A general theory of asymptotic consistency for subset selection with applications (Q761723)

The paper deals with the multiple decision problem of selecting a random nonempty subset from the set of k populations \(P_ 1,P_ 2,...,P_ k\) characterized by parameters \(\theta_ 1,\theta_ 2,...,\theta_ k\), respectively. \(X^ n_ i\) denotes the estimate of \(\theta_ i\) based on n observations, \(i=1,2,...,k\). \(\sigma\) is a nuisance parameter estimated by \(S^ n\). It is assumed that the probability model of \((X^ n,S^ n)\) is invariant under the group of permutations on \(\{\) 1,2,...,k\(\}\). A subset selection procedure is given by \(\delta_ n(a| X, S)=\Pr \{decision\) a \(|\) X, \(S\}\) where the decision space is \(A=\{a\subset \{1,2,...,k\}\}\). The decision a is interpreted as selecting the populations \(P_ i\), \(i\in a\). The loss function is assumed to be permutation-invariant on \(\{\) 1,2,...,k\(\}\). The consistency of the procedure \(\delta_ n\) is defined by the property that the risk of \(\delta_ n\) tends to minimum risk as n tends to infinity. Six types of loss functions are considered and conditions are given for pointwise and uniform consistency of the decision procedures. As an application, the problem of selecting the set containing the best normal population is discussed. It is shown that in the class of procedures given by Seal, only the Gupta procedure can be consistent for the loss functions. Conditions for consistency of the Gupta procedure for the loss functions are given. Some other well known procedures are also examined for consistency with respect to the loss functions.