The Ryszard Zielinski's works on nonparametric quantile estimators and their use in robust statistics

Tomasz Rychlik. "The Ryszard Zielinski's works on nonparametric quantile estimators and their use in robust statistics." Mathematica Applicanda 40.2 (2012): null. <http://eudml.org/doc/293182>.

@article{TomaszRychlik2012,
abstract = {This is a survey paper describing achievements of professor Ryszard Zieliński in the subject of nonparametric estimation of population quantiles based on samples of fixed size, and applications of the quantile estimators in the robust estimation of location parameter. Zielinski assumed that a finite sequence of independent identically distributed random variables X1, . . . ,Xn is observed, and their common distribution function F belongs to the family F of continuous and strictly increasing distribution functions. He considered the family T of randomized estimators XJ:n which are single order statistics based on X1, . . . ,Xn with a randomly determined number J. The random variable J is independent of the sample and has an arbitrary distribution on the numbers 1, . . . , n. It was proved that T is the maximal class of estimators which are functions of the complete and sufficient statistic (X1:n, . . . ,Xn:n), and are equivariant with respect to the strictly increasing transformations, i.e., satisfy T(φ(X1:n), . . . ,φ(Xn:n)) = φ(T(X1:n, . . . ,Xn:n)) for arbitrary strictly increasing φ. A number of examples showed that the estimators that do not belong to T are very inaccurate for some F€F.   For comparing estimators, there were used various accuracy criteria based on the difference F(T) - q, where 0 < q < 1 is the quantile order. They are invariant with respect to the strictly increasing transformations. Optimal estimators with respect to the mean absolute loss E|F(T)-q|, mean quadratic loss E(F(T)-q)2, expected LINEX loss E[exp(a[F(T)-q])-a[F(T)-q]-1], a≠0, and Pitman closeness measure were explicitly determined. Further, the best estimators in narrower classes of median-unbiased estimators U(q) = \{T€T : med(T, F) = F-1(q)\},  (where med(T, F) stands for the median of the distribution of estimator T when the parent distribution function is F), and F-unbiased estimators V(q) = \{T € T : EF(T) = q\} of quantiles F-1 (q), 0 < q < 1, are determined for some accuracy criteria. Also, random confidence intervals for F-1(q), F€F, of the form [XI:n,XJ:n] on a fixed confidence level 0 <  < 1, i.e. satisfying P(XI:n ≤F-1(q) ≤XJ:n)≥γ,  F € F, , and minimizing E(J - I), are described. Median-unbiased estimators of quantiles were applied by Zielinski in the robust estimation of location parameter. For the i.i.d. sample X1, . . . ,Xn from the location model Fμ(x) = F(x - μ), where μ€R and F is a known unimodal distribution function, and the ε-contamination of the model Z(μ) = \{G = (1 -ε)Fμ +εH : H - arbitrary distribution function\} for some fixed 0 <ε< 1/2 , the most robust translation equivariant estimator with respect to the median oscillation criterion bn(T, μ) = supG1,G2€Z(μ) |med(T,G1) - med(T,G2)| has the form XJ:n - F-1(q*), XJ:n  €U(q*). Number q*  is chosen so to minimize function (ε, 1 - ε)Э q→ F-1(q/(1-ε))-F-1((q-ε)/(1-ε)). If F is unimodal and symmetric, then q* = ½.. However, Zielinski also showed that a slight modification of the ε-contamination for symmetric unimodal F may imply that XJ:n - F-1(q*), XJ:n € U(q*), for some q*≠1/2 is the most robust estimator with respect to the median oscillation criterion. Celem tej przeglądowej pracy jest opis wyników profesora Ryszarda Zielińskiego dotyczącychnieparametrycznych estymatorów kwantyli w skończonych próbach oraz ich zastosowania w odpornej estymacjiparametru położenia. Główne przesłanie badań Zielińskiego było następujące:do estymacji kwantyli należy używać pojedynczych statystyk pozycyjnych, a już ich liniowekombinacje mogą być bardzo niedokładne w dużych modelach nieparametrycznych.Optymalny wybór statystyki pozycyjnej zależy od kryterium oceny błędu estymacji.},
author = {Tomasz Rychlik},
journal = {Mathematica Applicanda},
keywords = {quantile, order statistics, randomized estimator, equivariant estimator, confidence interval, most robust estimator, robustness.},
language = {eng},
number = {2},
pages = {null},
title = {The Ryszard Zielinski's works on nonparametric quantile estimators and their use in robust statistics},
url = {http://eudml.org/doc/293182},
volume = {40},
year = {2012},
}

