Processing math: 100%
Research article

Multiple robust estimation of parameters in varying-coefficient partially linear model with response missing at random

  • Received: 25 December 2021 Revised: 06 February 2022 Accepted: 10 February 2022 Published: 08 March 2022
  • In this paper, we consider the multiple robust estimation of the parameters in the varying-coefficient partially linear model with response missing at random. The multiple robust estimation method is proposed, and the multiple robustness of the proposed method is proved. Numerical simulations are conducted to investigate the finite sample performance of the proposed estimators compared with other competitors.

    Citation: Yaxin Zhao, Xiuli Wang. Multiple robust estimation of parameters in varying-coefficient partially linear model with response missing at random[J]. Mathematical Modelling and Control, 2022, 2(1): 24-33. doi: 10.3934/mmc.2022004

    Related Papers:

    [1] Marcelo Menezes Morato, Vladimir Stojanovic . A robust identification method for stochastic nonlinear parameter varying systems. Mathematical Modelling and Control, 2021, 1(1): 35-51. doi: 10.3934/mmc.2021004
    [2] Biresh Kumar Dakua, Bibhuti Bhusan Pati . A frequency domain-based loop shaping procedure for the parameter estimation of the fractional-order tilt integral derivative controller. Mathematical Modelling and Control, 2024, 4(4): 374-389. doi: 10.3934/mmc.2024030
    [3] Muhammad Nawaz Khan, Imtiaz Ahmad, Mehnaz Shakeel, Rashid Jan . Fractional calculus analysis: investigating Drinfeld-Sokolov-Wilson system and Harry Dym equations via meshless procedures. Mathematical Modelling and Control, 2024, 4(1): 86-100. doi: 10.3934/mmc.2024008
    [4] Yanchao He, Yuzhen Bai . Finite-time stability and applications of positive switched linear delayed impulsive systems. Mathematical Modelling and Control, 2024, 4(2): 178-194. doi: 10.3934/mmc.2024016
    [5] Weiwei Han, Zhipeng Zhang, Chengyi Xia . Modeling and analysis of networked finite state machine subject to random communication losses. Mathematical Modelling and Control, 2023, 3(1): 50-60. doi: 10.3934/mmc.2023005
    [6] Xiaoyu Ren, Ting Hou . Pareto optimal filter design with hybrid H2/H optimization. Mathematical Modelling and Control, 2023, 3(2): 80-87. doi: 10.3934/mmc.2023008
    [7] Hongyu Ma, Dadong Tian, Mei Li, Chao Zhang . Reachable set estimation for 2-D switched nonlinear positive systems with impulsive effects and bounded disturbances described by the Roesser model. Mathematical Modelling and Control, 2024, 4(2): 152-162. doi: 10.3934/mmc.2024014
    [8] Ruxin Zhang, Zhe Yin, Ailing Zhu . Numerical simulations of a mixed finite element method for damped plate vibration problems. Mathematical Modelling and Control, 2023, 3(1): 7-22. doi: 10.3934/mmc.2023002
    [9] Hassan Alsuhabi . The new Topp-Leone exponentied exponential model for modeling financial data. Mathematical Modelling and Control, 2024, 4(1): 44-63. doi: 10.3934/mmc.2024005
    [10] Vladimir Djordjevic, Ljubisa Dubonjic, Marcelo Menezes Morato, Dragan Prsic, Vladimir Stojanovic . Sensor fault estimation for hydraulic servo actuator based on sliding mode observer. Mathematical Modelling and Control, 2022, 2(1): 34-43. doi: 10.3934/mmc.2022005
  • In this paper, we consider the multiple robust estimation of the parameters in the varying-coefficient partially linear model with response missing at random. The multiple robust estimation method is proposed, and the multiple robustness of the proposed method is proved. Numerical simulations are conducted to investigate the finite sample performance of the proposed estimators compared with other competitors.



    The model considered in this paper is a classical semi-parametric model, varying-coefficient partially linear model, and it has the following form

    Y=XTθ(T)+ZTβ+ε, (1.1)

    where Y is the response variable, X, Z and T are qdimensional, pdimensional and one-dimensional covariates, respectively. β=(β1,,βp)T is a pdimensional unknown parameter vector, θ()=(θ1(),,θq())T is a qdimensional unknown non-parametric function vector, ε is the random error and satisfies E(ε|X,Z,T)=0. Model (1.1) has been well studied by many statisticians, see the literatures, for example, Fan and Huang [1], You and Zhou [2], Huang and Zhang [3], Zhao [4], Feng, Zhang and Lu [5] among others.

