ENBIS: European Network for Business and Industrial Statistics

Setting up experiments for mean and variance model estimation

22 September 2009, 14:25 – 14:45

Abstract

Submitted by

Peter Goos

Authors

Peter Goos, Marta Emmett, Eleanor C. Stillman

Abstract

Industrial researchers often want to find settings for the experimental factors that yield the desired response while simultaneously minimizing the variance about the response. This can be achieved by using a dual response approach, in which a regression model is fitted for the variance of the response as well as for its mean. These two regression models are referred to as the dispersion and location models, respectively. The dispersion model is based on either the residuals from the location model or sample variances. As dispersion models based on residuals are sensitive to miss-specification of the location model, the use of
sample variances is often appealing. How to construct informative experimental designs for the joint estimation of mean and variance functions, based on sample variances, is, however, an open research
question. We show how this problem can be tackled using a composite optimal experimental design criterion tailored to models involving sample variances.

Return to programme