Asymptotic expansions in Anscombe's theorem for repeated significance tests and estimation after sequential testing (Q1820535)

Let \(x_ 1,x_ 2,..\). be independent random variables with a common normal distribution N(\(\theta\),1). There are two hypotheses for the value of \(\theta\), the null hypothesis \(H_ 0: \theta =0\) and the alternative \(H_ 1: \theta \neq 0\). Define the stopping time \(\tau =\tau_{a,c}=\inf \{n\geq 1:| S_ n| \geq c_ n\}\) where \(S_ n=x_ 1+...+x_ n\), \(c_ n=\sqrt{2a(n+c)}\), \(n\geq 1\) for some fixed \(a>0\) and \(c\geq 0\). Then for a given positive integer \(N_ 0\), the repeated significance test for \(\theta\) rejects \(H_ 0\) in favor of \(H_ 1\) iff \(\tau \leq N_ 0\). Now if \(T=\min \{\tau,N_ 0\}\), then \(x_ 1,...,x_ T\) are the available data after the repeated significance test. Thus we come to the main problem treated in this paper: to construct confidence intervals for \(\theta\) based on the observations \(x_ 1,...,x_ T\). The author presents the solution to this problem deriving firstly a nice refinement of a result of F. Anscombe concerning the asymptotic normality of the ratio \((S_{\tau}-\theta \tau)/\sqrt{\tau}\) as \(a\to \infty\). Then he has constructed confidence intervals for \(\theta\) with a prescribed accuracy. Useful numerical results are given. Some related topics are also discussed.