    In practical applications, missing data problems are frequently encountered in almost all research areas, such as psychological sciences, medical studies, industrial and agricultural production. The complete-case (CC) method will lose the estimation efficiency due to the disregard of the information from the missing values, and may result in biased results if the data is not missing completely at random. For details, see Little and Rubin [6]. The inverse probability weighted (IPW) method is another frequently used method dated back to Horvitz and Thompson [7] that can be applied to the case of missing covariates. This method is to take the inverse of the selection probability as the weight to the fully observed data, and under missing at random (MAR) assumption this method is unbiased. It has attracted much attention in statistical analysis with missing data, but still doesn't make full use of the incomplete data. The imputation method is a popularly method to deal with missing responses in many studies which was introduced by Yates [8]. The concept of imputation is to fill in each missing data with a suitable value, and then use the observed value and the imputed value for statistical inference by the standard method. This method can improve the efficiency of the resulted estimators, see the literatures, for example, Cheng [9], Wang and Rao [10,11], Wang, Linton and Hardle [12], and so forth. In order to further improve the efficiency of estimation, Robins, Rotnitzky and Zhao [13] propose an augmented inverse probability weighted (AIPW) method. This method has the double robustness, that is, if the selection probability and the conditional expectation function are both correctly specified, the resulted estimator will reach the semi-parametric effective bound, and if either of the two assumed models is correctly specified, the estimator is consistent, see the details in Robins and Rotnitzky [14] and Scharfstein, Rotnitzky and Robins [15]. In subsequent ten years, the doubly robust estimation has been well studied, see for example, Kang and Schafer [16], Qin, Shao and Zhang [17], Cao, Tsiatis and Davidian [18], Han [19], and Rotnitzky et al. [20].

    However, double robustness does not provide sufficient protection for estimation consistency, since it allows only one model for the selection probability and one for the conditional expectation function. It is often risky to assume that one of these two models is correctly specified with an unknown data generating process. Noticed this, Han and Wang [21] propose multiple robust estimator for the population mean when the response variable is subject to ignorable missingness. They suggest multiple models for both the selection probability function and the outcome regression model, and the resulted estimator is consistent if any of the multiple models is correctly specified, and attains the semi-parametric efficiency bound when one selection probability and one outcome regression model are correctly specified, without requiring knowledge of which models are correct. For the details please resort to Han and Wang [21]. Subsequently, Han [22] studies the multiple robust estimator for the linear regression model. He discusses the numerical implementation of the proposed method through a modified Newton-Raphson algorithm, derives the asymptotic distribution of the resulted estimator and provides some ways to improve the estimation efficiency. Later, Sun, Wang and Han [23] propose multiple robust kernel estimating equations (MRKEEs) for nonparametric regression, demonstrate its multiple robustness, and show that the resulted estimator achieves the optimal efficiency within the class of augmented inverse propensity weighted (AIPW) kernel estimators when including correctly specified models for both the missingness mechanism and the outcome regression. Please refer to Sun, Wang and Han [23] for more discussion. In addition, the multiple robust estimation with nonignorably missing data has been studied recently, and here we just list some literatures, see for example, Han [24] and Li, Yang and Han [25].

    To the best of our knowledge, the multiple robust estimation for the parameters of the varying-coefficient partially linear model with response missing at random has not been studied. So in this paper, applying the idea of Han [22] and Sun, Wang and Han [23], we consider the multiple robust estimation method for the parameters of the varying-coefficient partially linear model with missing response, and the proposed method is demonstrated superior over the existing competitors via simulation studies.

    This paper is organized as follows. The proposed estimation technique and its multiple robustness are presented in Section 2. Numerical simulation studies are conducted in Section 3 in order to examine the performance of the proposed method. The technical proofs are also provided in Section 4. Conclusions are summarized in Section 5.

    Suppose the available incomplete data {(Ri,Yi,Xi,Zi,Ti),i=1,2,,n} is a random sample from model (1.1), that is

    Yi=XTiθ(Ti)+ZTiβ+εi, (2.1)

    where Ri is an indicator variable, when Yi can be observed, then Ri=1, and when Yi is missing with Ri=0. The covariate Xi,Zi and Ti are all observed. Following Han [22] and Sun, Wang and Han [23], we also suppose the auxiliary variables Si relate to (Ri,Yi,Xi,Zi,Ti) is available. Just as Han [22] points out that the auxiliary variables do not enter the regression model and are not of direct statistical interest, but they can reduce the impact of missing data on estimation and improve the estimation efficiency. Let Vi=(XTi,ZTi)T denote the covariates. The missing mechanism we assume in this paper is MAR mechanism that commonly used in practice. Specifically, given the covariates Vi, Ti and the available auxiliary variables Si, the missing of Yi is independent of Yi, that is,

    P{Ri=1|Yi,Vi,Ti,Si}=P{Ri=1|Vi,Ti,Si}ˆ=π(Vi,Si). (2.2)

    Here we assume that π() is only related to V and S.

    We first carry out the estimator of the varying coefficient functions θ(). For any t in a small neighborhood of t0, using the local linear fitting for θj(t),j=1,2,,q, we have

    θj(t)θj(t0)+θj(t0)(tt0)=aj+bj(tt0).

    Suppose the parameter β is known, and then minimizing the following objective function

    ni=1Ri{YiZTiβqj=1(aj+bj(Tit0))Xij}2Kh(Tit0)

    about (aj,bj),j=1,2,,q, we can obtain the estimator of θ(t) at t0, where Kh()=h1k(/h), k() is a kernel function, and h is the bandwidth. Let

    Dt0=(XT1h1(T1t0)XT1XTnh1(Tnt0)XTn),
    Wt0=diag(Kh(T1t0)R1,Kh(T2t0)R2,,Kh(Tnt0)Rn),

    and

    S(t0)=(Iq,0q)(DTt0Wt0Dt0)1DTt0Wt0=(S1(t0),S2(t0),,Sn(t0)),

    then the estimator of the coefficient functions θ(t) at t0 is given by

    ˜θ(t0)=nk=1Sk(t0)(YkZTkβ). (2.3)

    Substituting (2.3) into (2.1), we obtain

    ˜Yi=˜ZTiβ+εi, (2.4)

    where ˜Yi=YiXTiˆg(Ti), ˜Zi=ZiˆμT(Ti)Xi with ˆg(t)=nk=1Sk(t)Yk and ˆμ(t)=nk=1Sk(t)ZTk.

