Margin of Error
The margin of error is the half-width of a confidence interval: the distance from the point estimate out to either end of the interval, quoted as plus or minus. For a sample proportion at 95% confidence it equals 1.96 × √(p × (1 − p)/n), a critical value times a standard-error. A poll reporting 52% support with a margin of error of ±3 points is saying that sampling variability alone makes the interval from 49% to 55% a plausible range for the true share, with the procedure designed to capture the truth in 95% of repeated samples. It is the most widely seen expression of statistical uncertainty in public life, appearing under nearly every opinion poll.
The margin of error depends on only three things: the confidence level, the variability of the outcome, and the sample size. Because it contains the square root of n, quadrupling the sample halves the margin. At 95% confidence and the worst-case p = 0.5, a sample of 400 gives about ±4.9 points, 1,000 gives about ±3.1, 2,500 gives about ±2.0 and roughly 1,067 gives the familiar ±3. Raising the confidence level widens the margin (2.576 standard errors for 99% instead of 1.96) while lowering it narrows it, which is why a margin of error is meaningless unless the confidence level is stated, with 95% as the unstated default. Pollsters quote the worst-case p = 0.5 version because it is the largest, making a single number safe for every question in a survey.
The margin of error covers only random sampling error, which is the most dangerous misunderstanding about it. It says nothing about coverage bias from a flawed sampling frame, nonresponse, people who misreport, question wording or an unrepresentative panel, and these often dwarf the sampling error. The 1936 Literary Digest poll had millions of responses, so its nominal margin was tiny, and it still called the presidential election wrong, because it sampled the wrong people. Modern pollsters weight their data to match known demographics, which adds variability beyond the simple formula, so the honest uncertainty of a published poll is usually larger than the quoted margin, and polls with the same margin can differ in real quality by a wide gap.
Several interpretive traps follow from the formula. The quoted figure applies to the full sample, and subgroups, such as one age band or region, have fewer respondents and so a much larger margin; with 200 people in a subgroup, the 95% margin is about ±7 points. The margin on a difference or lead is not the margin on either number: because the two shares in a two-way race are negatively correlated, the margin on the gap between candidates is close to double the quoted one, which is why a 3-point lead with ±3 points looks like a tie but is not strictly one. And a 95% interval is not a statement that the true value has a 95% chance of lying in this particular interval in the frequentist reading; it is a statement about the long-run capture rate of the method, as treated in the confidence-interval entry.
Beyond polls, margins of error appear wherever an estimate is reported: clinical-trial effect sizes, A/B test lifts, quality-control averages, and machine learning accuracy computed on a test set. A classifier scoring 90% on 500 test examples has a margin of about ±2.6 points, so a rival at 91% is not clearly better, and a leaderboard difference smaller than the margin should be treated as noise. Reporting the margin alongside the estimate turns a bare point-estimate into a claim with its own honesty built in, and a requirement of a given margin (for example ±2 points) is also how required sample sizes are worked out before a study begins: solve the margin formula for n.
Read the margin of error as the plus-or-minus attached to an estimate for sampling noise alone: a critical value times a standard error, shrinking with the square root of the sample size, and a lower bound on honest uncertainty that says nothing about bias, nonresponse or any of the other ways a survey can be wrong.