Adaptive regression
Througout the book a linear functional regression model \[ Y=X\beta+ Z \] is considered, where \(Y= (Y_1,\dots, Y_n)'\) is a vector of observations, \(X= (x_1',\dots, x_n')\) is a design \((n\times p)\) matrix, \(\beta\) is a \((p\times 1)\) vector parameter, and \(z\) is a \((n\times 1)\) vector of i.i.d. r.v.s, with a symmetric density \(f\) belonging to a family \[ F= \{f: f(z)= s^{-1}\cdot f_0(z/s),\;s>0\}, \] and \(f_0\) is an unknown unit element of \(F\). The following four basic regression methods are involved: least absolute deviations (LAD), least squares (LS), trimmed least squares (TLS) and \(M\)-regression. The goal of the book is to develop some adaptive methods of estimation based on convex combination of pairs of the above procedures. Each of the basic estimation methods has its own competitive edge, but none is good for all purposes. Therefore the proposed adaptive methods will help the users to make an objective choice and to combine desirable properties of both estimators. The concept of adaptive estimator was proposed by the authors in 1987 at the First Int. Conf. at Neuchâtel on Statistical data analysis based on \(L_1\)-norm and related methods. The idea is transparent and explained in Chapter 1. Suppose that the scale characteristic \(s\) has a reasonable consistent estimator \(s_n\) based on the observations. Let \(T_n(\delta)\) be a solution of the minimization problem \[ \sum^n_{i=1} \rho ((Y_i- x_i't)/s_n)\to\min, \quad t\in \mathbb{R}^p, \tag{1} \] with \(\rho(z)= \delta\rho_1(z)+ (1-\delta) \rho_2(z)\), \(0\leq \delta\leq 1\), where \(\rho_1(z)\) and \(\rho_2(z)\) are symmetric discrepancy functions defining the respective estimators \(\widehat{\beta}_{(1)}\) and \(\widehat{\beta}_{(2)}\). Then \(\sqrt{n} (T_n(\delta)- \beta)\) has an asymptotically normal distribution \[ N_p(0, (X'X)^{-1} \sigma^2(\delta, \rho,f). \] Using \(\delta= \delta_0\), which minimizes the \(\sigma^2 (\delta,\rho,f)\) with respect to \(\delta\), we get an estimator \(T_n(\delta_0)\). Typically, \(\sigma^2(\delta, \rho,f)\) depends on \(f\) only through two moments of \(f_0\). Instead of \(\sigma^2(\delta, \rho, f)\) we then minimize its estimation with the two moments replaced by their estimators based on the data. Denote \(\widehat{\delta}_0\) the minimizing value, then \(T_n (\widehat{\delta}_0)\) is called an adaptive combination of \(\widehat{\beta}_{(1)}\) and \(\widehat{\beta}_{(2)}\). It attains the minimum possible asymptotic variance among all solutions of (1). Chapter 2 introduces two new reasonable estimators to the scale characteristic. Chapters 3 through 6 present four new adaptive estimators based on the combinations of two estimators: LAD with LS, LAD with TLS, LAD with \(M\)-estimation, and LS with TLS. Each estimator is illustrated by an interesting numerical example. Chapter 7 proposes an adaptive choice of trimming proportions in the TLS estimation, which is a key point for practical applications of the method. In Chapter 8 it is shown that \(\widehat{\delta}_n\), leading to the optimal adaptive combination of the two estimators, also leads to the optimal adaptive combination of two tests in the linear hypothesis. Chapter 9 studies the computational aspects of the procedures, and the reader will find information on how to run the computer programs in S-PLUS, available on the authors' website. Finally, an Appendix provides a mathematical background to the proposed methods, mostly based on asymptotic properties of studentized \(M\)-estimators. Remarks of the reviewer: 1. The significance of adaptive estimators is questionable, because the optimality properties of these estimators, in a wide class of regular estimators, are not known, while the optimality properties of the basic estimators are well-known. 2. The proofs in the Appendix are too short, and the bibliographical notes are absent in the Appendix. Therefore the reader has to search for the detailed proofs through a long reference list. 3. The material of subsection 4.2 is not delivered carefully, because the formulas (4.7) and (4.9) are note explained in the book.
- Nonparametric and robust methods. (Editorial)
- Combining different procedures for adaptive regression
- Adaptive \(L\)-estimation for linear models
- Computational aspects of adaptive combination of least squares and least absolute deviations estimators
- On adaptive combination of regression estimators
- Adaptive regression with Brownian path covariate
- Optimal pseudo-Gaussian and rank-based random coefficient detection in multiple regression
- A simple R-estimation method for semiparametric duration models
- Flexible \(L\)-estimation in the linear model
- Combined regression models
- Response-Adaptive Regression for Longitudinal Data
- On adaptive linear regression
- scientific article; zbMATH DE number 4143248 (Why is no real title available?)
- On Adaptive M-Regression
- scientific article; zbMATH DE number 5511108 (Why is no real title available?)
- ADAPTIVE ESTIMATION IN LINEAR REGRESSION
- Adaptive Regression by Mixing
- scientific article; zbMATH DE number 2051024 (Why is no real title available?)
- Best monotone M-estimators
- Adaptive L-estimation for regression slope under asymmetric error distributions
- R-estimation in semiparametric dynamic location-scale models
- Efficient pseudo-Gaussian and rank-based detection of random regression coefficients
- Efficient Fully Distribution-Free Center-Outward Rank Tests for Multiple-Output Regression and MANOVA
- Estimation and Testing in M‐quantile Regression with Applications to Small Area Estimation
- An image reconstruction algorithm based on the semiparametric model for electrical capacitance tomography
- Highly robust training of regularized radial basis function networks.
This page was built for publication: Adaptive regression
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4492939)
