Universal domination and stochastic domination: Estimation simultaneously under a broad class of loss functions (Q1065468): Difference between revisions

Let x be a multi-dimensional random vector with probability density function \(f_{\theta}(x)\), where \(\theta \in R^ b\) is unknown. For estimating \(\theta\) let the loss function be given by \(L((\theta -\delta (x))'D(\theta -\delta (x))),\) where D is a given non-negative definite matrix, \(\delta\) denotes any estimator of \(\theta\), and L(\(\cdot)\) is nondecreasing. An estimator \(\delta_ 1\) universally dominates \(\delta_ 2\) if for every \(\theta\) and every nondecreasing loss function L, \(E L(\theta - \delta_ 1(x))\leq E L(\theta -\delta_ 2(x))\) with strict inequality for some particular loss function. A (nonrandomized) estimator \(\delta_ 1\) stochastically dominates \(\delta_ 2\) if for every \(\theta\), \(L(\theta -\delta_ 1(x))\leq^{d}L(\theta -\delta_ 2(x)),\) with strict inequality for some \(\theta\), where \(Y\leq^{d}Z\) means \(P\{\) \(Y\geq c\}\leq P\{Z\geq c\}\) for every c. It is shown that the criterion of universal domination is equivalent to the criterion of stochastic domination. It is shown that when X-\(\theta\) has a p-variate t distribution \((p=1,2)\), there exists no estimator for \(\theta\) that universally dominates x. But for \(p\geq 3\), James-Stein positive part estimators that universally dominate x, are specified. It is also shown that when x is a p-variate normal random variable, then no James-Stein positive part estimator universally dominates x.