蔡擇林
(湖北師范學(xué)院 數(shù)學(xué)與統(tǒng)計(jì)學(xué)院, 湖北 黃石 435002)
混合系數(shù)線性模型的一類有偏估計(jì)
蔡擇林
(湖北師范學(xué)院 數(shù)學(xué)與統(tǒng)計(jì)學(xué)院, 湖北 黃石 435002)
研究了連續(xù)測量數(shù)據(jù)情況下的混合系數(shù)線性模型的參數(shù)估計(jì)問題. 利用壓縮估計(jì)方法給出了該模型的一類有偏估計(jì),并討論了該估計(jì)的一些優(yōu)良性質(zhì)和壓縮系數(shù)的選取問題.
混合系數(shù)線性模型;stein估計(jì);嶺估計(jì);均方誤差 .
考慮如下混合系數(shù)線性模型
z(t)=[x(t)]′a+[y(t)]′β
(1)
式中x(t)=(x1(t),x2(t),…,xp(t))′,y(t)=(y1(t),y2(t),…,yq(t))′,x1(t),x2(t),…,xp(t) 及y1(t),y2(t),…,yq(t)都是t的已知函數(shù),a是p×1的固定系數(shù)向量,β是q×1的隨機(jī)系數(shù)向量,且β~(b,∑).
現(xiàn)對m個樣品,分別在tij(ti1
(2)
這里的βi和εij分別是每個樣品的隨機(jī)系數(shù)和每次測量的誤差,βi與εij獨(dú)立,且
若記zi=(zi1,zi2,…,zini)′,εi=(εi1,εi2,…,εini)′
則可得
zi=Xia+Yiβi+εi
(3)
式中
設(shè)Ci=(Xi,Yi),d=(a′,b′)′,ei=Yi(βi-b)+εi, 則式(3)變?yōu)?/p>
(4)
進(jìn)一步,記
D=diag(D1,D2,…,Dm)
z=Cd+e,e~(0,D)
(5)
這里p≥0,q≥0. 顯然,當(dāng)p=0時模型化為完全隨機(jī)系數(shù)形式,當(dāng)q=0時模型化為一般的線性模型.這里還要求 rank(Xi)=p,rank(Yi)=q,rank(Xi,Yi)=p+qg.混合系數(shù)線性模型的結(jié)構(gòu)和性質(zhì)與一般的線性模型及完全隨機(jī)系數(shù)的線性模型有很大的差異,而在實(shí)際問題中又有著廣泛的應(yīng)用背景,如經(jīng)濟(jì)分析、可靠性退化分析以及生物學(xué)等領(lǐng)域,因此,對該模型的研究無論是從理論上還是從應(yīng)用角度都是十分重要和必要的.自莊東辰、茆詩松(文獻(xiàn)[4])提出了混合系數(shù)線性模型以后,許多學(xué)者研究了這種模型的參數(shù)估計(jì)(見文獻(xiàn)[5]-[18]),基于(5),文獻(xiàn)[4]給出了d的LS 估計(jì), LS估計(jì)雖然無偏,但當(dāng)系數(shù)陣接近病態(tài)時, LS估計(jì)的均方誤差過大,穩(wěn)定性不好,針對此情況,文獻(xiàn)[7]提出了一種有偏估計(jì)——Stein估計(jì).本文亦利用壓縮估計(jì)方法給出了該模型的一類有偏估計(jì),并討論了該估計(jì)的一些優(yōu)良性質(zhì)和壓縮系數(shù)的選取問題.
其中L=CQ,r=Q′d,Q′C′CQ=Λ=diag(λ1,λ2,…,λg)
文獻(xiàn)[7]給出了d的如下Stein 估計(jì)
(6)
其典則形式為
其中0≤k≤ 1 稱為壓縮系數(shù)
本文給出d的如下估計(jì)
(7)
其典則形式為
其中0≤K=diag(k1,k2,…,kg)≤I稱為壓縮系數(shù)。
顯然,當(dāng)K=kI,0≤k≤1 時即為普通stein估計(jì)(6),當(dāng)K=(Λ+S)-1Λ,S=diag(s1,s2,…,sg)≥0 時即為嶺估計(jì)。
證明 顯然
其中
lj=(L′ML)jj≥0
引理1 對于模型y=Xβ+e,e~(0,V),
假定X=I,則Ay~β的充要條件為1)AV對稱,2)A的所有特征根在[0,1]內(nèi).
