|
Activity Number:
|
290
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 8, 2006 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Physical and Engineering Sciences
|
| Abstract - #306892 |
|
Title:
|
A Comparison of Three Approaches to Modeling a Multivariate Response in a Designed Experiment
|
|
Author(s):
|
Steven LaLonde*+ and Peter Bajorski
|
|
Companies:
|
Rochester Institute of Technology and Rochester Institute of Technology
|
|
Address:
|
98 Lomb Memorial Drive, Rochester, NY, 14623-5604,
|
|
Keywords:
|
principal components ; multivariate regression ; designed experiment ; response curve modeling
|
|
Abstract:
|
This paper examines three approaches to modeling a response curve as a function of factors in a designed experiment. The response data is a 21 step density log exposure curve. The experiment is a 2 level full factorial with 5 variables. The first approach is multivariate regression of the original twenty-one density variables. The second approach is a three step procedure: 1)reduce the number of response variables using principal components, 2)model the principal components using regression, and 3)reconstruct original twenty-one density variables from predicted principal components. The third approach is similar to the second, except that the usual principal components method is replaced by a non-negative principal components method described by Bajorski and Hall (2006). The paper addresses such issues as noise reduction, interpretability of solutions, and variable selection.
|
