Approximations for marginal densities of M-estimators.
Date
1994
Authors
Fan, Rocky Yuk-Keung.
Journal Title
Journal ISSN
Volume Title
Publisher
Dalhousie University
Abstract
Description
In this thesis we present a finite sample approximation for the marginal densities of a multivariate M-estimator. The result is particularly useful in robust statistics where an estimator usually is defined implicitly and does not have a closed form, and for small sample problems where the asymptotic results may not be reliable.
Precisely, let Y$\sb1,\... ,Y\sb{n}$ be independent m-dimensional random observations such that each observation has a density function which is parameterized by a p-dimensional parameter $\eta$. Let $\\eta$ be an M-estimator of $\eta$, the solution of the system
$${1\over n}\Sigma\sbsp{l=1}{n} \Psi\sb{jl}(Y\sb{l},\\eta) = 0,\ j = 1,\...,p.$$Our primary objective is to derive an approximation for the marginal densities of a component in $\\eta$ under $\eta = \eta\sb0.$ The result is then extended to a real-valued function $\rho(\\eta),\ \rho : Re\sp\rho \to \Re$, and finally to a real-valued vector $\rho(\\eta) = \{\rho\sb1(\\eta),\...,\rho\kappa(\\eta)\},\ \rho : Re\sp\rho \to Re\sp{k},\ k \le p$.
We begin with an overview of the general problem and some background information. Then we derive the main results and discuss the relationship among our approach and some existing techniques for the problem. In addition, we implement the approximations for several location-scale and multiple regression examples. Finally, we discuss the limitation and some potential applications of our results.
Thesis (Ph.D.)--Dalhousie University (Canada), 1994.
Precisely, let Y$\sb1,\... ,Y\sb{n}$ be independent m-dimensional random observations such that each observation has a density function which is parameterized by a p-dimensional parameter $\eta$. Let $\\eta$ be an M-estimator of $\eta$, the solution of the system
$${1\over n}\Sigma\sbsp{l=1}{n} \Psi\sb{jl}(Y\sb{l},\\eta) = 0,\ j = 1,\...,p.$$Our primary objective is to derive an approximation for the marginal densities of a component in $\\eta$ under $\eta = \eta\sb0.$ The result is then extended to a real-valued function $\rho(\\eta),\ \rho : Re\sp\rho \to \Re$, and finally to a real-valued vector $\rho(\\eta) = \{\rho\sb1(\\eta),\...,\rho\kappa(\\eta)\},\ \rho : Re\sp\rho \to Re\sp{k},\ k \le p$.
We begin with an overview of the general problem and some background information. Then we derive the main results and discuss the relationship among our approach and some existing techniques for the problem. In addition, we implement the approximations for several location-scale and multiple regression examples. Finally, we discuss the limitation and some potential applications of our results.
Thesis (Ph.D.)--Dalhousie University (Canada), 1994.
Keywords
Statistics.