闕 燁
(南京理工大學(xué) 理學(xué)院,南京 210094)
非參數(shù)混合效應(yīng)模型的估計(jì)*
闕 燁
(南京理工大學(xué) 理學(xué)院,南京 210094)
針對(duì)非參數(shù)混合效應(yīng)模型提出估計(jì)方法,通過B樣條的方法估計(jì)非參數(shù)函數(shù),使用懲罰(非加權(quán))最小二乘方法估計(jì)隨機(jī)效應(yīng),然后通過構(gòu)造正態(tài)似然函數(shù)得到方差的估計(jì),并且證明了方差分量的相合性和函數(shù)部分的漸近正態(tài)性;最后給出數(shù)字模擬來展示所提出方法的估計(jì)效果,結(jié)果表明:該方法給出的估計(jì)效果良好,且在數(shù)值上是穩(wěn)定的。
非參數(shù)混合效應(yīng)模型;B樣條;懲罰最小二乘;正態(tài)似然函數(shù)
在分析相關(guān)數(shù)據(jù)的時(shí)候經(jīng)常用到混合效應(yīng)模型,因此此處考慮非參數(shù)混合效應(yīng)模型:
(1)
混合效應(yīng)模型對(duì)于分析縱向和重復(fù)測(cè)量數(shù)據(jù)來說是一個(gè)重要工具,近年來它引起了國(guó)內(nèi)外統(tǒng)計(jì)學(xué)者的極大關(guān)注。Cai T,etal.(2002)[1]研究了在集群故障時(shí)間數(shù)據(jù)下的半?yún)?shù)混合效應(yīng)模型,他們?cè)谖闹刑岢隽藢?duì)于這個(gè)隨機(jī)效應(yīng)模型的推斷和預(yù)測(cè);Chen Z和Dunson D(2003)[2]研究了線性混合效應(yīng)模型的隨機(jī)效應(yīng)選擇;Li W B和Xue L G(2014)[3]研究了廣義部分線性混合效應(yīng)模型的有效推斷問題,他們對(duì)參數(shù)和方差分量提出了一系列半?yún)?shù)估計(jì)值,然后使用局部線性光滑方法去展示非參數(shù)分量的估計(jì)值;Pang Z和Xue L G(2012)[4]研究了單指標(biāo)混合效應(yīng)模型,為了估計(jì)指標(biāo)系數(shù)和聯(lián)系函數(shù),提出了一系列新的估計(jì)方程去調(diào)整邊界效應(yīng),使用局部線性光滑方法去估計(jì)非參數(shù)函數(shù);Schimek M G(2000)[5]研究了部分線性模型在光滑樣條下的估計(jì)和推斷,描述了在光滑樣條方法下的廉價(jià)直接算法,光滑參數(shù)可以被選擇通過一個(gè)無偏風(fēng)險(xiǎn)標(biāo)準(zhǔn);Zhong X P,etal.(2003)[6]對(duì)帶有變量誤差的線性混合效應(yīng)模型的估計(jì)提出了統(tǒng)一分類法;Zhao H B和You J H(2011)[7]研究了帶有測(cè)量誤差的部分線性回歸模型的不同估計(jì),他們給出的估計(jì)是漸近無偏估計(jì)并且實(shí)現(xiàn)了非參數(shù)有效邊界。其他的關(guān)于隨機(jī)效應(yīng)的文獻(xiàn)可參看Li W B和Xue L G(2013)[8],Liang H(2009)[9],Lindstrom M J和Bates D M(2010)[10],Wu H和Zhang J T(2002)[11]。此處與已有文獻(xiàn)的不同之處是本文使用B樣條來估計(jì)函數(shù)部分,從而將非參數(shù)模型轉(zhuǎn)化為線性混合效應(yīng)模型。B樣條估計(jì)方法的優(yōu)點(diǎn)在于它有緊支撐,這使得計(jì)算速度加快,函數(shù)部分的擬合曲線會(huì)更光滑,這無疑在視覺上更吸引人。而隨機(jī)效應(yīng)項(xiàng)的處理借鑒了已有文獻(xiàn)的方法并在此基礎(chǔ)上進(jìn)行改進(jìn),請(qǐng)參考文獻(xiàn)Gu C和Ma P(2005)[12]。
2.1 非參數(shù)函數(shù)的估計(jì)
使用B樣條基函數(shù)將g(·)表示為
2.2 隨機(jī)效應(yīng)項(xiàng)b的估計(jì)
這里使用懲罰最小二乘法估計(jì)b,極小化
(2)
2.3 方差分量的估計(jì)
-n(n-1)log(σε2)-nlog(σε2+nσb2)-
首先,給出下列條件:
Ⅰc0表示一個(gè)常數(shù),有E(e2)≤c0<∞成立,這里e=Zb+ε=Y-g(U);
Ⅱ 對(duì)任意的i,協(xié)變量Ui是獨(dú)立同分布的變量,且Ui的分布是緊支撐集;
定理2 在條件(Ⅰ)—(Ⅲ)下,有
(3)
(4)
定理1的證明 定理1的證明類似于Huang J Z,etal.(2004)[13]中定理2的證明方法,這里忽略該定理的證明過程。
定理2的證明 先證明式(3)。
下面證明式(4),進(jìn)行一系列的演算后有:
N2=OP(n-1/2)
N3=OP(n-1/2)
通過利用式(3)就得到了式(4)的證明。
例1 考慮模型:Yi=cos(πUi)+b+εi,i=1,2,…,n,其中Ui服從區(qū)間(0,1)上的均勻分布,模型的隨機(jī)效應(yīng)項(xiàng)服從均值為0,方差為0.16的正態(tài)分布,誤差項(xiàng)服從均值為0,方差為0.04的正態(tài)分布的隨機(jī)變量。運(yùn)用B樣條估計(jì)函數(shù)項(xiàng),本例中樣條的節(jié)點(diǎn)由廣義交叉驗(yàn)證(GCV)方法得出節(jié)點(diǎn)個(gè)數(shù)為5個(gè),樣條階數(shù)為3。
圖1 函數(shù)部分的真實(shí)曲線和估計(jì)曲線(n=100)Fig.1 The real link function curve and the estimated link function curve(n=100)
圖2 函數(shù)估計(jì)值的箱線圖(n=100)Fig.2 The boxplot for the estimated link function(n=100)
[1] CAI T,CHENG S C,WEI L J.Semiparametric Mixed-effects Models for Clustered Failure Time Data[J].Journal of the American Statistical Association,2002,97(6):514-522
