Quick Answer: How Do You Read a Poll?
To read a poll, look beyond the headline percentage. Check who was surveyed, the sample size, field dates, sampling or recruitment method, weighting, and reported uncertainty. Margin of error describes sampling uncertainty under a statistical method, not every possible source of polling error. Larger samples usually reduce sampling variability. Higher confidence levels produce wider intervals. Subgroups typically carry more uncertainty because their sample sizes are smaller. No single poll should be treated as a final verdict.
How to Read a Poll in 10 Questions
Before interpreting any poll result, work through these ten questions in order. They move from the population being studied down to the sources of error the headline never mentions.
A poll-reading checklist
What Does a Poll Percentage Actually Estimate?
A poll estimates opinions, preferences, characteristics, or intentions in a target population using information collected from a sample of that population. The headline percentage is a sample-based estimate of a population quantity, not a direct measurement of it.
When a poll reports that 48% of respondents support a particular position, that number describes the weighted share of the sample who gave that response under that pollster's methodology. It is an estimate of a population proportion. It is not the population proportion itself, and it would be incorrect to treat it as a precise statement about what the full population believes.
Two concepts are worth separating clearly before going further.
Population
The population is the full group the researchers want to understand. A poll might target all adults in a country, all registered voters, all likely voters, or a specific demographic segment. The population definition matters enormously. A poll of all adults does not necessarily estimate the same thing as a poll of likely voters, even when the question is identical. Before comparing two polls, confirm that they share the same population definition.
Sample
The sample is the group of respondents whose answers are actually collected. The poll uses information from the sample to estimate a quantity in the population. Whether that estimate is useful depends on how the sample was selected, how many people it includes, how they were weighted, and how many refused to participate.
What Is Sample Size in a Poll?
Sample size, written n, is the number of respondents whose data goes into the estimate you are reading. This sounds simple, but there is a common trap: the overall sample size for a poll may differ from the sample size that applies to a particular result.
A poll of 1,200 adults has an overall n of 1,200. But if you are reading the result for 18-to-29-year-olds, the relevant n might be 160. The margin of error for that subgroup is much wider than the margin of error for the full sample. Always identify which sample size applies to the specific result you are interpreting.
Larger sample size does not automatically fix a biased poll
This point trips up a lot of readers. A bigger sample reduces sampling variability, all else equal. But a large sample collected through a biased process can still produce a badly wrong estimate. Poor coverage, systematic nonresponse, leading questions, or bad weighting can all introduce errors that a large n does not fix. A smaller probability-based sample, drawn and weighted carefully, often outperforms a much larger convenience sample.
What Is Margin of Error in a Poll?
Margin of error (MOE) quantifies sampling uncertainty under the assumptions of the statistical method used. It describes how far the sample estimate might plausibly fall from the true population value due to random variation in who ended up in the sample, not due to systematic bias or other nonsampling errors.
For a simple large-sample proportion interval, the margin of error follows from three pieces: the sample proportion, the sample size, and the confidence level chosen by the analyst.
The margin of error formula
Each term matters. p̂ is the estimated proportion from the sample. n is the sample size. z* is the critical value corresponding to the confidence level: approximately 1.96 for 95% and 2.576 for 99%. The formula above assumes a simple random sample. Polls with complex designs, stratification, clustering, or significant weighting require design-based variance estimation, and the textbook formula may understate or mischaracterize the true uncertainty.
Margin of error is stated in percentage points, not percent
When a poll reports 48% ± 3%, the 3 refers to three percentage points, not three percent of 48. Three percent of 48 would be 1.44 percentage points, which is a completely different and smaller quantity. The potential interval under the stated method runs from 45% to 51%. Confusing percentage points with percent is one of the most common misreads of polling data.
Worked Example 1: Reading a Margin of Error
A poll asks whether respondents support a policy. The result is 52% support, 43% oppose, 5% undecided. The sample size is 1,000. Assuming a simple random sample, the approximate worst-case MOE at 95% confidence is:
MOE ≈ 1.96 × √(0.52 × 0.48 / 1000) ≈ 1.96 × 0.0158 ≈ ±3.1 percentage points
The 95% confidence interval for support runs from approximately 48.9% to 55.1%. This means: if this sampling procedure were repeated many times under its assumptions, about 95% of the resulting intervals would contain the true population proportion.
