Erratum to: Estimating extreme quantiles under random truncation

Laurent Gardès, Gilles StupflerTest 2015 · DOI · OpenAlex

Introduction

In the field of mathematical statistics, particularly within extreme value theory, precision is paramount. This document serves as a formal erratum for the research paper "Estimating extreme quantiles under random truncation," originally published in the journal TEST in 2014. Such corrections are essential to maintain the integrity of statistical estimators and ensure the reproducibility of results in the scientific literature.

Background: Random Truncation and Extreme Quantiles

The original research addresses the estimation of extreme quantiles—values representing rare, high-impact events in the tails of a distribution—when data is subject to random truncation. Random truncation occurs when an observation $X$ is only recorded if it satisfies a condition relative to another random variable $T$.

A common example is found in insurance: a claim $X$ is only observed if it exceeds a deductible $T$. To accurately estimate the risk of extreme events (e.g., a "once-in-a-century" flood or a massive financial loss), statisticians must account for the data points missing due to this truncation threshold. The original paper proposed a method for this estimation, which relies on the accurate reconstruction of the distribution function from observed data.

The Mathematical Correction

The erratum identifies and corrects a typographical or definitional error on the third page of the original publication. The error concerns the fundamental relationship used to reconstruct the cumulative distribution function (CDF) from the estimated cumulative hazard function.

In the original text, the relationship between the estimated cumulative hazard function, denoted as $\hat{\Lambda}_N^F(y)$, and the estimated cumulative distribution function, $\hat{F}_N(y)$, was incorrectly defined. The corrected identity is:

$$\hat{F}_N(y) = 1 - \exp\left(-\hat{\Lambda}_N^F(y)\right)$$

This formula represents a standard identity in survival analysis and reliability engineering. It establishes that the survival function (the probability that a variable exceeds a certain value, $1 - F$) is the exponential of the negative cumulative hazard.

Implications for Statistical Modeling

In extreme value statistics, small errors in foundational formulas can lead to significant inaccuracies when extrapolating to the far tails of a distribution. By clarifying the link between the hazard function and the CDF, this erratum ensures that the "tail index" and subsequent quantile estimates derived using the authors' methodology remain mathematically sound.

The use of $\hat{\Lambda}_N^F(y)$ typically implies the application of a non-parametric approach, such as a variation of the Lynden-Bell estimator, which is specifically designed for truncated data. This correction ensures that researchers implementing these methods achieve the intended theoretical precision required for modeling extreme risks. While the erratum does not revisit the datasets or specific limitations discussed in the 2014 study, it provides the necessary mathematical foundation for their correct application.

Summary updated

BibTeX
@inproceedings{gardes2015erratum,
  author = {Laurent Gardès and Gilles Stupfler},
  title = {{Erratum to: Estimating extreme quantiles under random truncation}},
  booktitle = {Test},
  year = {2015},
  doi = {10.1007/s11749-015-0437-3},
}
Figure