    For model (2.4), using the complete data, the CC estimator of β can be obtained by solving the following estimation equation

    ni=1Riˆξi(β)=0, (2.5)

    where

    ˆξi(β)=˜Zi(˜Yi˜ZTiβ)=(ZiˆμT(Ti)Xi)[YiXTiˆg(Ti)(ZiˆμT(Ti)Xi)Tβ].

    From Little and Rubin [6] we know that the CC estimator maybe biased unless the missing mechanism is missing completely at random. So following the works of Robins, Rotnitzky and Zhao [13], the doubly robust estimator ˆβAIPW of β can be defined by

    1nni=1{Riˆπ(Vi,Si)ˆξi(β)Riˆπ(Vi,Si)ˆπ(Vi,Si)ηi(β)}=0, (2.6)

    where ˆπ(Vi,Si) is some estimated value of π(Vi,Si), ηi(β)=E[ˆξi(β)|Vi,Ti,Si]. ˆβAIPW has been improved in terms of consistency, but in practice it is still a great risk to assume that one of the two assumed models is correctly specified. So inspired by Han [22] and Sun, Wang and Han [23], next we shall give the multiple robust estimation for β.

    Suppose there are J and K models used to estimate π(V,S) and E(Y|V,T,S). Let P={πj(αj):j=1,,J} and F={ak(γk):k=1,,K} denote the set of these two models respectively, where αj and γk are the corresponding parameters.

    Let ˆαj, ˆγk be the estimator of αj, γk respectively. Usually, ˆαj can be obtained by maximizing the binomial likelihood

    ni=1{πji(αj)}Ri{1πji(αj)}1Ri.

    According to the property of MAR assumption, it can be seen that Y and R are conditionally independent with respect to (V,T,S), that is, E(Y|V,T,S)=E(Y|R=1,V,T,S). Therefore, using the complete observation data to fit the model ak(γk), we can obtain ˆγk. Let ˆβk be the solution of

    1nni=1{ZiˆμT(Ti)Xi}{RiYi+(1Ri)aki(ˆγk)XTi˜θ(Ti)ZTiβ}=0. (2.7)

    Obviously, ˆβk is an estimated value of β.

    Next, let m=ni=1Ri represents the number of the observable response variables. Without loss of generality, R1==Rm=1, Rm+1==Rn=0. Let ω(V,S)=1π(V,S), similar to Han [22], the following formulas hold

    E(ω(V,S)[πj(αj)E{πj(αj)}]|R=1)=0, (2.8)
    E(ω(V,S)[Uk(β,γk)E{Uk(β,γk)}]|R=1)=0, (2.9)

    where j=1,,J, k=1,,K, Uk(β,γk)={ZμT(T)X}{ak(γk)XTθ(T)ZTβ}. Therefore, the weights ωi,i=1,,m can be defined by

    ωi0,i=1,,m;mi=1ωi=1,mi=1ωi{πji(ˆαj)νj(ˆαj)}=0,j=1,,J,mi=1ωi{ˆUki(ˆβk,ˆγk)ηk(ˆβk,ˆγk)}=0,k=1,,K,

    where

    νj(ˆαj)=1nni=1πji(ˆαj),j=1,,J,
    ηk(ˆβk,ˆγk)=1nni=1ˆUki(ˆβk,ˆγk),k=1,,K,
    ˆUki(ˆβk,ˆγk)={ZiˆμT(Ti)Xi}{aki(ˆγk)XTi˜θ(Ti)ZTiˆβk}.

    Based on the empirical likelihood method, under the above constraints, the Lagrange multiplier method is used to solve the maximum value problem of mi=1ωi, and we use the solution as the weight ωi(i=1,,m) to estimate the parameter β. For ease of presentation, let ˆαT={(ˆα1)T,,(ˆαJ)T}, ˆβT={(ˆβ1)T,,(ˆβK)T}, ˆγT={(ˆγ1)T,,(ˆγK)T}, and ˆgi(ˆα,ˆβ,ˆγ)T=[π1i(ˆα1)ν1(ˆα1),,πJi(ˆαJ)νJ(ˆαJ), {ˆU1i(ˆβ1,ˆγ1)η1(ˆβ1,ˆγ1)}T,,{ˆUKi(ˆβK,ˆγK)ηK(ˆβK,ˆγK)}T]. Based on the empirical likelihood theory, we have

    ˆωi=1m11+ˆρTˆgi(ˆα,ˆβ,ˆγ),i=1,,m, (2.10)

    where ˆρT=(ˆρ1,,ˆρJ+pK) is the (J+pK)-dimension Lagrange multiplier, and is the solution of

    1mmi=1ˆgi(ˆα,ˆβ,ˆγ)1+ρTˆgi(ˆα,ˆβ,ˆγ)=0. (2.11)

    Due to the non-negativity of the weight ˆωi, ˆρ satisfies

    1+ˆρTˆgi(ˆα,ˆβ,ˆγ)>0,i=1,,m. (2.12)

    So we can solve the equation

    mi=1ˆωiˆξi(β)=0 (2.13)

    to obtain the multiple robust estimator of the parameter β, denoted by ˆβMR.

