Michael Hamada

Analysis of Designed Experiments with Complex Aliasing

Traditionally, Plackett-Burman (PB) designs have been used in screening experiments for identifying important main effects. The PB designs whose run sizes are not a power of two have been criticized for their complex aliasing patterns, which according t..

Using statistically designed experiments to improve reliability and to achieve robust reliability

This tutorial explains statistically designed experiments which provide a proactive means to improve reliability as advocated by Genichi Taguchi. That is, by systematic experimentation, the important parameters (factors) affecting reliability can be identified along with parameter values that yield reliability gains. In addition to improving reliability, Taguchis robust design can be used to achieve robust reliability; that is, to make a process or product reliability insensitive to factors which are hard or impossible to control. Robust design is also implemented using statistically designed experiments. This paper presents classes of experimental plans for reliability improvement and robust reliability. An important feature of the reliability data collected from such experiments is censoring which occurs when some of the experimental units have not failed by the end of the experiment. Consequently, the analysis methodology must account for these censored data which are likely to occur in light of the ever increasing reliability of todays products. Several appropriate methods are discussed briefly. These experimental plans and analysis methods are illustrated using three documented experiments which improved fluorescent lamp and industrial thermostat reliability and which achieved robust reliability for night-vision goggles. >

Detecting Spatial Effects From Factorial Experiments: An Application From Integrated-Circuit Manufacturing

In addition to the number of functional chips on a silicon wafer, spatial patterns of the nonfunctional chips can provide important information for improving integrated-circuit fabrication processes. In this article, we consider a binary (functional/nonfunctional) response for each chip and propose a method to detect spatial effects in such processes through factorial experimentation. By using a measure of spatial dependence, process factors (parameters) that influence the spatial clustering of functional or nonfunctional chips can be identified. The proposed method, which assumes that a wafer contains a moderately large number of chips and that at least one wafer is produced for each experimental setting, is demonstrated with data from an actual experiment.

A critical look at accumulation analysis and related methods

Industrial quality characteristics are often measured categorically rather than numerically, such as by recording a response as “slight,” “moderate,” or “extreme.” Accumulation analysis, a method proposed by Taguchi (1974) for analyzing ordered categorical data from industrial experiments, is used in Japanese industry and is becoming popular in the United States. Nair (1986) proposed using the first two components of the accumulation analysis statistic separately, as well as using simpler alternatives, to detect location and dispersion effects, respectively. We expose some problems with accumulation analysis in the multifactor setting, which is the usual industrial setting, since it is more efficient to investigate many factors simultaneously. Our results show that accumulation analysis detects spurious factor effects and reverses the order of factor importance. Furthermore, reanalysis of data from two real experiments reveals that these problems with accumulation analysis are realized in practice. We dem...

Analysis of censored data from highly fractionated experiments

Although censored (including interval-censored) data may be easier or less costly to collect than complete data, they contain less information and are harder to analyze. Existing methodology is inadequate for analyzing such data from highly fractionated experiments. We propose an iterative method that provides a simple and flexible way to consider many models simultaneously. The methods simplicity makes it easy to implement with existing software, results in computational savings, and promotes experimenter involvement. We demonstrate the procedure by reanalyzing data from two real experiments. A simulation study also demonstrates its superiority over some existing methods, including Taguchis minute accumulating analysis.

The Treatment of Related Experimental Factors by Sliding Levels

Two factors are related when one factors desirable range depends on the level of another factor. For example, in a chemical reaction, the desirable reaction-time range shifts downward as temperature increases. This is reflected in the experimental desi..

Continuous Process Improvement with Observational Studies

The essential feature of an observational study performed on a process is that data are collected without deliberately changing the process. Control charts, a type of observational study, have traditionally been used to establish and maintain process st..

An explanation and criticism of minute accumulating analysis

Minute accumulating analysis (MAA) is Taguchis method for analyzing life test data from designed experiments. Motivated by comparing survival curves, MAA analyzes sequences of binary data generated from the life test data. This paper explains MAA and e..

The costs of using incomplete response data for the exponential regression model

This paper investigates the asymptotic and small sample costs of using incomplete response data, Situations are identified where the information loss is substantial, Moreover, the small sample properties of the estimators are even worse than suggested by their asymptotic counterparts. These results provide the practitioner with guidance as to the severity of the costs he can incur, This is especially helpful when he cars choose the type of incomplete data that he observes.

The costs of using incomplete exponential data

Abstract Incomplete data, whose values are not known exactly, are often collected in reliability studies. We investigate the asymptotic and small sample costs of several types of incomplete exponential data. Situations are identified where the information loss is substantial. Moreover, small sample properties of the estimators are appreciably worse than suggested by their asymptotic counterparts for these same situations. These results are useful for planning reliability studies that collect such data.