Optimal Designs For Two Non-Interactive Treatment
Ralph A. Bradley,A. Hedayat
DOI: https://doi.org/10.1080/00401706.1977.10489498
1977-02-01
Technometrics
Abstract:Consider two treatments and a combination of them. Given n available experimental units, the problem is to determine the number of units n, to be assigned to the first treatment, n 2 to the second, and n, to the combination, n = Σ i , n i , n i ≥ 0, i = 1, 2, 3, subject to suitable optimality criteria. This study was motivated by an example of a weighing design with n = 18. n 1 = 2. n 2 = 9, n s = 7 given by Box and Tiao [2, Section 2.7.3]. The allocation seemed far from optimal leading to the present investigation. Four models are developed for two non-interactive treatments and the combination of them. They may be described as 1. A weighing model, 2. A sum of independent random variables model, 3. A mixtures model, and 4. A factorial model. Three optimality criteria are considered: 1. D-optimality, 2. A-optimality and 3. minimization of the variance of the difference of estimators of the two treatment means. It is shown for all criteria for Models 2, 3, 4 and for Criterion 3 for Model I that n s = 0, n 1 = n 2 = n/2 if n is even, n 1 = (n + 1)/2. n 2 = (n – 1)/2 if n is odd. But rather different optimal designs are obtained for Criteria 1 and 2 and Model I, different for the two criteria and different also from the design of the Box-Tiao example. The results presented here are of interest for two reasons. They have intrinsic interest showing that the choice of model and optimality criterion may lead to an intuitively unexpected design in a simple experimental situation. They have instructional interest providing elementary illustrations of modern design concepts that may be used as class examples or problems for solution.
statistics & probability