Reduced rank regression. With applications to quantitative structure-activity relationships (Q1901120)

This book considers the multivariate regression when the regression coefficient matrix has reduced rank. The reduced rank regression model can be written in matrix form as \(Y=1_n \mu'+ XM+E\) where \(M\) is a \(p\) by \(q\) matrix of coefficients of fixed rank \(r< \min (p,q)\), \(X\) is an \(n\) by \(p\) known fixed matrix of predictors, \(Y\) is an \(n\) by \(q\) random matrix of responses and \(E\) is an \(n\) by \(q\) random matrix of errors. After having considered several parametrizations of \(M\) the author recommends the following parametrization: \(M= AB'\), \(A'X'X A=I_r\), \(B\) arbitrary, where \(A\) is a \(p\) by \(r\) matrix, \(B\) is a \(q\) by \(r\) matrix, both \(A\) and \(B\) have rank \(r\). Under normality assumptions, maximum likelihood estimators of the unknown parameters are obtained for the following situations: (a) known error covariance matrix; (b) error covariance matrix proportional to the identity matrix; (c) unstructured error covariance matrix and (d) diagonal error covariance matrix. Prediction criteria for the rank selection and variable selection of reduced rank regression models are also discussed. In a separate chapter, a Bayesian analysis of reduced rank regression is given. Since an exact analytical derivation of the posterior distribution of the parameters is intractable, Gibbs sampling of the posterior distribution and Gibbs sampling of the predictive distribution are suggested. A numerical example in quantitative structure -- activity relationships, i.e., relationships between chemical structure and biological activity, is used to illustrate the methodology developed for reduced rank regression.