Abstract
This paper develops a robust and efficient estimation procedure for quantile partially linear additive models with longitudinal data, where the nonparametric components are approximated by B spline basis functions. The proposed approach can incorporate the correlation structure between repeated measures to improve estimation efficiency. Moreover, the new method is empirically shown to be much more efficient and robust than the popular generalized estimating equations method for non-normal correlated random errors. However, the proposed estimating functions are non-smooth and non-convex. In order to reduce computational burdens, we apply the induced smoothing method for fast and accurate computation of the parameter estimates and its asymptotic covariance. Under some regularity conditions, we establish the asymptotically normal distribution of the estimators for the parametric components and the convergence rate of the estimators for the nonparametric functions. Furthermore, a variable selection procedure based on smooth-threshold estimating equations is developed to simultaneously identify non-zero parametric and nonparametric components. Finally, simulation studies have been conducted to evaluate the finite sample performance of the proposed method, and a real data example is analyzed to illustrate the application of the proposed method.
Similar content being viewed by others
References
Brown BM, Wang YG (2005) Standard errors and covariance matrices for smoothed rank estimators. Biometrika 92:149–158
Carroll RJ, Maity A, Mammen E, Yu K (2009) Efficient semiparametric marginal estimationfor partially linear additive model for longitudinal/clustered data. Stat BioSci 1:10–31
Cheng G, Zhou L, Huang J (2014) Efficient semiparametric estimation in generalized partially linear additive models for longitudinal/clustered data. Bernoulli 20:141–163
Fan J, Li R (2001) Variable selection via nonconcave penalized likelihood and its oracle properties. J Am Stat Assoc 96:1348–1360
Fu L, Wang YG (2012) Quantile regression for longitudinal data with a working correlation model. Comput Stat Data Anal 56:2526–2538
Fu L, Wang YG, Zhu M (2015) A Gaussian pseudolikelihood approach for quantile regression with repeated measurements. Comput Stat Data Anal 84:41–53
Guo J, Tang M, Tian M, Zhu K (2013) Variable selection in high-dimensional partially linear additive models for composite quantile regression. Comput Stat Data Anal 66:56–67
Härdle W, Liang H, Gao J (2000) Partially linear models. Springer, Heidelberg
Härdle WK, Müller M, Sperlich S, Werwatz A (2004) Nonparametric and semiparametric models. Springer, New York
Henry K, Erice A, Tierney C, Balfour HH Jr, Fischl MA, Kmack A, Liou SH, Kenton A, Hirsch MS, Phair J, Martinez A, Kahn JO (1998) A randomized, controlled, double-blind study comparing the survival benefit of four different reverse transcriptase inhibitor therapies (three-drug, two-drug, and alternating drug) for the treatment of advanced AIDS. J Acquir Immune Defic Syndr Hum Retrovirol19:339–349
Jung SH (1996) Quasi-likelihood for median regression models. J Am Stat Assoc 91:251–257
Koenker R, Bassett G Jr (1978) Regression quantiles. Econometrica 46:33–55
Lai P, Wang Q, Lian H (2012) Bias-corrected GEE estimation and smooth-threshold GEE variable selection for single-index models with clustered data. J Multivar Anal 105:422–432
Li G, Lian H, Feng S, Zhu L (2013) Automatic variable selection for longitudinal generalized linear models. Comput Stat Data Anal 61:174–186
Lian H (2012) Variable selection in high-dimensional partly linear additive models. J Nonparametr Stat 24:825–839
Lian H, Liang H, Wang L (2014) Generalized additive partial linear models for clustered data with diverging number of covariates using GEE. Stat Sin 24:173–196
Liu X, Wang L, Liang H (2011) Variable selection and estimation for semiparametric additive partial linear models. Stat Sin 21:1225–1248
Lu X, Fan Z (2015) Weighted quantile regression for longitudinal data. Comput Stat. 30:569–592. doi:10.1007/s00180-014-0550-x
Ma S, Yang L (2011) Spline-backfitted kernel smoothing of partially linear additive model. J Stat Plann Inference 141:204–219
Ma S, Song Q, Wang L (2013) Simultaneous variable selection and estimation in semiparametric modeling of longitudinal/clustered data. Bernoulli 19:252–274
Pollard D (1990) Empirical processes: theory and applications. In: NSF-CBMS regional conference series in probability and statistics. Institute of Mathematical Statistics, Hayward, CA
Schumaker L (1981) Spline functions: basic theory. Wiley, New York
Tang CY, Leng C (2011) Empirical likelihood and quantile regression in longitudinal data analysis. Biometrika 98:1001–1006
Tibshirani R (1996) Regression shrinkage and selection via the LASSO. J R Stat Soc B 58:267–288
Ueki M (2009) A note on automatic variable selection using smooth-threshold estimating equations. Biometrika 96:1005–1011
Wang L, Liu X, Liang H, Carroll RJ (2011) Estimation and variable selection for generalized additive partial linear models. Ann Stat 39:1827–1851
Yin G, Cai J (2005) Quantile regression models with multivariate failure time data. Biometrics 61:151–161
Zhao P, Li G (2013) Modified SEE variable selection for varying coefficient instrumental variable models. Stat Methodol 12:60–70
Zou H (2006) The adaptive LASSO and its oracle properties. J Am Stat Assoc 101:1418–1429
Acknowledgments
We are grateful for the insightful comments from the anonymous reviewers and editors, which have greatly helped improve the quality of this paper. This work is supported by the Chongqing University Postgraduates’ Innovation Project, the National Natural Science Foundation of China (Grant No. 11171361) and Ph.D. Programs Foundation of Ministry of Education of China (Grant No. 20110191110033).
