Sample Size and Sampling Methods

Sample size is the number of observations you need so that an estimate or a test is precise enough for the decision. It depends on four things: the confidence level (or α), the margin of error or the smallest difference worth detecting, the variation in the data (σ, or p for proportions), and, for tests, the power.

For a mean, n = (zσ/E)²; for a proportion, n = z²p(1 − p)/E². The sample also has to be representative, which is a question of how you sample, not only how many.

An original chart computed from n = z²p(1 − p)/E². Precision is expensive: going from ±5% to ±2.5% takes the sample from 385 to 1,537.

Population and sample

The population is every unit you want to draw conclusions about, such as all orders shipped this quarter or all parts a process will make. A sample is the subset you actually measure. Population values (μ, σ, p) are parameters; sample values (x̄, s, p̂) are statistics that estimate them. In Six Sigma the population is often a process that keeps producing, so a sample should cover the conditions the process really runs under: shifts, machines, operators, and suppliers.

Which sampling method should you use?

Method	How it works	Use when
Simple random	Every unit has an equal chance of selection	You have a list of the population and no important subgroups
Stratified	Split into strata (shift, product, region), then sample randomly within each	Subgroups differ and each must be represented
Systematic	Every kth unit after a random start	Units flow in sequence; watch for cycles that match k
Cluster	Randomly choose whole groups (boxes, days, branches) and measure all units in them	A full list is impractical and groups are similar
Rational subgroups	Small groups made close together in time, taken at intervals	Building control charts: variation within groups is short-term, between groups shows shifts

Avoid convenience samples, such as the first 30 parts of the morning or only the complaints that reached a manager. They are easy to collect but rarely represent the process.

How do you calculate sample size?

Goal	Formula	Inputs
Estimate a mean within ±E	n = (z·σ / E)²	z = 1.96 for 95% confidence; σ from history or a pilot
Estimate a proportion within ±E	n = z²·p(1 − p) / E²	Use p = 0.5 if unknown (largest n)
Detect a difference Δ between two means	n per group ≈ 2[(z_α/2 + z_β)·σ / Δ]²	z_β = 0.84 for 80% power, 1.28 for 90%
Small finite population N	n = n₀ / [1 + (n₀ − 1)/N]	n₀ from the formulas above

Always round up. If σ is a guess, use a pilot sample or the upper end of past values, because an underestimated σ gives a sample that is too small.

Worked examples

Mean. Cycle time has σ ≈ 2.0 minutes and you want the mean within ±0.5 minutes at 95% confidence: n = (1.96 × 2.0 / 0.5)² = 61.5, so 62.
Proportion. To estimate a defect rate near 10% within ±3 points: n = 1.96² × 0.10 × 0.90 / 0.03² = 384.1, so 385. With no idea of p, use 0.5: ±5 points also needs 385.
Finite population. If only 1,000 invoices exist, the 385 shrinks to 385 / (1 + 384/1,000) = 278.2, so 279.
Comparing two means. To detect a 1.0-minute difference with σ = 2.0, α = 0.05, and 80% power: 2 × [(1.96 + 0.84) × 2.0 / 1.0]² = 62.8, so 63 per group by the normal approximation. The exact t-based calculation gives 64, which is what software reports.

Power and sample size for tests

In a hypothesis test, power is the chance of detecting a real difference of a given size. Too small a sample makes Type II errors likely, so “no significant difference” means little. Required n rises with smaller differences (n ∝ 1/Δ²), more variation, a smaller α, and higher power. Decide the smallest difference that matters to the business before collecting data, then size the sample for it.

Building a data collection plan

State the question and the metric, with an operational definition everyone measures the same way.
Confirm the measurement system is adequate with MSA.
Choose the sampling method and size from the sections above, including the strata to capture.
Decide who collects, where, when, and on what form, such as a check sheet.
Record the stratification factors (shift, machine, operator, and so on) so you can slice the data later.
Run a short pilot, fix the form and the definitions, then collect.

Frequently asked questions

Is a sample of 30 always enough?

No. Thirty is a rough point where the t distribution is close to normal and averages start to look normal for moderately skewed data. Whether 30 is enough depends on the precision you need, the variation, and the size of difference you must detect. Use the formulas.

Why does halving the margin of error need four times the sample?

Because precision improves with the square root of n: the standard error is σ/√n. To halve the error you must make √n twice as large, which means n four times as large.

Does a bigger population need a bigger sample?

Only slightly, and only when the sample is a large share of the population. Beyond a few thousand units, the required sample hardly changes, which is why the finite population correction matters only for small populations.

What is the difference between a sample size for estimation and one for a test?

Estimation sizes the sample for a margin of error around one value. A test sizes it to detect a difference with a stated power, so it also needs the smallest important difference and the power you want, usually 80% or 90%.

Last updated: November 11, 2026 · Published by Lean Six Sigma Training Guide (sixsigmatrainingfree.com)