Confidence intervals: repeated-sample coverage and what the endpoints mean

Identify the parameter being estimated

Suppose a fictional process produces independent normally distributed measurements with unknown mean μ and known standard deviation 10 units. The target is the process mean, not the next individual measurement. The NIST guide to confidence limits for a mean explains why an interval accompanies an estimate that varies across samples. A confidence statement must identify both the parameter and the procedure. Changing the target from a mean to a proportion, variance or future observation requires a corresponding method rather than reusing the same formula automatically.

Read the confidence level as a method property

For a 95% procedure with correct coverage, repeated independent samples processed by the same rule would produce intervals containing the fixed true mean about 95% of the time. Once one sample has been observed, its interval either contains that mean or it does not. The frequentist confidence level is not ordinarily a posterior probability assigned to the fixed parameter inside that specific interval. A Bayesian credible interval can express a different probability statement under a prior and model. The terminology matters because the two interpretations rely on different frameworks.

Calculate an illustrative normal interval

The common sample mean is held at 50 to isolate the effect of sample size; these are not four actual observed samples. The critical value 1.96 comes from the standard normal reference table. The calculation is exact for the stated known-variance normal model, apart from the displayed rounding.

Know when the formula changes

If the population standard deviation is unknown, an appropriate one-sample normal model uses the sample standard deviation and a Student t critical value with n−1 degrees of freedom. The NIST mean-interval guide presents that form. Dependence, heavy tails, small samples or a different sampling design can require further changes. A familiar plus-or-minus expression does not validate itself. Record why the interval method matches the collected data, and avoid reporting a mechanically calculated interval as though its nominal coverage survived every possible departure from the model.

Separate precision from individual spread

The NIST discussion of scale describes standard deviation as a measure of observation variability. In the table, increasing sample size narrows uncertainty about the mean while the process standard deviation remains ten. Individual measurements can therefore remain widely dispersed even when the mean is estimated precisely. A prediction interval for a new observation generally includes additional variability beyond uncertainty in the mean. Do not present a narrow confidence interval as the range containing 95% of individual outcomes; that is a different question.

Do not use overlap as a universal comparison test

Two confidence intervals that overlap can still be consistent with a statistically detectable difference under an appropriate direct comparison. Conversely, interval separation alone does not supply all the design information needed for a causal claim. If the scientific question concerns a difference, estimate that difference and its uncertainty directly, accounting for pairing or dependence where relevant. Comparing two separate intervals by eye can ignore the covariance between estimates. State the estimand clearly instead of treating a visual overlap rule as a universal substitute for analysis.

Report assumptions and practical scale

A very narrow interval around a small effect can indicate precision without showing practical importance. A wide interval can include effects with substantially different implications, even when a single point estimate looks reassuring. Present units, sample size, collection design and the interval method. Explain relevant thresholds only when they are justified for the substantive question. If many intervals are selected from a larger search, multiplicity and selection can alter the interpretation of nominal coverage; report that search rather than displaying only the most striking interval.

Confidence-interval checklist

• Name the parameter and target population.
• State the sampling and distribution assumptions.
• Identify known or estimated variability.
• Show the estimate, standard error and critical value.
• Interpret confidence as repeated-method coverage.
• Separate mean uncertainty from individual prediction.
• Report selection, multiplicity and practical units.

Retain the unrounded estimate and endpoints in the calculation record. Rounding too early can alter a threshold comparison or make two independently calculated values appear inconsistent. A good interval report gives readers enough information to understand both the arithmetic and the assumptions that make its confidence level meaningful.

Sources and verification

• Confidence limits for the mean — NIST/SEMATECH e-Handbook — 2026-10-23
• Standard normal cumulative distribution table — NIST/SEMATECH e-Handbook — 2026-10-23
• Measures of scale — NIST/SEMATECH e-Handbook — 2026-10-23