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
To establish the asymptotic properties of the proposed estimators, the following regularity conditions are needed in this paper.
-
(C1)
\(E{g_l}({z_l}) = 0\) and \({g _l} \in {\mathscr {H}_r}\), \(l=1,\ldots ,d\) for some \(r>1/2\), where \(\mathscr {H}_r\) is the collection of all functions on [0, 1] whose \(\rho \)th order derivative satisfies the H\(\ddot{o}\)lder condition of the order v with \(r \equiv \rho + v\) and \(0<v\le 1\).
-
(C2)
There exist constants \(\delta _1\) and \(\delta _2\) such that the marginal density \(f_l(z_l)\) of \(Z_l\) satisfies \(0< \delta _1 \le f_l(z_l) \le \delta _2 < \infty \) on [0, 1] for every \(l=1,\ldots ,d\). The joint density \(f_{ll'}(z_l,z_{l'})\) of \((Z_{l},Z_{l'})\) satisfies \(0< \delta _1 \le f_{ll'}(z_l,z_{l'})\le \delta _2 < \infty \) for all \((z_l,z_{l'})\in [0,1]^2\), \(1\le l\ne {l'}\le d\).
-
(C3)
The dimensions p, d of \(\varvec{x}_{ij}\) and \(\varvec{z}_{ij}\) are fixed and \(\max n_i\) is bounded when \(m\rightarrow \infty \). The distribution functions \(F_{ij}(t)=p( {y_{ij}} - \varvec{x}_{ij}^T{\varvec{\beta } } - \sum \nolimits _{l = 1}^d {{g_l}({z_{ijl}}) \le t\left| {{\varvec{x}_{ij}},{\varvec{z}_{ij}}} \right. } )\) are absolutely continuous, with continuous densities \(f_{ij}\) uniformly bounded, and its first derivative \({f'_{ij}}( \cdot )\) uniformly bounded away from 0 and \(\infty \) at the points \(0, i=1,\ldots ,m, j=1,\ldots ,n_i\).
-
(C4)
For any positive definite matrix \(\varvec{W}_i\), \(\mathrm{{ }}{m^{ - 1}}\sum \nolimits _{i = 1}^m {\varvec{H}_i^T{\varvec{\Lambda }_i}{\varvec{W}_i}} {\varvec{\Lambda }_i}{\varvec{H}_i}\) converges to a positive definite matrix, where \(\varvec{\Lambda } _i\) is an \(n_i \times n_i\) diagonal matrix with the jth diagonal element \(f_{ij}(0)\) and \({\sup _i}\Vert {{\varvec{H}_i}}\Vert < + \infty \).
-
(C5)
Matrix \(\varvec{\Omega }\) is positive definite and \(\varvec{\Omega }=O(\frac{1}{m})\).
-
(C6)
The parameter space \(\Theta \) is compact, and the true parameter vector \({\varvec{\theta } _0}\) is an interior point in \(\Theta \).
-
(C7)
The differentiation of \({{\tilde{U}}_\tau }(\varvec{\theta } )\), \(\frac{{\partial {{\tilde{U}}_\tau }(\varvec{\theta } )}}{{\partial \varvec{\theta } }}\) is positive definite with probability 1.
