RANK INFERENCES FOR THE ACCELERATED FAILURE TIME MODELS ZHOU FANG NATIONAL UNIVERSITY OF SINGAPORE 2014 RANK INFERENCES FOR THE ACCELERATED FAILURE TIME MODELS ZHOU FANG (B.Sc. Wuhan University) A THESIS SUBMITTED FOR THE DEGREE OF DOCTOR OF PHILOSOPHY DEPARTMENT OF STATISTICS AND APPLIED PROBABILITY NATIONAL UNIVERSITY OF SINGAPORE 2014 ii ACKNOWLEDGEMENTS I am so grateful that I have Dr. Xu Jinfeng and Professor Chen Zehua as my supervisors. They are truly great mentors not only in statistics but also in daily life. I would like to thank them for their guidance, encouragement, time, and endless patience. Next, special acknowledgement goes to the faculties and staff of DSAP, especially Associate Professor Li Jialiang and Mr. Zhang Rong. Anytime I encountered difficulties and tried to seek help from them, I was always warmly welcomed. I also thank all my colleges who helped me to make life easier as a graduate student. I wish to express my gratitude to the university and the department for supporting me through NUS Graduate Research Scholarship. Finally, I will thank my family for their love and support. iii CONTENTS Acknowledgements Summary vii List of Tables x List of Figures Chapter Introduction 1.1 1.2 ii xii Introduction to Survival Analysis . . . . . . . . . . . . . . . . . . . 1.1.1 Survival Data and Right Censoring . . . . . . . . . . . . . . 1.1.2 The Cox Proportional Hazards Model . . . . . . . . . . . . . 1.1.3 The Censored Accelerated Failure Time Model . . . . . . . . Semi-parametric Models . . . . . . . . . . . . . . . . . . . . . . . . CONTENTS iv 1.2.1 Partially Linear Model . . . . . . . . . . . . . . . . . . . . . 1.2.2 Varying Coefficients Model . . . . . . . . . . . . . . . . . . . 1.3 Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.4 Objectives and Organization . . . . . . . . . . . . . . . . . . . . . . 12 Chapter Partially Linear Accelerated Failure Time Model 15 2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.2 Partially Linear Model for Uncensored Data . . . . . . . . . . . . . 17 2.3 Existing Rank-based Methods for the Censored Partially Linear Model 19 2.4 Proposed Local Gehan Method . . . . . . . . . . . . . . . . . . . . 22 2.4.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.4.2 Asymptotic Properties . . . . . . . . . . . . . . . . . . . . . 26 2.4.3 Optimal Bandwidth . . . . . . . . . . . . . . . . . . . . . . . 30 2.4.4 Estimation of Limiting Covariance Matrix . . . . . . . . . . 31 Numerical Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.5.1 Computation Algorithm . . . . . . . . . . . . . . . . . . . . 33 2.5.2 Bandwidth Selection . . . . . . . . . . . . . . . . . . . . . . 34 2.5.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.5 2.6 Chapter Varying-coefficient Accelerated Failure Time Model 46 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.2 Extended Local Gehan Procedure . . . . . . . . . . . . . . . . . . . 50 3.2.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.2.2 Asymptotic Properties . . . . . . . . . . . . . . . . . . . . . 51 CONTENTS 3.3 3.4 v 3.2.3 Optimal Bandwidth . . . . . . . . . . . . . . . . . . . . . . . 56 3.2.4 Estimation of Limiting Covariance Matrix . . . . . . . . . . 56 Numeric Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.3.1 Computation algorithm . . . . . . . . . . . . . . . . . . . . . 58 3.3.2 Bandwidth selection . . . . . . . . . . . . . . . . . . . . . . 58 3.3.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 Chapter Variable Selection in the Partially Linear Accelerated Failure Time Model 74 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.2.1 Penalized global Gehan estimator . . . . . . . . . . . . . . . 