    In calculation of the weight ˆωi, the Lagrange multiplier ˆρ is essential. The calculation algorithm we used is similar to Han [22], for the details please refer to Han [22], here we omit.

    The multiple robustness of ˆβMR is given by the following theorem.

    Theorem 2.1. Suppose that the conditions C1–C5 in Section 4 hold, and if P contains a model that correctly specifies π(V,S), or F contains a correctly specified model for E(Y|V,T,S), then mi=1ˆωiˆξi(β)P0 with n.

    In this section, we conduct some numerical simulations to evaluate the feasibility of the above method and the finite sample performance of the proposed estimator ˆβMR. Several indices of multiple robust estimates, inverse probability weighted estimates, and augmented inverse probability weighted estimates are compared and analyzed under different sample sizes.

    We consider five mutually independent covariates, namely: XN(0,1),TU(0,1),Z1N(1,5),Z2B(0.5,1),Z3N(0,1). The response variable is generated by the model Y=XTθ(T)+ZTβ+ε, where θ(t)=sin(πt) and β=(1,1,2)T. In addition, We consider three auxiliary variable, namely S(1)=1+Z(1)Z(2)+ε1, S(2)=I{S(1)+0.4ε2>2.8}, S(3)=exp[{S(1)/9}2]+ε3, where I() is an indicator function. (ε,ε1,ε2,ε3)TN(0,Σ). The diagonal elements of the matrix Σ are 1,0.5,1,2, the elements at positions (1,2) and (2,1) are 0.5, and the remaining elements are all 0. The probability of selection is logit{π(V,S)}=3.55S(2), under which there are approximately 34% of the subjects with missing Y. The models for correctly estimating π(V,S) and E(Y|V,T,S) are logit{π1(α1)}=α11+α12S(2) and a1(γ1)=XTθ(T)+γ11Z1+γ12Z2+γ13Z3+γ14S(3) respectively. In addition, we also use two incorrect models in the simulation process, namely logit{π2(α2)}=α21+α22Z1+α23Z2+α24Z3, a2(γ2)=XT(4T2+4T)+γ21Z1+γ22Z2+γ23Z3+γ24S(3). For simplicity, we use the Rule of Thumb method to obtain the optimal bandwidth when estimating the nonparametric functions, that is, h=1.06{min(qr,sig)}n1/5, where sig is the standard deviation of covariate T, qr=(Q3Q1)/1.34, Q1 and Q3 are the first and third quartile, respectively. In simulation, we generate random samples with n=200 and n=500 respectively, and repeat the process 500 times to calculate the average biases, mean squared errors (MSEs), the root of mean squared errors (RMSEs) and median absolute error (MAEs).

    In order to verify the superiority of the multiple robust estimation method, we give the calculated indices of the parameter β under different estimation methods, which are the inverse probability weighted estimates ˆβIPW, and the augmented inverse probability weighted estimates ˆβAIPW and multiple robust estimates ˆβMR. To distinguish all the estimators constructed based on different methods and models, each estimator is assigned a name with the form "Method-0000", where each digit of the four-digit number, from left to right, indicates whether π1(α1),π2(α2),a1(γ1),a2(γ2) is used in the construction (1 means yes, 0 means no), respectively. The simulation results are reported in Table 1 and Table 2 with the sample size n=200 and n=500.

    Table 1.  The biases, MSEs, RMSEs and MAEs (multiplied by 102) of different estimators for parameter β when sample size n=200.
    Method β1 β2 β3
    Bias MSE RMSE MAE Bias MSE RMSE MAE Bias MSE RMSE MAE
    IPW-1000 0.071 0.083 2.872 2.296 1.451 2.645 16.26 12.96 0.003 1.360 11.66 9.160
    IPW-0100 0.329 0.233 4.826 3.677 0.021 4.591 21.43 16.04 2.113 3.915 19.79 14.65
    AIPW-1010 0.030 0.078 2.799 2.211 1.324 2.654 16.29 12.88 0.073 1.402 11.84 9.207
    AIPW-1001 0.113 0.074 2.724 2.206 0.644 2.629 16.22 13.18 0.181 1.546 12.44 9.692
    AIPW-0110 0.065 0.069 2.625 2.125 0.387 2.283 15.11 12.17 0.232 1.551 12.45 9.778
    AIPW-0101 -4.930 130.1 114.1 33.09 -5.591 212.9 145.9 51.96 -1.457 447.1 211.4 59.17
    MR-1111 0.066 0.022 2.479 2.193 -0.921 2.341 15.46 12.00 -0.563 0.591 11.18 9.120
    MR-1110 0.030 0.021 2.532 2.208 0.637 2.207 15.59 12.08 0.487 0.572 11.23 9.510
    MR-1101 0.031 0.023 2.574 2.140 0.602 2.206 15.13 13.26 0.487 0.573 12.01 9.165
    MR-1011 0.068 0.020 2.608 2.361 -0.927 2.334 16.35 12.31 -0.544 0.592 12.84 9.997
    MR-1010 0.031 0.022 2.427 2.072 0.639 2.207 16.01 12.24 0.501 0.570 11.81 9.772
    MR-1001 0.032 0.023 2.899 2.508 0.603 2.208 16.37 13.06 0.501 0.571 12.92 9.328
    MR-0111 0.065 0.022 2.623 2.283 -0.909 2.341 15.24 13.52 -0.555 0.590 12.12 9.808
    MR-0110 0.029 0.024 2.487 2.904 0.629 2.203 16.20 12.73 0.492 0.573 11.42 9.629
    MR-0101 0.121 0.106 3.458 3.140 0.560 5.210 18.53 15.87 -1.064 1.371 16.54 12.61