What you cannot conclude: You cannot say "there is a 95% probability the true support is between 48.9% and 55.1%." Under the standard frequentist interpretation, the true value is fixed; it either falls in this interval or it does not. You also cannot say this interval captures every possible source of error. Question wording, nonresponse, and weighting are not inside the MOE.
Simplified example assuming simple random sampling and a standard large-sample interval. Professional polls require design-based variance estimates.
What Does a 95% Confidence Level Mean?
Confidence level is one of the most misread numbers in polling. The standard frequentist interpretation goes like this: a 95% confidence procedure is designed so that, if you repeated the same sampling process many times and built a confidence interval each time, approximately 95% of those intervals would contain the true population parameter under the stated assumptions.
It does not mean there is a 95% probability that the true value lies in this specific interval. It does not mean the poll is 95% accurate. Once the interval is calculated from one sample, the true population value either falls inside it or it does not. The 95% describes the long-run behavior of the procedure, not the probability for any one result.
Confidence level vs. confidence interval
These two terms describe different things. The confidence level (95%) is a property of the statistical procedure. The confidence interval (45% to 51%, for example) is the specific numerical range calculated from one sample. Keeping them separate matters, because readers sometimes read "95% confidence level" as though it were a property of the interval itself, which it is not.
Higher confidence means a wider interval
For a fixed dataset and method, increasing the confidence level requires a larger critical value z*, which makes the interval wider. A 99% interval is wider than a 95% interval for the same data. A 90% interval is narrower. The tradeoff is real: more confidence in the procedure comes at the cost of less precision in the stated range.
| Confidence level | Critical value (z*) | Standard error | Margin of error | Interval (estimate 50%) |
|---|---|---|---|---|
| 90% | 1.645 | 0.0158 | ±2.6 pp | 47.4% to 52.6% |
| 95% | 1.960 | 0.0158 | ±3.1 pp | 46.9% to 53.1% |
| 99% | 2.576 | 0.0158 | ±4.1 pp | 45.9% to 54.1% |
Illustrative SRS approximation. Professional polls use design-based variance estimates that may differ from these figures.
How Sample Size Affects Margin of Error
The relationship between sample size and margin of error follows a square-root pattern. Standard error decreases approximately as 1/√n under simple sampling assumptions. That has a counterintuitive practical consequence: to cut the margin of error roughly in half, the sample size must be multiplied by approximately four, not two.
Worked Example 2: Sample Size Comparison
| Sample size (n) | Standard error | Approx. MOE | Interval width |
|---|---|---|---|
| 100 | 0.0500 | ±9.8 pp | 19.6 pp |
| 400 | 0.0250 | ±4.9 pp | 9.8 pp |
| 1,000 | 0.0158 | ±3.1 pp | 6.2 pp |
| 1,600 | 0.0125 | ±2.5 pp | 4.9 pp |
| 2,500 | 0.0100 | ±2.0 pp | 3.9 pp |
Going from n = 400 to n = 1,600 (a fourfold increase) cuts the MOE from roughly ±4.9 to ±2.5 percentage points. The gain from growing a sample of 1,600 to 2,500 is much smaller, and the diminishing returns become more pronounced as n grows further.
Illustrative SRS approximation. Not a universal rule for professionally weighted polls.
The diminishing-returns curve also explains why national polls rarely interview more than 1,500 to 2,000 respondents for their main estimate. Beyond that range, additional interviews reduce sampling error only slightly, while the cost of reaching respondents keeps rising. Resources are often better spent on methodology quality than on raw volume.
Why Margin of Error Is Not Total Polling Error
This is the most consequential point in any discussion of how to read a poll. A margin of error of ±3 percentage points does not mean a poll is guaranteed to be within 3 points of reality. Sampling uncertainty is only one component of total survey error, and it is often not the largest component.
A reported MOE generally does not capture coverage error, nonresponse bias, question wording effects, mode differences, interviewer effects, likely-voter modeling uncertainty, weighting model choices, or data-processing mistakes. A poll can have a narrow MOE and still be systematically wrong because of problems the MOE was never designed to measure.