77 Numerical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4.3.1 Tuning parameter selection . . . . . . . . . . . . . . . . . . 78 4.3.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 4.3 4.4 Chapter Conclusion and Discussion 86 Chapter A Theoretical Proof of Chapter 89 A.1 Lemma A.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 A.2 Lemma A.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 A.3 Lemma A.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 A.4 Proof of Lemma 2.4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . 94 A.5 Proof of Theorem 2.4.2 . . . . . . . . . . . . . . . . . . . . . . . . . 94 CONTENTS vi A.6 Proof of asymptotic normality of α ˆ . . . . . . . . . . . . . . . . . . 101 A.7 Lemma A.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 A.8 Proof of Theorem 2.4.3 . . . . . . . . . . . . . . . . . . . . . . . . . 103 Chapter B Theoretical Proof of Chapter 107 B.1 Lemma B.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 B.2 Lemma B.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 B.3 Lemma B.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 B.4 Proof of Lemma 3.2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . 112 B.5 Proof of Theorem 3.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . 113 B.6 Lemma B.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 B.7 Proof of Theorem 3.2.3 . . . . . . . . . . . . . . . . . . . . . . . . . 118 Bibliography 123 vii SUMMARY In many biomedical studies, the response of interest is the time until some event to occur. Such data is usually referred to as lifetime data, failure time data or survival data. By linearly relating the logarithm of survival time to the covariates, a semiparametric accelerated failure time model is often used to examine the covariate effect, providing an easy and direct interpretation. In some applications, the assumption that the covariate effects are linear and constant may be too restrictive. Hence it is desirable to develop more flexible models incorporating nonlinear or varying covariate effects. We consider the partially linear and varying coefficients accelerated failure time models for the analysis of right censored survival data. Rank-based inferential procedures along with the kernel smoothing method are proposed. Firstly, in the censored partially linear accelerated failure time model, we propose a local Summary viii Gehan loss function-based estimation procedure using the kernel smoothing method. The estimation can be obtained through standard ’quantreg packages’ available in R. Under mild regularity conditions, we establish the asymptotical normality of the local Gehan estimator. A resampling procedure is also developed to estimate the limiting covariance matrix. We then extend the local Gehan estimator to two global Gehan estimators. One is obtained by averaging the local ones at all observed points. Another is the minimizer of profile Gehan loss function. Without considering local smoothing, a global Gehan estimator based on piecewise linear approximation to the nonparametric term is also proposed. Simulation results suggest their favorable performance in terms of bias and variance. Compared with the existing methods such as the stratified method and the spline method, the proposed methods exhibit certain advantages. Real data applications are also conducted to illustrate the practical utilities of the proposed methods. Secondly, we extend the local Gehan procedure to the censored varying coefficients accelerated failure time model. The varying coefficients model allows extra dynamics of covariate effects and includes many aforementioned models as special