TY - JOUR
AU - Tomasz Rychlik
TI - The Ryszard Zielinski's works on nonparametric quantile estimators and their use in robust statistics
JO - Mathematica Applicanda
PY - 2012
VL - 40
IS - 2
SP - null
AB - This is a survey paper describing achievements of professor Ryszard Zieliński in the subject of nonparametric estimation of population quantiles based on samples of fixed size, and applications of the quantile estimators in the robust estimation of location parameter. Zielinski assumed that a finite sequence of independent identically distributed random variables X1, . . . ,Xn is observed, and their common distribution function F belongs to the family F of continuous and strictly increasing distribution functions. He considered the family T of randomized estimators XJ:n which are single order statistics based on X1, . . . ,Xn with a randomly determined number J. The random variable J is independent of the sample and has an arbitrary distribution on the numbers 1, . . . , n. It was proved that T is the maximal class of estimators which are functions of the complete and sufficient statistic (X1:n, . . . ,Xn:n), and are equivariant with respect to the strictly increasing transformations, i.e., satisfy T(φ(X1:n), . . . ,φ(Xn:n)) = φ(T(X1:n, . . . ,Xn:n)) for arbitrary strictly increasing φ. A number of examples showed that the estimators that do not belong to T are very inaccurate for some F€F.   For comparing estimators, there were used various accuracy criteria based on the difference F(T) - q, where 0 < q < 1 is the quantile order. They are invariant with respect to the strictly increasing transformations. Optimal estimators with respect to the mean absolute loss E|F(T)-q|, mean quadratic loss E(F(T)-q)2, expected LINEX loss E[exp(a[F(T)-q])-a[F(T)-q]-1], a≠0, and Pitman closeness measure were explicitly determined. Further, the best estimators in narrower classes of median-unbiased estimators U(q) = {T€T : med(T, F) = F-1(q)},  (where med(T, F) stands for the median of the distribution of estimator T when the parent distribution function is F), and F-unbiased estimators V(q) = {T € T : EF(T) = q} of quantiles F-1 (q), 0 < q < 1, are determined for some accuracy criteria. Also, random confidence intervals for F-1(q), F€F, of the form [XI:n,XJ:n] on a fixed confidence level 0 <  < 1, i.e. satisfying P(XI:n ≤F-1(q) ≤XJ:n)≥γ,  F € F, , and minimizing E(J - I), are described. Median-unbiased estimators of quantiles were applied by Zielinski in the robust estimation of location parameter. For the i.i.d. sample X1, . . . ,Xn from the location model Fμ(x) = F(x - μ), where μ€R and F is a known unimodal distribution function, and the ε-contamination of the model Z(μ) = {G = (1 -ε)Fμ +εH : H - arbitrary distribution function} for some fixed 0 <ε< 1/2 , the most robust translation equivariant estimator with respect to the median oscillation criterion bn(T, μ) = supG1,G2€Z(μ) |med(T,G1) - med(T,G2)| has the form XJ:n - F-1(q*), XJ:n  €U(q*). Number q*  is chosen so to minimize function (ε, 1 - ε)Э q→ F-1(q/(1-ε))-F-1((q-ε)/(1-ε)). If F is unimodal and symmetric, then q* = ½.. However, Zielinski also showed that a slight modification of the ε-contamination for symmetric unimodal F may imply that XJ:n - F-1(q*), XJ:n € U(q*), for some q*≠1/2 is the most robust estimator with respect to the median oscillation criterion. Celem tej przeglądowej pracy jest opis wyników profesora Ryszarda Zielińskiego dotyczącychnieparametrycznych estymatorów kwantyli w skończonych próbach oraz ich zastosowania w odpornej estymacjiparametru położenia. Główne przesłanie badań Zielińskiego było następujące:do estymacji kwantyli należy używać pojedynczych statystyk pozycyjnych, a już ich liniowekombinacje mogą być bardzo niedokładne w dużych modelach nieparametrycznych.Optymalny wybór statystyki pozycyjnej zależy od kryterium oceny błędu estymacji.
LA - eng
KW - quantile, order statistics, randomized estimator, equivariant estimator, confidence interval, most robust estimator, robustness.
UR - http://eudml.org/doc/293182
ER -