Resources What's New Support Download Contact Sales: Arthur-Industrial@outlook.com
Home › Resources › Capability for non-normal data

Capability for non-normal data

Why normal-based Cpk misleads for skewed data, and the percentile, distribution-fit and transformation methods.

For skewed characteristics (flatness, roundness, runout, surface finish, cycle time) a normal-based Cpk can be badly wrong: it usually overstates defects in the short tail and understates them in the long tail.

Approaches

  • Fit a distribution (Weibull, lognormal, gamma, extreme value) and calculate the expected fraction outside each limit. Express it as an equivalent Ppk or as ppm.
  • Percentile (ISO) method: replace by the distance between the 0.135th and 99.865th percentiles of the fitted distribution, and by the distance from the median to those percentiles.
  • Transform the data (Box-Cox or Johnson) to normal, transform the spec limits the same way, and compute capability on the transformed scale.

Things to check first

  • Mixed populations (two cavities, two machines) can look non-normal. Separate them.
  • A natural boundary (zero for flatness) produces skew. A distribution with a bound, or the half-normal, may fit best.
  • Coarse measurement resolution produces few distinct values and fails normality tests. Fix the gauge.

Questions and answers

Which method should I report to my customer?

Use the one your customer's quality manual asks for. Many automotive customers accept distribution-fit or percentile methods. State the method and the distribution in the report.

In RS-1: Quality > Non-Normal Cpk; Data > Box-Cox.

← All articles