cases. Under mild regularity conditions, we prove that the local Gehan loss-based estimator continues to enjoy good properties. Theoretical properties are established and numerical examples are given for the illustration. In computation, a censored version cross validation method is also proposed to choose the smoothing parameter. In parallel with the partially linear model, a resampling method by random perturbation is proposed for inferential purposes. Finally, we study the problem of variable selection in the censored partially linear accelerated failure time model. Combined with the penalty, the global Gehan loss function with piecewise linear approximation offers a convenient tool for simultaneous estimation and variable selection. Extensive simulation studies were conducted to investigate the properties of the proposed variable selection procedures in terms of both the reduced Summary model error and the probability of identifying the correct model. ix B.5 Proof of Theorem 3.2.2 B.5 113 Proof of Theorem 3.2.2 By Lemma 3.2.1, γn−1 Q∗n (v1 , v2 ) = Bn (v1 , v2 ) + rn (v1 , v2 ). where rn (v1 , v2 ) →p ∗ )T , uniformly over any bounded set. Note that γn−1 Q∗n (v1 , v2 ) is minimized by (θˆn∗T , α ˆ 2n and Bn (v1 , v2 ) is minimized by ∗ T (θ˜n∗T , α ˜ 2n ) = −γn−2 A−1 (SnT1 (0, 0), Sn2 (0, 0))T . By similar argument as in the proof of Theorem 2.4.2, we can obtain ∗ ∗ T (θˆn∗T , α ˆ 2n ) = (θ˜n∗T , α ˜ 2n ) + op (1). This implies the asymptotic representation (3.9). We next show the asymptotic normality ˆ(z0 ). From (3.9), we have of a √ nh(ˆ a(z0 ) − a(z0 )) = −γn−2 [4τ υf (z0 )pΣ(z0 )]−1 Sn11 (0, 0) + op (1), (B.1) where Sn11 (0, 0) = 2γn [n(n − 1)]−1 ∆i I(εi + δi (z0 ) ≤ εj + δj (z0 ))(Xi − Xj )Kh (Zi − z0 ) i=j Kh (Zj − z0 ) + ∆j I(εj + δj (z0 ) ≤ εi + δi (z0 ))(Xj − Xi )Kh (Zj − z0 )Kh (Zi − z0 ) Thus, we can write −γn−2 Sn11 (0, 0) = Sna1 (0, 0) + Sna2 (0, 0), where Sna1 (0, 0) = 2γn−1 [n(n − 1)]−1 ∆i [I(εi ≤ εj )](Xj − Xi )Kh (Zi − z0 )Kh (Zj − z0 ) i=j +∆j [I(εj ≤ εi )](Xi − Xj )Kh (Zj − z0 )Kh (Zi − z0 ) B.5 Proof of Theorem 3.2.2 114 Sna2 (0, 0) = 2γn−1 [n(n − 1)]−1 ∆i [I(εi + δi (z0 ) ≤ εj + δj (z0 )) − I(εi ≤ εj )] i=j (Xj − Xi )Kh (Zi − z0 )Kh (Zj − z0 ) + ∆j [I(εj + δj (z0 ) ≤ εi + δi (z0 )) −I(εj ≤ εi )](Xi − Xj )Kh (Zj − z0 )Kh (Zi − z0 ) We next prove that Sna1 (0, 0) → N (0, υν0 f (z0 )Σ(z0 )) (B.2) where Σ(z0 ) = E(Xi XTi |Zi = z0 ). Note that we can write √ Sna1 (0, 0) = n[n(n − 1)]−1 hn (Di , Dj ) i=j where hn (Di , Dj ) = wn (Di , Dj ) + wn (Dj , Di ) is symmetric with wn (Di , Dj ) = h−3/2 ∆i [I(εi ≤ εj )](Xj − Xi )K( Z j − z0 Z i − z0 )K( ) h h Similarly to the arguments in the proof of Lemma B.3, it can be shown that E[ hn (Di , Dj ) ] = o(n) By Lemma B.1, this implies that hn (Di , Dj ) = E[hn (Di , Dj )|Di ] + op (1), Sna1 (0, 0) = √ n(2n−1 n i=1 rn (Di ) + op (1)) since it is easy to check that r¯n = 0. We have rn (Di ) = E[hn (Di , Dj )|Di ] = E(wn (Di , Dj )|Di ) + E(wn (Dj , Di )|Di ) Zj − z Zi − z )∆i E{(Xj − Xi )K( )|Xi , Zi , Ui∗ , εi } h h Zj − z Z i − z0 +h−3/2 G(εi )K( )E{∆j (Xi − Xj )K( )|Xi , Zi , Ui∗ , εi } h h Z i − z0 = h−1/2 (G(εi ) − 1)∆i K( )[( K(t)f (z0 + th)dt)Xi h = h−3/2 (1 − G(εi ))K( B.5 Proof of Theorem 3.2.2 − E(Xj |Zj = z0 + th)K(t)f (z0 + th)dt] h−1/2 G(εi ) − 115 G(u)h(u)duK( Zi − z )[( h K(t)f (z0 + th)dt)Xi E(Xj |Zj = z0 + th)K(t)f (z0 + th)dt] Under condition (C3), E(Xi |Zi = z0 ) = 0, rn (Di ) → h−1/2 (G(εi ) − 1)K( +h−1/2 G(εi ) Zi − z )[( h G(u)h(u)duK( K(t)f (z0 + th)dt)Xi ∆i ] Z i − z0 )[( h K(t)f (z0 + th)dt)Xi ] Furthermore, E[rn (Di )rn (Di )T ] → E{h−1 (G(εi ) − 1)2 K ( Zi − z )[( h K(t)f (z0 + th)dt)Xi ∆i ] [( K(t)f (z0 + th)dt)Xi ∆i ]T + h−1 