Sample size and power
How many units you need to estimate a value or detect a difference, and the inputs every sample size calculation needs.
A sample size calculation balances cost against the risk of a wrong conclusion. Every calculation needs the same four ingredients.
The inputs
- Significance level α (false alarm risk), usually 0.05.
- Power (1 − β), usually 0.80 or 0.90.
- The difference that matters (for example, a shift of 0.01 mm), chosen from engineering, not statistics.
- The process standard deviation, from past data or a pilot study.
Common formulas
With α = 0.05 (two-sided) and 80% power, . To detect a difference of one standard deviation you need about 16 units per group; for half a standard deviation, about 63.
Practical guidance
- Control chart limits: 20 to 25 subgroups (about 100 values).
- Capability study: at least 100 values over representative conditions; 30 gives a very wide interval on Cpk.
- Gage R&R: commonly 10 parts, 3 operators, 2 or 3 trials.
- Reliability demonstration with zero failures: .
Questions and answers
I can only afford 10 parts. Is that useless?
It can still detect large differences. Use a power calculation to see what difference 10 parts can detect, and report the confidence interval honestly.
In RS-1: Quality > Sample Size and Reliability > Sample Size.
