Sponsoring Section/Society: ASA Quality and Productivity
Session Slot: 10:30-12:20 Wednesday
Estimated Audience Size: xx-xxx
AudioVisual Request: xxx
Session Title: Planned & Inadvertent Split-Plotting in Industrial Experiments
Theme Session: No
Applied Session: Yes
Session Organizer: Lucas, James M. J M Lucas & Associates
Address: 5120 New Kent Rd. Wilmington, DE 19808
Phone: 302-368-1214
Fax: 302-456-4013
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 0 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 10 minutes Floor Discussion - 10 minutes
Session Chair: Saccucci, Michael S. Domain Manufacturing Corporation
Address: 3502 Keswick Way Westchester, PA 19382
Phone:
Fax:
1. Split-Plot Response Surface Designs
Lucas, James M., J M Lucas & Associates
Address: 5120 New Kent Rd. Wilmington, DE 19808
Phone: 302-368-1214
Fax: 302-456-4013
Hazel, Malcolm C., Campbell Soup Co.
Abstract: We run a Response Surface Experiment in two different ways; as a (completely) randomized experiment and as a split-plot experiment. Our experiment was motivated by the need to emphasize the difference between a randomized run order and a randomized experiment. The error structure and significance of many of the minor variables changes markedly. The effects of the most significant variables and the overall results of the experiment are similar for both ways of running the Response Surface Experiment because of the dominant effect of a few variables. Analysis procedures are shown which extract all the information from the data. We show how standard regression programs can be used, not only to get the correct tests for Split-Plot Experiments, but also to get good confidence and prediction intervals.
2. Randomizing Run Order Without Resetting Factors
Webb, Derek, Montana St. University
Address: Montana St. University Department of Mathematical Sciences Bozeman, MT 59715
Phone: 406-994-5361
Fax:
Email: webb@math.montana.edu
Lucas, James M., J M Lucas & Associates
Borkowski, John J., Montana St. University
Abstract: Many experiments are run using a random run order, yet if successive runs have the same level of a factor, that factor is not reset. Therefore (complete) randomization is not achieved, and the errors will not be independent. We develop the expected covariance matrix, and calculate the expected bias for significance tests in such experiments. Advantages and disadvantages of using a random run order instead of complete randomization are discussed. In many cases it is appropriate to use a random run order instead of (completely) randomizing the experiment. This is because the maximum variance increases (and the G-efficiency drops) more slowly than the cost drops. This gives justification for the procedure that is so often used in practice. However, it is very important to recognize and understand the difference between a random run order and complete randomization so that the appropriate experimental procedure is used.Often the experimental goal is to find improved operating conditions or to model how a process works over the experimental region. In many such experimental situations the signal to noise ratio is known to be large. In such situations a random run order is often appropriate. When the purpose of the experiment is scientific understanding (and hypothesis tests), then (complete) randomization is required. When there is a single factor that is hard-to-change then the appropriate split-plot experiment should be considered to get a super-efficient experiment (Anbari and Lucas 1994).
3. Strategies for Designing Complex Split-Plot Experiments
Bancroft, Diccon, W.L. Gore & Associates
Address: 297 Blue Ball Road Elkton, MD 21921
Phone: 302-239-0727
Fax:
Email: dbancrof@wlgore.com
Abstract: Split-plot experiments are blocked experiments in which some experimental factors are applied at the block level. Some strategies for finding efficient split-plot experiments are described. These are particularly useful when the block factors are not orthogonal to subplot treatments. Examples given include a class of strip-plot experiments useful for the estimation of variance components, and an unbalanced factorial split-plot.
Discussant: Vining, G. Geoffrey University of Florida
Address: University of Florida Department of Statistics Gainesville, FL 32611
Phone: 352-392-1941 x204
Fax:
Email: vining@stat.ufl.edu