Standard Error: Formula, Calculation, Examples, and Interpretation

What Is Standard Error?

Picture a population — all the adults in a country, every widget from a factory, every patient in a database. You cannot measure all of them, so you draw a random sample. Your sample gives you a mean, a proportion, or some other number you want to use to represent the population. That number is an estimate, and like any estimate, it carries uncertainty.

Standard error quantifies that uncertainty. Specifically, it answers: if you repeated the same study with a fresh random sample of the same size, how much would your statistic move around from sample to sample?

Suppose a researcher draws repeated samples of 40 students from a large university and records the mean GPA each time:

Each sample mean differs slightly. The standard deviation of those sample means — across many such repetitions — is the standard error of the mean. In practice, you estimate it from a single sample using the formula SE = s / √n.

Standard Error Formula

The most common form — the standard error of the sample mean — divides the standard deviation by the square root of the sample size.

In most real-world research, the population standard deviation σ is unknown, so researchers use s — the sample standard deviation — and obtain the estimated SE. The population version is used when σ is genuinely known (rare in practice, common in textbook exercises).

Formula components at a glance

How to Calculate Standard Error

The calculation has three steps regardless of the dataset size.

Worked Examples

Example 1 — Summary statistics given

Example 2 — Calculated from raw data

Why Sample Size Reduces Standard Error

SE = s / √n. As n grows, √n grows too, making SE smaller — assuming the underlying variability in the data stays roughly the same. This is not a coincidence; it reflects a mathematical fact about how averaging works: more data points dampen the influence of any single unusual observation.

The relationship follows a square-root rule with a diminishing-return character. Quadrupling the sample size halves the SE:

Standard Error vs Standard Deviation

This is the single most common source of confusion in applied statistics. Here is the clearest way to keep them straight: SD asks "how spread out is this dataset?" while SE asks "how precisely can I estimate this statistic?"

A concrete example makes this stick. Suppose you measure the heights of 10,000 adults. The SD of those heights might be around 7 cm — that number does not shrink just because you have a large sample. It reflects genuine human variation. But the SE of the sample mean would be 7 / √10,000 = 0.07 cm, telling you the sample mean is an extremely precise estimate of the population mean height. The two numbers describe entirely different things.

Standard Error and Sampling Distributions

Every statistic has a sampling distribution — the theoretical distribution of that statistic across all possible samples of the same size from the same population. For the sample mean, this distribution has a standard deviation equal to σ / √n, which is the population-level standard error.

When σ is unknown (the usual case), you estimate the sampling distribution's standard deviation using s / √n. That estimate is what most people mean when they say "standard error." The two ideas — the theoretical standard deviation of the sampling distribution and the estimated SE from your data — converge as the sample grows.

Standard Error and the Central Limit Theorem

The Central Limit Theorem (CLT) states that, under appropriate conditions, the sampling distribution of the sample mean becomes approximately normal as sample size increases — regardless of the shape of the underlying population distribution. The standard deviation of that normal distribution is SE = σ / √n.

This is why SE appears in the denominators of z- and t-tests. When the CLT applies, you know the shape (approximately normal) and the spread (SE) of the sampling distribution, which is exactly the information you need to construct p-values and confidence intervals.

Standard Error and Confidence Intervals

SE is the building block of every standard confidence interval. The general structure is:

Choosing between z* and t*

When σ is known, use z* (e.g. z* = 1.960 for 95% confidence). When σ is unknown — the typical situation — use t* from the t-distribution with n − 1 degrees of freedom. For large samples (n ≥ 100 or so) the t* and z* values are nearly identical, but for small samples the t* value is noticeably larger, which appropriately widens the interval to account for extra uncertainty.

Standard Error and Margin of Error

The margin of error is simply the critical value multiplied by the SE. It is half the width of a confidence interval.

The three concepts are related but not interchangeable. You cannot swap one for another — reporting SD where SE is expected, for instance, typically inflates the perceived uncertainty. A useful check: if the SE and SD are equal in your output, your n = 1 (or something has gone wrong).

Standard Error and Hypothesis Testing

In a one-sample hypothesis test, SE appears in the denominator of the test statistic, scaling the difference between the sample estimate and the null-hypothesised value by the estimated sampling variability:

A larger SE means the same difference between x̄ and μ₀ produces a smaller test statistic — and a less extreme p-value. This is why underpowered studies with small n (and large SE) often fail to detect effects that exist. The test statistic and the resulting p-value are both functions of SE, which is itself a function of both SD and n.

