Confidence intervals with fixed proportional accuracy (Q1082764)

Let \(X_ 1,X_ 2,..\). be independent and normally distributed with unknown mean \(\theta\) and unit variance. Let \(\bar X{}_ n=(X_ 1+...+X_ n)/n\) and define estimators of \(\theta\) of the form \({\hat \theta}_ n=\bar X_ n-n^{-1} b_ n(\bar X_ n)\) where \(\{b_ n\), \(n\geq 1\}\) is a convergent sequence of absolutely continuous functions. For a proportionality function \(\Delta\), let \({\hat \Delta}_ n=\Delta (\bar X_ n)\) denote the maximum likelihood estimators of \(\Delta\) (\(\theta)\). Let \(\{g_ n\), \(n\geq 1\}\) denote a convergent sequence of continuous functions on (-\(\infty,\infty)\) for which \(g_ n>-n\) for all n and let \(c_ n=\{1+n^{-1} g(\bar X_ n)\}^{1/2}\). Then the sequential procedure \(t=t_ h\) is defined by \[ t=\inf \{n\geq m;\quad n>c^ 2_ n/h^ 2 {\hat \Delta}_ n\}. \] For the sequential procedure, let \(\gamma_ h(t:\theta)=P_{\theta}\{| {\hat \theta}_ t-\theta | \leq h{\hat \Delta}_ t^{1/2}\}\) and define the average confidence level of t under density \(\xi\) by \(\gamma_ h(t;\xi)=\int \gamma_ h(t;\theta)\xi (\theta)d\theta.\) In this paper, as in the author's paper, Ann. Stat. 14, 1049-1067 (1986; see the foregoing review, Zbl 0603.62089), the author obtained asymptotic expansions for \(\gamma_ h(t;\xi)\) and the expected sample size of this sequential procedure. The results provide a simple example of a related problem in which the estimator and the stopping time are much more highly dependent, but a weak asymptotic expansion is still possible.