Using Split-Plot Designs for Effecient Experimentation

• Activity Number: 524
• Type: Roundtables
• Date/Time: Wednesday, August 6, 2014 : 12:30 PM to 1:50 PM
• Sponsor: Quality and Productivity Section
• Author(s): Brooks Henderson*+
• Companies: Stat-Ease
• Keywords: design of experiments ; hard to change ; ANOVA ; split-plot ; DOE ; randomization

Abstract

As a practical matter, engineers cannot always accommodate complete randomization for their design of experiments (DOE). Often, it needs to be sorted for ease of experimentation by those factors that are hard to change (HTC). This restriction in randomization creates a split-plot design . To obtain meaningful p-values, the statistical analysis must take into account the resulting multiple error terms. Split-plot designs group runs by specific levels of the HTC factors. For example, if an experiment calls for treating samples at multiple levels of temperature in an oven, it is far more efficient to treat all the samples at one level of oven temperature in one group than to run just one sample at a time and change temperatures between each run. In fact, many experimenters routinely sort their designs, unknowingly creating a split-plot design and thus invalidating p-values computed as if the runs had been completely randomized. This round table is for DOE advisors who want to help experimenters handle HTC factors, understand the trade-offs in power versus complete randomization, and apply the proper statistical analysis.