Optimal simultaneous confidence bounds (Q797940): Difference between revisions

Let the real random variable \(y=y(x)\), \(x\in X\subset R^ k\), be modeled as \(y=m(x)+e\), in which \(Ee=0\), Var e\(=\sigma^ 2\), unknown, and m is an unknown member of a given family of functions on X. It is assumed that as a result of observations on y at \(x_ 1,...,x_ n\in X\) there is available an estimator \^m(x) of m(x), \(x\in X\), and an independent estimator S of \(\sigma\), in such a way that \(U_ x\equiv (m(x)-\hat m(x))S^{-1}\) has a known distribution for every \(x\in X\). This applies for instance to the normal multiple linear regression model, where each \(U_ x\) has essentially a t-distribution. The object is to construct simultaneous confidence intervals for all m(x) of the form \[ (*)\quad\hat m(x)+Sp_ 1(x)\leq m(x)\leq\hat m(x)+Sp_ 2(x),\quad x\in X. \] For given \(0<\alpha <1\), the usual coverage probability criterion requires (*) to hold simultaneously for all \(x\in X\) with probability at least 1-\(\alpha\). Since this is an inconvenient restriction from the point of view of deriving optimal confidence bands the author proposes a new criterion that is easier to handle. Let \(\mu\) be a measure on X and define the expected coverage measure (ECM) as the expected \(\mu\) -measure of the x-set on which (*) holds. Then require EC\(M\geq 1-\alpha\), and subject to this condition minimize a given functional \(L(p_ 2-p_ 1)\). It is assumed that L has the form \(L(h)=\int\ell_ x(h(x))\mu (dx) (\ell_ x(\cdot)\geq 0\) and twice continuously differentiable), and that \(U_ x\) has a unimodal density \(f_ x\) with mode \(\theta\) (x). Under some additional conditions, the most important being \(\ell '\!_ x\geq 0\) and \(\ell ''\!_ x\geq 0\), an optimal solution has the following form: choose \(c>0\), then for \(\mu\) - a.e. \(x\in X\) take the \(p_ i(x)\) so that \(f_ x(p_ i(x))=c\ell '\!_ x(p_ 2(x)-p_ 1(x)), i=1,2\), if this is possible; if not possible then take \(p_ i(x)=\theta (x)\), \(i=1,2\). Consequences for the linear regression model are explored. Analogous results for one-sided confidence bounds are also stated.