-
(C8)
Let \(\lambda _{max}=\mathrm{max}(\lambda _1,\lambda _2)\) and \(\lambda _{min}=\mathrm{min}(\lambda _1,\lambda _2)\). The tuning parameters \(\lambda _{min}\) and \(\lambda _{max}\) satisfy \(m^{r/(2r+1)}\lambda _{max} \rightarrow 0\), and \(m^{(1+\kappa )r/(2r+1)}\lambda _{min} \rightarrow \infty \), as \(m \rightarrow \infty \).
Proof of Theorem 1
Let \(\varvec{G}_i^T = \varvec{H}_i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}\), so \({U_\tau }(\varvec{\theta } ) = - \sum \nolimits _{i = 1}^m {\varvec{G}_i^T} {\varvec{S}_i}\). Let \({{\bar{U}}_\tau }(\varvec{\theta } )= - \sum \nolimits _{i = 1}^m {\varvec{G}_i^T} {\varvec{P}_i}\) with \({\varvec{P}_i} = {(\tau - p({y_{i1}} - \varvec{h}_{i1}^T\varvec{\theta } \le 0),\ldots ,\tau - p({y_{i{n_i}}} - \varvec{h}_{i{n_i}}^T\varvec{\theta } \le 0))^T}\). We can obtain
where \(\varvec{g}_{ij}\) is a \((p+dK)\times 1\) vector and \(\varvec{G}_i^T = ({\varvec{g}_{i1}},\ldots ,{\varvec{g}_{i{n_i}}})\). Under condition (C4) and from the uniform strong law of large numbers (Pollard 1990), we have
Therefore, \(\mathop {\sup }\limits _{\varvec{\theta } \in \Theta } \Vert {m^{ - 1}}\{ {{{\bar{U}}_\tau }(\varvec{\theta } ) - {U_\tau }(\varvec{\theta } )} \}\Vert = o(m^{-1/2})\) a.s.. Now focusing on \({{\bar{U}}_\tau }(\varvec{\theta } )\),
is positive definite. Since \(p({y_{ij}} - \varvec{h}_{ij}^T{\varvec{\theta } _0} \le 0) = \tau \) by condition (C3), \({\varvec{\theta } _0}\) is the unique solution of the equation \({{\bar{U}}_\tau }(\varvec{\theta } )=\varvec{0}\). Due to \(\varvec{\hat{\theta }}\) is the solution of the equation \( {U_\tau }(\varvec{\theta } )=\varvec{0}\) together with condition (C4), hence \(\varvec{\hat{\theta }} \rightarrow {\varvec{\theta } _0}\) as \(m\rightarrow +\infty \). For any \(\varvec{\theta }\) satisfying \(\Vert {\varvec{\theta } - {\varvec{\theta } _0}}\Vert \) \(<\) cm\(^{-1/3}\),
The first term can be written as
According to Lemma 3 of Jung (1996), we have
Then the first term
From the law of large numbers (Pollard 1990) the second term
Therefore, \({U_\tau }(\varvec{\theta } ) - {U_\tau }({\varvec{\theta } _0}) = {{\bar{U}}_\tau }(\varvec{\theta } ) + {o_p}(\sqrt{m} )\). By Taylor’s expansion of \({{\bar{U}}_\tau }(\varvec{\theta } ) \) together with \({{\bar{U}}_\tau }(\varvec{\theta }_0 )=0\), we can obtain
Because \(\varvec{\hat{\theta }}\) is in the \(m^{-1/3}\) neighborhood of \(\varvec{\theta }_0\) and \({U_\tau }(\varvec{\hat{\theta }} ) = \varvec{0}\), we have
To obtain the closed form expression of \(\varvec{\hat{\beta }}\), similar to Ma et al. (2013), we write the inverse of \(\varvec{D}_\tau ({\varvec{\theta } _0})=\sum \nolimits _{i = 1}^m {\varvec{H}_i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}{\varvec{\Lambda } _i}} {\varvec{H}_i}\) as the following block form:
