|
Statistisk forsøksdesign baserer seg på at dataene er tilnærmet
normalfordelte. En ofte nyttet måte å komme dit for «unormale» data er ved en
Box-Cox transformasjon. Artikkelen dokumenterer hvordan transformasjonen
samspiller med selve analysen, og viser at en god analyses må baseres på et
samspill mellom transformasjon, modellenes enkelthet og normalfordelingsforutsetning.
Publisert i “Quality Engineering”, tidsskrift for «American Society for Quality”.
The Box-Cox transformation was evaluated with reference to a six-factor full-factorial (26) data set with 64 runs. The data were used to determine the optimal operating conditions for a milling machine with respect to surface finish. A suitable transformation was determined by minimizing the mean square errors, evaluating
the size of the effect significances and the normal probability plots of the estimated effects, Shapiro-Wilk tests, and the model residuals. The achievement
of both normality with constant variance and a simple model came about as a result of a trade-off between several different criteria.
DOI:10.1080/08982112.2011.616150E.
http://www.tandfonline.com/doi/abs/10.1080/08982112.2011.616150
https://www.cristin.no/as/WebObjects/cristin.woa/wo/0.Profil.29.25.2.3.15.1.0.3
