The core idea
A confidence interval combines an estimate with a margin based on its standard error and a chosen coverage level. A 95% procedure captures the fixed population parameter in about 95% of repeated samples when its model and sampling assumptions hold.
1. State the target before estimating it
A population parameter is a fixed feature of a defined population, such as its mean μ or the proportion p with a specified characteristic. A sample statistic, such as x̄ or p̂, changes with the selected sample. Define who or what belongs to the population and the period being studied. “All enrolled learners this term” differs from “people visiting the library today”. A random selection process gives a basis for assessing sampling variation. A voluntary online poll can overrepresent people with time, devices or strong opinions, even if thousands respond. Missing responses and inaccurate measurements create other problems. The simple formulas here assume independent observations from a common population model. Clustered classes, repeated measurements of one person, unequal selection probabilities or a large sampling fraction need methods reflecting that design.
Sources: OpenStax: Data, Sampling, and Variation in Data and Sampling ↗
2. Standard deviation and standard error describe different variation
The population standard deviation σ describes variation among individual observations. The standard error of a sample mean describes variation among sample means from repeated samples of the same size. For n independent observations with common finite variance, SE(x̄) = σ/√n. When σ is unknown, s/√n estimates this quantity using the sample standard deviation s. With s = 12 and n = 36, the estimated standard error is 2, not 12 and not 12/36. Increasing n fourfold halves the standard error if the underlying variability stays the same. It does not make individual people or objects less variable. Under suitable conditions the central limit theorem makes the distribution of sample means approximately normal for large n. It does not make the original observations normal, and there is no sample-size rule that cures every heavy-tailed or dependent dataset.
Sources: OpenStax: The Central Limit Theorem for Sample Means ↗
3. Confidence belongs to the procedure
Imagine repeatedly drawing a fresh random sample and constructing an interval by the same rule. A procedure with 95% coverage includes the true fixed parameter in about 95% of those intervals over the long run. Particular batches of 100 intervals need not contain exactly 95 successes. After one interval is calculated, it either contains the parameter or it does not. In this frequentist interpretation, 95% is not a probability assigned to that fixed parameter being inside the already fixed interval. Nor does it say that 95% of observations lie there, or that the estimate has a 95% chance of being exactly correct. Wider intervals generally provide greater coverage when using the same model and data. The selected confidence level is a feature of the method, not a score for the quality of the study.
Sources: NIST: Confidence Limits for the Mean ↗ · OpenStax: A Single Population Mean using the Normal Distribution ↗
4. Worked case: a mean interval using t
An invented quality-control sample contains 36 independently selected components. Their signed mass deviations from a target have mean 52 mg and sample standard deviation 12 mg. For this teaching calculation, assume these deviations follow a normal population model with unknown mean and variance. A two-sided 95% t interval is x̄ ± t* s/√n, with n−1 = 35 degrees of freedom. The relevant t quantile is about 2.0301. Thus SE = 12/6 = 2 mg and the margin is 2.0301×2 ≈ 4.0602 mg. The interval is approximately [47.94, 56.06] mg for the population mean deviation. It is not an interval for the next component’s deviation. The t distribution accounts for estimating σ with s; its heavier tails make the critical value slightly larger than the normal value 1.96.
Sources: OpenStax: A Single Population Mean using the Student t Distribution ↗ · NIST: Confidence Limits for the Mean ↗
5. Worked case: an approximate proportion interval
In a fictional random sample of 400 learners, 240 report access to a study desk. Assume independent yes/no observations with one common probability p, a representative sampling frame and accurate answers. Then p̂ = 240/400 = 0.60. The simple normal-approximation standard error is √(p̂(1−p̂)/n) = √(0.24/400) ≈ 0.02449. A nominal 95% interval is 0.60 ± 1.96×0.02449, or approximately [0.5520, 0.6480]: 55.2% to 64.8%. The margin is about 4.8 percentage points. Here both observed counts, 240 yes and 160 no, are comfortably large. This familiar approximation can behave poorly with small samples or proportions near zero or one; Wilson or other appropriate binomial intervals are preferable in such settings. The calculation cannot correct exclusion of learners who lacked a way to respond.
An interval targets a population quantity
| Model and target | Estimate | 95% interval |
|---|---|---|
| Normal population mean deviation, n = 36 | 52 mg; SE = 2 mg | 47.94–56.06 mg (t) |
| Common yes/no proportion, n = 400 | 0.60; SE ≈ 0.02449 | 55.2%–64.8% (approx.) |
Sources: OpenStax: A Population Proportion ↗ · NIST: Confidence Intervals for Proportions ↗
6. Interpret precision alongside possible bias
Random error concerns variation due to which observations happened to enter the sample. Bias concerns a systematic tendency of the process to miss the target. A very large convenience sample can give a narrow interval around the wrong population quantity. State the estimate, interval, sample size, target population, selection process and assumptions together. If responses are clustered within a few schools, treating every learner as independent can understate uncertainty. If sampling without replacement takes a substantial share of a finite population, an appropriate finite-population correction may matter. Weighted surveys require design-aware analysis. These are changes to the probability model, not reasons to add arbitrary extra decimals or an invented safety margin. A useful conclusion explains what the interval addresses and which measurement, coverage and modelling uncertainties remain outside it.
Sources: OpenStax: Data, Sampling, and Variation in Data and Sampling ↗ · NIST: Confidence Limits for the Mean ↗
PUT IT INTO PRACTICE
Practice: how sample size changes a known-σ interval
- Assume independent normal observations with known population standard deviation σ = 10 units. A sample of n = 100 has mean 80. Use a normal critical value of 1.96.
- Calculate the standard error, margin and two-sided 95% interval. Explain why the known σ permits this normal calculation.
- For comparison, suppose n = 400 and the sample mean remains 80. Calculate the new margin and compare interval widths.
- Solution: at n = 100, SE = 1 and margin = 1.96, giving [78.04, 81.96]. At n = 400, SE = 0.5 and margin = 0.98, giving [79.02, 80.98]. Four times as many independent observations halve the width. The unchanged mean is a comparison assumption; a real new sample need not reproduce 80.
Check your understanding
Does the standard error describe the spread of individual observations?
No. It describes the sampling variability of an estimator. Standard deviation describes the spread of the observations themselves.
Does 95% confidence mean 95% of learners are inside the interval?
No. The interval targets a population parameter, such as a mean or proportion, not a range containing individuals.
Why use n−1 degrees of freedom in the t example?
Estimating the mean constrains the sample deviations to sum to zero. Only n−1 of those deviations can vary freely.
Will increasing a biased sample remove its bias?
Not automatically. It may reduce random variation while preserving the same systematic exclusion or measurement problem.
Is a margin of 0.048 a 4.8% relative error here?
It is 4.8 percentage points on the proportion scale. Relative to 0.60, it is about 8%, a different comparison.
Can a confidence interval validate the sampling assumptions?
No. Its calculation presupposes a model. Selection, dependence, measurement and coverage need separate investigation.