| There has been considerable interest recently on the use of statistically designed experiments to identify both location and dispersion effects for quality improvement. Analysis of dispersion effects usually requires replications which can be expensive or time consuming. Thus, there have been several papers recently that have considered identification of both location and dispersion effects from unreplicated fractional factorial experiments. This dissertation provides a systematic and comprehensive study of the properties of various methods that are commonly used or have been proposed recently and demonstrates that all the non-iterative methods suffer from a problem referred to as “structural bias.” This bias can be serious and can lead to detection of spurious effects or, more importantly, missing of active effects. Both theoretical and simulation results are used to characterize situations when these methods will suffer from these biases. Some of the methods are intrinsically flawed in that they remain biased as the number of observations goes to infinity. Based on these analyses, iterative strategies are proposed for model selection and estimation. A real example as well as simulations are used to illustrate the results. |
