LimitLine

Learn · updated 2026-10-28

Sample Size and Power: How Many Parts Do You Need?

By Nicholas Williford, Lean Six Sigma Black Belt · About LimitLine

Statistical power is the chance a test detects a difference that really exists; the sample size is chosen so that power reaches a target (usually 80%) for the smallest difference that matters.

Too few parts and a real difference hides in the noise; the test says "not shown" and the project dies. Too many and you waste a week of production. Four numbers decide it, and you choose three of them before you collect anything.

Effect size dthe difference you need to detect ÷ σαfalse-alarm risk you accept, usually 0.05Powerchance of detecting a real difference, usually 0.80nparts per group: what you solve for

Cohen's Table (2-Sample T Test, α = 0.05 Two-Sided, Power 0.80)

Effect size dMeansParts per group
0.2 (small)Shift of 0.2σ394
0.5 (medium)Shift of 0.5σ64
0.8 (large)Shift of 0.8σ26

LimitLine's exact calculation (noncentral t) gives 64 and 26, the numbers in Cohen's table; for d = 0.2 it gives 394, where Cohen's approximation printed 393.

0.00.20.40.60.81.0d 0.8: 26d 0.5: 642255075100parts per group
Power rises with sample size. A large effect reaches 80% power at 26 parts per group, a medium one at 64.

Worked Example

A new fixture should cut the bore diameter's mean by 0.005 mm, and σ is 0.010 mm, so d = 0.005 ÷ 0.010 = 0.5. To detect that with 80% power at α = 0.05 you need 64 parts with the old fixture and 64 with the new. If you can only make 20 of each, the power is 34%: most of the time the trial would miss a real improvement.

Three Ways to Need Fewer Parts

  1. Pair the data. Measure the same parts before and after; the paired test removes part-to-part variation (see hypothesis tests).
  2. Reduce σ. A better gauge or tighter control of other factors makes the same shift a bigger d; check the gauge with a Gauge R&R first.
  3. Ask for a bigger difference. If only a 0.010 mm change is worth the cost, design for d = 1.0, not 0.5.
Work out a sample size in LimitLinePower and sample size for t tests, proportions and ANOVA (a Pro tool; Free users can try it on the example).

Sources

  1. J. Cohen (1988), Statistical Power Analysis for the Behavioral Sciences, 2nd edition, Lawrence Erlbaum (Table 2.4.1)
  2. NIST/SEMATECH e-Handbook, §7.2.2.2 Sample sizes required

The LimitLine Newsletter

About one email a month: new guides, tool updates and one practical SPC tip. Unsubscribe in one click.

We'll email you once to confirm before anything else is sent. Privacy