The vast majority of response surface methods used in practice to develop, assess, and compare experimental designs focus on variance properties of designs. Because response surface models only approximate the true unknown relationships, models are subject to bias errors as well as variance errors. Beginning with the seminal paper of Box and Draper (1959) and over the subsequent 50 years, methods that consider bias and mean-squared-error (variance and bias) properties of designs have been presented in the literature. However, these methods are not widely implemented in software and are not routinely used to develop, assess, and compare experimental designs in practice. Methods for developing, assessing, and comparing response surface designs that account for variance properties are reviewed. Brief synopses of publications that consider bias or mean-squared-error properties are provided. The difficulties and approaches for addressing bias properties of designs are summarized. Perspectives on experimental design methods that account for bias and/or variance properties and on future needs are presented.
Revised: December 30, 2010 |
Published: December 1, 2010
Citation
Piepel G.F. 2010.Perspectives on Prediction Variance and Bias in Developing, Assessing, and Comparing Experimental Designs.ASQ Statistics Division Newsletter 28, no. 1:16-25.PNNL-SA-69664.