where \({\varvec{D}^{11}} = {(\varvec{D_{XX}} - \varvec{D_{X\Psi }}\varvec{D}_{\Psi \Psi }^{ - 1} \varvec{D_{\Psi X}})^{ - 1}},{\varvec{D}^{22}} = {(\varvec{D_{\Psi \Psi }} - \varvec{D_{\Psi X}}\varvec{D_{XX}}^{ - 1}\varvec{D_{X\Psi }})^{ - 1}}, {\varvec{D}^{12}} = - {\varvec{D}^{11}}\varvec{D_{X\Psi }}\varvec{D_{\Psi \Psi } }^{ - 1}\) and \({\varvec{D}^{21}} = - {\varvec{D}^{22}}\varvec{D_{\Psi X}}\varvec{D_{XX}}^{ - 1}.\) Furthermore, let \({U_\tau }({\varvec{\theta } _0}) = {(\varvec{U}_1^T,\varvec{U}_2^T)^T},{\varvec{U}_1} = \sum \nolimits _{i = 1}^m {-\varvec{X}_i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}{\varvec{S}_i}({\varvec{\theta } _0})} ,{\varvec{U}_2} = \sum \nolimits _{i = 1}^m -\varvec{\Psi } _i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}{\varvec{S}_i}({\varvec{\theta } _0}) \). Then
That is,
Let \(\varvec{\tilde{P}} = \varvec{X({X^T}}\varvec{\Lambda } {\varvec{V^{ - 1}}\varvec{\Lambda }\varvec{X}})^{-1}{\varvec{X}^T}\varvec{\Lambda }{\varvec{V^{ - 1}}}\varvec{\Lambda }\), \( \varvec{\tilde{\Psi }} = (\varvec{I - \tilde{P}})\varvec{\Psi }\) with \(\varvec{\tilde{\Psi } }= {(\varvec{\tilde{\Psi }_1}^T,\ldots ,\varvec{\tilde{\Psi }_m}^T)^T}\), \({\varvec{D_\Psi }} = \mathop {\lim }\nolimits _{m \rightarrow \infty } \frac{1}{m}\sum \nolimits _{i = 1}^m {\varvec{\tilde{\Psi }_i}^T{\varvec{\Lambda _i}}\varvec{V_i^{ - 1}}{\varvec{\Lambda _i}}} {{\varvec{\tilde{\Psi }}_i}}\), \({\varvec{\Sigma _\Psi }} = \mathop {\lim }\nolimits _{m \rightarrow \infty } \frac{1}{m}\sum \nolimits _{i = 1}^m \varvec{\tilde{\Psi }_i}^T{\varvec{\Lambda _i}}\varvec{V_i^{ - 1}}\varvec{\Sigma }_i \varvec{V_i^{ - 1}}{\varvec{\Lambda _i}} {{\varvec{\tilde{\Psi }}_i}}\). Because \(\varvec{S}_i\) are independent random variables with mean zero, and
The multivariate central limit theorem implies that
Moreover, we have
by the law of large numbers. Then, by using Slutskys theorem, it follows that
This complete the proof of part (i). Now, we prove part (ii).
In order to obtain the convergence rate of \(\hat{g}_l(\cdot )\), let \(u_{ij} = \sum \nolimits _{l = 1}^d {g _{0l}}(z_{ijl}) -\varvec{\psi }_{ij}^T\varvec{\gamma }_{0}\). By the conditions (C1), (C2) and Corollary 6.21 in Schumaker (1981), we have \(|u_{ij}|=O\left( {{K^{ - r}}} \right) \). From the asymptotic normality of \(\varvec{\hat{\gamma } }\), we have
The triangular inequality implies that
This completes the proof. \(\square \)
Proof of Theorem 2
Because \({{\tilde{S}}_{ij}} - {S_{ij}} = {\mathop {\mathrm{sgn}}} ( - {d_{ij}})\Phi ( - \left| {{d_{ij}}} \right| ),\) where \(sgn(\cdot )\) is the sign function and \(d_{ij}=(\varepsilon _{ij}+u_{ij})/r_{ij}\). We can obtain
where \(\varvec{g}_{ij}\) is the jth column of \(\varvec{H}_i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}\). Because
where \(\zeta (t)\) is between 0 and \({r _{ij}}t-u_{ij}\). Since \(\int _{ - \infty }^{ + \infty } {\Phi ( - \left| t \right| )} \{ {2I(t \le 0) - 1}\}dt = 0\), and by condition (C3), there exists a constant C such that \({\sup _{ij}}\left| {{f'_{ij}}(\zeta (t))} \right| \le C\). Then
