Standard Error Calculator

The standard error tells you how much a sample mean is likely to differ from the true population mean. A sample of 10 patients gives a wobbly estimate; a sample of 10,000 gives a steady one. The standard error is the number that quantifies that difference, and it appears in every confidence interval, every t-test, every hypothesis-testing procedure ever invented.

The formula:

SE = s / √n

Where s is the sample standard deviation and n is the sample size. The standard error has the same units as the original data (e.g., kilograms, dollars, points), and it shrinks as 1/√n. To halve the standard error, you need 4× the sample size. To cut it by 10, you need 100× the sample size. This square-root scaling is why precise measurements of small effects are so expensive.

Where the formula comes from:

If you draw repeated samples of size n from a population and compute the mean of each sample, those sample means form a distribution. The Central Limit Theorem says that for any reasonable population, this distribution of sample means is approximately Normal, centered on the true population mean μ, with standard deviation σ/√n. The standard error is just our best estimate of σ/√n, substituting the sample standard deviation s when σ is unknown (which it almost always is).

SE vs SD (standard deviation):

These two are constantly confused. They are not the same.

• Standard deviation (s or σ): how spread out individual data points are. A sample of adult heights has s ≈ 8 cm regardless of sample size. That is intrinsic variability.
• Standard error (SE): how spread out sample means are. Same population, SE shrinks with √n. That is sampling precision.

Use SD to describe the data. Use SE to describe how well you estimate the mean. Confidence intervals around a mean are built from SE, not SD.

Confidence interval construction:

Once you have x̄ (sample mean) and SE, the confidence interval is:

CI = x̄ ± z · SE (or t · SE for small n)

Common z values for known population variance or large n (typically n ≥ 30):

For small samples (n < 30), use the t-distribution with n−1 degrees of freedom; the constants are similar but slightly larger.

Worked example, lab measurements:

A chemist takes n = 36 measurements of a compound concentration. The sample mean is x̄ = 78 mg/L and the sample standard deviation is s = 12 mg/L.

SE = s / √n = 12 / √36 = 12 / 6 = 2.0 mg/L

The 95% confidence interval is:

CI = 78 ± 1.96 × 2.0 = 78 ± 3.92 = (74.1, 81.9) mg/L

So the true mean concentration is somewhere between 74.1 and 81.9 mg/L with 95% confidence.

Worked example, sample-size impact:

Same chemist runs only n = 9 measurements with the same s = 12 mg/L.

SE = 12 / √9 = 12 / 3 = 4.0 mg/L (twice as large)

95% CI = 78 ± 1.96 × 4.0 = (70.2, 85.8) — a much wider interval. To halve the CI width from this 9-sample estimate, the chemist would need to run n = 36 (a 4× increase in samples gives 2× precision).

Standard error of a proportion:

For binary outcomes (yes/no, success/failure), the standard error has a different form:

SE_p = √(p · (1 − p) / n)

Where p is the sample proportion. This is what political polls use. A 1,000-respondent poll with 50% support has SE = √(0.5·0.5/1000) = 0.0158 ≈ 1.6%, giving the typical ±3% margin of error you see in published results.

When the formula fails:

• Small n (under about 30): the t-distribution should be used instead of the Normal. Multiply by t_(n-1) rather than z. For n = 5 at 95%, t = 2.776 rather than 1.96.
• Highly skewed distributions: the Central Limit Theorem still applies for reasonable n, but convergence is slow. Bootstrap methods are safer for heavy-tailed data.
• Dependent observations: SE = s/√n assumes independent samples. Time-series data, cluster-sampled surveys, and repeated measures violate this — design-effect or hierarchical models are needed.
• Stratified or weighted samples: the formula needs modification to account for the sampling design.

Practical advice:

Always report SE with the sample mean. “Mean = 78 mg/L, SE = 2.0 mg/L” is informative; “Mean = 78” without SE is useless because there’s no way to know whether the true value is 77 or 87.

Comparing two means? The standard error of the difference is √(SE₁² + SE₂²), not SE₁ + SE₂. This shows up in every t-test.

SEM in research papers:

Many published studies show “mean ± SEM” rather than “mean ± SD”. The argument for SEM: it lets the reader see the precision of the estimate, which is what matters for inference. The argument for SD: it lets the reader see the variability in the data, which matters for understanding the phenomenon. Both are useful; the difference is the factor √n. Always read which one is plotted.