     | Show Table
    DownLoad: CSV
    Table 2.  The biases, MSEs, RMSEs and MAEs (multiplied by 102) of different estimators for parameter β when sample size n=500.
    Method β1 β2 β3
    Bias MSE RMSE MAE Bias MSE RMSE MAE Bias MSE RMSE MAE
    IPW-1000 -0.071 0.030 1.732 1.401 0.202 1.023 10.11 8.104 -0.243 0.649 8.054 6.333
    IPW-0100 0.208 0.135 3.677 2.869 0.034 4.570 21.38 15.17 0.264 3.559 18.86 13.43
    AIPW-1010 0.014 0.027 1.649 1.334 -0.426 1.013 10.06 7.950 0.126 0.611 7.815 5.072
    AIPW-1001 -0.009 0.029 1.692 1.334 -0.173 0.964 9.820 7.813 0.364 0.556 7.457 5.692
    AIPW-0110 0.017 0.031 1.762 1.380 -0.026 0.870 9.328 7.529 -0.945 0.567 7.527 6.027
    AIPW-0101 3.621 165.3 128.6 38.69 -1.019 294.4 171.6 63.55 15.32 123.2 351.0 69.06
    MR-1111 0.040 0.018 1.665 1.308 0.257 0.950 9.379 7.060 -0.248 0.387 6.919 4.147
    MR-1110 -0.031 0.022 1.689 1.347 0.219 0.913 9.828 7.301 0.114 0.404 6.484 4.991
    MR-1101 0.039 0.021 1.669 1.386 -0.460 1.081 10.33 7.582 0.228 0.396 6.873 4.136
    MR-1011 0.057 0.020 1.670 1.302 0.265 0.893 9.785 8.108 0.220 0.382 6.622 4.618
    MR-1010 0.027 0.015 1.537 1.256 -0.071 0.860 9.603 7.096 0.265 0.359 6.140 5.054
    MR-1001 0.044 0.026 1.714 1.350 0.441 1.039 10.36 8.224 0.326 0.462 7.427 6.043
    MR-0111 -0.028 0.031 1.746 1.371 0.329 1.014 9.751 7.301 0.164 0.406 7.520 5.150
    MR-0110 0.034 0.029 1.657 1.395 0.480 1.076 9.560 8.061 0.326 0.441 7.293 5.875
    MR-0101 1.027 0.103 2.853 2.331 0.537 4.031 11.39 10.57 0.931 1.230 11.36 9.480

     | Show Table
    DownLoad: CSV

    It can be seen from the two tables that regardless of the estimation method, the larger the sample size, the better the estimation effect. And when the models for estimating the selection probability and the conditional expectation are all specified correctly, the estimated results obtained by the multiple robust estimation method, the inverse probability weighted estimation method and the augmented inverse probability weighted estimation method are not much different, but the effect of multiple robust estimation is better in terms of MSE. When all the models for estimating the selection probability and the conditional expectation are specified incorrectly, the AIPW0101 has unsatisfactory effects, the resulted estimators have larger deviations, but our proposed MRE0101, despite using two incorrect models, can generate better estimators. The interesting observation that ˆβMR seems to still provide a reasonable (at least not too bad) estimate of β even if there is no model correctly specified is similar to Han [22]. In a word, it is obvious that our proposed multiple robust estimation method is better than the two competitors.

    Before we give the proof of Theorem 2.1, some notations and interpretations are presented firstly.

    Let Φ(t)=E[RXZT|T=t], Ψ(t)=E[RXXT|T=t], then

    θ(Ti)={Ψ(Ti)}1{E[RiXiYi|Ti]Φ(Ti)β}. (4.1)

    Substituting (4.1) into (2.1), we obtain

    ˇYi=ˇZTiβ+εi, (4.2)

    where ˇYi=YiXTig(Ti), ˇZi=ZiμT(Ti)Xi, with g(Ti)={Ψ(Ti)}1E[RiXiYi|Ti], μ(Ti)={Ψ(Ti)}1Φ(Ti). From model (4.2), using the complete data, the CC estimator of β can be obtained by solving the following estimation equation

    ni=1Riξi(β)=0,

    where ξi(β)=ˇZi(ˇYiˇZTiβ)=(ZiμT(Ti)Xi)[YiXTig(Ti)(ZiμT(Ti)Xi)Tβ], and E[ξi(β)]=0.

    Suppose C to be a positive constant which can represent different values, and assume the following conditions C1–C5 hold.

    C1 The bandwidth h satisfies h=Cn1/5, that is h0 and nh as n, where C>0 is a given positive constant.

    C2 The kernel function K() is a symmetric probability kernel function, and t2K(t)dt0, t4K(t)dt<.

