Analysis Strategies for Constrained Mixture and Mixture Process Experiments Using JMP(R) Pro 14 ( 2019-EU-45MP-068 )
Feb 11, 2019 4:33 PM
| Last Modified: Mar 28, 2019 7:05 AM
Level: Intermediate Job Function: Analyst / Scientist / Engineer Philip Ramsey, Owner and Professor, North Haven Group and University of New Hampshire
Best Contributed Paper Finalist
Although important to process/product development or improvement, constrained mixture and mixture process experiments have been challenging to analyze due to large effect correlations caused by regional and linear constraints. Often small underfit models with lack of fit or large overfit models with inflated prediction variance have been selected. With the advanced model selection methods in the Generalized Regression platform of JMP Pro, it is now possible to fit and evaluate large sets of potential mixture effects. Model selection techniques such as pruned forward, LASSO, or all possible models allow one to find a best subset of effects from a large pool of candidates even when the number of potential effects exceeds the number of experimental trials (supersaturation). Using a 10-component mixture experiment with both regional and linear constraints, we demonstrate how Generalized Regression can evaluate even large mixture models (e.g., Scheffe full cubic) and identify the important subset of effects. The autovalidation method of Gotwalt and Ramsey (2018) to prevent overfitting is illustrated. We use the same methods to analyze a complicated mixture process factor experiment. The talk is primarily live JMP analyses and discussion. Generating mixture or mixture process experiments in Custom Design is also discussed.