Design Selection for Model Discrimination

Abstract Fractional factorial designs are investigated in terms of how well they partition the mean response space. This partitioning is fundamental to the capacity of a design to distinguish between competing models, that is, to the problem of model discrimination. A new proof of a classic result due to J. N. Srivastava, based solely on first moment properties, provides generalization while clarifying its central role in design selection. The 2-level orthogonal arrays are then examined for 16 and 20 runs. Determining the distribution of the non-isomorphic sub-OAs for each competing orthogonal array allows those that are best for discriminating across a range of models to be identified.