Abstract
In modern statistical applications, the dimension of covariates can be much larger than the sample size, and extensive research has been done on screening methods which can effectively reduce the dimensionality. However, the existing feature screening procedure can not be used to handle the ultrahigh-dimensional survival data problems when failure indicators are missing at random. This motivates us to develop a feature screening procedure to handle this case. In this paper, we propose a feature screening procedure by sieved nonparametric maximum likelihood technique for ultrahigh-dimensional survival data with failure indicators missing at random. The proposed method has several desirable advantages. First, it does not rely on any model assumption and works well for nonlinear survival regression models. Second, it can be used to handle the incomplete survival data with failure indicators missing at random. Third, the proposed method is invariant under the monotone transformation of the response and satisfies the sure screening property. Simulation studies are conducted to examine the performance of our approach, and a real data example is also presented for illustration.
Similar content being viewed by others
References
Bitouzé D, Laurent B, Massart P (1999) A Dvoretzky-Kiefer-Wolfowitz type inequality for the Kaplan-Meier estimator. Annals de I’Institut Henri Poincare B 35:735–763
Candes E, Tao T (2007) The Dantzig selector: statistical estimation when \(p\) is much larger than \(n\). Ann Stat 35:2313–2351
Fan J, Li R (2001) Variable selection via nonconcave penalized likelihood and its oracle properties. J Am Stat Assoc 96:1348–1360
Fan J, Lv J (2008) Sure independence screening for ultrahigh dimensional feature space. J R Stat Soc Ser B 35:2313–2351
Fan J, Song R (2010) Sure independence screening in generalized linear models with NP-dimensionality. J Am Stat Assoc 38:3567–3604
Fan J, Feng Y, Wu Y (2010) High-dimensional variable selection for Cox’s proportional hazards model. Statistics 2:70–86
Gill R (1981) Testing with replacement and the product limit estimator. Ann Stat 9:853–860
Gill R (1983) Large sample behaviour of the product-limit estimator on the whole line. Ann Stat 11:49–58
González S, Rueda M, Arcos A (2008) An improved estimator to analyse missing data. Stat Pap 49:791–792
He X, Wang L, Hong H (2013) Quantile-adaptive model-free variable screening for high-dimensional heterogeneous data. Ann Stat 41:342–369
Li R, Zhong W, Zhu L (2012) Feature screening via distance correlation learning. J Am Stat Assoc 107:1129–1140
Li G, Peng H, Zhang J, Zhu L (2012) Robust rank correlation based screening. Ann Stat 40:846–877
Lin W, Lv J (2013) High-dimensional sparse additive hazards regression. J Am Stat Assoc 108:247–264
Little R, Rubin D (2002) Statistical analysis with missing data. Wiley, Hoboken
Qin J, Shao J, Zhang B (2008) ANOVA for longitudinal data with missing values. J Am Stat Assoc 103:797–810
Rosenwald A, Wright G, Wiestner A, Chan W et al (2003) The proliferation gene expression signature is a quantitative integrator of oncogenic events that predicts survival in mantle cell lymphoma. Cancer Cell 3:185–197
Shen Y, Liang H (2018) Quantile regression and its empirical likelihood with missing response at random. Stat Pap 59:685–707
Song R, Lu W, Ma S, Jeng X (2014) Censored rank independence screening for high-dimensional survival data. Biometrika 101:799–814
Tibshirani R (1997) Regression shrinkage and selection via the Lasso. J R Stat Soc Ser B 58:267–288
van der Laan M (1996) Efficient estimation in the bivariate censoring model and repairing NPMLE. Ann Stat 24:596–627
van der Laan M, Mckeague I (1998) Efficient estimation from right-censored data when failure indicators are missing at random. Ann Stat 26:164–182
Wang J (1987) A note on the uniform consistency of the Kaplan-Meier estimator. Ann Stat 15:1313–1316
Wang Q, Rao J (2002) Empirical likelihood-based inference under imputation for missing response data. Ann Stat 30:894–924
Wu Y, Yin G (2015) Conditional quantile screening in ultrahigh-dimensional heterogeneous data. Biometrika 102:65–76
Zhang H, Lu W (2007) Adaptive Lasso for Cox’s proportional hazards model. Biometrika 94:691–703
Zhao S, Li Y (2012) Principled sure independence screening for Cox models with ultra-high-dimensional covariates. J Multivar Anal 105:397–411
Zhang J, Liu Y, Wu Y (2017) Correlation rank screening for ultrahigh-dimensional survival data. Comput Stat Data Anal 108:121–132
Zhu L, Li L, Li R, Zhu L (2011) Model-free feature screening for ultrahigh-dimensional data. J Am Stat Assoc 106:1464–1475
Acknowledgements
Fang’s research is supported by Project supported by Provincial Natural Science Foundation of Hunan (Grant No. 2018JJ2078) and Scientific Research Fund of Hunan Provincial Education Department (Grant No. 17C0392).
