On empirical Bayes estimation of multivariate regression coefficient (Q2469012)

Summary: We investigate the empirical Bayes estimation problem of multivariate regression coefficients under the squared error loss function. In particular, we consider the regression model \(Y=X\beta+\varepsilon\), where \(Y\) is an \(m\)-vector of observations, \(X\) is a known \(m\times k\) matrix, \(\beta\) is an unknown \(k\)-vector, and \(\varepsilon\) is an \(m\)-vector of unobservable random variables. The problem is squared error loss estimation of \(\beta\) based on some ``previous'' data \(Y_1,\dots,Y_n\) as well as the ``current'' data vector \(Y\) when \(\beta\) is distributed according to some unknown distribution \(G\), where \(Y_i\) satisfies \(Y_i=X\beta_i+\varepsilon_i\), \(i=1,\dots,n\). We construct a new empirical Bayes estimator of \(\beta\) when \(\varepsilon_i\sim N(0,\sigma^2I_m)\), \(i=1,\dots,n\). The performance of the proposed empirical Bayes estimator is measured using the mean squared error. The rates of convergence of the mean squared error are obtained.