Normality checks and transformations
How to check normality, what to do when data are not normal, and the Box-Cox and Johnson transformations.
Many methods assume normal data. When data are clearly not normal, you have several options.
Check first
- Plot a histogram and a normal probability plot.
- Use the Anderson-Darling test as a guide, not a verdict.
- Look for causes: mixed populations (two machines), outliers, a physical boundary near the data (zero, a hard stop), or measurement resolution (only a few distinct values).
Options when data are not normal
- Fix the cause if the non-normality comes from mixing streams. Stratify the data.
- Fit a better distribution (Weibull, lognormal, gamma) and use it for capability.
- Transform the data so they become normal, analyze, then transform limits back.
- Use nonparametric methods for tests.
Box-Cox transformation
Raises the data to a power λ (λ = 0 means the natural log). The best λ is chosen by maximum likelihood; common rounded values are 0.5 (square root), 0 (log) and −1 (reciprocal). Requires positive data. Introduced by Box and Cox in 1964.
Johnson transformation
Chooses from three families of transformations to normalize a wider range of shapes, including data with negative values.
Questions and answers
Do I transform my specification limits too?
Yes. Apply the same transformation to the data and the spec limits before computing capability. RS-1 does this for you.
The Box-Cox interval for λ includes 1. What does that mean?
λ = 1 means no transformation. The data are close enough to normal to leave them as they are.
