TABLE OF CONTENTS "Experience does not ever err; it is only your judgment that errs in promising itself results, which are not caused by your experiments."
LEONARDO DA VINCI
Much as we'd like to begin building properly designed experiments geared for response surface methods, it is necessary to first dispel the notion that you can save time and money by simply gathering data from historical records. To some extent, this chapter serves a similar purpose to the lectures given to teens on the dangers of drugs, tobacco, sex, etc. Regression of happenstance data may be every bit as addicting as these substances and activities. However, we're all adults now, so a better reason for discussing happenstance regression may be to slip in some new statistics that can be extremely helpful for diagnosing lousy layouts of input factors.
Learning by bad example is often much more interesting and therefore effective than displaying perfected demonstrations of technical virtuosity. We hope this chapter will generate thought on your part as to how better to design RSM experiments and, more importantly, induce you to vow never to try working around this by collecting historical data and running it through a regression or neural network package.
The mathematical method of ordinary least squares (OLS), attributed by himself and others to Gauss (1795), set the stage for creating multilinear predictive models from historical data under the banner of "regression" a term coined by Galton in 1885 (Stigler, 1986). These models are typically rated on...
