Description
Judge which effects in an unreplicated factorial are real by comparing them with the spread expected if most effects are noise, graphically (Daniel) or numerically (Lenth).
A factorial or fractional factorial run once has no degrees of freedom for error, so significance must rest on effect sparsity: most estimated effects are noise. Daniel's half-normal plot orders the absolute effects against half-normal quantiles; noise effects fall on a line through the origin, the 68 percent point gives an approximate standard error, and effects beyond guardrails are declared active, after which they are removed and the plot redrawn. A full normal plot of signed effects, with quantiles at (i - 3/8)/(k + 1/4), works the same way. Lenth's method gives a numeric version: s0 = 1.5 x median absolute effect, the pseudo standard error PSE = 1.5 x median of absolute effects below 2.5 s0, a margin of error ME = t(0.975, m/3 df) x PSE, and a simultaneous margin SME. Effects are drawn as bars with both lines.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Goal
- Prioritise, Compare designs
- Stage
- Evaluation
- Data
- Quantitative
- Effort
- Hours
Steps
- Estimate all m effects from the unreplicated two-level design.
- Sort the absolute effects and plot them against half-normal quantiles at probabilities (i - 1/2)/m.
- Read an approximate standard error at the 68 percent point and flag effects that lie well off the line through the origin.
- Remove clearly active and borderline effects and replot to re-estimate the standard error.
- For Lenth's method, compute s0, PSE, ME and SME and draw effects as bars with ME and SME lines.
- Treat effects beyond SME as active, those between ME and SME as possibly active, and confirm them in a follow-up run.
Sources
- Design and Analysis of Experiments, Vol. 2 (Hinkelmann & Kempthorne, 2005)
- Statistical Design and Analysis of Experiments (Mason et al., 2003)
Also known as
- Daniel plot
- Half-normal probability plot of effects
- Lenth's method
- Pseudo standard error (PSE)
- Effect sparsity screening
- Normal quantile plot of factor effects
Categories
- Analytics & Behaviour
- Statistics & Modelling