Multivariate reduced-rank regression (Q1271110)

Consider the general multivariate linear model \[ Y_k= CX_k+ \varepsilon_k, \quad k=1,\dots, T, \tag{1} \] where \(Y_k\) is an \(m\times 1\) vector of observed response variables, \(X_k\) is an \(n\times 1\) vector of observed predictor variables, \(C\) is the \(m\times n\) repression coefficient matrix to be estimated, and \(\varepsilon_k\) is an \(m\times 1\) vector of centered random errors with \(\text{Cov} (\varepsilon_k)= \Sigma_{\varepsilon\varepsilon}\) being a nonsingular matrix. The \(\varepsilon_k\) are assumed to be independent for different \(k\). This book is devoted to the theory and applications of the multivariate reduced-rank regression (RRR), which is characterized by the condition \[ \text{rank}(C)= r<\min(m,n). \tag{2} \] The aim of this assumption is to reduce the number of parameters in the model (1), which is important in various practical situations. The book covers three areas: regression modeling for general multi-response data, for multivariate time series data, and for longitudinal data. Chapter 1 gives basic results for the classical model (1). Chapter 2 contains important developments and results for the analysis of model (1), (2). The regression coefficient matrix is expressed as \(C=AB\) where \(A\) is an \(m\times r\) matrix, \(B\) is an \(r\times n\) matrix, and both have rank \(r\). For the case of normal errors, the MLE of these matrices are constructed and asymptotic normality of the MLE is proved. The likelihood ratio test statistic is built for identification of \(\text{rank}(C)\). These methods are extended in Chapter 3 to the RRR with two sets of regressors. Chapters 4 and 5 focus on modeling of multivariate time series data. They deal with autocorrelated errors, a situation that is widely used in econometrics. The nested RRR methods are applied to multiple autoregressive time series. It is shown how unit-root nonstationarity can be investigated through RRR methods and associated canonical correlation methods. Chapter 6 deals with the analysis of balanced longitudinal data, especially through use of the popular growth curve or generalized multivariate analysis of variance (GMANOVA) model. A collection of flexible models is obtained that can be useful for more parsimonious modeling of growth curve data. In Chapter 7, the seemingly unrelated regression (SUR) equations models are considered. These models are more general than the model (1) because they allow the different response variables \(y_{ik}\) to have different predictor variables \(X_{ik}= (x_{i1k},\dots, x_{ink})'\) for different \(i\), so that \(y_{ik}= X_{ik}' C_{(i)}+\varepsilon_{ik}\), \(i=1,\dots, m\), and the errors \(\varepsilon_{ik}\) are contemporaneously correlated across the different response variables. In Chapter 8, the use of the RRR methods in financial economics is discussed. Here several applications of the RRR methods arise in a fairly natural way from economic theories. The last chapter highlights some other techniques associated with recent developments in multivariate regression modeling. Each chapter, besides of theoretical results, contains the details on computational procedures, illustrated with numerical examples from biochemistry, marketing, and finance. The connection of the RRR methods to multivariate analysis of variance (MANOVA), discriminant analysis, principal components, canonical correlation analysis, and errors-in-variables models is also discussed. This monograph is a textbook and a handbook for both practitioners and researchers who deal with moderate or high-dimensional multivariate data. It is not an easy reading; the reader must be familiar with baisc matrix theory, with multivariate methods and univariate linear regression modeling. It will be useful for the advanced graduate students in statistics, econometrics, business and engineering.