节点文献

相依数据的若干统计模型及分析

【作者】 林路

【导师】 张润楚;

【作者基本信息】 南开大学 , 概率论与数理统计, 2001, 博士

【摘要】 在一些近代科学研究中,如生命科学和信息科学的研究,人们获得的数据往往具有量大、高维和相依等特点。于是,关于相依随机变量的研究,已引起人们的重视,取得一些研究成果(如陆传荣,林正炎(1997)及其它所列文献),并提出了一些研究和加工相依变量的有效方法,如Bootstrap、分块Bootstrap、Jackknife和分块Jackknife(Blocks of blocks jackknife)等方法(Lahiri(1999),Demitris and Joseph(1992))。然而,对各种具体的相依数据的统计模型(如相依数据的非线性回归和相依数据的非参数回归模型)的研究还不够充分和不够完备。为了进一步发展相依随机变量及相关统计模型的理论,本文研究相依数据的若干统计模型及其分析,包括相依数据的线性回归模型、非线性回归模型、半参数及非参数回归模型中参数估计和函数拟合等。 我们在第一章研究了噪声为弱相依过程的固定设计或随机设计非参数回归模型 Yi=g(Xi)+εi,i=1,…,n,其中作{ε1,ε2,…}是一均值为0方差为σ2的宽平稳随机过程。在自协方差R(k)=E(εiεi+k)满足一些较弱的条件下,我们定义了固定设计模型函数g(·)的两种类型的Bootstrap小波估计同时定义了随机设计模型函数g(·)的Bootstrap小波估计其中Yi*,ti*,Xi*和Ai*分别是观测值、设计点和设计区间的Bootstrap样本。在自协方差R(k)=E(εiεi+k)满足一些更弱的条件下,我们定义了固定设计模型函数g(·)的两种类型的分块Bootstrap小波估计tv同时定义了随机设计模型函数g(·)的分块Bmtst*p /J’波估计 hi j(t)二(hi)‘二 K”Em(t,X;)/f(t), i二1其中X*,允X:和人分别是观测值、设计点和设计区间的分块Bootstra。样本,l是数据块宽.我们得到固定设计模型中 BOOtstr叩小波和分块 BOOtSnap/J’波估计量g*一和g*o的偏差和方差的渐近界及近似渐近正态性,建立了随机设计模型中的Bootstrap /J’波和分块Boot。。rap /J’波估计量3*的依概率收敛性及偏差的渐近界,提出了选择数据块大小的原则.本文的理论结果和数据模拟表明,在较弱的数据假设下,非参数回归模型中的B皿tstmp小波和分块Bootstrap /J’波方法是有效的. 第二章研究了相依数据的线性回归模型中分块拟回归方法.为了减少计算复杂性,在线性回归的计算机试验设计中,最近人们提出了一种新的统计方法~拟回归(O侧叫2000),AMM O一(2001)).在独立同分布的线性回归模型中,拟回归不仅能提高计算速度,而且有较好的统计性质。然而,对相关数据模型,这种方法的统计性质并不好.针对这一问题。本文提出回归系数凡的分块拟回归估计: Q ITh 6。一 4 3T}。i二 0、l…·.D—1.其中Ti为适当的观测数据块,Q为数据块的个数.我们得到分块拟回归的小样本和大样本性质,如无偏性、均方收敛性、强收敛性和渐近正态性,并讨论了曲线拟合的性质.这些结果表明,分块拟回归比原拟回归渐近有效.同时,我们还指出了分块拟回归(包括原拟回归)在高维问题中的缺陷.为改善曲线拟合,我们提出一种修正分块拟回归.修正分块拟回归只需将PO的估计调整为 IQM r_lP一1 。_.上 歹7风下7飞/。NJ 下7风。_』。,_。。2。_ 山二7了了}二_三_\1一7刁工工二人z“\c门一1)L+S厂I叭i一I〕L+S)5其中驯和闪分别为观测变量和设计变量,/为基函数.本文的理论结果和数据模拟表明,在高维问题中,修正分块拟回归比原分块拟回归(包括原拟回归)渐近有效. S v 我们在第三章研究了弱相依数据的半参模型中的分块Euclidean经验似然方法.如果对问题的背景所知甚少,仅仅知道某无偏估计函数(如一二矩),人们有时称此为半参模型.在独立同分布数据的半参模型种,人们用经验似然进行参数估计(Owen(1988,1990),Qin。nd L娜less(1994)).在弱相依数据的半参模型中,本文引入一种分块E皿lid阴n经验似然方法.这种方法结构简单,其对数似然比为: l。。(01=HT(01’SZ‘(01T(01。 ”HH\”’一2“\”}“B’”)“\”)’其中穴0)和S以0)分别为

