Margin of error and confidence bounds

A margin of error is the distance from an estimate to either confidence bound under a stated method and confidence level. The interval is estimate minus MOE through estimate plus MOE. It summarizes sampling uncertainty, not every source of error.

For a sample proportion p̂, the familiar normal standard error is √[p̂(1−p̂)/n]. For a mean, it is s/√n when s is the sample standard deviation. This calculator multiplies by √design effect and, when a valid finite N is entered, by √[(N−n)/(N−1)]. The selected normal critical value then produces the MOE.

Worked proportion example

A simple random sample of 1,000 people estimates a proportion of 0.52 from a population of 100,000. The uncorrected standard error is √(0.52 × 0.48 / 1000). Applying the finite-population factor of about 0.9950 gives SE ≈ 0.015720.

At 95% confidence, multiply by 1.959964 to obtain MOE ≈ 0.030810. In percentage terms, report 52.00% ± 3.08 percentage points, with bounds about 48.92% and 55.08%. Do not call the MOE 3.08 percent of the estimate; it is an absolute percentage-point distance.

Relative MOE divides 0.030810 by 0.52, giving about 5.92%. This can help compare precision across positive estimates, but it becomes unstable near zero and is undefined when the estimate equals zero.

Proportion mode versus mean mode

Small-sample mean intervals normally use a t critical value rather than a normal z value. This calculator intentionally uses the selected normal critical value for transparent planning and large-sample approximation. For exact publication work, match the estimator and critical distribution to the analysis protocol.

Finite population correction

Sampling without replacement reduces uncertainty when the sample consumes a meaningful share of a finite population. The correction equals one when n is tiny relative to N and approaches zero when nearly the entire population is observed. Enter zero for N to ignore it.

The correction assumes a simple random sample without replacement from a known fixed population. It is not appropriate merely because a city, customer list, or voter roll is finite. If selection is with replacement, probabilities are unequal, or the target population is conceptual or changing, use the design’s variance estimator.

A census can still have nonsampling error even though sampling MOE approaches zero. Coverage gaps, nonresponse, measurement errors, coding mistakes, and processing problems are not removed by observing more units under a flawed process.

Design effect and effective sample size

Design effect compares an estimator’s variance under the actual design with variance under a simple random sample of the same nominal size. A design effect above one inflates SE by its square root. The approximate effective sample size is n divided by design effect.

Clusters often increase design effect because people within the same cluster resemble one another. Stratification can sometimes reduce variance. Survey weights can increase variability when weights differ substantially. A single guessed design effect is only a planning device; final analysis should use the sample design and supplied replicate weights or variance strata.

Do not multiply MOE by design effect directly. Variance multiplies by design effect, so standard error and MOE multiply by its square root.

Reverse sample-size planning

For a target proportion MOE e, the large-population starting size is z²p(1−p)DEFF/e². If no expected p is available, 0.5 gives the largest Bernoulli variance and a conservative simple-plan size. For a mean, replace p(1−p) with an expected variance s².

The calculator then applies n = Nn₀/(N+n₀−1) for a finite population and rounds up. The result is a number of completed usable responses, not invitations. If response rate is expected to be 40%, divide completed-sample need by 0.40 to plan contacts, then consider ineligibility and missing items.

Sample-size formulas do not guarantee representativeness. A large convenience sample may have very small nominal MOE and substantial selection bias. Probability sampling, coverage, recruitment, and follow-up determine whether the interval method is defensible.

U.S. Census margins of error

The American Community Survey publishes margins of error at the 90% confidence level. A published ACS MOE should not be interpreted as 95% unless converted using the appropriate critical-value ratio. Dividing a 90% MOE by 1.645 estimates its standard error; multiplying by 1.96 gives an approximate 95% MOE.

Use the Census Bureau’s provided margins and guidance for ACS estimates rather than reconstructing them from a displayed count as if the survey were a simple random sample. ACS uses a complex design and specialized variance methods. When comparing two estimates, the uncertainty of their difference also matters; overlapping individual intervals are not a universal significance test.

Converting a published ACS MOE

Suppose an ACS table reports an estimate of 12,000 with a 90% MOE of 1,645. An approximate standard error is 1,645/1.645 = 1,000. A corresponding 95% normal MOE is 1.959964 × 1,000 ≈ 1,960, giving an approximate 95% interval from 10,040 through 13,960.

Convert the MOE, not the endpoints independently. Keep the estimate unchanged, recover the SE with the original confidence factor, and apply the new factor. This transformation assumes the published MOE follows that normal-factor relationship and does not replace Census guidance for special estimates, derived ratios, or collapsed categories.

When combining ACS estimates, the MOE of a sum is not normally the sum of the component MOEs. Approximate variance rules require converting each MOE to an SE and accounting for covariance assumptions. Census documentation provides formulas and cautions for derived estimates.

Comparing two independent estimates

To study a difference, uncertainty belongs to the difference itself. If two estimates are independent with standard errors SE₁ and SE₂, the difference SE is √(SE₁² + SE₂²). Multiply by the appropriate critical value for its MOE. If estimates overlap in samples, categories, time, or geography, covariance changes this formula.

Two individual 95% confidence intervals can overlap while their difference is statistically distinguishable, and nonoverlap is more conservative than a direct test in many settings. Do not use visual overlap as a universal rule. Convert both published MOEs to compatible confidence levels before comparing.

The practical importance of a difference is separate from its statistical precision. A tiny, precisely estimated change may be unimportant; a large but imprecise change may warrant more data. Report the point difference, its interval, units, and context.

Planning contacts and sensitivity scenarios

The reverse result is completed analyzable cases. If 1,055 completions are needed and the expected response rate is 50%, an initial invitation target is at least 2,110 before accounting for invalid contact information or ineligibility. Response-rate assumptions should come from comparable prior work, not optimism.

For proportion planning without a prior estimate, use p = 0.5 because p(1−p) is largest there. If a credible prior value is used, run sensitivity cases around it. A design effect is also uncertain before fieldwork; planning with values such as 1.0, 1.5, and 2.0 makes the cost-precision tradeoff visible.

Rounding sample size down fails the requested algebraic target, so the calculator rounds up. Operational strata may need separate minimums, and unequal allocation can increase total sample beyond the simple overall calculation. Ethical, legal, or service-access decisions may require precision for subgroups, not only the aggregate.

Interpretation and reporting traps

A 95% confidence interval is not a promise that this one realized interval has a 95% probability of containing a fixed parameter under the standard frequentist interpretation. The long-run procedure produces intervals covering the target at the stated rate when assumptions hold.

Higher confidence produces a wider MOE at fixed n. A larger sample reduces MOE roughly with the square root of n, so halving MOE generally requires about four times the sample before finite-population effects. More decimal places do not create more information.

Always report the confidence level beside the MOE. The notation “52% ± 3%” is ambiguous about whether 3 means percentage points or a relative percentage and about which confidence factor was used. A clearer statement is “52.0%, 95% margin of error ±3.1 percentage points.”

Accuracy boundary: MOE describes estimated sampling error. It does not include bias, bad questions, fraud, nonresponse, frame omissions, data processing mistakes, or model misspecification.

Frequently asked questions

References

U.S. Census Bureau — margin of error definition

U.S. Census Bureau — standard errors and confidence intervals

U.S. Census Bureau — ACS sampling error and 90% MOEs