    C3 For each t(0,1), f(t),Φ(t),Ψ(t) and θ(t) are twice continuous differentiable at point t, where f(t) is the density function of the variable T.

    C4 sup0t1E[ε4i|Ti=t]<, sup0t1E[X4ir|Ti=t]<, and they are continuous about t, where Xir is the r-th component of Xi, i=1,,n, r=1,,q.

    C5 For a given t, Ψ(t) is a positive definite matrix.

    Next, a Lemma is needed in proof of Theorem 2.1, and the proof can be found in Zhao [4].

    Lemma 4.1. Suppose conditions C1–C5 hold, then we have

    sup0<t<1ˆμ(t)Ψ(t)1Φ(t)=Op(Cn),
    sup0<t<1ˆg(t)Ψ(t)1Φ(t)βθ(t)=Op(Cn),

    where Cn=h2+(log(1/h)nh)1/2.

    Proof of Theorem 2.1: Assuming that P contains a model that correctly specifies π(V,S), without loss of generality, let π1(α1) be the model, α10 represents the truth value of α1, that is π1(α10)=π(V,S). Next, we combine the theory of empirical likelihood to prove that ˆβMR is a consistent estimator of β.

    Referring to the method in Han [22] to establish the relationship between the weight ˆωi and the empirical likelihood on the biased sample. Let pi represent the conditional empirical probability on the biased sample (Yi,Xi,Zi,Ti,Si),Ri=1,i=1,,m, based on (2.8),(2.9) and ω(V,S)=1π1(α10), a more reasonable value of pi can be given by the following constrained optimization problem:

    maxp1,,pmmi=1pi;pi0,i=1,,m;mi=1pi=1,
    mi=1pi{πij(ˆαj)νj(ˆαj)}/π1i(ˆα1)=0,j=1,,J,
    mi=1pi{ˆUki(ˆβk,ˆγk)ηk(ˆβk,ˆγk)}/π1i(ˆα1)=0,k=1,,K.

    Using the Lagrange multiplier method again, we get

    ˆpi=1m11+ˆλTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1),i=1,,m,

    where ˆλT=(ˆλ1,,ˆλJ+pK) is the (J+pK)-dimensional Lagrange multiplier, and satisfies

    1mmi=1ˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)1+λTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)=0.

    Due to the non-negativity of ˆpi, ˆλ satisfies 1+ˆλTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)>0,i=1,,m. Since

    1mmi=1ˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)1+λTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)=1ν1(ˆα1)1mmi=1ˆgi(ˆα,ˆβ,ˆγ)1+π1i(ˆα1)ν1(ˆα1)ν1(ˆα1)+{λν1(ˆα1)}Tˆgi(ˆα,ˆβ,ˆγ)=1ν1(ˆα1)1mmi=1ˆgi(ˆα,ˆβ,ˆγ)1+{λ1+1ν1(ˆα1),λ2ν1(ˆα1),,λJ+pKν1(ˆα1)}ˆgi(ˆα,ˆβ,ˆγ),

    then the solution of (2.11), ˆρ, can be written as ˆρ1=(ˆλ1+1)/ν1(ˆα1) and ˆρl=ˆλl/ν1(ˆα1),l=2,,J+pK. Therefore

    ˆωi=1mν1(ˆα1)/π1i(ˆα1)1+ˆλTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)=ˆpiν1(ˆα1)π1i(ˆα1).

    Just like White [24], let αj, βk and γk are the minimum points of the corresponding Kullback-Leibler distance respectively, then we have ˆαjPαj,ˆβkPβk,ˆγkPγk, and n1/2(ˆαjαj), n1/2(ˆβkβk) and n1/2(ˆγkγk) are bounded by probability. At the same time, νj(ˆαj)Pνj, ηk(ˆβk,ˆλk)Pμk, where νj=E[πj(αj)], μk=E[Uk(βk,γk)]. Generally speaking, when the model πj(αj) for π(V,S) is correctly specified, we have πj(αj)=π(V,S), and when the model ak(γk) for E(Y|V,T,S) is correctly specified, we have ak(γk)=E(Y|V,T,S). Let αT={(α1)T,,(αJ)T}, βT={(β1)T,,(βK)T}, γT={(γ1)T,,(γK)T}, and suppose ˆρPρ.

    Based on the empirical likelihood theory, it can be known that ˆλP0. According to the appendix in Han [22], ˆλ=Op(n1/2) holds. Since the model π1(α1) is correct, then we have mnPν1, and

    mi=1ˆωiˆξi(β)=1mni=1Riν1(ˆα1)/π1i(ˆα1)1+ˆλTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)ˆξi(β)=ν1(ˆα1)mni=1Ri/π1i(ˆα1)1+ˆλTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)ˆξi(β)=ν1mni=1Ri/π1i(ˆα1)1+ˆλTˆgi(ˆα,ˆβ,ˆγ)/π1i(ˆα1)ˆξi(β)=1nni=1Riπ1i(α1)ˆξi(β)+op(1).

