Group adaptive stein estimation of normal means
Sarel J. Steel · Communications in Statistics - Simulation and Computation · 1993
Stein estimators achieve substantial risk reduction only if the target towards which shrinkage is done is chosen correctly or nearly correctly. In the context of estimating normal means which may have a grouped structure, Stein estimators are studied in which the target is chosen data dependently by maximizing an estimate of the risk reduction associated with given targets. A group adaptive Stein estimator is found whose maximal risk is only slightly above the minimax risk and whose risk is automatically as low as may reasonably be expected when grouping is present.