G2 (εi )( [( K(t)f (z0 + th)dt)Xi ][( K 2( Zi − z )[( h Z i − z0 )[( h K(t)f (z0 + th)dt)Xi ]T } K(t)f (z0 + th)dt)Xi ∆i ] [( K(t)f (z0 + th)dt)Xi ∆i ]T } + h−1 EG2 (εi )E( [( K(t)f (z0 + th)dt)Xi ][( E{K ( Zi − z )[( h Z i − z0 ) h K(t)f (z0 + th)dt)Xi ]T + 2h−1 G(εi )(G(εi ) − 1) K(t)f (z0 + th)dt)Xi ∆i ][( → h−1 E(G(εi ) − 1)2 E{K ( G(u)h(u)du)2 K ( G(u)h(u)du)2 E{K ( Z i − z0 ) h K(t)f (z0 + th)dt)Xi ]T } + 2h−1 E[G(εi )(G(εi ) − 1)] K(t)f (z0 + th)dt)Xi ∆i ][( K(t)f (z0 + th)dt)Xi ]T } B.5 Proof of Theorem 3.2.2 116 Since G(εi ) ∼ U (0, 1), EG(εi ) = 12 , E(G(εi ))2 = 13 , var(G(εi )) = 12 , E(G(εi ) − 1)2 = EG2 (εi ) − 2EG(εi ) + = 1/3, then f (z0 ) E(∆2i Xi XTi |Zi = z0 + th)f (z0 + th)K (t)dt + υf (z0 ) E(Xi XTi |Zi = z0 + th)f (z0 + th)K (t)dt − f (z0 ) E(∆i Xi XTi |Zi = z0 + th)f (z0 + th)K (t)dt 3 → f (z0 )E(∆2i Xi XTi |Zi = z0 )ν0 + υf (z0 )E(Xi XTi |Zi = z0 )ν0 − f (z0 )E(∆i Xi XTi |Zi = z0 + th)ν0 = υν0 f (z0 )Σ(z0 ). E(rn (Di )rnT (Di )) → So, Sna1 (0, 0) → N (0, υν0 f (z0 )Σ(z0 )) To prove the asymptotic normality of Sna1 (0, 0), it is sufficient to check the LindebergFeller condition: ∀ε > 0, n−1 n T i=1 E{rn (Di )rn (Di ) I( √ rn (Di ) > ε n)} → 0. This can be easily verified by applying the dominated convergence theorem. However, the asymptotic representation of Sna2 (0, 0) is slightly different with that in the proof of Theorem 2.4.2. Next we show that Sna2 (0, 0) = 2h2 [τ υf (z0 )µ2 Σ(z0 )a (z0 ) + o(1)] + op (1). γn We may write Sna2 (0, 0) = 2[n(n − 1)]−1 h∗n (Di , Dj ) i=j (B.3) B.5 Proof of Theorem 3.2.2 117 where h∗n (Di , Dj ) = wn∗ (Di , Dj ) + wn∗ (Dj , Di ) is symmetric with wn∗ (Di , Dj ) = nh−1 γn [∆i (I(εi +δi (z0 ) ≤ εj +δj (z0 ))−I(εi ≤ εj ))](Xj −Xi )K( Z j − z0 Z i − z0 )K( ). h h By applying Lemma B.1, it can be shown that Sna2 (0, 0) = E[h∗n (Di , Dj )] + op (1). Note that δj (z0 ) − δi (z0 ) = 1 [(Zj − z0 )2 XTj − (Zi − z0 )2 XTi ]a (z0 ) + [(Zj − z0 )2 − (Zi − z0 )2 ]φ (z0 ) 2 = o((Zj − z0 )2 ) + o((Zj − z0 )2 ). It follows by using the same arguments in the proof of Lemma B.1 that E[h∗n (Di , Dj )] = 2nh−1 γn E{ (Xj − Xi )K( [G(ε + δj (z0 ) − δi (z0 )) − G(ε)]g(ε)dε Z j − z0 Zi − z )K( )} h h = 2nh−1 γn [τ υ + O(h)]E[(δj (z0 ) − δi (z0 ))(Xj − Xi )K( = G(u)h(u)du Z j − z0 Z i − z0 )K( )](1 + o(1)) h h 2h2 [τ υf (z0 )µ2 Σ(z0 )a (z0 ) + o(1)]. γn This proves (B.3). By combining (B.2) and (B.3) and using the approximation given in (B.1), we obtain (3.10). B.6 Lemma B.4 B.6 118 Lemma B.4 If E[ Hn (Di , Dj ) ] = O(h−2 ), then √ ˆn ) = o(1) almost surely and Un = n(Un − U r¯n + o(1) a.s. Proof : The proof of Powell, Stock and Stoker (1989) for Lemma A.1 suggests that ˆn ] = O(n−2 h−2 ). By theorem 1.3.5 of Serfling (1980), E[ Un − U n i=1 E[ ˆn ] = Un − U ˆn = o(1) almost surely. The second result O(n−1 h−2 ) < ∞. This implies that Un − U ˆn . follows by an application of the strong law of large numbers to U B.7 Proof of Theorem 3.2.3 Let θ∗ and α∗ be defined the same as before. We introduce the reparametrized objective function n n ¯ ∗ (θ∗ ; α∗ ) = 2/[n(n − 1)] Q n ∆i [(εi − γn α2∗ (Zi − z0 )/h − γn θ∗T Ui + δi (z0 )) i=1 j=1 −(εj − γn α2∗ (Zj − z0 )/h − γn θ∗T Uj + δj (z0 ))]− Kh (Zi − z0 )Kh (Zj − z0 )(Wi + Wj ). +∆j [(εj − γn α2∗ (Zj − z0 )/h − γn θ∗T Uj + δj (z0 )) −(εi − γn α2∗ (Zi − z0 )/h − γn θ∗T Ui + δi (z0 ))]− Kh (Zj − z0 )Kh (Zi − z0 )(Wj + Wi ). T (θ ∗ , α∗ ), S ¯n2 (θ∗ , α∗ ))T =( Let S¯n (θ∗ , α2∗ ) = (S¯n1 2 ¯ ∗T ∗ ∗ θ∗ Qn (θ , α2 ), ¯∗ ∗ ∗ T α∗2 Qn (θ , α2 )) , we can show that S¯n (θ∗ , α2∗ ) has a similar local linear approximation as stated in Lemma B.3. B.7 Proof of Theorem 3.2.3 119 T (θ ∗ , α∗ ), where To make the proof concise, we prove this for S¯n1 