Decomposable three-way layouts (Q689403)

The authors introduce an apparently new and important subclass of three- way layouts or two-way elimination of heterogeneity designs specified by the full information matrices \(S_{3.12}\) being decomposable in the following way: \[ S_{3.12}=\xi_ 1S_{3.1}+\xi_ 2S_{3.2}-\xi_ 0S_{3.0};\qquad \xi_ i\geq 0, \quad 0\leq i\leq 2; \] the information matrices \(S_{3.1}\), \(S_{3.2}\) and \(S_{3.0}\) corresponding, respectively, to the treatment-row subdesign, the treatment-column subdesign, and the model in which both rows and columns are ignored. The special case when \(\xi_ 1=\xi_ 2=\xi_ 0=1\) comprises designs for which the study of relationships between properties of the three-way design itself and the corresponding properties of its treatment-row and treatment-column subdesigns is simplified. As the study of block designs is more straightforward than that of three-way designs, it can be seen that the level of difficulty is here reduced from one three-way level to two-way level designs. The authors believe that this decomposability property identifies the most general subclass of three-way layouts or two-way elimination of heterogeneity designs for which certain results on connectedness, orthogonality and balance hold. A problem which seems not yet to have been completely solved concerns the relationship between connectedness for treatments in a two-way elimination of heterogeneity design and connectedness in its treatment- row and treatment-column subdesigns. \textit{D. Raghavarao} and \textit{W. T. Federer} [Ann. Stat. 3, 730-735 (1975; Zbl 0303.62063)] showed that if a two-way elimination of heterogeneity design is connected for treatments, then the treatment-row and treatment-column subdesigns are also connected (the row-column subdesign need not, however, be connected). However, the converse of this statement is not generally true as was shown by \textit{K. R. Shah} and \textit{C. G. Khatri} [Commun. Stat. 2. 571- 573 (1973; Zbl 0271.62097)]. Raghavarao and Federer (1975) showed that for equireplicate row-column designs satisfying the condition \(N_{13}N_{32}=ke^{(n)}e^{(n)'}\) connectedness of the treatment- row and treatment-column subdesigns leads to treatment-connectedness. This result was first strengthened by \textit{L. L. Sia} [Commun. Stat., Theory Methods A6, 1165-1169 (1977; Zbl 0394.62061)], who showed that when \(S_{3.1}\) and \(S_{3.2}\) commute in an equireplicate row-column design (or equivalently when \(N_{31}N_{13}\) and \(N_{32}N_{23}\) commute and \(N_{12}=e^{(r)}e^{(c)'})\), then connectedness of the treatment-row and treatment-column subdesigns implies treatment- connectedness if and only if the sums of the eigenvalues of \(S_{3.1}\) and \(S_{3.2}\) corresponding to the same eigenvectors are different from \(k\), the number of replications of each treatment. The commutativity of \(S_{3.1}\) and \(S_{3.2}\) by itself is not sufficient for this result to still hold, as was shown by Shah and Khatri (1973), where \(S_{3.1}\) and \(S_{3.2}\) commute. The equireplicate condition was relaxed by \textit{J. K. Baksalary} and \textit{R. Kala} [Biom. J. 22, 105-109 (1980; Zbl 0462.62063)], where the more general commutativity condition \(A_{3.1}A_{3.2}=A_{3.2}A_{3.1}\) was considered. They gave an extension for two-way elimination of heterogeneity designs with equal row sizes and equal column sizes, satisfying the generalized decomposability property. The proof follows that of Baksalary and Kala.