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
Firstly, we introduce the following Lemma which is useful for proving Theorem 2.1.
Lemma 4.1
Under Assumptions 1–5, for any positive \(\varepsilon \), we have
and
Proof
Let \(B^{0}\) denote a Brownian bridge. With the representation of that process from a Brownian bridge given by Bitouzé et al. (1999), we have
Let
Based on Theorem 5.1 in van der Laan (1996), the estimator \(\tilde{G}\) of G is efficient for the reduced data. Therefore, the result of Theorem 1.1 in Gill (1983) imply that
where \({\mathop {\rightarrow }\limits ^{d}}\) represents the convergence in distribution, and D[0, L] is the cadlag function space of real-valued functions on [0, L] endowed with the supremum norm. Hence it follows from (10) that
Because \(|[\{1-G(t)\}\{1-V(t)\}]/\{1-K(t)\}|\le 1\), combining (9), we have
The first part of Lemma 4.1 is proved, and then we begin to prove the second part of it. By simple calculation, we have
and
Because the \(\tilde{V}(t)\) is obtained by switching the role of failure time and censoring time in (2), similar to the proof of (7), we have
Moreover, we also have
If we can show
where \({\mathop {\rightarrow }\limits ^{p}}\) represents the convergence in probability. Then, we can obtain, according to Lemma 2.6 in Gill (1983), that \(\sup _{t\in (0,L]}|(\tilde{G}(t)-G(t))/(1-G(t))|\) is bounded in probability. Therefore, it follows by combining (7) and (11)–(14)
Next, we begin to prove (15). For the distribution function G(t), the cumulative hazard function H(t) can be defined as
and its estimator \(\hat{H}_{n}(t)\) is given by
By simple calculation, we have
Based on Lemma 4.2 in van der Laan (1996) and Assumption 2, we can obtain
Moreover, similar to the proof of Theorem 1 in Wang (1987), we can show that
It follows form Assumptions 3 and 6 that \(1/((1-F_{Y}(y))(1-F_{Y,n}(y)))\) is bounded. Therefore, we have
and
Now by using Lemma 2 in Gill (1981) we get
So (15) follows. The proof of Lemma 4.1 is completed. \(\square \)
Proof of Theorem 2.1
Let \(\omega ^{*}_{k}=\{1/n\sum _{i=1}^{n}X_{ki}u(Y_{i})G_{k}(Y_{i})\}^{2}\), \(k=1,\cdots ,p\). we have
Combining the strong law of large numbers and Assumption 1, we can obtain that
where \({\mathop {\longrightarrow }\limits ^{a.s.}}\) represents almost sure convergence. Based on the Cauchy−Schwarz inequality and the boundedness of \(\tilde{G}(t)\), G(t), \(\tilde{u}(t)\) and u(t), combining Assumption 4, we can obtain that
There exists positive constants \(c_{1}\), \(c_{2}\) and \(c_{3}\) such that
and
Based on \(V(t)=1-u(t)\) and \(\tilde{V}(t)=1-\tilde{u}(t)\), we have
Because the estimator \(\tilde{V}(t)\) is obtained by switching the role of failure time and censoring time and combining (2), we can show that, by using (8) in Lemma 4.1 and switching the role of \(\tilde{V}(t)\) and \(\tilde{G}(t)\),
Without loss of generality, when n is large enough, we can show that
Combining Assumption 6 and the Cauchy–Schwarz inequality, we have
where \(c_{5}\) is a positive constant. Therefore, based on (7) in Lemma 4.1, we can obtain that
Similarly, there exists a positive constant \(c_{6}\) such that
Note that \(\{X_{ki}u(Y_{i})G(Y_{i}):~i=1,\cdots ,n\}\) are independent identically sample form \(X_{k}u(Y)G(Y)\), then, based on Hoeffding’s inequality, we have
According to (16), it is easy to show that
Therefore, from (17)–(24), there exists a positive constant \(\gamma \) such that
Immediately, we have
On the other hand, similar to the proof of the second part of Theorem 2.1 in Zhang et al. (2017), we can prove that
where \(d=|\mathscr {A}|\) is the cardinality of \(\mathscr {A}\). The proof of Theorem 2.1 is completed. \(\square \)
Proof of Theorem 2.2
Let \(\eta =\min \limits _{k\in \mathscr {A}}\{\omega _{k}\}-\max \limits _{k\not \in \mathscr {A}}\{\omega _{k}\}\). It follows from Assumption 5 that \(\eta >0\). Similar to the proof of (24) and (25) in Theorem 2.1, we can show that
Therefore, combining (26) and (27), when n is large enough, we can obtain that
and
Thus, the proof of Theorem 2.2 is completed \(\square \)
Rights and permissions
About this article
Cite this article
Fang, J. Feature screening for ultrahigh-dimensional survival data when failure indicators are missing at random. Stat Papers 62, 1141–1166 (2021). https://doi.org/10.1007/s00362-019-01128-5
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00362-019-01128-5