Optimal Blocked and Split-Plot Designs Ensuring Precise Pure-Error Estimation of the Variance Components. Issue 1 (2nd January 2020)

• Record Type: Journal Article
• Authors: Mylona, Kalliopi; Gilmour, Steven G.; Goos, Peter
• Is Part Of: Technometrics. Volume 62:Issue 1(2020)
• Page Start: 57
• Page End: 70
• Publication Date: 2020-01-02
• Subjects: Model-independent variance component estimates -- Restricted or residual maximum likelihood (REML) -- Restricted randomization -- Treatment model; Statistical physics -- Periodicals; Chemistry -- Statistical methods -- Periodicals; Engineering -- Statistical methods -- Periodicals; 519.5
• DOI: 10.1080/00401706.2019.1595153
• Languages: English
• ISSNs: 0040-1706

Abstract

Abstract: Textbooks on response surface methodology generally stress the importance of lack-of-fit tests and estimation of pure error. For lack-of-fit tests to be possible and other inference to be unbiased, experiments should allow for pure-error estimation. Therefore, they should involve replicated treatments. While most textbooks focus on lack-of-fit testing in the context of completely randomized designs, many response surface experiments are not completely randomized and require block or split-plot structures. The analysis of data from blocked or split-plot experiments is generally based on a mixed regression model with two variance components instead of one. In this article, we present a novel approach to designing blocked and split-plot experiments which ensures that the two variance components can be efficiently estimated from pure error and guarantees a precise estimation of the response surface model. Our novel approach involves a new Bayesian compound D-optimal design criterion which pays attention to both the variance components and the fixed treatment effects. One part of the compound criterion (the part concerned with the treatment effects) is based on the response surface model of interest, while the other part (which is concerned with pure-error estimates of the variance components) is based on the full treatment model. We demonstrate that our new criterion yields split-plot designs that outperform existing designs from the literature both in terms of the precision of the pure-error estimates and the precision of the estimates of the factor effects.