273
Some Guidelines for Designing an Industrial Experiment
a mean or a standard deviation.
Gauge capability (or measurement error) is also an important factor. While some individuals have claimed that unless a certain minimum level of gauge error has been achieved designed experiments cannot be used, we do not support this view. If gauge capability is poor, then only relatively large factor effects will be detected by the experiment or additional replication will be required. Another possibility is to make repeat measurements on the same experimental unit to reduce the effect of measurement error. These repeat measurements are not replicates, and it is fairly typical to analyze the average of these repeat measurements when their purpose has been to combat problems with gauge capability. Designed experiments can be of considerable value in studying the performance of measurement systems. For example, see Montgomery and Runger (1993 a, b), and the references therein.
Choice of Experimental Design
If the first three steps are done correctly, this step is relatively easy. Choice of design involves selecting a particular test plan, such as a 2k p fractional factorial as suggested earlier, a central composite design or an augmented simplex-lattice design. It also involves the consideration of sample size (number of replicates), the selection of a suitable run order for the experimental trials, and the determination of whether or not blocking or other randomization restrictions are involved. Box, Hunter, and Hunter (1978) and Montgomery (1991) discuss some of the morÄ™ important types of experimental designs, and these books can, in a sense, be used much as we would use a catalog for selecting an appropriate experimental design for a wide variety of problems.
There has been considerable interest in recent years in interfacing the Computer into the process of planning, conducting, and analyzing data from designed experiments. There are over 40 companies supplying various types of "DOX" software. Computer-aided experimental design is, in some ways, a very good thing. For example, the Computer can easily and effectively generate the test matrix, once the type of design has been chosen. It is interesting to consider the possibility of a package that solicits all of the input de taił and background and then using some knowledge-based (or expert system) process, will recommend a design. Because of the relatively high degree of human interaction that occurs on well-designed experiments, I don't think it is likely that an effective product of this type will appear soon. There are already some products that produce a "recommended" design, using some "black-box" method that is masked from the user. I have always been very skeptical about