Because \(\int _{ - \infty }^{ + \infty } {\Phi ( - \left| t \right| )} \left| t \right| dt = {1 / 2}\), thus
Under conditions (C4) and (C5), as \(m \rightarrow \infty \),
In addition,
By Cauchy-Schwartz inequality,
For each \(j=1,\ldots ,n_i\),
where \(\Delta \) is a positive value, and \(\xi \) lies between \((-r_{ij}\Delta -u_{ij},r_{ij}\Delta -u_{ij})\). Let \(\Delta =m^{1/3}\). Under condition (C5), since \(r_{ij}=O(m^{-1/2})\), then \(r_{ij}\Delta =O(m^{-1/6})\). As \(m\rightarrow \infty \), both \({\Phi ^2}( - \Delta ) \) and \( {r _{ij}}\Delta {f_{ij}}(\xi )\) go to 0. By conditions (C4) and (C6), it is easy to obtain \({m^{ - 1}}{\mathop {\mathrm{var}}} \left[ {{{\tilde{U}}_\tau }(\varvec{\theta } ) - {U_\tau }(\varvec{\theta } )} \right] =o(1)\). Therefore, we have \(m^{ - 1/2}\left[ {{{\tilde{U}}_\tau }(\varvec{\theta } ) - {U_\tau }(\varvec{\theta } )} \right] \rightarrow 0\) as \(m\rightarrow \infty \) for any \(\theta \). We complete the proof of Theorem 2. \(\square \)
Proof of Theorem 3
Since \({\sup _{\varvec{\theta } \in \Theta }}\left\| {{m^{ - 1}}\left[ {{{\bar{U}}_\tau }(\varvec{\theta } ) - {U_\tau }(\varvec{\theta } )} \right] } \right\| = o({m^{{{ - 1} / 2}}} )\) a.s. and Theorem 2, we can obtain \({\sup _{\varvec{\theta } \in \Theta }}\left\| {{m^{ - 1}}\left[ {{{\tilde{U}}_\tau }(\varvec{\theta } ) - {{\bar{U}}_\tau }(\varvec{\theta } )} \right] } \right\| = o({m^{{{ - 1} / 2}}} )\) by the triangle inequality. Since \(\varvec{\theta }_0\) is the unique solution of equation \({{\bar{U}}_\tau }(\varvec{\theta } )=0\). This together with the definite of \(\varvec{\tilde{\theta }}\) implies \(\varvec{\tilde{\theta }}\rightarrow \varvec{\theta }_0\) as \(m\rightarrow \infty \). In order to prove the normality of \(\varvec{\tilde{\theta }}\), we first prove that \({m^{ - 1}}\{ {{\tilde{\varvec{D}}_\tau }({\varvec{\theta } _0}) - {\varvec{D}_\tau }({\varvec{\theta } _0})}\}\mathop \rightarrow \limits ^p 0\), where \({\tilde{\varvec{D}}_\tau }( {{\varvec{\theta } _0}}) = {{\partial {{\tilde{U}}_\tau }( \varvec{\theta })} / {\partial \varvec{\theta } }}{|_{\varvec{\theta } = {\varvec{\theta } _0}}} = \sum \nolimits _{i = 1}^m {\varvec{H}_i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}{\varvec{\tilde{\Lambda } }_i}} {\varvec{H}_i}\). If we denote \(\varvec{G}_i^T = \varvec{H}_i^T{\varvec{\Lambda } _i}\varvec{V}_i^{ - 1}=(\varvec{g}_{i1},\ldots ,\varvec{g}_{in_i})\), where \(\varvec{g}_{ij}\) is a \((p+dK)\times 1\) vector, we can obtain that
Because
where \(\zeta _t\) lies between 0 and \(r _{ij}t-u_{ij}\). By condition (C3), \(f'_{ij}(\cdot )\) is uniformly bounded, hence there exists a constant C satisfying \(\left| {{f'_{ij}}({\zeta _t})} \right| \le C\), and by condition (C5), we have
By the strong law of large number, we know that \({m^{ - 1}}{\tilde{\varvec{D}}_\tau }({\varvec{\theta } _0}) \rightarrow E\{ {{m^{ - 1}}{\tilde{\varvec{D}}_\tau }({\varvec{\theta } _0})}\}\). Using the triangle inequality, we have
