Standard error

On this page 7

Standard error is the standard deviation of a sampling distribution: it measures how much a statistic computed from a sample, most often the mean, would vary if the same study were repeated many times with new random samples. In statistics, it is the basic measure of estimate precision, and it sits underneath confidence intervals, hypothesis tests and the uncertainty reported for regression coefficients.

The idea separates two kinds of spread. Standard deviation describes how individual observations differ from one another, while standard error describes how much the estimate itself is uncertain. Because averaging cancels out part of the random variation, the standard error is always smaller than the standard deviation of the data and shrinks as the sample grows.

How the standard error of the mean is calculated

For the sample mean, the standard error equals the population standard deviation divided by the square root of the sample size, written as SE = σ / √n. Because σ is rarely known, analysts substitute the sample standard deviation s and report the estimated standard error s / √n.

The square root has a practical consequence that surprises many newcomers: to halve the standard error, the sample must be four times larger. Precision therefore improves quickly for small samples and slowly for large ones, which is why survey and experiment budgets show strongly diminishing returns.

The table shows how the standard error of a mean changes with sample size when the data have a standard deviation of 20 units.

Sample sizeSquare root of nStandard errorApproximate 95% margin
2554.0±7.8
100102.0±3.9
400201.0±2.0
1,600400.5±1.0
10,0001000.2±0.4
Standard error of the mean for a standard deviation of 20

When the sample is a large share of a finite population, a finite population correction of √((N − n) / (N − 1)) reduces the standard error. It is usually ignored when the sample is below about 5 percent of the population, because the correction is then negligible.

Standard error of other statistics

Every estimator has its own standard error, because each statistic has its own sampling distribution. The mean is simply the most familiar case; proportions, differences between groups and regression slopes all come with their own formulas.

The most frequently used formulas are collected below.

StatisticStandard error formulaKey assumption
Sample means / √nIndependent observations
Sample proportion√(p(1 − p) / n)Enough successes and failures
Difference of two means√(s₁²/n₁ + s₂²/n₂)Independent groups
Difference of two proportions√(p₁(1 − p₁)/n₁ + p₂(1 − p₂)/n₂)Independent samples
Regression slopeResidual SD / (sx · √(n − 1))Constant residual variance
Common standard error formulas

For statistics without a convenient formula, such as the median or a ratio of two means, practitioners use the bootstrap: they resample the data with replacement thousands of times and take the standard deviation of the recomputed statistic as its empirical standard error.

Standard error and standard deviation are related but answer different questions, and confusing them is one of the most common reporting mistakes in applied research. Standard deviation describes the data; standard error describes the reliability of a number derived from the data.

The comparison below separates standard error from the measures it is most often mistaken for.

MeasureWhat it describesEffect of a larger sampleTypical use
Standard deviationSpread of individual valuesStabilizes, does not shrinkDescribing variability
VarianceSquared spread of valuesStabilizes, does not shrinkModeling and derivations
Standard errorSpread of an estimate across samplesShrinks with √nPrecision of estimates
Confidence intervalPlausible range for a parameterBecomes narrowerReporting uncertainty
Margin of errorHalf-width of a confidence intervalBecomes smallerSurvey results
How standard error relates to neighboring concepts

The link to inference is direct. A 95 percent confidence interval for a mean is approximately the estimate plus or minus 1.96 standard errors for large samples, and a test statistic such as a z-score or t-statistic is the estimated effect divided by its standard error.

Using standard error correctly in analysis

Standard error should be reported whenever the goal is to show how precise an estimate is, and standard deviation whenever the goal is to show how varied the observations are. Good practice states which one is shown, together with the sample size.

The decision table summarizes the usual choices.

SituationRecommended measureReason
Describing patient ages in a studyStandard deviationReaders need the spread of values
Reporting an average treatment effectStandard error or confidence intervalReaders need estimate precision
Small sample under 30Standard error with t-distributionNormal approximation is too optimistic
Clustered or repeated dataCluster-robust standard errorObservations are not independent
Unequal residual variance in regressionHeteroskedasticity-robust standard errorClassic formula is biased
Choosing what to report

With small samples, the estimated standard error is itself uncertain, so intervals use the Student t-distribution with n − 1 degrees of freedom instead of the normal value of 1.96. For a sample of 10, the multiplier rises to about 2.26.

Common misconceptions and failure modes

The standard error only captures random sampling variability, and most errors in its use come from forgetting that boundary. It says nothing about systematic problems in how the data were collected.

  • Small standard error does not mean an unbiased estimate: sampling bias or measurement error survives any sample size.
  • Error bars showing standard error look deceptively tight: one standard error corresponds to roughly a 68 percent interval, not 95 percent.
  • Independence is essential: clustered, time-series or repeated observations make naive standard errors too small.
  • Overlapping error bars do not prove no difference: the comparison requires the standard error of the difference itself.
  • Huge samples make trivial effects significant: a tiny standard error can turn a negligible difference into a statistically significant one.

Put simply

Imagine asking 100 random people how many hours they sleep. You get an average. If you asked a different 100 people, the average would come out a little different. Standard error tells you how much that average would typically wobble from one group to the next.

A small standard error means your average is probably close to the true value. Asking more people makes the wobble smaller. However, you need four times as many people to cut it in half.

In practice

An online retailer runs an experiment on its checkout page. Group A sees the current design and Group B sees a simplified one, with 8,000 visitors in each group. The conversion rate is 4.0 percent in Group A and 4.6 percent in Group B.

The analyst computes the standard error of each proportion, about 0.22 percentage points for A and 0.23 for B, and combines them into a standard error of the difference of about 0.32 points. The observed difference of 0.6 points divided by 0.32 gives a z-value near 1.9, and the 95 percent confidence interval runs from roughly 0.0 to 1.2 percentage points.

Because the interval barely reaches zero, the team does not declare a clear winner yet. It extends the test to about 16,000 visitors per group, which lowers the standard error by roughly 30 percent and gives a more decisive answer before the new design is rolled out.