The Total Survey Error framework
Survey methodologists use a framework called Total Survey Error to think about all the ways a survey estimate can differ from the true population value. The major sources include:
| Error source | What it means | Captured in typical MOE? |
|---|---|---|
| Sampling error | Random variation from using a sample rather than the whole population | Yes |
| Coverage error | Parts of the target population missing from the sampling frame or recruitment process | No |
| Nonresponse error | People who were contacted but did not participate differ systematically from those who did | No |
| Measurement error | Question wording, order, or mode affects how people answer | No |
| Adjustment error | Weighting or modeling choices introduce systematic differences | Partially, in some designs |
| Processing error | Data entry, coding, or analysis mistakes | No |
These sources of error cannot simply be added together into one known total error figure, because most of them are not directly observable. What the framework gives you is a way of thinking about why two polls with identical reported margins of error can produce different results, and why a poll's true uncertainty is larger than any number printed in the methodology notes.
How Poll Weighting Works
Pollsters rarely draw a perfect probability sample. Even well-designed samples tend to overrepresent some groups and underrepresent others because of who picks up the phone, who opens an email invitation, and who chooses to respond. Weighting adjusts each respondent's contribution to the estimate so that the weighted sample better reflects known or modeled population characteristics.
What variables get weighted
Common weighting variables include age, gender, education, region, and race or ethnicity depending on the data source and population. For election polls, pollsters may also weight on past vote recall, party registration, or turnout propensity from a likely-voter model. The specific variables used, and the benchmarks used to calibrate them, are a methodological choice that can affect results.
Weighting can reduce bias but increase variance
This tradeoff surprises people. When weights are unequal, some respondents count for more in the final estimate than others. Highly unequal weights effectively reduce the amount of independent statistical information in the sample, so the true sampling uncertainty may be larger than a simple formula based on raw respondent count would suggest.
Survey designers use a concept called design effect to capture this. If a poll's design effect is 1.5, then the effective sample size is approximately n divided by 1.5, not n itself. A weighted sample of 1,200 respondents with a design effect of 1.5 carries roughly the same sampling precision as an unweighted sample of 800 under simple random sampling assumptions.
Weighting is not manipulation. It is a standard adjustment intended to align the sample with population benchmarks. But weighting on the wrong variables, using outdated benchmarks, or using extreme weights can introduce its own errors. Inspecting a poll's methodology disclosure for its weighting approach is part of reading a poll responsibly.
Why Subgroup Results Carry More Uncertainty
Polls routinely publish crosstabulations that break results down by age, education, region, or other characteristics. These subgroup results are interesting, but they come with substantially more uncertainty than the overall topline, because the relevant sample size is much smaller.
Worked Example 3: Overall vs Subgroup Precision
A poll of 1,200 adults includes approximately 180 respondents aged 18 to 29. Applying the simplified SRS formula:
| Group | n | Approx. MOE |
|---|---|---|
| Full sample | 1,200 | ±2.8 pp |
| Ages 18 to 29 | 180 | ±7.3 pp |
The subgroup margin of error is more than 2.5 times as wide as the overall figure. A five-point difference in support between this subgroup and the full sample would fall well within the subgroup's uncertainty range. The overall poll MOE does not apply to this subgroup result.
Simplified SRS illustration. Real weighted subgroup MOEs may differ due to design effects and unequal weights.
Readers should approach cross-tab data with particular care. A two-point difference between demographic groups within one poll is often noise rather than signal. When a subgroup result matters to your interpretation, look for whether the pollster published a specific MOE for that subgroup, or better yet, check whether the same pattern appears across multiple polls or across multiple waves of the same poll.
What Does a Statistical Tie Mean?
You will often see headlines describing a race as a "statistical tie" or saying a candidate leads "within the margin of error." These phrases are widely used but frequently misapplied.
The correct question when evaluating a lead is: how much uncertainty surrounds the difference between the two estimates? That is not the same question as: is each estimate's individual confidence interval wide?
Worked Example 4: Interpreting a Two-Candidate Lead
The observed lead is 3 percentage points. The overall poll MOE is about ±3.1 pp. Does this mean it is a statistical tie?
Not necessarily. In a two-candidate poll where respondents fall into one category or the other, the two proportions are negatively correlated: if Candidate A's share goes up, Candidate B's goes down. This negative covariance means the standard error of the difference is not simply double the individual standard error.
SE(A − B) = √[(p̂A(1−p̂A) + p̂B(1−p̂B) + 2p̂Ap̂B) / n]
Under SRS assumptions with p̂A = 0.49 and p̂B = 0.46:
SE(A − B) ≈ √[(0.49 × 0.51 + 0.46 × 0.54 + 2 × 0.49 × 0.46) / 1000] ≈ 0.0316
The 95% MOE for the difference is roughly 1.96 × 0.0316 ≈ ±6.2 percentage points. The observed 3-point lead falls within that range, so this specific result does not rule out a lead of zero at the 95% level. But applying the overall individual MOE of ±3.1 pp directly to the difference and calling it a definitive tie or a definitive lead would both be wrong.
Simplified SRS model. Real poll design effects and weighting alter the true variance of the difference.
Overlapping individual confidence intervals for two estimates do not by themselves prove that the estimates are statistically indistinguishable. Testing whether a difference is distinguishable from zero requires examining the uncertainty in the difference itself, which depends on both variances and their covariance. This is one of the most commonly misunderstood points in public polling coverage.
Other Factors That Shape Poll Results
Question wording and order
The exact words a question uses can shift responses substantially. A question that asks "do you support allowing gay couples to marry" and one that asks "do you believe marriage should be defined as between a man and a woman" are both polling on the same policy but are likely to produce different numbers. Question order also matters: an earlier question can prime respondents to think about a topic in a particular way before they answer a later one. When comparing polls on the same topic, read the actual question text before assuming the results should match.
Field dates and timing
A poll is a snapshot of opinion during a specific window of time. Publication date and field dates are different things. A poll published Monday may have interviewed respondents the prior Wednesday through Friday. If a significant event occurred between the field period and publication, the poll cannot reflect it. Examining field dates before interpreting results is especially important during fast-moving news cycles.
Mode effects
Results sometimes differ depending on how the poll was conducted. Live telephone interviews, automated telephone scripts, online panels, text message surveys, and face-to-face interviews can produce different responses to the same question. Reasons include interviewer effects (people give socially desirable answers when talking to a human), how response options are presented visually versus verbally, and who is reachable by each method. A mode difference between two otherwise similar polls can explain a gap that looks like a real opinion difference.
Nonresponse bias
Modern polls, whether telephone or online, reach only a fraction of the people they attempt to contact, and a fraction of those actually complete the interview. If the people who respond differ systematically from those who do not on the topic being measured, the result can be biased in ways that weighting cannot fully correct. Response rate alone does not determine poll quality. A low response rate combined with good weighting can outperform a higher response rate from a poorly matched sample. But high nonresponse is a legitimate methodological concern worth noting.
Probability vs. nonprobability samples
Classical polling drew random samples where each member of the population had a known probability of being selected. Many modern online polls use nonprobability panels, where respondents opt in to participate over time and are recruited through various channels. Selection probabilities in such designs are not known in the classical sense, and the textbook margin of error formula does not apply directly without additional assumptions. Some organizations report "credibility intervals" or design-adjusted figures for nonprobability polls, which serve a similar function but rest on different assumptions. When reading a poll's methodology, check what kind of sample was drawn.
Likely-voter models
For election polls, many organizations apply a likely-voter screen that attempts to estimate which registered voters will actually cast a ballot. Different pollsters use different models, and those models can produce meaningfully different topline numbers from identical raw data. The likely-voter model's assumptions are usually not fully captured in the reported margin of error, so differences between polls may reflect modeling choices rather than genuine opinion differences.
How to Interpret Undecided Voters
Many polls include an undecided or "none/other" category that receives its own share of the responses. If a poll reports Candidate A at 46%, Candidate B at 44%, and undecided at 10%, the 2-point gap between A and B is a gap among all respondents, not just those who picked one of the two candidates.
To find the share among decided respondents only, you would divide each candidate's share by 90% (the sum of the two candidates' shares): A would have 46/90 ≈ 51%, B would have 44/90 ≈ 49% of decided voters. That is a different number from 46% and 44%, and it matters for how you think about where undecideds might break.
Some pollsters also ask undecided respondents which option they lean toward. Results may be reported with and without leaners. When comparing two polls, confirm which measure each one is using before drawing conclusions about movement.
How to Compare Two Polls Responsibly
Two polls asking what appears to be the same question can return different numbers for reasons that have nothing to do with opinion change. Before concluding that public opinion has shifted, work through this checklist:
- Same population? A poll of all adults and a poll of likely voters are not measuring the same thing.
