Subjective hierarchical Bayes estimation of a multivariate normal mean: On the frequentist interface (Q2640275)

This long paper deals with the interface of two approaches to shrinkage estimation of the multivariate normal mean, such as the hierarchical or empirical Bayes approach and the minimax approach. The hierarchical Bayes setup considered here is quite general, allowing the use of second-stage prior distributions. First-stage priors are used as usual to model suspected relations among the means. There are four sections in all. Section 1 gives an introduction with the examples 1 and 2 from the first author's book, ``Statistical decision theory and Bayesian analysis''. 2n ed. (1985; Zbl 0572.62008). These examples emphasize the richness of the structures that can be modelled within the hierarchical Bayesian framework. Section 2 takes up the hierarchical Bayes estimator with two-stage priors. Examples 3 and 4 are given to indicate the diverse possibilities for the choice of the second-stage prior. This section also gives three lemmas which are all proved. These lemmas keep track of when the Bayes estimator actually exists. Section 3 deals with the estimated accuracy and loss and also the question, what error measures are to be associated with the hierarchical Bayes estimator? Two types of measures are considered, such as the Bayesian posterior measure and an unbiased estimate of loss and accuracy. Both these measures are compared with. Two numerical examples and a subsequent discussion are also given. The last section 4 gives minimaxity of the hierarchical Bayes estimator. Risks and various estimates of risks or losses of the hierarchical Bayes estimator are considered. An interesting minimax result is established, which shows that minimaxity holds for any second-stage prior on the mean. The authors also state that the numerical verification of minimaxity is far simpler than the analytic verification. Finally, Appendix I gives a generalization of examples 3 and 4, while Appendices II and III prove two lemmas related to sections 2 and 3.