Missing data imputation using the multivariate \(t\) distribution (Q1893357)

This paper presents monotone data augmentation (MDA) using the multivariate \({\mathbf t}\) distribution. The description of monotone patterns and the associated multivariate \({\mathbf t}\) model are given in Section 2. We shall see that a multivariate \({\mathbf t}\) distributed data set consisting of \(n\) observations of \(p\) variables can be represented as a multivariate normal distributed data set with a common mean vector and different covariance matrices that are proportional to each other. The proportionality coefficents, called weights, are independently and identically distributed with \(\Gamma (\nu/2, \nu/2)\), where \(\nu\) is called the degrees of freedom of the \({\mathbf t}\) distribution. Section 3 extends our former results in ibid. 46, No. 2, 198-206 (1993; Zbl 0778.62004), to the case where the distributions of the observations are normal with a common mean vector and different covariance matrices that are proportional to each other with known proportionality coefficients. Section 4 presents MDA using the \({\mathbf t}\) distribution with known and unknown weights, with monotone and nonmonotone missing data, and with known and unknown degrees of freedom. Two numerical examples are given in Section 5 to illustrate the methodology, to compare results obtained using the multivariate \({\mathbf t}\) distribution with results obtained using the normal distribution, and to compare the rate of convergence of MDA with the rate of convergence of DA. By the term DA, here we mean the data augmentation algorithm applied to the whole rectangular data set rather than a created monotone pattern. The proof of the main results in Section 3 is given in the Appendix.