KL′ML=L′MLK
證明 將模型(5)變?yōu)?/p>
(C′C)-1C′Z=d+(C′C)-1C′e,(C′C)-1C′e~(0,(C′C)-1C′DC(C′C)-1)
其中A=QKQ′,y=(C′C)-1C′Z,V=(C′C)-1C′DC(C′C)-1
故AV=(AV)′?QKQ′(C′C)-1C′DC(C′C)-1=(C′C)-1C′DC(C′C)-1QKQ′ ?KΛ-1L′DLΛ-1=Λ-1L′DLΛ-1K?KΛ-1L′MLΛ-1+σ2KΛ-1=Λ-1L′MLΛ-1K+σ2Λ-1K?KL′ML=L′MLK
A=QKQ′的特征根顯然均在[0,1]內(nèi),由引理1知,定理得證。
1)極小化均方誤差法
2)極小化均方誤差的無偏估計(jì)法
[1]Rao C R.The theory of least squares when the parameters are stochastic and its application to the analysis of Growth curves[J]. Biomitrika, 1965,52: 447~458.
[2]Swamy P A V B.Statistical inference in random coefficient regression model[J]. New York:Springer-Verlage, 1971.
[3]Johanson S.Asumptotic inference on random coefficient regression models[J]. Scand J Statist, 1982,9:201~207.
[4]Zhuang dongchen, Mao shisong. Estimation of parameters for mixed-effect coefficient liear model [J]. Journal of Applied Probability and Statistics, 1996, 12(1): 81~87.
[5]Chen jing. Ridge estimator for mixed-effect coefficient linear model[J]. Journal of Qingdao Univ, 2007,20(2):37~41.
[6]Liu xiaomao, Zhang jun. The root squaring estimation of parameters in a linear model with mixed coefficients[J]. J Huazhong Univ of Sci &Tech, 1997,25(3):111~112.
[7]Liu xiaomao, Mao shisong. Stein shrunken estimates of parameters for mixed-effect coefficient linear model[J]. Acta Mathematica Scientia, 2001,21A(4):453~457.
[8]Liu xiaomao, Zhang jun. The estimation of parameters and distributions of estimates for a linear model with mixed coefficients[J]. J Huazhong Univ of Sci &Tech,1997,25(6):103~105.
[9]Li hui, Liu jianzhou. Some results on parameter estimation in a linear model with mixed coefficents[J]. Journal of Jishou Univ, 2004,25(1):16~21.
[10]Pan jinxiao, Xue yakui. A new estimate of parameters in the linear model of mixed coefficient[J]. Journal of North China Institute of Technology, 1998,19(4):283~285.
[11]Liu weiqi, Wang feng. An estimate of parameters in the linear model mixed coefficient with constraint[J]. Journal of Engineering Mathematics, 2000,17(4):103~106.
[12]Li na. An estimate of parameters in the linear model of mixed coefficient[J]. Advanced Mathematics Bulletin, 2005,54(3):64~66.
[13]Li na. Local stein shrunken estimates of parameters for mixed-effect coefficient linear model[J]. Journal of Yangtze Univ, 2008,5(3):124~126.
[14]Li na, Wang lei. An admissible estimate of parameter in the linear model of mixed coefficient[J].Guangxi Sciences, 2009,16(4):403~405.
[15]Li na. Comparison between a new estimation and least square estimation of parameter in a linear model with mixed coefficients[J].Journal of China Three Gorges Univ, 2009,31(1): 110~112.
[16]Liu hongling. The improvement of estimation of parameters for mixed-effect coefficient linear model[J].Journal of Mathematics, 2006,26(5):513~518.
[17]Chen jing. Local root square estimation of parameters in a linear model with mixed coefficients[J].Journal of Qingdao Univ, 2005,18(1):7~11.
[18]Cai zelin. Almost unbiased ridge estimator for mixed-effect coefficient linear model [J].Journal of Mathematics, 2013,26(5):513~518.
Keywords:mixed-effect coefficient linear model;stein estimator;ridge estimator;mean square error
Anewbiasedestimatorforamixed-effectcoefficientlinearmodel
CAI Ze-lin
(College of Mathematics and Statistics, Hubei Normal University, Huangshi 435002, China)
In the paper, we consider a mixed-effect coefficient linear model with the repeatedly measured data. Using the shrunken method, a new biased estimator of the model are given and its optimal property is shown. And the optimal value for the shrunken coefficient is established.
2013—09—15
湖北省教育廳科學(xué)技術(shù)研究資助項(xiàng)目(Q20122203)
蔡擇林(1963— ),男,湖北浠水人,教授,主要研究方向?yàn)閿?shù)理統(tǒng)計(jì).
O212
A
1009-2714(2014)01- 0005- 05
10.3969/j.issn.1009-2714.2014.01.002