【Abstract】 The data gathered from the study of some modern sciences such as biology and information science is usually large, high dimensional and dependent. So, the theory of dependent random variables is an important subject in statistics. It is well known that one has obtained some theoretical and practical achievements in the field of dependent random variables recently (Lu and Lin (1997) and the literature listed in the reference). And some powerful tools for treating dependent random variables, for example, Bootstrap, blockwise Bootstrap, Jackknife and blocks of blocks jackknife, have been introduced in recent literature (Lahiri (1999), Demitris and Joseph (1992)). So far, however, the theory that studies specially the statistical models with dependent data such as nonlinear regression and nonparametric regression models is not very satisfactory. In order to develop and unprove the theory of dependent random variables and related statistical models, this paper mainly focuses on the statistical theory of parameter estimation and fitting of function in the statistical models with dependent data involving linear regression model, nonlinear regression model, semi-parametric model and nonparametric regression model.In Chapter 1, we consider the following fixed design or random design nonparametric model with dependent noises:where {c, e2,.?.} is a stationary stochastic process with mean 0 and variaiice a2. Under some weak assumptions on auto-covariance R(k) = E(ici), we define two Bootstrap wavelet estimators of gQ) in the fixed design model as1 NNg*(t) t)n(t and gt = Y7 J Em(t, s)dsi=1j=iand a Bootstrap wavelet estimator of g(.) in the random design model asi= 1where Y7, t, X, and A are respectively the Bootstrap samples of observations, design points and design intervals. Under weaker assumptions on auto-covariance than that oflXdata i1l fixed desigIl model, we define two blockwise BootstraP wave1et estimators of g(.)in the fixed design model as1 + K*Em(t, t:)K(t:) and g*(t) = f K* l Em(t, s)d8g*(t) = ai F K*Em(t, t:)K(t:) and g*(t) = Z K* l* Em(t, s)d8l=1 i=l iA:and a blockwise Bootstrap wavelet estimators of g(.) in the random design model asblj(t) = (bl)--’ Z K*Em(t, X;)/i(t),i=lwhere K*, t: 3 X; 5 and A: are re8pectively the blockwise Bootstfap samples of observations,design points and design interals, l is the width of data block. In the fixed design model,the asymptotic bounds of the biases and variances fOr the Bootstrap wavelet and blockwiseBootstrap wave1et estimators g*(t) and g*(t) are obtained, their asymptotic normality fora modified version are establi8hed. In the random design model, the consistency and theasymptotic bound of the bias for the Bootstrap wavelet and blockwise Bootstrap waveletestimator g(t) are proved and a principle of selecting the width of data block is given.These re8ults show that both the BootstraP wavelet and blockwise BootstraP waveletmethods are valid in the models with weakly dependent processes.In Chapter 2, we study the blockwise quasi--regression in the linear models with weaklydependent data. Quasi-regression, a new method motivated by the problerns arising incomputer experiments, mainly focuses on speeding up evaluation (Owen (2000), An andOwen (2001)). For i.i.d. observations, the quasi-regression method is valid not only i11speeding l1p evaluation, but also in parametric statistical inference and fitting of function.However, fOr dePendent observations, this method is invalid. To improve quasi--regression,in this ChaPter, a blockwise quasi--regression estimator of the regression coefficient fij isdefined asQ1,QPj = 6 gTf5j = 0,1,...,P-- l,where Ti is the block of observations and Q is the flumber of the blocks. Some small sam-ple and large sample superiority of this estimate 8uch as the unbiasedness, convergence,asymptotic norlnality and the property of simulation, are obtained. We also find soIne de-fects of blockwise quasi-r

  • 【网络出版投稿人】 南开大学
  • 【网络出版年期】2002年 01期
  • 【分类号】O212.1;O211.67
  • 【被引频次】2
  • 【下载频次】816
  • 攻读期成果
节点文献中: 

本文链接的文献网络图示:

本文的引文网络