Learn · updated 2026-10-28
Process Capability for Non-Normal Data
Non-normal capability analysis measures Cpk or Ppk on data that don't follow a bell curve, by fitting the right distribution (or transforming the data) and using its percentiles instead of mean ± 3σ.
The usual Cpk and Ppk assume a bell curve. Flatness, runout, roughness, leak rate and cycle time often aren't: they bunch up near zero and trail off to the right, exactly where the upper spec limit sits. On such data the normal formula can promise a capability the process doesn't have.
1. TestAnderson-Darling normality test (p < 0.05: not normal)2. Fitlognormal, Weibull or a Box-Cox transform3. IndexPpu = (USL − median) ÷ (99.865th percentile − median); Ppl = (median − LSL) ÷ (median − 0.135th percentile); Ppk = the smallerWhythe 0.135% and 99.865% points replace mean ± 3σ
Worked Example: Flatness, Upper Spec 5
Thirty flatness readings (an illustrative example). Flatness can't go below 0, and most parts are near 1 with a few up to 4.
| Method | Ppk | Expected out of spec |
|---|---|---|
| Normal formula: (USL − mean) ÷ 3s | 1.56 | 1.5 ppm |
| Lognormal fit (percentile method) | 1.00 | 1337 ppm |
The lognormal is fitted by maximum likelihood: the mean and standard deviation of ln(x), dividing by n. The Anderson-Darling test rejects normality (p < 0.0001); the lognormal fits (AD 0.41). By the normal formula this process looks comfortably capable at 1.56. The lognormal fit says 1.00 (with only a USL, Ppk is Ppu; and with 30 parts, treat it as approximate): about 1.3 parts per thousand above the spec, roughly a thousand times more than the normal formula predicts, because the long right tail is where the bad parts are.
Which Fit to Use
- Lognormal: positive data with a long right tail (flatness, runout, particle counts).
- Weibull: strength, life and time to failure.
- Box-Cox: a power transform that makes the data normal; report the index of the transformed data and transform the spec the same way.
- Pick the fit that passes the goodness-of-fit test and makes physical sense; with fewer than about 50 parts, the tail estimate is rough.
Sources
- ISO 22514-2:2017, Statistical methods in process management: capability and performance, Part 2 (the percentile method)
- J. A. Clements (1989), Process Capability Calculations for Non-Normal Distributions, Quality Progress 22(9), 95–100
- NIST/SEMATECH e-Handbook, §6.1.6 What is Process Capability? (non-normal data)
The LimitLine Newsletter
About one email a month: new guides, tool updates and one practical SPC tip. Unsubscribe in one click.