    Refer to Zhao [4], since

    ˆξi(β)=[ZiˆμT(Ti)Xi]εi+[μ(Ti)ˆμ(Ti)]TXiεi+[ZiˆμT(Ti)Xi]XTi[θ(Ti)ˆg(Ti)+ˆμ(Ti)β]+[μ(Ti)ˆμ(Ti)]TXiXTi[θ(Ti)ˆg(Ti)+ˆμ(Ti)β],ξi(β)=[ZiμT(Ti)Xi][ZiμT(Ti)Xi]Tβ+[ZiμT(Ti)Xi]εi,

    and E[Xiεi]=0, E[(ZiμT(Ti)Xi)Xi)T]=0, we have

    ˆξi(β)ξi(β)=[μ(Ti)ˆμ(Ti)]TXiεi+[μ(Ti)ˆμ(Ti)]TXiXTi[θ(Ti)ˆg(Ti)+ˆμ(Ti)β]+[ZiμT(Ti)Xi][XTiθ(Ti)XTiˆg(Ti)+XTiˆμ(Ti)βZTiβ+XTiμ(Ti)β].

    Combine conditions C1, C4, C5 and Lemma 4.1, we have

    1nni=1Riπ1i(α1)ˆξi(β)1nni=1Riπ1i(α1)ξi(β)P0.

    That is, 1nni=1Riπ1i(α1)ˆξi(β)P1nni=1Riπ1i(α1)ξi(β). Then we have

    mi=1ˆωiˆξi(β)=1nni=1Riπ1i(α1)ˆξi(β)+op(1)=1nni=1Riπ1i(α1)ξi(β)+op(1)PE[Rπ(V,S)ξ(β)]=0.

    Therefore, when n, β is the solution of the formula (2.13), which shows that ˆβMR is a consistent estimator of β.

    Next, suppose that F contains a model that correctly specifies E(Y|V,T,S). Without loss of generality, let a1(γ1) be the true model and γ10 be the true value of γ1, that is a1(γ10)=E(Y|V,T,S), and γ1=γ10. A previous constraint is actually

    mi=1ˆωiˆU1i(ˆβ1,ˆγ1)=1nni=1ˆU1i(ˆβ1,ˆγ1),

    and ˆβ1Pβ1=β, so we get 1nni=1ˆU1i(ˆβ1,ˆγ1)P0.

    Let g(α,β,γ)T=[π1(α1)ν1,,πJ(αJ)νJ,{U1(β1,γ1)η1}T,,{UK(βK,γK)ηK}T], due to 1nni=1ˆU1i(ˆβ1,ˆγ1)P1nni=1U1i(β,γ10),1nni=1[ˆξi(β)ξi(β)]P0, and E[U1(β,γ10)]=0, then we have

    mi=1ˆωiˆξi(β)=mi=1ˆωi{ˆξi(β)ˆU1i(ˆβ1,ˆγ1)}+1nni=1ˆU1i(ˆβ1,ˆγ1)=1mni=1Riˆξi(β)ˆU1i(ˆβ1,ˆγ1)1+ˆρTˆgi(ˆα,ˆβ,ˆγ)+E[U1(β,γ10)]+op(1)=1P(R=1)E[Rξ(β)U1(β,γ10)1+ρTg(α,β,γ)]+op(1)=1P(R=1)E{E[Rξ(β)U1(β,γ10)1+ρTg(α,β,γ)|Y,V,T,S]}+op(1)P0.

    This shows that ˆβMR is a consistent estimator of β.

    So the proof of Theorem 2.1 is completed.

    In this article, we have proposed the multiple robust estimators for parameters in varying-coefficient partially linear model with missing response at random, and the multiple robustness of our proposals has been shown theoretically under some regular conditions. Our simulation studies fully demonstrate the superiority of our multiple robust estimation method through Table 1 and Table 2. Finally, we point out some problems for the future researches. First, we only discuss the multiple robust estimation process of parameters, and the fitting of nonparametric function curves can be expanded in the future studies. Next, based on the model in this article, if the missing mechanism is nonignorable missing, how to obtain the robust estimation of parameters is also worth studying.

    The research is supported by NSF projects (ZR2021MA077 and ZR2019MA016) of Shandong Province of China.

    The authors declare that they have no conflicts of interest to this work.