which is equivalent to \({{m^{ - 1}}\{ {{\tilde{\varvec{D}}_\tau }({\varvec{\theta } _0}) - {\varvec{D}_\tau }({\varvec{\theta } _0})}\}}\mathop \rightarrow \limits ^p 0.\) By Taylor series expansion of \({{{\tilde{U}}_\tau }({\varvec{\theta } })}\) around \(\varvec{\theta }_0\), we have
where \(\varvec{\theta } ^*\) lies between \(\varvec{\theta } \) and \(\varvec{\theta } _0\). Let \(\varvec{\theta }=\varvec{\tilde{\theta }}\). Because \({{\tilde{U}}_\tau }(\varvec{\tilde{\theta } }) = 0\) and \(\varvec{\tilde{\theta } } \rightarrow {\varvec{\theta } _0}\), we therefore obtain \(\varvec{\theta } ^* \rightarrow {\varvec{\theta } _0}\) and \({\tilde{\varvec{D}}_\tau }({\varvec{\theta } ^*}) \rightarrow {\tilde{\varvec{D}}_\tau }({\varvec{\theta } _0})\). By Theorem 2 and \({m^{ - 1}}\{ {{\tilde{\varvec{D}}_\tau }({\varvec{\theta } _0}) - {\varvec{D}_\tau }({\varvec{\theta } _0})}\} = o_p(1)\), we have
Applying Theorem 1, we can easily complete the proof Theorem 3. \(\square \)
Proof of Theorem 4
Let \({\delta _m}=m^{-r/(2r+1)}\), \(\varvec{\theta }={{\varvec{\theta }}_0}+{\delta _m}{\varvec{T}}\). Define \({ S_m}(\varvec{\theta } ) =({\varvec{I}_{p+dK}} - \varvec{\hat{\Delta }} ){\tilde{U}_\tau }(\varvec{\theta } )+\varvec{\hat{\Delta }}\varvec{\theta }\). We want to prove that \(\forall \varepsilon > 0\), there exists a constant \(C > 0\), such that
for m large enough. This will imply with probability at least \(1 - \varepsilon \) that there exists a solution of the equation \({S_m}(\varvec{\theta } )=\varvec{0}\) such that \(\Vert \varvec{\bar{\theta }} - {\varvec{\theta } _0}\Vert = {O_p}({\delta _m})\). The key of the proof is to evaluate the sign of \({\delta _m}{\varvec{T}^T}{S_m}({\varvec{\theta } _0} + {\delta _m}\varvec{T}) \) on \(\{ {{\varvec{\theta } _0} + {\delta _m}\varvec{T}:\Vert \varvec{T}\Vert = C}\}\). Note that
where \({\varvec{\tilde{\theta }} }\) is between \(\varvec{\theta }_0\) and \({{\varvec{\theta } _0} +{\delta _ m}\varvec{T}}\). Next we will consider \(I_{m1}\) and \(I_{m2}\) respectively. For \(I_{m1}\), by some basic calculations, we have
Using the Cauchy-Schwarz inequality, we have
Since \(\mathop {\min {{\hat{\delta } }}}\nolimits _{j \in \mathscr {A}} (\lambda _1 ,\kappa ) \le \mathop {\min {{\hat{\delta } }}}\nolimits _{j \in {\mathscr {A}_0}} (\lambda _1 ,\kappa )\) and \(\mathop {\min {{\hat{\delta } }}}\nolimits _{j \in {\mathscr {B}}} (\lambda _2 ,\kappa ) \le \mathop {\min {{\hat{\delta } }}}\nolimits _{j \in {{\mathscr {B}}_0}} (\lambda _2 ,\kappa )\), we only need to gain the convergence rate of \(\mathop {\min {{\hat{\delta } }}}\nolimits _{j \in {\mathscr {A}_0}} (\lambda _1 ,\kappa )\) and \(\mathop {\min {{\hat{\delta } }}}\nolimits _{j \in {{\mathscr {B}}_0}} (\lambda _2 ,\kappa )\). By Theorem 3, we know that the initial estimator \({\varvec{{\tilde{\gamma }} }}\) is \(m^{r/(2r+1)}\)-consistent. By using the condition \(m^{r/(2r+1)}\lambda _{max} \rightarrow 0\), for \(\varepsilon >0\) and \({j \in {{\mathscr {B}}_0}}\), we can derive that
