scientific article; zbMATH DE number 4133308

From MaRDI portal
Publication:3034671





[For the entire collection see Zbl 0638.00032.] Let \(X_ 1,...,X_ k\) be independent normal random variables with unknown means \(\theta_ 1,...,\theta_ k\) and known variances \(\sigma^ 2_ 1,...,\sigma^ 2_ k\), respectively. It is of interest to choose \(X_ i\) associated with \(\theta_{\max}\). This paper considers the properties of the natural selection rule \(d^ N\), which chooses \(X_{\max}.\) First, using the 0-1 loss function, it is shown that \(d^ N\) is minimax iff \(\sigma^ 2_ 1=...=\sigma^ 2_ k\). If \(\theta_ 1,...,\theta_ k\) are sufficiently close, and if \(\sigma^ 2_{(k)}\) (associated with \(\theta_{\max})\) is small in comparison to \(\sigma^ 2_ i\) for \(i\neq (k)\) then \(d^ N\) performs worse than the no data rule. However, with respect to the average risk function (where the average is taken over all k! permutations of the \(\theta\)-vector), \(d^ N\) is minimax, although it is not the only rule having this property. Since the minimax principle does not lead to reasonable alternatives to \(d^ N\), Bayes rules are derived under exchangeable priors. It is seen that \(d^ N\) is an extended Bayes rule.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3034671)