    [1] J. Fan, T. Huang, Profile Likelihood Inferences on Semiparametric Varying-Coefficient Partially Linear Models, Bernoulli, 11 (2005), 1031–1057. https://doi.org/10.3150/bj/1137421639 doi: 10.3150/bj/1137421639
    [2] J. You, Y. Zhou, Empirical likelihood for semiparametric varying-coefficient partially linear regression models, Stat. Probabil. Lett., 76 (2006), 412–422. https://doi.org/10.1016/j.spl.2005.08.029 doi: 10.1016/j.spl.2005.08.029
    [3] Z. Huang, R. Zhang, Empirical likelihood for nonparametric parts in semiparametric varying-coefficient partially linear models, Stat. Probabil. Lett., 79 (2009), 1798–1808. https://doi.org/10.1016/j.spl.2009.05.008 doi: 10.1016/j.spl.2009.05.008
    [4] P. Zhao, Infrerence for semiparametric varying coefficient partially linear models, doctoral dissertation, Beijing: Beijing University of Technology, 2010.
    [5] J. Feng, R. Zhang, Y. Lu, Inference on varying-coefficient partially linear regression model, Acta Math. Appl. Sin-E, 31 (2015), 139–156. https://doi.org/10.1007/s10255-015-0457-5 doi: 10.1007/s10255-015-0457-5
    [6] R. Little, D. Rubin, Statistical analysis with missing data, 1 Eds., New York: Wiley Press, 1986.
    [7] D. Horvitz, D. Thompson, A generalization of sampling without replacement from a finite universe, J. Am. Stat. Assoc., 47 (1952), 663–685. https://doi.org/10.1080/01621459.1952.10483446 doi: 10.1080/01621459.1952.10483446
    [8] F. Yates, The analysis of replicated experiments when the field results are incomplete, Exp. Agr., 1 (1933), 129–142.
    [9] P. Cheng, Nonparametric estimation of mean functionals with data missing at random, J. Am. Stat. Assoc., 89 (1994), 81–87. https://doi.org/10.1080/01621459.1994.10476448 doi: 10.1080/01621459.1994.10476448
    [10] Q. Wang, J. Rao, Empirical likelihood for linear regression models under imputation for missing responses, Can. J. Stat., 29 (2001), 596–608. https://doi.org/10.2307/3316009 doi: 10.2307/3316009
    [11] Q. Wang, J. Rao, Empirical likelihood-based inference under imputation for missing response data, Ann. Stat., 30 (2002), 896–924. https://doi.org/10.1214/aos/1028674845 doi: 10.1214/aos/1028674845
    [12] Q. Wang, O. Linton, W. Hardle, Semi-parametric regression analysis with missing response at random, J. Am. Stat. Assoc., 99 (2004), 334–345. https://doi.org/10.1198/016214504000000449 doi: 10.1198/016214504000000449
    [13] J. Robins, A. Rotnitzky, L. Zhao, Estimation of Regression Coefficients When Some Regressors Are Not Always Observed, J. Am. Stat. Assoc., 89 (1994), 846–866. https://doi.org/10.1080/01621459.1994.10476818 doi: 10.1080/01621459.1994.10476818
    [14] J. Robins, A. Rotnitzky, Semiparametric efficiency in multivariate regression models with missing data, J. Am. Stat. Assoc., 90 (1995), 122–129. https://doi.org/10.1080/01621459.1995.10476494 doi: 10.1080/01621459.1995.10476494
    [15] D. Scharfstein, A. Rotnitzky, R. Robins, Adjusting for Nonignorable Drop-Out Using Semiparametric Nonresponse Models, J. Am. Stat. Assoc., 94 (1999), 1096–1120. https://doi.org/10.1080/01621459.1999.10473862 doi: 10.1080/01621459.1999.10473862
    [16] J. Kang, J. Schafer, Demystifying double robustness: A comparison of alternative strategies for estimating a population mean from incomplete data (with Discussion), Statistics, 22 (2007), 523–539. https://doi.org/10.1214/07-STS227 doi: 10.1214/07-STS227
    [17] J. Qin, J. Shao, B. Zhang, Efficient and doubly robust imputation for covariate-dependent missing responses, J. Am. Stat. Assoc., 103 (2008), 797–810. https://doi.org/10.1198/016214508000000238 doi: 10.1198/016214508000000238
    [18] W. Cao, A. Tsiatis, M. Davidian, Improving efficiency and robustness of the doubly robust estimator for a population mean with incomplete data, Biometrika, 96 (2009), 723–734. https://doi.org/10.1093/biomet/asp033 doi: 10.1093/biomet/asp033
    [19] P. Han, A note on improving the efficiency of inverse probability weighted estimator using the augmentation term, Stat. Probabil. Lett., 82 (2012), 2221–2228. https://doi.org/10.1016/j.spl.2012.08.005 doi: 10.1016/j.spl.2012.08.005
    [20] A. Rotnitzky, Q. Lei, M. Sued, J. Robins, Improved double-robust estimation in missing data and causal inference models, Biometrika, 99 (2012), 439–456. https://doi.org/10.1093/biomet/ass013 doi: 10.1093/biomet/ass013
    [21] P. Han, L. Wang, Estimation with missing data: Beyond double robustness, Biometrika, 100 (2013), 417–430. https://doi.org/10.1093/biomet/ass087 doi: 10.1093/biomet/ass087
    [22] P. Han, Multiply Robust Estimation in Regression Analysis With Missing Data, J. Am. Stat. Assoc., 109 (2014), 1159–1173. https://doi.org/10.1080/01621459.2014.880058 doi: 10.1080/01621459.2014.880058
    [23] Y. Sun, L. Wang, P. Han, Multiply robust estimation in nonparametric regression with missing data, J. Nonparametr. Stat., 32 (2020), 73–92. https://doi.org/10.1080/10485252.2019.1700254 doi: 10.1080/10485252.2019.1700254
    [24] P. Han, Calibration and multiple robustness when data are missing not at random, Stat. Sinica, 109 (2018), 1725–1740. https://doi.org/10.5705/ss.202015.0408 doi: 10.5705/ss.202015.0408
    [25] W. Li, S. Yang, P. Han, Robust estimation for moment condition models with data missing not at random, J. Stat. Plan. Infer., 207 (2020), 246–254. https://doi.org/10.1016/j.jspi.2020.01.001 doi: 10.1016/j.jspi.2020.01.001
    [26] H. White, Maximum likelihood estimation of misspecified models, Econometrica, 50 (1982), 1–25. https://doi.org/10.2307/1912526 doi: 10.2307/1912526
  • Reader Comments
  • © 2022 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2815) PDF downloads(70) Cited by(0)

Figures and Tables

Tables(2)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog