A poll’s margin of error tells you how far its numbers could be from the true figure simply because the pollster interviewed a sample instead of everyone. A poll of about 1,000 people has a margin of roughly ±3 points at 95% confidence, so a candidate shown at 48% is probably somewhere between 45% and 51%. The margin on the gap between two candidates is about double that, and that’s the part most headlines get wrong.
The margin also leaves out every kind of error that isn’t random sampling, which in modern polling is often the bigger problem.
How is the margin of error calculated? #
For a simple random sample, the 95% margin of error for a percentage near 50% is:
MoE ≈ 1.96 × √(p × (1 − p) / n)
where n is the sample size and p is the share giving an answer. Pollsters use p = 0.5 because it gives the widest margin, and the formula simplifies to a handy shortcut: MoE ≈ 98 ÷ √n, in percentage points.
| Sample size | Margin of error (95%) |
|---|---|
| 200 | ±6.9 |
| 400 | ±4.9 |
| 600 | ±4.0 |
| 800 | ±3.5 |
| 1,000 | ±3.1 |
| 1,500 | ±2.5 |
| 2,000 | ±2.2 |
To cut the margin in half you need four times as many interviews. That’s why most national polls stop around 1,000 to 1,500 people and most state polls land between 500 and 1,000. Our guide to what makes a good sample size covers how pollsters choose.
The size of the population barely matters once it’s large. A random sample of 1,000 gives about the same margin for Wyoming as for the whole country.
What does “95% confidence” mean? #
If a pollster could repeat the same survey many times with fresh random samples, about 95 of every 100 results would land within the margin of error of the true value. About 5 would land outside it by chance alone.
That has a practical consequence. A competitive Senate race might get 20 or 30 polls in the last month. Even if every one were conducted perfectly, you’d expect one or two to fall outside their stated margin. Those are the “shock polls” that go viral.
Is a lead within the margin of error a tie? #
Not quite, and the margin on the lead is bigger than most people think.
When one candidate’s share goes up, the other’s usually goes down, so the uncertainty on the gap between them is roughly twice the poll’s stated margin. Pew Research Center’s explainer puts it simply: a 3-point margin for each candidate becomes about a 6-point margin for the difference.
Take a poll with Candidate A at 48%, Candidate B at 45% and a ±3 margin:
- Candidate A’s range: 45% to 51%.
- Candidate B’s range: 42% to 48%.
- The 3-point lead itself: anywhere from B up 3 to A up 9.
So a lead needs to be larger than about twice the stated margin before one poll can tell you, with 95% confidence, who’s ahead. Being inside the margin doesn’t make it a coin flip, though. A 3-point lead in that poll still makes Candidate A more likely to be ahead than behind. It just isn’t settled.
Why is the real margin bigger than the reported one? #
The published number assumes a perfect random sample. Real polls fall short of that in several ways.
Weighting adds a design effect #
Pollsters weight their samples so the mix of age, education, race and region matches the population. That fixes bias but increases variability. Statisticians call the increase the design effect, and Pew notes it makes the true margin larger than the simple formula suggests. Good pollsters report a margin that already includes it.
Non-sampling error isn’t in the number at all #
- Non-response: if the people who answer differ politically from those who don’t, the poll is off no matter how big it is. See non-response bias.
- Coverage: some groups are hard to reach by phone or missing from online panels.
- Likely-voter models: in a midterm, guessing who will turn out can move a result several points.
- Timing: a poll finished three weeks ago measures three weeks ago.
Opt-in online polls work differently #
Polls drawn from opt-in online panels aren’t random samples, so a classical margin of error doesn’t strictly apply. Many report a “credibility interval” or a modeled margin instead. Treat those as a rough guide to precision.
Why do subgroups have much bigger margins? #
The headline margin applies to the whole sample. A subgroup has its own, much larger one. In Pew’s example, a Hispanic subgroup of about 15% of a 1,067-person poll carried a margin of about ±8 points for each candidate and ±16 points for the gap. If a story leans on “young voters swung 10 points,” check how many young voters were interviewed. Our guide to reading poll crosstabs shows how.
Do polling averages fix the margin of error? #
They fix part of it. Averaging several independent polls shrinks random sampling error, which is why an average moves less than any single poll. An average can’t fix a bias that most polls share. If nearly every pollster under-reaches the same kind of voter, the average inherits that miss. That happened in several recent presidential elections, when averages understated Donald Trump’s support in many states. Our beginner’s guide to polling averages explains how different averages are built.
Election Tracker shows the sample size and population on every poll card, plus each race’s 30-day average and a lead chip. A useful habit: if the average lead is smaller than about twice a typical poll’s margin (roughly 6 points for 1,000-person polls, 8 to 9 for 500-person polls), treat the race as competitive and watch the trend line rather than any single release.
How to read a poll’s margin of error in 30 seconds #
- Find the sample size and population. Likely voters, registered voters or adults.
- Estimate the margin with 98 ÷ √n if it isn’t printed.
- Double it for the lead. If the lead is smaller than that, one poll can’t call the race.
- Check the field dates. Anything older than a couple of weeks may be stale in a moving race.
- Look at the average. One poll is a data point; the average of recent polls is the better estimate.
Frequently asked questions #
What is a good margin of error for a political poll? #
For a statewide or national poll, ±3 to ±4 points is typical and respectable. District polls often run ±4.5 to ±5 because their samples are smaller. Anything wider than about ±6 is fine for spotting a landslide but can’t separate a close race.
Does the margin of error apply to each candidate or to the lead? #
The published margin applies to each candidate’s percentage. The margin on the difference between two candidates is roughly double. Most confusion about “statistical ties” comes from mixing these up.
Why don’t pollsters just interview more people? #
Cost and diminishing returns. Going from 1,000 to 4,000 interviews only cuts the margin from about ±3.1 to ±1.5, and it roughly quadruples the cost. It also does nothing about non-response or a bad likely-voter model, which are often larger sources of error.
Do online polls have a margin of error? #
Probability-based online panels, recruited by mail or phone from a random sample, can report a standard margin of error. Opt-in panels can’t in the strict sense, so they often publish a modeled margin or credibility interval instead. Check the methodology note to see which kind of panel a poll used.