Margin of Error
Also known as: MOE, sampling error, polling error, survey accuracy, statistical margin, sampling uncertainty
Margin of error (MOE) is the maximum expected difference between a sample estimate and the true population value, at a specified confidence level (almost always 95%).
If a poll shows 52% support for a candidate with a margin of error of +-3%, the true support is likely between 49% and 55%. Margin of error is determined primarily by sample size (larger samples have smaller margins of error) and population variability. It quantifies the uncertainty introduced by surveying a sample rather than the entire population. Margin of error is the public communication of what statisticians call the half-width of the confidence interval.
Margin of Error, explained at three levels
At home
When you see a political poll that says '52% support Candidate A, margin of error +-3%,' the margin of error tells you how accurate the 52% estimate is likely to be. The pollsters didn't survey everyone -- they surveyed a sample of, say, 1,000 people. Because of this, their estimate might be off by up to 3 percentage points (either way) from the real support level. So the true support is probably somewhere between 49% and 55%. If Candidate A has 52% and Candidate B has 51%, and the margin of error is +-3%, the race is essentially tied -- we can't tell from this poll who is actually ahead.
At work
Margin of error appears in any research based on a sample: customer satisfaction surveys ('78% satisfaction, +-5% margin of error'), market research ('60% of consumers would consider the product, +-4% MOE'), and employee engagement surveys ('82% engaged, +-3% MOE'). The margin of error quantifies how much you should trust the specific number reported: a 78% satisfaction result with +-5% MOE means the true satisfaction is likely between 73% and 83% -- meaningful for business decisions. If the margin of error is very large relative to the number, the survey didn't include enough responses to produce a reliable estimate.
At home
The margin of error formula for proportions: MOE = z * sqrt(p*(1-p)/n), where z is the z-score for the desired confidence level (1.96 for 95%), p is the estimated proportion, and n is the sample size. For a 50/50 proportion (maximum variability), a sample of 1,000 gives a MOE of about +-3.1% at 95% confidence. The margin of error shrinks with the square root of sample size: doubling from 1,000 to 2,000 reduces MOE from +-3.1% to about +-2.2% (not by half -- by about 30%). Cutting MOE in half requires quadrupling the sample size. This is why very precise survey estimates are expensive: the returns diminish quickly.
At work
Subgroup analysis and margin of error: the reported margin of error applies to the overall survey estimate, but subgroup analyses (how did women respond vs. men? How did customers in region A respond vs. region B?) have larger margins of error because they use smaller subsamples. A survey with 1,000 total respondents split equally across 4 regions gives only 250 per region -- with a margin of error of about +-6.2% per region rather than +-3.1% for the total. Many subgroup comparisons that appear meaningful in survey reports are within the margin of error for the subgroups and should not be interpreted as real differences.
At home
Margin of error vs. total survey error: margin of error captures sampling error (the uncertainty from using a sample instead of a census) but not the much larger category of non-sampling errors: coverage error (not reaching everyone in the target population -- online surveys miss offline populations), non-response bias (those who respond may differ from those who don't), measurement error (question wording, order effects, social desirability bias affecting responses), and processing errors. Political polling failures in recent US elections were primarily due to non-response bias and social desirability bias -- systematic differences between who completed the poll and the full electorate -- not sampling error quantified by the margin of error. The margin of error provides false precision if non-sampling errors are larger.
At work
Design effect in complex surveys: the standard margin of error formula assumes simple random sampling (every person in the population has an equal probability of being selected). Most real surveys use more complex designs: stratified sampling (dividing the population into groups and sampling each separately), cluster sampling (sampling groups like households rather than individuals), and probability-proportional-to-size sampling. These designs affect the effective sample size and therefore the margin of error. The design effect (DEFF) is the ratio of actual variance to the variance you would have gotten with simple random sampling of the same size. Cluster sampling typically has DEFF > 1 (worse than simple random sampling); stratified sampling typically has DEFF < 1 (better). Reporting margins of error without accounting for design effects produces overconfident precision claims for clustered survey designs.
Real-world examples
At home
Lena ran a neighborhood association and surveyed 50 of the 400 households on a proposed local policy. The survey showed 60% support. She calculated the margin of error: +-14% (because 50 was a small sample). She reported honestly at the meeting: 'Our survey showed about 60% support, but with 50 responses from 400 households, the margin of error is +-14% -- so the true support is likely somewhere between 46% and 74%. We don't have a precise enough estimate to be confident about majority support. We should survey more households or proceed cautiously.' The margin of error framing prevented a false sense of certainty.
At work
Jerome conducted a customer satisfaction survey for a retail chain. Total respondents: 800 across 40 stores (20 per store). The overall satisfaction score was 78% with a +-3.5% margin of error. His report correctly noted: the per-store analysis (based on 20 respondents each) had a +-22% margin of error per store -- far too wide to reliably compare individual stores. He recommended against using store-level satisfaction rankings for performance management with the current survey design, instead proposing to oversample lower-traffic stores (50+ respondents each) to reduce per-store margins of error to +-14% -- still wide but usable for identifying clearly under-performing stores.
Common confusions
- Many people assume the margin of error accounts for all sources of polling and survey error — Margin of error only quantifies random sampling error -- the uncertainty from using a sample instead of surveying everyone. It does not account for systematic biases: who you can reach with your survey method (coverage error), who chooses to respond vs. ignore the survey (non-response bias), how question wording influences answers (measurement error), or whether respondents answer honestly (social desirability bias). In political polling, non-response bias (certain groups less likely to respond) has historically been a larger source of polling error than sampling error. The margin of error provides a precise quantification of one component of uncertainty while leaving other potentially larger components unquantified.
- Some assume that a lead within the margin of error means two candidates are in a statistical tie — A lead within the margin of error doesn't mean the race is literally tied -- it means the poll cannot reliably determine who is ahead. If candidate A has 52% and candidate B has 48% with a +-3% margin of error, the most likely outcome is still that A leads (52% is more likely to be the true value than 48%). But the uncertainty is large enough that B could plausibly be ahead. 'Statistical tie' is an oversimplification: any difference close to the margin of error should be interpreted as 'inconclusive' rather than as a literal tie, which implies exactly equal support.
For analysts
Margin of error is fundamental to helping clients interpret survey research appropriately. Common client errors: treating small-sample survey results as precise facts (a 65% satisfaction result from 25 respondents has a +-20% margin of error and is nearly meaningless as a precise estimate); comparing subgroup results without recognizing that subgroup margins of error are much wider than overall margins of error; and ignoring non-sampling errors that may systematically bias survey results regardless of sample size. When reviewing client survey methodologies, always check: what is the actual margin of error for the key estimate? For important subgroup comparisons? Are non-sampling errors (response bias, coverage) likely to be significant? Providing this analysis prevents clients from over-interpreting survey noise as signal.
History & origin
The margin of error concept became widely known through political polling in the 20th century. George Gallup pioneered scientific political polling in the 1930s and began communicating sampling uncertainty to audiences. The +-3% margin of error became a cultural shorthand for poll reliability, familiar to any reader of newspaper political coverage. The 1948 US presidential election famously illustrated polling error: the Dewey vs. Truman polls showed Dewey ahead, but Truman won -- an early lesson in the difference between sampling error (margin of error) and non-sampling errors (biases in sampling method, question wording, response rates). Modern discussions of polling accuracy increasingly distinguish margin of error (quantified sampling uncertainty) from broader polling error (systematic biases in who responds, how, and truthfully).