Lower quantile estimation within an artificially censored framework

Abstract

Quantile estimation is a vital aspect of statistical analyses in a variety of fields. For example, lower quantile estimation is crucial to ensure the safety and reliability of wood-built structures. Various statistical tech-niques, which include parametric, non-parametric and mixture modelling are available for estimation of lower quantiles. An intuitive approach would be to consider models that ˝t the tail of the sample instead of the entire range. Quantiles of interest can be estimated by arti˝cially censoring observations beyond a chosen threshold. The choice of threshold is crucial to ensure e°cient and unbiased quantile estimates, and usually the 10th empirical percentile is chosen as the threshold. [16] proposes a bootstrap approach in order to ob-tain a better threshold for the censored Weibull MLE, however, this approach is computationally expensive. A new threshold selection technique is proposed that makes use of a standardised-weighted adjusted trun-cated Kolmogorov-Smirnov test (SWAKS-MLE). The SWAKS-MLE outperforms in the bootstrap threshold censored Weibull MLE method, in addition to being vastly less computationally intensive.