Standard Error Calculator

Standard Error of a Proportion

When your statistic is a proportion rather than a mean — the fraction of survey respondents who agree, the percentage of parts that fail quality control — the SE formula changes:

Standard Error of the Difference Between Two Means

When you compare two independent groups — a treatment and a control, males and females, two product variants — the SE of the difference between sample means combines the within-group variability of each:

This formula underpins the two-sample t-test. Welch's version (shown above) does not assume the two groups have equal variances, which makes it the safer default in practice. If equal variances can be justified, a pooled-variance version is available, but Welch's test is recommended unless you have good reason to assume equal variances.

Standard Error in Regression

In regression output, every coefficient estimate comes with its own SE. The SE of a coefficient reflects how precisely that coefficient is estimated from the data, given the model structure and assumptions:

A coefficient with a smaller SE is estimated more precisely, all else equal. This t-statistic tests whether the coefficient differs significantly from zero. Confidence intervals for coefficients also use the same structure: β̂ ± t* × SE(β̂).

The phrase "standard error of the estimate" (sometimes written SEE) refers to something distinct: it measures the typical residual error — how far predictions stray from observed values. Do not confuse SEE with the SE of a coefficient.

Calculating Standard Error in Software

Excel

Excel has no single built-in SE function, but the calculation takes one line once you have your data in a column (e.g. A2:A101):

-- Sample standard deviation
=STDEV.S(A2:A101)

-- Sample size
=COUNT(A2:A101)

-- Standard error (one formula)
=STDEV.S(A2:A101)/SQRT(COUNT(A2:A101))

-- 95% CI lower bound (uses t critical value)
=AVERAGE(A2:A101) - CONFIDENCE.T(0.05, STDEV.S(A2:A101), COUNT(A2:A101))

-- 95% CI upper bound
=AVERAGE(A2:A101) + CONFIDENCE.T(0.05, STDEV.S(A2:A101), COUNT(A2:A101))

Use STDEV.S (sample SD) rather than STDEV.P (population SD) unless you genuinely have the full population. CONFIDENCE.T uses the t-distribution; CONFIDENCE.NORM uses the normal distribution (appropriate when σ is known or n is very large).

R

# Manual calculation
x <- c(12, 15, 14, 18, 16, 13, 17, 15)
n <- length(x)
s <- sd(x)           # sample SD
se <- s / sqrt(n)    # standard error
cat("SE =", round(se, 4))

# From a t-test (gives SE automatically)
t.test(x)

# From regression output
model <- lm(y ~ x, data = df)
summary(model)  # Std. Error column shows SE for each coefficient

Python

import numpy as np
from scipy import stats

x = [12, 15, 14, 18, 16, 13, 17, 15]

# Manual
n = len(x)
s = np.std(x, ddof=1)   # ddof=1 for sample SD
se = s / np.sqrt(n)
print(f"SE = {se:.4f}")

# Using scipy — stats.sem() uses ddof=1 by default
se_scipy = stats.sem(x)
print(f"SE (scipy) = {se_scipy:.4f}")

# Note: np.std(ddof=0) is the population SD; use ddof=1 for sample

SPSS

In SPSS, SE appears automatically in several procedures. The most direct route for descriptive SE is:

Analyze → Descriptive Statistics → Explore — tick "Statistics" and ensure "Standard error of mean" is selected. For regression, Analyze → Regression → Linear shows SE for each coefficient in the Coefficients table.

Precision vs Accuracy

A small SE means the estimator is precise — repeated samples would produce similar values. It does not mean the estimator is accurate — close to the true population value. An estimator can be consistently wrong if there is systematic bias in the data collection.

Increasing the sample size will steadily reduce SE toward zero, but if the sample is drawn using a flawed method — voluntary response, convenience sampling, non-response bias — the resulting estimate may remain far from the population parameter regardless of how small SE becomes.

Small Samples and Non-Normal Data

The SE formula s / √n is always calculable, but the inferential procedures built around it (t-tests, z-intervals) rest on assumptions that need scrutiny when samples are small or data are non-normal.

With small n, the t-distribution with n − 1 degrees of freedom accounts for the extra uncertainty from estimating σ from few data points — this is why t* values are larger than z* values for small samples. Severe skewness, heavy tails, or the presence of influential outliers can make the CLT approximation unreliable even at moderate n, and in those cases alternatives such as bootstrap intervals or non-parametric tests may be more appropriate.

Dependent and Clustered Data

The formula SE = s / √n assumes observations are drawn independently. When observations are correlated — repeated measures on the same person, students clustered within schools, patients within hospitals — the effective amount of independent information is less than n, and the standard SE formula can substantially underestimate sampling uncertainty.

Appropriate methods in these contexts include mixed-effects models, cluster-robust standard errors, and survey-design-adjusted estimators. Reporting naive SE from a clustered design can produce intervals that are far too narrow and p-values that are far too small.

Common Mistakes

Formula Cheat Sheet

Frequently Asked Questions

Related Statistics Resources