Generalized order statistics, Bahadur representations, and sequential nonparametric fixed-width confidence intervals (Q1113589): Difference between revisions

Let \(X_ 1,X_ 2,..\). be a sequence of i.i.d. random variables with unknown distribution F and h a given real valued function on \(R^ m\) (m\(\geq 1)\). For each integer \(n\geq m\) let \(W_{n1}\leq...\leq W_{n,n(m)}\) be the ordered values of the \(h(X_{i_ 1},...,X_{i_ m})\) taken over all n(m) m-tuples \((i_ 1,...,i_ m)\) of distinct elements from \(\{\) 1,...,n\(\}\). (The W's are called ``generalized order statistics''.) Let \(H_ F\) be the distribution function of \(h(X_ 1,...,X_ m)\) and \(\xi_ p=H_ F^{-1}(p)\) its p th quantile \((0<p<1).\) A Bahadur-type representation is given for the sample quantile \(W_{n,k(n)}\), where the integer valued random variable k(n) (depending on \(X_ 1,...,X_ n)\) satisfies k(n)/n\(\to p\) a.s. (as \(n\to \infty)\) at a prescribed rate. The results are used to construct a sequential fixed width confidence interval for \(\xi_ p\) (asymptotic as the width \(d\to 0)\) of the form \((W_{N,a(N)},W_{N,b(N)})\) with suitably chosen stopping time N and sequences of integers a(n), b(n) (random since a certain nuisance parameter has to be replaced by a strongly consistent estimator). Several properties of this confidence interval are proved.