[2] CHEN Z,DUNSON D.Random Effects Selection in Linear Mixed Models[J].Biometrics,2003,59(4):762-769
[3] LI W B,XUE L G.Efficient Inference in a Generalized Partially Linear Model with Random Effect for Longitudinal Data[J].Communication in Statistics Theory and Methods,2014,44(2):241-260
[4] PANG Z,XUE L G.Estimation for the Single-index Models with Random Effects[J].Computational Statistics and Data Analysis,2012,56(6):1837-1853
[5] SCHIMEK M G.Estimation and Inference in Partially Linear Models with Smoothing Splines[J].Journal of Statistical Planning and Inference,2000,91(2):525-540
[6] ZHONG X P,FUNG W K,WEI B C.Estimation in Linear Models with Random Effects and Errors-in-variables[J].Annals of the Institute of Statistical Mathematics,2002,54(3):595-606
[7] ZHAO H B,YOU J H.Difference Based Estimation for Partially Linear Regression Models with Measurement Errors[J].Journal of Multivariate Analysis,2011,102(10):1321-1338
[8] LI W B,XUE L G.Efficient Inference about the Partially Linear Varying Coefficient Model with Random Effect for Longitudinal Data[J].Communications in Computer and Information Science,2013(391):573-583
[9] LIANG H.Generalized Partially Linear Mixed-effects Models Incorporating Mismeasured Covariates[J].Annals of the Institute of Statistical Mathematics,2009,61(1):27-46
[10] LINDSTROM M J,BATES D M.Nonlinear Mixed-effects Models for Repeated Measures Data[J].Biometrics,2010,46(46):673-687
[11] WU H,ZHANG J T.Local Polynomial Mixed-effects Models for Longitudinal Data[J].Journal of the American Statistical Association,2002,97(459):883-897
[12] GU C,MA P.Optimal Smoothing in Nonparametric Mixed-effect Models[J].Annals of Statistics,2005,33(3):1357-1379
[13] HUANG J Z,WU C O,ZHOU L.Polynomial Spline Estimation and Inference for Varying Coefficient Models with Longitudinal Data[J].Statistica Sinica,2004,14(3):763-788
責(zé)任編輯:李翠薇
The Estimation for the Nonparametric Mixed Effects Model
QUE Ye
(School of Science, Nanjing University of Science and Technology, Nanjing 210094, China)
The estimation methodology for the nonparametric mixed effects model is proposed. For this, we use the B-splines methods to estimate the function, and employ the penalized least square method to obtain the estimator of the random effects. Further, we construct the normal likelihood function to estimate the variance components. And we also prove the consistency of variance components and the asymptotic normality of the link function.A simulation study is carried out to show the estimation effect of our proposed methodology.It shows our proposed methodology performs well.Our algorithm is stable numerically.
nonparametric mixed effects model; B-splines method; penalized least square method; normal likelihood function
10.16055/j.issn.1672-058X.2017.0000.003
2016-05-19;修回時(shí)間:2016-06-24.
江蘇省自然科學(xué)基金(BK20131345).
闕燁(1992-),女,安徽淮南人,助教,碩士研究生,從事非參數(shù)統(tǒng)計(jì)及應(yīng)用研究.
O212.7
A
1672-058X(2017)01-0010-04