Generalized Bayesian D criterion for single-stratum and multistratum designs

TL;DR

This paper extends the Bayesian D criterion to multistratum experimental designs, enabling optimal design selection in complex, multi-layered experimental structures beyond the original single-stratum framework.

Contribution

The paper introduces a generalized Bayesian D criterion that applies to both single- and multistratum designs, broadening the scope of Bayesian optimal design methods.

Findings

Successfully extended Bayesian D criterion to multistratum designs

Provides a unified approach for single- and multistratum experimental design selection

Enhances design efficiency in complex industrial experiments

Abstract

DuMouchel and Jones (1994) proposed the Bayesian D criterion by modifying the D-optimality approach to reduce dependence of the selected design on an assumed model. This criterion has been applied to select various single-stratum designs for completely randomized experiments when the number of effects is greater than the sample size. In many industrial experiments, complete randomization is sometimes expensive or infeasible and, hence, designs used for the experiments often have multistratum structures. However, the original Bayesian D criterion was developed under the framework of single-stratum structures and cannot be applied to select multistratum designs. In this paper, we study how to extend the Bayesian approach for more complicated experiments and develop the generalized Bayesian D criterion, which generalizes the original Bayesian D criterion and can be applied to select single-stratum and multistratum designs for various experiments when the number of effects is greater than the rank of the model matrix.