Extended optimality of sequential probability ratio tests (Q800674): Difference between revisions

Let \(P_ 0\) and \(P_ 1\) be probability measures on (\(\Omega\),F). Let \(\{F_ t\}\) be an increasing family of sub-\(\sigma\) algebras of F, for \(t\in Z\) or \(t\in R\). Let \(D_ T\) be the likelihood ratio \(dP_ 1/dP_ 0\) at a stopping time T. It is shown that if \(\alpha\) and \(\beta\) are the error probabilities of a sequential probability ratio test which stops on its boundaries and a competing test has error probabilities \(\alpha\) '\(\leq \alpha\), \(\beta\) '\(\leq \beta\), and \(P_ 1(T'<\infty)=1\), then \(E_ 0(g(D_{T'}))\geq E_ 0(g(D_ T))\) for any non-negative, convex g such that \(E_ 0(g(D_ t))<\infty\) for all t and g(x)/x\(\to \infty\) for \(x\to \infty.\) It can be shown that in the case where the SPRT hits the boundary this implies the weak admissibility of the SPRT in the general model in the sense of the reviewer, \textit{B. K. Ghosh} and \textit{G. Simons} [ibid. 4, 237-251 (1976; Zbl 0323.62053)]. The boundary condition is often satisfied in continuous time and in that case it extends their result. Moreover, when the increasing component of the Doob-Meyer decomposition of \(g(D_ t)\) is non-random this gives the stronger conclusion that there do not exist alternative tests with \(P_ 0(T'\leq t)\geq P_ 0(T\leq t)\) for all t with strictly better error probabilities. In the case where \(P_ 0\) and \(P_ 1\) represent the measures for the Brownian motion and the Brownian motion with constant drift, respectively, the Wald-Wolfowitz optimality of the SPRT follows as well.