T S¯n1 (θ∗ , α2∗ ) n n −1 ∆i I[(εi − γn α2∗ (Zi − z0 )/h − γn θ∗T Ui + δi (z0 )) = 2γn [n(n − 1)] i=1 j=1 ≤ (εj − γn α2∗ (Zj − z0 )/h − γn θ∗T Uj + δj (z0 ))](Ui − Uj )Kh (Zi − z0 )Kh (Zj − z0 ) (Wi + Wj ) + ∆j I[(εj − γn α2∗ (Zj − z0 )/h − γn θ∗T Uj + δj (z0 )) ≤ (εi − γn α2∗ (Zi − z0 )/h − γn θ∗T Ui + δi (z0 ))](Uj − Ui ) Kh (Zj − z0 )Kh (Zi − z0 )(Wj + Wi ) Let Un = γn−1 [S¯n1 (θ∗ , α2∗ )−S¯n1 (0, 0)] = 2[n(n−1)]−1 i=j (Wi +Wj )Mn (Di , Dj , θ ∗ , α∗ ), where Mn (Di , Dj , θ∗ , α2∗ ) = [mn (Di , Dj , θ∗ , α2∗ ) + mn (Dj , Di , θ∗ , α2∗ )] and mn (Di , Dj , θ∗ , α2∗ ) = ∆i [I(εi − γn α2∗ (Zi − z0 )/h − γn θ∗T Ui + δi (z0 ) ≤ εj − γn α2∗ (Zj − z0 )/h − γn θ∗T Uj + δj (z0 )) −I(εi + δi (z0 ) ≤ εj + δj (z0 ))] (Ui − Uj )Kh (Zi − z0 )Kh (Zj − z0 ) Note that Un = 4n−1 n i=1 Wi [(n n ∗ ∗ j=1,j=i Mn (Di , Dj , θ , α2 )], − 1)−1 conditional on {Di }ni=1 , this is a weighted average of Wi . Note that n E(Un |{Di }ni=1 ) = 4n−1 E( n [(n − 1)−1 i=1 j=1,j=i n n = 4n−1 Mn (Di , Dj , θ∗ , α2∗ )]Wi |{Di }ni=1 ) [(n − 1)−1 i=1 Mn (Di , Dj , θ∗ , α2∗ )] j=1,j=i = 2[n(n − 1)]−1 Mn (Di , Dj , θ∗ , α2∗ ). i=j n Var(Un |{Di }ni=1 ) = 2n−2 n [(n − 1)−1 i=1 Mn (Di , Dj , θ∗ , α2∗ )]2 . j=1,j=i B.7 Proof of Theorem 3.2.3 120 By Lemma B.4, it can be shown that 2[n(n − 1)]−1 Mn (Di , Dj , θ∗ , α2∗ ) = E(Mn (Di , Dj , θ∗ , α2∗ )) + o(1) = γn A∗ θ∗ + o(1) i=j almost surely, where A∗ = 4τ υf (z0 )diag(Ip , µ2 Ip ) It is also easy to check that 2n−2 n i=1 [(n − 1)−1 Σ(z0 ). n ∗ ∗ j=1,j=i Mn (Di , Dj , θ , α2 )] = o(1) almost surely. Thus for almost surely every sequence {Di }ni=1 , Un = γn A∗ θ∗ + op (1), where op (1) is in the probability space generated by {Wi }ni=1 . Similar to the proofs of Lemma 3.2.1 and the asymptotic representation in Theorem 3.2.2, we can show that for almost surely every sequence {Di }ni=1 , √ ˆ ¯(z0 ) − a(z0 )) = −γn−2 [4τ υf (z0 )Σ(z0 )]−1 S¯n11 (0, 0) + op (1), nh(a (B.4) where op (1) is in the probability space generated by {Wi }ni=1 , and S¯n11 (0, 0) = 2γn [n(n − 1)]−1 (Wi + Wj )[(∆i I(εi + δi (z0 ) ≤ εj + δj (z0 )))(Xi − Xj ) i=j +(∆j I(εj + δj (z0 ) ≤ εi + δi (z0 )))(Xj − Xi )]Kh (Zj − z0 )Kh (Zi − z0 ) The approximation (B.1) can be strengthened to almost surely convergence, i.e., √ nh(ˆ a(z0 ) − a(z0 )) = −γn−2 [4τ υf (z0 )Σ(z0 )]−1 Sn11 (0, 0) + op (1), a.s. (B.5) Combining (B.4) and (B.5), we have that for almost surely every sequence {Di }ni=1 , √ ˆ ¯(z0 ) − a ˆ(z0 )) = −γn−2 [4τ υf (z0 )Σ(z0 )]−1 [S¯n11 (0, 0) − Sn11 (0, 0)] + op (1). nh(a Note that γn−2 [S¯n11 (0, 0) − Sn11 (0, 0)] B.7 Proof of Theorem 3.2.3 = 2γn−1 [n(n − 1)]−1 i=j 121 1 [(Wi − ) + (Wj − )][(∆i I(εi + δi (z0 ) ≤ εj + δj (z0 )))(Xi − Xj ) 2 +(∆j I(εj + δj (z0 ) ≤ εi + δi (z0 )))(Xj − Xi )]Kh (Zi − z0 )Kh (Zj − z0 ) n = 4γn−1 n−1 i=1 (Wi − ){(n − 1)−1 [∆i (I(εi + δi (z0 ) ≤ εj + δj (z0 )))(Xi − Xj ) j=1,j=i +(∆j I(εj + δj (z0 ) ≤ εi + δi (z0 )))(Xj − Xi )]Kh (Zi − z0 )Kh (Zj − z0 ). And E{γn−2 [S¯n11 (0, 0) − Sn11 (0, 0)]|{Di }ni=1 } = 0. We have Var{γn−2 [S¯n11 (0, 0) − Sn11 (0, 0)]|{Di }ni=1 } n = 16γn−2 n−2 (n − 1)−2 [∆i I(εi + δi (z0 ) ≤ εj + δj (z0 ))(Xi − Xj ) { i=1 j=1,j=i +∆j I(εj + δj (z0 ) ≤ εi + δi (z0 ))(Xj − Xi )]Kh (Zi − z0 )Kh (Zj − z0 )}2 = V1 + V2 + V3 . where n V1 = 16γn−2 n−2 (n − 1)−2 h−4 [∆i [I(εi + δi (z0 ) ≤ εj + δj (z0 ))]2 i=1 j=1,j=i T (Xi − Xj )(Xi − Xj ) K ((Zi − z0 )/h)K ((Zj − z0 )/h), n V2 = 16γn−2 n−2 (n −2 −4 − 1) [∆j [I(εj + δj (z0 ) ≤ εi + δi (z0 ))]2 h i=1 j=1,j=i T (Xj − Xi )(Xj − Xi ) K ((Zj − z0 )/h)K ((Zi − z0 )/h), n V3 = 16γn−2 n−2 (n − 1)−2 h−4 i=1 j1 =i j2 =i [∆i I(εi + δi (z0 ) ≤ εj1 + δj1 (z0 ))(Xi − Xj1 ) +∆j1 I(εj1 + δj1 (z0 ) ≤ εi + δi (z0 ))(Xj1 − Xi )] [∆i I(εi + δi (z0 ) ≤ εj2 + δj2 (z0 ))(Xi − Xj2 ) B.7 Proof of Theorem 3.2.3 +∆j2 I(εj2 + δj2 (z0 ) ≤ εi + δi (z0 ))(Xj2 − Xi )] K ((Zi − z0 )/h)K((Zj1 − z0 )/h)K((Zj2 − z0 )/h). Lemma B.4 can be used to show that V1 = o(1) almost surely and V2 = o(1) almost surely; and a minor extension of Lemma B.4 to third-order U-statistic can be used to show that V3 = 34 f (z0 )ν0 υΣ(z0 )+o(1) a.s The asymptotic normality of γn−2 [S¯n11 (0, 0)− Sn11 (0, 0)] follows by showing that the condition of Lindeberg-Feller central limit theorem for triangular arrays holds almost surely. We have, for almost surely every sequence {Di }ni=1 , γn−2 [S¯n11 (0, 0) − Sn11 (0, 0)] → N (0, f (z0 )ν0 υΣ(z0 )) in distribution. This proves the Theorem 3.2.3. 122 123 Bibliography [1] Buckley, I. V. and James, I. (1979). Linear regression with censored data. Biometrika 66, 3, 429-436. [2] Cai, J. W., Fan, J. Q., Jiang, J. C. and Zhou, H. B. (2008). Partially linear hazard regression with varying coefficients for multivariate survival data. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 70, 1, 141-158. [3] Cai, J. W., Fan, J. Q., Zhou, H. B. and Zhou, Y. (2007). Marginal hazard models with varying-coefficients for multivariate failure time data. The Annals of Statistics, 35, 324-354. [4] Chen, K., Shen, J. and Ying, Z. (2005). Rank esimation in partial linear model with censored data. Statistica Sinica, 15, 767-779. [5] Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society B, 34, 187-220. [6] Cox, D. R. (1975). Partial likelihood. Biometrika 62, 2, 269-276. Bibliography [7] David W. Hosmer, Jr. and Stanley Lemeshow. (1999) Applied survival analysis: regression modeling of time to event data. Wiley series in probability and statistics. Wiley, New York. [8] Eilers, P. and Marx, B. (1996). Flexible smoothing with B-spline and penalties. Statistical Science, 11, 89-121. [9] Engle, R. F., Granger, C. W. J., Rice, J. and Weiss, A. (1986). Semiparametric estimates of the relation between weather and electricity sales. Journal of the American Statistical Association, 81, 394, 310-320. [10] Eubank, R. L., Kambour, E. L., Kim, J. T., Klipple, K, Reese, C. S. and Schimek, M. (1998). Estimation in partially linear models. Computational Statistics and Data Analysis, 29, 1, 27-34. [11] Fan, J. and Gijbels, I. (1996). Local Polynomial Modelling and Its Applications. Chapman and Hall, London. [12] Fan, J. and Zhang, W. (1999). Statistical estimation in varying coefficient models. Annals of Statistics, 27, 5, 1491-1518. [13] Fan, J. and Li, R. (2001) Variable selection via nonconcave penalized likelihood and its oracle properties. Journal of the American Statistical Association, 96, 456, 1348-1360. [14] Fan, J. and Gijbels, I. (1992). Variable bandwidth and local linear regression smoothers. The Annals of Statistics, 20, 2008-2036. [15] Fan, J. (1993). Local linear regression smoothers and their minimax efficiency. The Annals of Statistics, 21, 196-216. [16] Fan, J. Q. and Gijbels, I. (1996). Local polynomial modelling and its applications: monographs on statistics and applied probability 66. Chapman and Hall, London. [17] Fan, J. Q., Lin, H. Z. and Zhou, Y (2006). Local partial-likelihood estimation for lifetime data. The Annals of Statistics, 34, 1, 290-325. [18] Fan, J. Q. and Zhang, W. Y. (2008). Statistical methods with varying coefficient models. Statistics and Its Interface, 1, 1, 179-195. [19] Gehan, E. (1965). A generalized wilcoxon test for comparing arbitrarily singlecensored samples. Biometrika, 52, 203-223. 