What is the margin of error in a poll?

The margin of error is a range that expresses the uncertainty in a poll's estimate due to random sampling. It is calculated as 2 × 1.96 × √(p(1−p)/n) for a two-candidate margin, which reflects 95% confidence based on sampling variability alone.

Why is the real polling error about twice the reported margin of error?

The reported margin of error only captures random sampling error. It does not capture non-sampling error — coverage error (who is in the sample frame), non-response bias (who agrees to take the poll), mode effects (phone vs. online), late deciders, and likely-voter modeling. When you include these, historical polls show average error roughly 2x the stated margin.

How often do polls fall outside their stated margin of error?

In a sample of 9,000+ general-election polls since 1980, roughly 25% of polls fell outside their stated margin of error — far more than the 5% you would expect if the stated margin were accurate.

Why is the margin of error for a two-candidate margin twice the margin for a single proportion?

When you subtract two candidates' shares, the uncertainties add, so the variance of the margin is roughly double the variance of either single share. The standard error of a margin is therefore about 2x the standard error of a single proportion at the same sample size.

How should I read a poll given the real uncertainty?

Treat the reported margin of error as a lower bound, not a ceiling. Discount small leads that fall within the true (roughly doubled) uncertainty. Prefer poll averages over single polls, and weight them by pollster track record and recency.

Polling Uncertainty, Explained

When pollsters report a margin of error of ±3 points, the real-world uncertainty is closer to ±6 — roughly double. The reported margin of error captures sampling error only. It does not capture non-sampling error: coverage error, non-response, mode effects, late deciders, and likely-voter modeling.

The data

This piece is built on more than 9,000 general-election polls (presidential, Senate, and gubernatorial) conducted since 1980. For each poll, the theoretical margin of error is computed as 2 × 1.96 × √(p(1−p)/n) and compared to the actual error against the certified result. The error-ratio distribution (actual error ÷ theoretical margin of error) has a median near 1.5 with a substantial right tail.

What's in the tool

MoE calculator — shows that the margin of error for a two-candidate margin is roughly 2× the margin of error for a single proportion at the same sample size.
Error comparison scatter — poll margin vs. actual margin, with the stated MOE band overlaid so you can see how often polls fall outside their own reported uncertainty.
Sources of error — coverage, non-response, mode, late deciders, likely-voter modeling, partisan non-response.
Practical takeaways — how to read polls given the true uncertainty.

Related projects

By G. Elliott Morris, author of Strength In Numbers: How Polls Work and Why We Need Them.