Asymptotic distributions in the testing and estimation of the missing- data multivariate normal linear patterned mean and correlation matrix (Q1067715)

The paper considers a random p-vector X with the N(\(\mu\),\(\Sigma)\) distribution. The structures on \(\mu\) and \(\Sigma\) are taken to be linear with \(\mu =\sum^{r}_{j=1}Z_ j\beta_ j\) where \(Z_ 1,Z_ 2,...,Z_ r\) are known vectors and \(\Sigma =DPD'\) where \(D=\sum^{t}_{j=1}\sigma_ j\Lambda_ j\), \(\Lambda_ 1,...,\Lambda_ t\) are known diagonal matrices and \(P=P^*+\sum^{m}_{k=1}\rho_ kG_ k\) with \(P^*\) a known symmetric matrix with ones along the diagonal, \(G_ 1,...,G_ m\) known symmetric matrices with zero along the diagonal. The problem attacked is threefold, consisting of Votaw's hypotheses: (i) \(H_{01}:\mu_ 1=\mu_ 2=...=\mu_ k\), \(\Sigma_ 1=\Sigma_ 2=...\Sigma_ j\); against all alternatives; (ii) \(H_{02}:\Sigma_ 1=\Sigma_ 2=...=\Sigma_ k\) against all alternatives; and (iii) \(H_{03}:\mu_ 1=\mu_ 2=...\mu_ k\), \(\Sigma_ 1=...\Sigma_ k\) against \(H_ A:\Sigma_ 1=\Sigma_ 2=...=\Sigma_ k.\) The author's earlier results for complete data in this problem [Ann. Stat. 7, 823-837 (1979; Zbl 0413.62032) and ibid. 11, 947-958 (1983; Zbl 0524.62050)] are utilized to derive the results when some of the data are missing. The main outcome of the study here is the derivation of the maximum likelihood equations and algorithms that solve the equations iteratively. Asymptotic distributions of the ML estimators and the likelihood ratio statistic (LRS) are also derived. It should be mentioned here that the results given in this paper are restricted to linearly patterned correlation matrices. This is evident from the covariance structure and the knowledge of the scale factor therein, as given above.