which implies that \({{\hat{\delta } }_j}(\lambda _2 ,\kappa ) ={o_p}( {{m^{{{ - r} / {(2r + 1)}}}}})\) for \({j \in {{\mathscr {B}}_0}}\). Therefore, we can get \({\min _{j \in {\mathscr {B}}}}{{\hat{\delta } }_j}(\lambda _2 ,\kappa ) = {o_p}( {{m^{{{ - r} / {(2r + 1)}}}}})\). With the similar argument, we can obtain that \({\min _{j \in \mathscr {A}}}{{\hat{\delta } }_j}(\lambda ,\kappa ) = {o_p}( {{m^{{{ - r} / {(2r + 1)}}}}})\). Thus, \(|I_{m11}|= {O_p}(\sqrt{m}{\delta _m})\Vert \varvec{T } \Vert -{o_p}({\sqrt{m} \delta _m^2})\Vert \varvec{T }\Vert \). In addition, \({{\hat{\delta } }_j} = \min ( {1,{\lambda _1 / {{{\left| {{{\tilde{\beta } }_j}} \right| }^{1 + \kappa }}}}})\), \(j = 1,\ldots ,p\), and \({{\hat{\delta } }_j} = \min ( {1,{\lambda _2 / {\Vert {{\varvec{\tilde{\gamma } }_j}} \Vert _{\varvec{W}_j}^{1 + \kappa }}}})\), \(j= p + 1,\ldots ,p + d\). Thus, we have \(\Vert {{I_{m12}}}\Vert \le {\delta _m}\Vert \varvec{T}\Vert \Vert {{\varvec{\theta } _0}}\Vert =O_p(\delta _m\Vert \varvec{T}\Vert )\). Therefore, \(| {{I_{m1}}}\Vert = {O_p}(\sqrt{m}{\delta _m})\Vert \varvec{T }\Vert ={o_p}({m}{\delta _m^2})\Vert \varvec{T }\Vert \). For \(I_{m2}\), we can derive that
With the same argument, it’s easy to prove that \({I_{m22}} = {O_p}( \delta _m^2 ){\Vert \varvec{T}\Vert ^2}=o(m\delta _m^2\Vert \varvec{T} \Vert ^2)\). Thus \({\delta _m}{\varvec{T}^T}{S_m}({\varvec{\theta } _0} + {\delta _m}\varvec{T}) \) is asymptotically dominated in probability by \(I_{m21}\) on \(\{ {{\varvec{\theta } _0} + {\delta _m}\varvec{T}:\Vert \varvec{T}\Vert = C}\}\) for the sufficiently large C, and \(I_{m21}>0\) by condition (C7). This implies \(\Vert {\varvec{\bar{\theta } }- {\varvec{\theta } _0}}\Vert =O_p({m^{{{ - r} / {(2r + 1)}}}})\), which in turn immediately implies that \( {| \sum \nolimits _j{ \bar{g} _{j}}(z_j) - \varvec{\psi }_{ij}^T\varvec{\gamma }_0 |} =O_p({m^{{{ - r} / {(2r + 1)}}}})\) and \({|\sum \nolimits _j{ g _{0j}}(z_j) - \varvec{\psi }_{ij}^T\varvec{\gamma }_0| } = O_p({m^{{{ - r} / {(2r + 1)}}}})\). This complete the proof of part (i).
Now, we prove part (ii). For any given \(j \in {{\mathscr {B}}_0^c}\), we have \(\Vert {\varvec{\tilde{\gamma }}_j}\Vert _{\varvec{W}_j} = {O_p}(m^{-r/(2r+1)})\), together with \(m^{(1+\kappa )r/(2r+1)} \lambda _{min} \rightarrow \infty \), we can derive that
This implies that
On the other hand, by the condition \( m^{r/(2r+1)}\lambda _{max}\rightarrow 0\), for \(\varepsilon >0\) and \(j \in {\mathscr {B}}_0\), we have
which implies that \({{\hat{\delta } }_j} = {o_p}(m^{-r/(2r+1)})\) for each \(j \in {{\mathscr {B}}_0}\). Therefore, we prove that \(P\{ {{{\hat{\delta } }_j} < 1~~for~all~j \in {\mathscr {B}}_0}\} \rightarrow 1.\) Thus, we complete the proof of (ii).
