Design & Factor Selection
Design Types & Categories
Decide on the potential factors to be used in the design; assign names and establish the number of levels of each that are necessary or feasible or that naturally exist. If you expect that the response will be a nonlinear function of the factors, at least 3 levels per factor must be used. If a Central Composite or Mixture design is being considered, the designs available depend on the number of levels.
Is this design and analysis intended to get information about the process response(s) or about response variation or both? If about variation, should a Subsidiary design be used? If not, what kind of replication should be used?
Are all the Factors now being controlled or are they all considered essential, that is , all Main Factors?
Are some Factors not now being controlled but could be controlled for the experiment? Should they also be considered as Main Factors or they be grouped as Subsidiary Factors?
Are there some Factors that cannot be controlled but may be measured during the experiment? They should be included as Casual Factors.
Are the Factors quantitative or qualitative? Can or should qualitative factors be converted to be quantitative? Qualitative factors may not be used in any form of response surface or nonlinear model.
How might the factors interact? Include any prior experience, intuition or technical judgment! Factors that interact must all be Main factors. Resolution 3 Screening designs may be used to show lack of significance but factor significance may be masked by interactions. Use Resolution 4 designs as a minimum if at all possible.
What and how are the Responses to be measured? Responses must be quantitative even if they are based on subjective evaluations.
How many runs are affordable? How many are needed to estimate the factors, interactions and experimental error?
Are environmental and other factors which may be categorized (even when not measurable) expected to be constant (homogeneous) for all the runs? If not, should a recognizable source of variation be included as a blocking factor? Should a stronger effort be made to classify or measure the factor so that it might be included as main or subsidiary factor?
Experiment with the design! Change the design requirements and examine the effect on the cost of the experiment before making a final choice.
Use plots to review the results. Look for trends and anomalies. Expect that your understanding of the experiment is correct, but look for the unexpected.
Remember that trends in plots or in regressions using qualitative factors may be misleading.
When doing regressions, iterate. Use both Half-Normal and Parameter Significance, with your own experience and judgment, as guides to removing or retaining regression parameters. Remember that a Factor may be physically significant without showing statistical significance in a particular experiment.
When an interaction is significant but not all of its main factors are significant, substitute interactions that use significant factors to possibly improve the model..
Interactions CAN be more significant than any of their factors.
When higher order terms (quadratics and interactions) appear to be as or more significant than the linear terms of factors, that might be an indication that a transform is needed.
Plot Residuals against the Response to look for trends. Perhaps a variance stabilizing transform would be useful to improve the regression model.
Learn more about the DOE tools for designed experiments in Six Sigma Demystified (2011, McGraw-Hill) by Paul Keller, in his online Intro. to DOE short course (only $99) or online Advanced Topics in DOE short course (only $139), or his online Black Belt certification training course ($875).