Calibrated bootstrap and saddlepoint approximations of finite population \(L\)-statistics (Q2010115)

Consider a finite population of $N$ units $U= (U_1, U_2,\dots,U_N)$. Let $X$ be a study variable taking values $X_i$ on $U_i$. For a simple random sample of size \(n\) drawn without replacement from $U$ define the \(L\)-statistic, $L =L_n(X)$ as the weighted average of the order statistics $(x_{(1)}, x_{(2)},\dots,x_{(n)})$, based on the sample. The problem considered here is the estimation of the distribution function (d.f.) of the studentized \(L\)-statistic in which the jackknife estimate of the variance of $L$ is used. It is seen that normal approximations to the data from small or moderate samples are not accurate enough. An improvement is to obtain the one-term Edgeworth expansion. Further, using the jackknife estimators of the parameters, empirical Edgeworth expansions (EEE) have been proposed and several variants of EEE followed in the literature. In this paper, the authors suggest two alternative methods to approximate the desired d.f. Using information on all the units of the population for an auxiliary variable $Z$ correlated with the study variate $X$, the first method gives a modification of nonparametric bootstrap approximation. The errors in this calibrated nonparametric bootstrap approximation to d.f. are found to be similar to those of EEE's. Next, an empirical saddle point approximation is proposed. The technique adapted is linearization of the \(L\)-statistic and then application of saddle point approximation to the d.f. of the linear part. Two such approximations based on the bootstrap, with and without auxiliary information are considered. Simulation results lead to the conclusion that the proposed approximations fare well when the variables $Z$ and $X$ are highly correlated. Numerical examples indicate that the proposed approximations adapt better to estimation of extreme quantiles of the distribution than the EEE with calibrated parameters.