Over the last two decades industrial usage of statistically designed experiments has increased substantially. As the applications have expanded, industrial experi- menters have frequently encountered nonnormal response variables. Generalized linear models (GLM) are a useful alternative to the traditional methods for analyzing such data based on transformations. This paper presents three examples of designed experiments with nonnormal responses. For each example, models are built using classical least squares methods applied to the appropriately transformed data and models are also built using the GLM. Comparisons between the modeling techniques are made, and the impact on interpretation of results from an engineering or scientific viewpoint is discussed. We show that GLM is an excellent alternative to the transformation approach.
Revised: April 18, 2002 |
Published: February 1, 2002
Citation
Borror C.M., A. Heredia-Langner, and D.C. Montgomery. 2002.Generalized linear models in the analysis of industrial experiments.Journal of Propagations in Probability and Statistics 2, no. 2:127-143.PNNL-SA-36087.