Join the Statistics Division on Friday, February 26, 2016 at 1:00PM Eastern Time for a webinar to be given by Willis Jensen, W.L. Gore & Associates.
Throughout the literature on design of experiments (DOE), the need to perform confirmation runs has been repeatedly emphasized. We provide a high level framework for the sequential nature of DOE and show how the confirmation phase fits within that high level framework. We provide a review of the literature that addresses confirmation runs and show that the guidance given in the literature is inadequate.
Specifically, we address the following four questions that have not been well addressed in the literature:
- Why is confirmation needed in practice? We believe that this is becoming increasingly important as processes become more complex and there is a greater desire for robust conclusions and process capability demonstration.
- How do I know that the confirmation is successful? We review some of the methods and show that some of the common recommended methods do not work very well. A simulation study shows that simultaneous prediction intervals are the best way to assess confirmation runs.
- How many confirmation runs are needed? There is a lot of variety in the literature recommendations ranging anywhere from 1 to 15 confirmation runs, which appear to be loosely based on experience. We show some preliminary results that gives some more specific guidance on how many confirmation runs to perform.
- What if I fail the confirmation? We give some reasons why a confirmation can fail and suggest some possible remedies based on our own experiences.