Sample Size Calculator — Mean, Proportion & Capability Studies

Every Measure-phase decision rests on a sample, and a sample can mislead in two directions. Collect too little data and the study is underpowered: a real improvement — a genuine two-minute reduction in handle time, a true drop in defect rate — hides inside the noise, the hypothesis test comes back "not significant", and a working solution gets abandoned. Collect too much and you waste measurement effort: operators pulled off the line, destructive tests consumed, weeks of data collection to answer a question that fifty observations would have settled.

Sample size planning is the fix. Before collecting anything, you state the smallest difference worth detecting, the risk of a false alarm you can live with (α), and the probability of catching a real effect that you demand (power). The formulas below then tell you exactly how much data buys that guarantee — no more, no less.

For a continuous measurement (a mean), the large-sample normal approximation gives:

one-sample:   n = (zα + zβ)² · σ² / Δ²
two-sample:   n = 2 · (zα + zβ)² · σ² / Δ²   (per group)

where σ is the process standard deviation, Δ is the smallest difference you want to detect, z_α is the two-sided critical value for your significance level (1.96 at α = 0.05), and z_β is the value for your power (0.84 at 80%).

For a proportion (pass/fail data), the one-sample estimation form and the two-sample comparison form are:

one-sample:   n = z² · p(1 − p) / E²
two-sample:   n = (zα + zβ)² · [p₁(1 − p₁) + p₂(1 − p₂)] / (p₁ − p₂)²

with E the margin of error. Note that p(1 − p) peaks at p = 0.5, which is why 0.5 is the conservative default when you have no prior estimate.

When you sample from a small bounded population of size N (a batch, a fixed roster) rather than an ongoing stream, the finite population correction shrinks the requirement:

n_adj = n ÷ (1 + (n − 1) / N)

A call-center project wants to prove a new script cuts average handle time by at least 2 minutes. Historical σ is 5 minutes; the team will compare a before group and an after group (two-sample) at α = 0.05 with 80% power.

n = 2 · (zα + zβ)² · σ² / Δ²
  = 2 · (1.96 + 0.842)² · 5² / 2²
  = 2 · 7.849 · 25 / 4
  ≈ 98.1  →  round up: n = 99 per group  (198 calls total)

The raw formula lands at about 98 per group — the familiar textbook figure when the table values 1.96 and 0.84 are used — and rounding up (sample sizes always round up, never down) gives 99 calls per group. If the team could only measure 60 calls per group, the achieved power drops to roughly 60%: a two-in-five chance of missing a script that genuinely works. That is the conversation this calculator is built for.

The defaults — α = 0.05 and power = 0.80 — are the convention in Six Sigma training and in Minitab, and they are a sensible starting point: a 5% false-alarm risk and an 80% chance of catching a real effect. Tighten α to 0.01 when a false positive triggers an expensive change; raise power to 0.90 or 0.95 when missing a real effect means abandoning a good solution.

The input that dominates everything, though, is Δ. Because Δ is squared in the denominator, sample size grows quadratically as the detectable difference shrinks: halving Δ quadruples n, and chasing a difference one-tenth the size needs one hundred times the data. So resist the urge to ask for exquisite sensitivity — set Δ to the smallest difference that would actually change a decision, not the smallest difference you can imagine. The same logic applies to σ: since it is also squared, an honest estimate from recent data matters. When in doubt, use a slightly pessimistic (larger) σ and accept a modestly larger study over an underpowered one.

Capability indices need a different kind of answer, because Cp and Cpk are ratios of an estimated spread — and spread estimates converge slowly. Two industry rules apply. The AIAG SPC manual asks for at least 25 subgroups at your planned subgroup size (25 × 5 = 125 measurements at the typical subgroup of five), collected across enough time to capture real process variation. ISO 22514 asks for at least 50 individual measurements. This calculator surfaces both and recommends the more demanding of the two, which is what the LeanProjax toolkit records into a project. A Cpk quoted from 15 measurements is a guess wearing a decimal point — the confidence interval around it is wide enough to span "incapable" and "world-class" at once.