The article was interesting, but seemed to make the mistake of too narrow of a view, maybe based on only certain kinds of experience. Specifically, consider the following quote from the article:
"Testing theories is not usually found in industrial DOE, however. The relevant factors are nearly always known in advance. There is no issue about the importance of the factors."
The first problem is what is meant by the term "industrial DOE". Although I work at a national lab, I still consider the physical and engineering sciences experimental design problems I work on as "industrial DOE". Second, it goes way too far to say that relevant factors are NEARLY ALWAYS known in advance. Sometimes they may be known, sometimes not. In this day and age with technology advancing so quickly, there are often many factors that may influence something, and if it is new technology then very little or only partial information may be known about what is important. As one example, for many years I have applied mixture experiment designs and response surface methods to nuclear and chemical waste glass disposal problems. These wastes often have 10, 20, 30, or even 60 components. Many "major" as well as "minor" components may affect different properties/responses of interest. Most of the existing glass knowledge and experience is for relatively simple commercial glasses without troublesome minor components. Hence, there is little basis for knowing for sure which components are relevant and which are not, especially with many responses of interest. While this may not be a typical industrial problem, I am aware from friends and colleagues that there are many industrial DOE problems where the importance of factors is not known in advance. This is likely to be the case for at least one response in cases where there are many responses.
Also, over the years I have been involved in situations where empirical response surface models are pitted against theoretical models where coefficients have meaning. Often the theoretical models are semi-empirical. It is possible to treat the theoretical coefficients as unknown (along with the empirical coefficients), and use regression methods to estimate the theoretical coefficients from data. This provides a way to assess whether the theory agrees with data, and to modify incorrect theory when needed. Hence, there are cases where looking at coefficients from response surface/regression methods is very useful.
In summary, I suggest avoiding overly broad statements based on perspective and experience that may not included the different perspectives and experience of people with different kinds of "industrial statistics" problems.
The mis-guided statements Box, Hunter, Myers and others have made about optimal design over the years have generally been from too narrow of a perspective or range of experiences on their parts. I remember a conversation I had with Stu Hunter many years ago at a Gordon Conference, where I learned he seldom had been faced with designing an experiment for an irregularly shaped experimental region. Tt is easier for someone to dismiss optimal design methods if he/she has never had a need for them. Better to keep an open mind and remember different people have different problems, and not make claims or recommendations from too narrow a viewpoint.
Still, the article did have some valid points to make. I expect it was written to be a bit controversial.
Greg Piepel, Ph.D.
Statistics Resources, K5-12
Pacific Northwest National Laboratory
P.O. Box 999
Richland, WA 99352
U.S.A.
Voice: 509-375-6911
Fax: 509-375-2604
E-mail: greg.piepel@pnl.gov
PNNL Statistics Group: http://www.pnl.gov/statistics/
Greg is one of the leading experts in experiment design and is particularly well-known for his expertise in mixtures. We are flattered have his input and would grant that his points are well taken. As Greg easily detected, Ron was out to get the juices flowing.
Yes, the article was intended to be controversial. Looks like it was a success.