How to read an election poll

A single poll showing a candidate up three points is not a prediction; it is one noisy measurement of a moving target, built on a chain of assumptions about who will respond and who will vote. Reading a poll well means checking that chain before repeating the topline number. This is a methods guide — how polling works and how to judge it, not financial or electoral advice, and not a forecast of any specific race.

What the margin of error actually is

The margin of error describes sampling error only: the uncertainty from asking a sample instead of everyone. A poll of 1,000 people with a 3-point margin of error means that if the poll were repeated many times with different random samples, about 95% of those samples would land within 3 points of the true value — assuming everything else about the poll is unbiased.

Two consequences people routinely miss:

• The margin applies to each number separately, and roughly doubles for the gap between two candidates. A race reported as 48%–45% with a 3-point margin is not "outside the margin of error" — the gap itself carries something closer to a 4–5 point margin, so a 3-point lead is well within noise.
• It says nothing about non-sampling error. Bad weighting, a skewed sample frame, or a flawed likely-voter model do not show up in the stated margin at all, and in most elections they matter more than sampling noise does.

Weighting decides more than the sample size

Raw survey respondents are never a perfect random draw from the electorate — some groups answer surveys more than others. survey weighting adjusts the raw data so it matches known population figures (age, gender, region, education, past vote, and others) before the topline is calculated. Two pollsters interviewing statistically identical raw samples can publish different toplines purely because they weight on different variables, or weight education differently — a factor that became decisive in several recent US elections once its correlation with vote choice grew.

Look for what a pollster discloses about weighting, not just the sample size. A pollster who publishes only "n=1,000, MoE ±3%" is telling you less than one who lists the variables weighted and the source population figures used.

Likely-voter models are a second layer of guessing

Registered-voter polls ask everyone eligible; likely-voter models try to predict who will actually show up, usually from self-reported enthusiasm, past voting history, or both. This model is where a large share of polling error originates, because turnout itself is unpredictable and models built on past elections can miss a shift in who turns out this time. Two pollsters can survey the identical group of people and produce different toplines purely from different likely-voter screens.

The specific ways a poll goes wrong

• sampling bias. The people contacted are not representative of the electorate — for example, a phone poll systematically under-reaching people who do not answer unknown numbers, or an online-panel poll over-representing people who join panels for pay.
• nonresponse bias. Even from a representative initial sample, if the people who agree to respond differ systematically from those who don't (more politically engaged, more distrustful of institutions, whatever the pattern), the responses skew even after weighting corrects for demographics.
• selection bias in self-selected samples. A poll where anyone can opt in — a website click poll, a call-in survey — is entertainment, not measurement; it tells you about the people who chose to participate, not the population.
• House effects. Individual pollsters show consistent, small, directional leans across many polls, from methodology choices that are not necessarily errors, just consistent assumptions.

Read one poll skeptically, read many together

No single poll should move your assessment much. poll aggregation — averaging or model-weighting many polls together — cancels out a good share of each poll's individual noise and dampens the effect of any one pollster's house lean, which is why aggregators generally outperform any single poll and why a poll's news value often lies more in "did it shift the average" than in its own topline. When you see a poll reported, ask where it sits relative to the current polling average, not just what it says on its own.

Exit polls, conducted with voters as they leave the polling place on election day itself, answer a different question than pre-election polls — not "who will win" but "who voted and why" — and carry their own biases, particularly around which voters are willing to stop and answer on the way out. AP VoteCast and similar large-scale surveys were built partly to address exit polling's specific weaknesses by combining phone, text, mail, and online panel data with the voter file.

A short checklist before you share a poll number

1. Who conducted and paid for it, and what is their track record — is this pollster rated or scored by an independent evaluator?
2. What is the sample size and margin of error, and does the margin roughly double for a head-to-head gap?
3. Is it registered voters or likely voters, and how is "likely" defined?
4. What is disclosed about weighting?
5. Where does it sit relative to the current polling average, not in isolation?
6. Was the survey mode (phone, online panel, text-to-web) disclosed, and does that mode reach the population being studied?

Where the data comes from

Large media and academic partnerships such as AP VoteCast combine multiple modes and the voter file specifically to reduce the biases any single mode carries. Firms running proprietary standing panels, such as YouGov, can track the same or similar respondents over time, which helps with trend measurement but means their panel's composition is itself a methodology choice worth understanding. Standing tracker polls, such as Civiqs, trade the one-off study's precision on a single day for a continuous read on trend — useful for direction, less useful as a single-point prediction.

Common mistakes

• Treating a 2-point shift between two polls from different pollsters as a real movement in opinion, when it may just be the polls' individual noise or house effects.
• Reporting the raw sample size without checking whether the number that matters — likely voters, or a specific subgroup crosstab — has a much smaller, noisier base.
• Ignoring "undecided" and "third-party" shares, which can be larger than the margin between the leading two candidates and which pollsters allocate differently.
• Comparing polls across different modes or vendors as if methodology were held constant.

For survey research method more broadly — questionnaire design, weighting technique, and analysis after the data comes in — see how to analyze survey data. For the tools that run campaign-side voter analytics rather than public polling, see every tool in this category.