- Same question wording? Even small wording differences can shift results.
- Same mode? An online poll and a telephone poll may not be directly comparable.
- Overlapping field dates? If polls were in the field at different times, some of the gap may be timing, not opinion.
- Similar methodology? Different likely-voter screens or weighting benchmarks shift numbers independently of true opinion.
- Uncertainty around the difference? A 3-point swing between two polls with ±3 pp MOEs each could easily be sampling noise.
The most reliable signal comes from consistent movement across multiple polls from different organizations, not from any single result. That is why poll averages, when carefully assembled, can be more informative than any one poll by itself.
Common Poll-Reading Mistakes
Treating MOE as total error
Margin of error covers sampling variability. It leaves out nonresponse, coverage, question wording, and other systematic sources of difference.
Misreading 95% confidence
"There is a 95% chance the true value is in this interval" is the wrong interpretation. The 95% describes the procedure's long-run performance, not one interval's probability.
Applying overall MOE to subgroups
A subgroup with 150 respondents has a much wider margin of error than the full sample. The overall ±3 pp does not apply to it.
Confusing percentage points and percent
±3% in a poll typically means three percentage points, not three percent of the headline figure. These are very different quantities.
Assuming bigger sample fixes bias
A large convenience sample can be more biased than a small probability sample. Volume does not compensate for systematic error in how respondents were recruited.
Ignoring field dates
A poll conducted before a significant event cannot tell you what people think after it. Publication date and field dates are different.
Using individual MOEs to judge a lead
Whether a lead is statistically distinguishable from zero depends on the uncertainty in the difference, not on each candidate's individual interval.
Treating poll support as win probability
A candidate polling at 52% does not have a 52% chance of winning. These are completely different quantities from different types of analysis.
Overreacting to one poll
Any single poll contains sampling variability. One result showing a 4-point swing may be noise. A sustained pattern across multiple polls is more meaningful.
Treating web opt-in polls as scientific
Open online polls where anyone can click to vote have severe self-selection bias. 100,000 responses still does not make them representative.
Poll Anatomy Table: What Each Element Means
| Poll element | What it means | What it does NOT mean |
|---|---|---|
| Estimate (e.g., 48%) | The weighted sample proportion under the pollster's methodology; an estimate of the population proportion | The exact population proportion, or what the population "actually" thinks |
| Sample size (n) | The number of respondents whose data produced this estimate | A guarantee of accuracy; a large n does not fix systematic bias |
| Margin of error (±3 pp) | Sampling uncertainty under the stated statistical method and confidence level | The total error in the poll, including nonresponse, coverage, and measurement error |
| Confidence level (95%) | Long-run coverage rate of the interval procedure under its assumptions | The probability that this specific interval contains the truth; a measure of poll quality |
| Field dates | The period during which interviews were conducted | The date the article about the poll was published |
| Weighting | Adjustments to align sample composition with population benchmarks | Manipulation; also not a perfect correction for all sources of bias |
| Subgroup result | The estimate for a subset of respondents, based on a smaller n | Something to interpret using the overall poll MOE; subgroup MOE is wider |
Frequently Asked Questions
Key Takeaways
- A poll percentage is a sample estimate, not a direct measurement of what the population believes.
- Margin of error covers sampling uncertainty only. Nonresponse, coverage, and question effects are additional, not included.
- A 95% confidence level describes the long-run performance of the interval procedure, not the probability that any one specific interval is correct.
- To halve the sampling margin of error, you need roughly four times the sample, not twice.
- Subgroup results have wider uncertainty than the full-sample result. The overall MOE does not apply to them.
- Weighting is a standard adjustment, not manipulation, but it carries its own assumptions and tradeoffs.
- Evaluating whether a lead is distinguishable from zero requires examining the uncertainty in the difference, not just each candidate's individual MOE.
- A candidate's poll share is not their probability of winning. Those are different quantities from different analyses.
- Field dates and publication dates are not the same. A poll cannot reflect events that happened after its field period ended.
- Consistent patterns across multiple polls from different organizations carry more weight than any single poll result.
Sources and Methodological References
Related Resources on Statistics Fundamentals
Try the Margin of Error Calculator with a few different sample sizes to see how the curve behaves. Then check the Confidence Interval Visualizer to build intuition for what "95% confidence" looks like across repeated samples.