124 Bibliography [20] Green, P., Jennison, C. and Seheult, A. (1985). Analysis of field experiments by least squares smoothing. Journal of the Royal Statistical Society. Series B (Methodological), 47, 299-315. [21] Green, P. J. and Silverman, B. W. (1994) Nonparametric regression and generalized linear models: a roughness penalty approach. Chapman and Hall, London. [22] Hamilton, S. A. and Truong, Y. K. (1997). Local linear estimation in partly linear models. Journal of Multivariate Analysis, 60, 1, 1-19. [23] Hardle, W., Liang, H. and Gao, J. (2000). Partially linear models. Springer, New York. [24] Hastie, T. and Tibshirani, R. (1990). Exploring the nature of covariate effects in the proportional hazards model. Biometrics, 46, 1005-1016. [25] Hastie, T. and Tibshirani, R. (1993) Varying-coefficient models. Journal of the Royal Statistical Society. Series B (Methodological), 757-796. [26] Heckman, N. E. (1986). Spline smoothing in a partly linear model. Journal of the Royal Statistical Society. Series B (Methodological), 48, 2, 244-248. [27] Hettmansperger, T. P. and McKean, J. W. (1988). Robust nonparametric statistical methods. Arnold, London. [28] Hjort, N. L. and Pollard, D. (1993). Asymptotics for minimisers of convex processes, Preprint. [29] Hoerl, A. E. and Kennard, R. W. (1970) Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12, 1, 55-67. [30] Hosmer Jr. and Stanley Lemeshow (1999) Applied survival analysis: regression modeling of time to event data. Wiley, New York. [31] Jin, Z., Lin, D., Wei, L. J. and Ying, Z. (2003). Rank-based inference for the accelerated failure time model. Biometrika, 90, 341-353. [32] Jin, Z., Ying, Z. and Wei, L. J. (2001). A simple resampling method by perturbing the minimand. Biometrika, 88, 2, 381-390. [33] Johnson, B. A. (2009). Rank-based estimation in the l1-regularized partly linear model for censored outcomes with application to integrated analyses of clinical predictors and gene expression data. Biostatistics, 0, 0, 1-8. 125 Bibliography [34] Kalbfleisch, J. D. and Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data, 2rd ed. Wiley, New York. [35] Koenker, R. and D’ Orey, V. (1987). Algorithm AS 229: Computing regression quantiles. Journal of the Royal Statistical Society. Series C (Applied Statistics), 36, 3, 383-393. [36] Krall, J. M., Uthoff, V. A. and Harley, J. B. (1975) A step-up procedure for selecting variables associated with survival. Biometrics, 31, 1, 49-57. [37] Lai, T. and Ying, Z. (1991). Large sample theory of a modified Buckley-James estimator for regression analysis with censored data. The Annals of Statistics 19, 3, 1370-1402. [38] Lai, T. and Ying, Z. (1991). Rank regression method for left-truncated and right censored data. The Annals of Statistics, 19, 2, 531-556. [39] Li, Y. and Ruppert, D. (2008). On the asymptotics of penalized splines. Biometrika, 95, 415-436. [40] Long, Q., Chung, M., Moreno, C. S. and Johnson, B. A. (2011). Risk prediction for prostate cancer recurrence through regularized estimation with simultaneous adjustment for nonlinear clinical effects. The Annals of Applied Statistics, 5, 3, 2003-2023. [41] Morris, C. N., Norton, E. C. and Zhou, X. H. (1994). Parametric duration analysis of nursing home usage. Case Studies in Biometry, 231-248. [42] Miller, R. and Halpern, J. (1982). Regression with censored data. Biometrika 69, 3, 521-31. [43] Parzen, M. I., Wei, L. J. and Ying, Z. (1994). A resampling method based on pivotal estimating functions. Biometrika, 81, 2, 341-350. [44] Powell, J. L., Stock, J. H. and Stoker, T. M. (1989). Semiparametric estimation of index coefficients. Econometrica, 57, 1403-1430. [45] Prentice, R. L. (1978). Linear rank tests with right censored data. Biometrika 65, 167-179. [46] Rao, C. R. and Zhao, L. C. (1992). Approximation to the distribution of Mestimates in linear models by randomly weighted bootstrap. The Indian Journal of Statistics, Series A, 54, 323-331. 