is on the ballot. The contests that decide control of Congress are in individual states and districts, where polling is typically sparser and less well funded than national polling. A national generic-ballot number can be informative about the overall mood, but it is a different measurement from a poll of any one district.

# What averages can and can't do

Poll averages are usually better than any single poll, because random noise in polls tends to cancel out. But an average cannot fix an error that every poll shares. If most pollsters make the same weighting choice or miss the same group of voters, the average inherits the same blind spot, and it can look more certain than it deserves to.

Averages also depend on which polls are included and how much weight each one receives. Some aggregators account for a poll’s sample size, quality, recency, and historical performance; others use different methods. A poll average is therefore not a direct count of voters or a guarantee of the final result. It is an estimate built from other estimates.

When looking at an average, check whether it includes polls of the right population—national, state, or district—and whether the surveys are recent enough to reflect current opinion. It can also help to look at the range of results, not just the average. If polls are tightly clustered, they tell a more consistent story; if they vary widely, that disagreement is worth noting.

The best way to read polling is to treat each result as one piece of evidence. Look at who was surveyed, how the poll was conducted, when it was fielded, and what population it represents. A poll average can help reveal the broader pattern, but it cannot remove uncertainty or tell us exactly what will happen on Election Day.

---

No poll can tell us exactly what will happen on Election Day, but polls can still offer valuable evidence about voters’ views and the choices shaping a race. Read them as estimates rather than guarantees: consider who was surveyed, how and when the poll was conducted, and whether it represents the voters or area you care about. 

[Signup](#/portal/signup/free)

## A Checklist for Critiquing Any Poll

Nine questions to ask before trusting a poll's headline numbers.

Who paid for it, and who conducted it? 

Sponsorship doesn't automatically invalidate a poll. However, knowing who funded it and who conducted the research helps readers identify potential conflicts of interest and understand the context behind the findings.

When was it in the field? 

Check the dates when respondents were interviewed. Compare them with major news events, campaign developments, debates, and other events that might have influenced public opinion. Also consider how close the polling period was to Election Day.

Who was surveyed? 

Was the poll conducted among all adults, registered voters, or likely voters? These are three different populations and can produce different results. If the poll surveyed likely voters, look for an explanation of how the pollster defined and identified them.

How were people reached, and how many responded? 

Was the survey conducted through live telephone interviews, text messages, an online panel, or mail? Each method has strengths and limitations. Look for the sample size and, where available, information about response rates and recruitment methods.

What was the weighting? 

Pollsters often adjust their samples to better reflect the population they're trying to measure. At a minimum, look for weighting by age, race, sex, region, and education. Missing or poorly calibrated demographic groups can distort the results.

What exactly was asked, and in what order? 

Question wording matters. Small differences in phrasing can influence how respondents interpret a question. The order of questions can also shape answers by introducing particular topics or ideas beforehand. Look for the complete questionnaire, not just the headline question.

Is the gap larger than the margin of error? 

As a rough rule, a difference between candidates should be larger than approximately twice the poll's margin of error before treating it as a clearly distinguishable lead. Remember that this is a rule of thumb, not a guarantee, and that other sources of polling error may not be captured by the stated margin.

Does it agree with other recent polls? 

Compare the results with other recent, credible polls that use transparent methods and survey comparable populations. One poll that differs sharply from the broader pattern may be an outlier. Look for consistent trends rather than putting too much weight on a single result.

Does the pollster publish its methodology? 

A credible pollster should provide enough information for readers and researchers to evaluate its work. Look for sample details, field dates, survey mode, weighting procedures, question wording, and information about how the sample was recruited. If these details are missing, treat the findings with additional caution.

**The bottom line** 

A poll is only as useful as the methodology behind it. Always look beyond the headline number.