University of North Texas
Novel Simultaneous Estimation of CVF MEANs and NEF-QVF Variance Functions
Abstract
dc:descriptionStein (1956) proved a counter-intuitive theorem (a famous result known as Stein's paradox) showing that the sample mean (the usual unbiased estimator) for estimating the mean of a multivariate normal distribution is inadmissible because a better estimator exists: a biased shrinkage estimator (James and Stein, 1961) can have uniformly smaller risk compared to the ordinary maximum likelihood estimate. This seminal work has inspired a broad interest in the study of shrinkage estimators and in applications where multiple parameters need to be estimated simultaneously, especially in scenarios with high dimensionality or limited data per parameter. Among the recent research work on shrinkage estimators, a significant contribution is the paper by Xie, Kou and Brown (Ann. Stat., 2016), in which they proposed a class of semiparametric/parametric shrinkage estimators for a family of distributions with quadratic variance function and establish their asymptotic optimality. However, their method is not applicable to the distributions with cubic variance function such as the parametric family of the inverse Gaussian (IG) distribution on the positive real line. In this talk, we propose a new family of symmetric inverse Gaussian (SIG) distributions with cubic variance function. One of the novelties of our method is that by symmetrizing the regular IG distribution, we obtain the variance function as a cubic polynomial of the mean with a non-zero quadratic term, which leads to different asymptotic results depending on the dominant term and allows for variance function misspecication. Since an unbiased risk estimate (URE) under the sum of squared-error loss is not known to exist for the cubic variance function, Xie et al.'s method of minimizing an unbiased estimate of the risk is not applicable. We propose to minimize a biased estimate of the expected URE to obtain an asymptotically optimal shrinkage estimator. Under some mild conditions, we show that our biased estimate of the risk is uniformly close to the actual loss in L1 and in probability asymptotically. Furthermore, we prove that our shrinkage estimator is asymptotically optimal within the class of any other shrinkage estimators satisfying certain regularity conditions. Finally, we present some simulation results to demonstrate the performance of our method and apply our shrinkage estimator to the baseball data from Brown (2008) to predict batting averages.
Degree
thesis:*- Grantor dc:publisher
- University of North Texas
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Khan, Zubair Muhammad
- Contributors dc:contributor
-
- Song, Kai-Sheng
- Iaia, Joseph A.
- Trang, Nam
Subjects
dc:subject × 3Rights
dc:rights- Statement dc:rights
-
- Public
- Khan, Zubair Muhammad
- Copyright
- Copyright is held by the author, unless otherwise noted. All rights Reserved.
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
-
local-cont-no: submission_4572
ark: ark:/67531/metadc2566451