Robust Designs.
Abstract
An experimental design for fitting the regression model Y = X beta + epsilon can be judged good by various criteria, fourteen of which are enumerated. One of these is that the design should be insensitive to wild observations. This situation is discussed from the points of view of parameter estimation and response estimation. It is shown that a suitable measure of design sensitivity to wild observations can be defined which is closely related to r, the sum of squares of diagonal elements of the matrix R = X(X'X) sup(-1)X', so that the quantity r can be used as a basic design criterion when guarding against wild observations. The use of r is illustrated by applying it to families of composite designs to select the best design in stated circumstances. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1974
- Accession Number
- AD0779442
Entities
People
- George E. P. Box
- Norman Richard Draper
Organizations
- University of Wisconsin–Madison