For part (iii), apply the similar techniques as in part (ii), we have, with probability tending to 1, that \({{\hat{\delta } }_j} = 1\) for \( j\in \mathscr {A}_0^c\) and \({{\hat{\delta } }_j} <1\) for \( j\in \mathscr {A}_0\). \(\square \)
Proof of Theorem 5
As shown in (ii) and (iii) in Theorem 4, \({{\bar{\beta } }_j} = 0\) for \(j \in {\mathscr {A}_0^c}\) and \({\varvec{\bar{\gamma }}_j} = 0\) for \(j \in {{\mathscr {B}}_0^c}\) with probability tending to 1. Meanwhile, with probability tending to 1, \(({\varvec{\bar{\beta }}_{\mathscr {A}_0}},\varvec{\bar{\gamma }_{{\mathscr {B}}_0}}) \) satisfies the RSGEE
where \(\varvec{\theta }_{{\mathscr {A}_0},{{\mathscr {B}}_0}}=(\varvec{\beta }_{\mathscr {A}_0}^T,\varvec{\gamma }_{{\mathscr {B}}_0}^T)^T\), \(\varvec{\hat{S}}_{{\mathscr {A}_0},{{\mathscr {B}}_0}}=({\varvec{I}_{s+tK}} - {\varvec{\hat{\Delta }}}_{{\mathscr {A}_0},{{\mathscr {B}}_0}} )^{-1}\varvec{\hat{\Delta }}_{{\mathscr {A}_0},{{\mathscr {B}}_0}}\) and \(\varvec{\hat{\Delta }}_{{\mathscr {A}_0},{{\mathscr {B}}_0}} = {\mathrm{diag}}\{ {{{\hat{\delta } }_1},\ldots ,{{\hat{\delta } }_s},\underbrace{{{\hat{\delta } }_{p + 1}},\ldots ,{{\hat{\delta } }_{p + 1}}}_{K},\ldots ,\underbrace{{{\hat{\delta } }_{p + t}},\ldots ,{{\hat{\delta } }_{p + t}}}_{K}} \}\). Moreover, let \(\varvec{\hat{\Delta }}_{{\mathscr {A}_0}}={\mathrm{diag}}\{ {{{\hat{\delta } }_1},\ldots ,{{\hat{\delta } }_s}}\}\) and \(\varvec{\hat{\Delta }}_{{{\mathscr {B}}_0}}={\mathrm{diag}}\{ {\underbrace{{{\hat{\delta } }_{p + 1}},\ldots ,{{\hat{\delta } }_{p + 1}}}_{K},\ldots ,\underbrace{{{\hat{\delta } }_{p + t}},\ldots ,{{\hat{\delta } }_{p + t}}}_{K}} \}\). Define \(\varvec{\hat{S}}_{{\mathscr {A}_0}}=({\varvec{I}_{s}} - {\varvec{\hat{\Delta }}}_{{\mathscr {A}_0}} )^{-1}\varvec{\hat{\Delta }}_{{\mathscr {A}_0}}\) and \(\varvec{\hat{S}}_{{{\mathscr {B}}_0}}=({\varvec{I}_{tK}} - {\varvec{\hat{\Delta }}}_{{{\mathscr {B}}_0}} )^{-1}\varvec{\hat{\Delta }}_{{{\mathscr {B}}_0}}\). Then \({\Vert {\varvec{\hat{S}}_{{\mathscr {A}_0},{{\mathscr {B}}_0}}\varvec{\theta }_{{\mathscr {A}_0},{{\mathscr {B}}_0}} }\Vert ^2} ={\Vert {\left( \begin{array}{l} \varvec{\hat{S}}_{\mathscr {A}_0}\varvec{\beta }_{\mathscr {A}_0} \\ \varvec{\hat{S}}_{{\mathscr {B}}_0}\varvec{\gamma }_{{\mathscr {B}}_0}\\ \end{array} \right) }\Vert ^2} = {\Vert { {\varvec{\hat{S}}_{\mathscr {A}_0}\varvec{\beta }_{\mathscr {A}_0} } }\Vert ^2}+{\Vert { { \varvec{\hat{S}}_{{\mathscr {B}}_0}\varvec{\gamma }_{{\mathscr {B}}_0}} }\Vert ^2}\), where
where \(\nu = \mathop {\min }\nolimits _{j \in {\mathscr {A}_0}} \left| {\tilde{\beta } _j} \right| \). Similarly, we have \({\Vert {\varvec{\hat{S}}_{{\mathscr {B}}_0}\varvec{\gamma }_{{\mathscr {B}}_0} } \Vert ^2}={o_p}(m^{-2r/(2r+1)})\). Applying Theorem 3, we can easily complete the proof Theorem 5. \(\square \)
Rights and permissions
About this article
Cite this article
Lv, J., Yang, H. & Guo, C. Smoothing combined generalized estimating equations in quantile partially linear additive models with longitudinal data. Comput Stat 31, 1203–1234 (2016). https://doi.org/10.1007/s00180-015-0612-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00180-015-0612-8