Election polls can inform voters and news audiences, but only when their methods and limits are understood. This guide explains how to assess sample size, margin of error, confidence intervals, weighting, field dates, likely-voter models, and polling averages before deciding how much significance to attach to a result.
Overview
An election poll is a survey of people who are asked about their voting intentions, preferences, or views on an issue. It is a measurement taken at a particular time, not a guarantee of what will happen on election day. A poll showing one candidate ahead may describe the surveyed population accurately while still failing to identify the eventual winner because opinions can change, turnout can differ, or the poll may not fully represent the electorate.
The first step in reading any poll is to identify exactly what it measures. Is it a national vote, a state or regional contest, or a single constituency? Does it survey all registered voters, adults, or people judged likely to vote? Is the result a snapshot from several days of interviews or a single-day measurement? These details often matter more than the headline number.
It is also important to separate a poll from a poll average. An individual poll is one survey conducted by one organization. A polling average combines several polls using a stated method. Averages can reduce the influence of unusual results, but they do not eliminate problems shared by many surveys, such as a flawed turnout assumption or a rapid change in public opinion.
What to track
Sample size and margin of error
Sample size is the number of people included in the survey, or the number included in a particular subgroup. A larger sample generally produces a smaller margin of error when the sample is properly selected and the responses are otherwise comparable. However, a large sample does not automatically make a poll representative. If important groups are missing or the sample is weighted inappropriately, more interviews may simply produce a more precise version of a biased result.
The margin of error describes the amount of sampling uncertainty expected under a specified statistical model. It is commonly reported as a plus-or-minus figure. If a poll reports a candidate at 48% with a margin of error of 3 percentage points, the measured share should not be read as an exact value. The stated margin is an estimate of sampling error, not a promise that the true result must fall inside that range.
Margins of error usually apply to the full sample and may be larger for smaller groups. A poll of all voters may have a different level of uncertainty from the same poll’s subgroup of younger voters, rural respondents, or people in a particular constituency. Check whether the published margin applies to the number being discussed.
Confidence intervals
A confidence interval presents a range associated with a chosen confidence level. In practical terms, it shows how much uncertainty surrounds an estimate under the poll’s assumptions. It should not be interpreted as the probability that a candidate will win. If two candidates’ estimates overlap, that usually means the poll does not establish a clear lead by itself. Small differences within the stated uncertainty should be treated cautiously.
Who was surveyed?
Registered-voter polls ask people who are registered to vote, while likely-voter polls use questions or models intended to identify people who are more likely to participate. Neither approach is automatically superior. Registered-voter surveys may include people who ultimately do not vote, while likely-voter models can miss new or irregular voters if their participation is difficult to predict.
Online surveys may use probability-based panels, opt-in panels, or other recruitment methods. Telephone surveys may contact people through mobile phones, landlines, or a combination. The collection method matters, but the more useful questions are how respondents were recruited, how the pollster weighted the data, and whether the sample resembles the electorate the poll is intended to describe.
| Poll feature | What it tells you | What to watch for |
|---|---|---|
| Likely-voter poll | Estimates preferences among people judged likely to vote | Turnout model, screening questions, and assumptions about participation |
| Registered-voter poll | Measures views among registered voters | It may include people who do not vote and may differ from the eventual electorate |
| Online survey | Collects responses digitally through a stated panel or sample source | Recruitment, weighting, duplicate responses, and sample composition |
| Telephone survey | Collects responses through calls or automated systems | Response rates, mobile and landline coverage, and screening practices |
| National poll | Measures opinion across a country | It may not predict seats or constituencies in systems where regional results determine the outcome |
| Constituency or regional poll | Measures a specific local contest | Smaller samples can create wider uncertainty and less stable results |
| Polling average | Combines multiple polls to show a broader trend | Included polls, recency rules, house effects, and whether different methods are being combined |
Field dates, weighting, and undecided voters
Field dates show when interviews were conducted. A poll completed before a debate, major event, policy announcement, or news development may not reflect views afterward. The publication date can therefore be less important than the interview dates.
Weighting adjusts the sample so that its characteristics better match the population the poll aims to represent. Pollsters may weight by factors such as age, education, geography, gender, or past voting behavior. Read the methodology to see which variables were used and whether the poll reports an unweighted sample size as well as a weighted one.
Undecided voters require particular care. Some polls show only decided preferences, while others include undecided respondents in the base. A candidate’s percentage can appear different depending on whether undecided voters are displayed separately, allocated to a candidate, or excluded from the calculation.
Cadence and checkpoints
When following election news, record polls in a simple table rather than relying on isolated headlines. For each release, note the pollster, sponsor, field dates, sample type, sample size, geographic coverage, candidate wording, undecided share, and reported margin of error. This makes it easier to compare like with like.
At a monthly or quarterly checkpoint, review whether the poll average is moving consistently or whether the apparent change comes from one new survey. During a fast-moving campaign, a shorter review cycle may be useful. Compare polls conducted during similar periods and avoid treating a poll from an earlier phase of the campaign as directly equivalent to a newly conducted survey.
Also check the election calendar. Registration deadlines, early voting, debates, party selections, legal changes, and major local events can affect who is reachable, who intends to vote, and how respondents answer. The relevant calendar differs by country, state, province, or constituency, so avoid applying national assumptions to a local race.
How to interpret changes
A change in one poll is not necessarily a change in public opinion. Sampling variation, questionnaire wording, weighting, interviewer effects, and the timing of interviews can all produce differences. Look for a pattern across several independent polls before describing a trend as durable.
Compare the size of the reported change with the poll’s uncertainty. A movement of one or two percentage points may be less meaningful than a sustained shift that appears across polls using different methods. Even a consistent movement does not establish causation. If a poll changes after a major event, the survey can show that attitudes differ, but it may not prove why they differ.
National and constituency polls answer different questions. A national lead does not necessarily translate into control of a legislature where seats are won locally. Conversely, a party can trail nationally while performing strongly in the places that determine representation. For local races, inspect the geographic boundaries, sample size, and whether the survey reflects the actual electorate in that district.
Polling averages are most useful as trend indicators, not as election-night forecasts. Read the average’s methodology: how old a poll can be before it is excluded, whether polls are weighted by sample size, how pollster quality is handled, and whether multiple surveys from the same organization receive special treatment. Averages can also lag when opinion changes quickly.
When to revisit
Return to this analysis whenever a new poll is released, especially if a headline describes a major change or a decisive lead. Use the following checklist before sharing or citing it:
- What are the field dates, and did a significant event occur during or after them?
- Who was surveyed: adults, registered voters, likely voters, or another defined group?
- How many respondents were included, and what is the margin of error for the relevant result?
- Is the poll national, regional, statewide, or constituency-level?
- How were respondents recruited, and what weighting was applied?
- Who sponsored and conducted the poll, and is the full methodology available?
- How were undecided voters handled?
- Does the result fit a wider pattern in the polling average?
Editors should refresh examples, election-calendar references, terminology, and links to reputable polling aggregators as methods and election conditions change. Readers can use the same checklist throughout a campaign and after each major release. The most reliable habit is to treat polls as evidence with limits: examine the design, compare several measurements, and keep the distinction clear between a current estimate and a prediction of the final result.
