First Order Rotatable Designs with Correlated Errors (Fordwce)

Abstract

Rotatability as a desirable condition for fitting a responss surface was formally introduced by Box and Hunter (1957) who also derived the rotatability conditions (on the design matrix), assuming the errors to be homoscedastic. However, it is not uncommon to come across practical situations where tho errors are correlated, violating the usual assumptions. In this paper we confine to a first order (linear) regression model with correlated errors. We examine the concept of rotatability of this model and emphasize on properties such as weak rotatability of underlying designs. Various FORDs are examined and their robustness studied. Cost consideration also leads to interesting comparisons.