On the asymptotic properties of a kernel type quantile estimator from censored samples (Q1085915)

Let i.i.d. survival times \(X^ 0_ 1,...,X^ 0_ n\) be given which are censored on the right at \(U_ 1,..,U_ n\) (i.i.d. r.v's again, independent from \((X^ 0_ j))\) resulting in the data set \(\{\min (X^ 0_ j,U_ j),\Delta_ j\}^ n_{j=1}\), where \(\Delta_ j=1\) if \(X_ j\leq U_ j\) and 0 elsewhere. Based on this data set the distribution F of \(X^ 0_ 1\) can be estimated by the product-limit estimator \(\hat F_ n\), introduced by \textit{E. L. Kaplan} and \textit{P. Meier} [J. Am. Stat. Assoc. 53, 457-481 (1958; Zbl 0089.148)]. The authors estimate the quantile function \(F^{-1}(p)=\inf \{t; F(t)\geq p\}\) by kernel smoothed versions of \(\hat F_ n^{-1}\), that is by: \[ Q_ n(p)=h_ n^{-1}\int^{1}_{0}\hat F_ n^{-1}(t)\quad K((t- p)/h_ n)dt \] or by an approximating sum \(Q^*_ n\) of \(Q_ n\). As usual K is a probability density and the bandwith \(h_ n\) is a null- sequence. It is proven that \(\sqrt{n}(Q_ n(p)-F^{-1}(p))\) is asymptotically normal under certain conditions on F, K and \(h_ n\), and that \(Q_ n\) and \(Q^*_ n\) are asymptotically uniformly equivalent w.r. to the mean-square-error.