Sequential shrinkage estimation (Q1099912)

Let \(X_ 1,X_ 2,\ldots\) \((p\times 1)\) be i.i.d. \(N(\theta,\sigma^2V)\), with \(\theta\), \(\sigma\) unknown and \(V\) a known \(p\times p\) positive definite matrix. If it is decided to stop at stage \(n\) and \(\theta\) is estimated by \(\delta_ n=\delta_ n(X_ 1,\ldots,X_ n)\), then the loss will be \(L(\theta,\delta_ n)'Q(\delta_ n-\theta)+cn\), with \(Q\) a given positive definite matrix and \(c>0\) the cost per observation. A stopping rule is proposed along the ideas in \textit{Y. S. Chow} and \textit{H. Robbins}, Ann. Math. Stat. 36, 457--462 (1965; Zbl 0142.15601). It is shown that a James-Stein-type shrinkage estimator \(\delta_ n\) has a smaller risk than the conventional estimator \(\bar X_ n\). Asymptotic formulas (as \(c\to 0)\) for both estimators are developed and numerical results from computer simulation are presented.