Specifying Design Parameters

Before beginning a design, the factors that will be evaluated must be selected. All the recognized factors must be first classified into controllable, controllable with special conditions, and not controllable by a known or feasible method. There may also be unrecognized factors which may influence the response. These unrecognized factors are often the major source of variability in the response.

Of the recognized factors, it may be desirable to hold some at a constant level for the purpose of the experiment. It must then be realized that the results are constrained by the chosen constant factor-levels.

Of the remaining controllable factors, some may be considered as essential for the process and as amenable for routine control. Others may be too expensive to control routinely. Of the latter group, a choice can be made between purposeful (although non-routine) control at fixed levels and allowing the factor-levels to vary randomly (a casual factor). The decision to allow random variation should have some basis in the characteristics of variation. Random variation will be sensitive to the frequency of occurrence; however, a small number of runs may not represent the true distribution of the parameter and its average effect on the response. In addition, the limited time span of an experiment may not be long enough to allow the various sources of random variation to operate.

Setting the factor-levels at fixed levels for the experiment hides any information about the distribution of those levels, but facilitates ANOVA or regression analysis.

The non-routinely controlled group of factors may be considered as a source of noise (aka subsidiary factors or outer factors). When subsidiary factors are excluded from an ANOVA or regression, the remaining factors may be used to form subgroups. The mean and variance of each subgroup is calculated from the different values attributable to the varying levels of the subsidiary factors, as well as variation from unrecognized factors and inherent variation.

It is important that the amount of variation in the response caused by the subsidiary factor-levels should be small compared to the value of the response. If it is not small, one or more of the "noise" factors may be an important factor. If important, there may be justification of promoting it to "main factor" status with a commitment to establish control over that factor.