126 Bibliography [47] Rice, J. (1986). Convergence rates for partially splined models. Statistics and Probability Letters, 4, 4, 203-208. [48] Ritov, Y. (1990). Estimation in a linear regression model with censored data. The Annals of Statistics 18, 303-328. [49] Ruppert, D. and Carroll, R. (1997). Penalized regression splines. Unpublished Technical Report. [50] Serfling, R. (1980). Approximation theorems of mathematical statistics. Wiley, New York. [51] Shiau, J., Wahba, G. and Johnson, D. R. (1986). Partial spline models for the inclusion of tropopause and frontal boundary information in otherwise smooth two and three dimensional objective analysis. Journal of Atmospheric and Oceanic Technology, 3, 4, 714-725. [52] Speckman, P. (1988). Kernel smoothing in partial linear models. Journal of the Royal Statistical Society. Series B (Methodological), 50, 3, 413-436. [53] Tian, L., Zucker, D. and Wei, L. J. (2005). On the Cox model with time-varying regression coefficients. The Journal of American Statistical Association, 100, 172183. [54] Tibshirani, R. (1996) Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological), 267-288. [55] Tsiatis, A. A. (1990). Estimating regression parameters using linear rank tests for censored data. The Annals of Statistics 18, 354-372. [56] Wahba, G. (1984). Partial spline models for the semiparametric estimation of functions of several variables. Statistical analysis of time series, 319, 329. [57] Wang, L., Kai, B. and Li, R. Z. (2009). Local rank inference for varying coefficient models. The Journal of American Statistic Association, 104, 1631-1645. [58] Wei, L. J. (1992). The accelerated failure time model: a useful alternative to the Cox regression model in survival analysis. Statistics in Medicine 11, 14-15, 18711879. [59] Xu, J. (2005) Parameter estimation, model selection and inferences in L1 -based linear regression. Ph.D. Thesis. 127 Bibliography [60] Xu, J. F., Leng, C. L., Ying, Z. L. (2010). Rank-based variable selection with censored data. Statistics and computing 20, 2, 165-176. [61] Ying, Z. (1993). A large sample study of rank estimation for censored regression data. The Annals of Statistics, 21, 76-99. [62] Zou, H. (2006) The adaptive lasso and its oracle properties. Journal of the American statistical association, 101, 476, 1418-1429. 128 [...]... Analysis 5 and requires weaker assumptions for censored regression than the former, it provides a more powerful tool for the study of the model (1.5) in practice Therefore, our research mainly focuses on the rank- based inferences for the censored accelerated failure time model On the basis of Jin et al (2003), for the accelerated failure time model (1.5) with the ˜ observed censored data (1.1), we can... of non-parametric terms for the partially linear accelerated failure time model The remainder of this thesis is organized in three main chapters In chapter 2, we study the estimation problem for the accelerated failure time partially linear model We discuss the merits and drawbacks of existing rank- based methods for the model fitting Noticing these drawbacks, we propose a local rank procedure along with... are still surviving at the time when the study is terminated and their true survival times were not recorded This was the case with Stanford heart transplant data, analyzed by Miller and Halpern (1982), etc What was observed for each of the subjects was the censoring indicator taking value 1 for death and 0 for surviving, and the minimum of the survival time and the censoring time See, Fan and Gijbels... CHAPTER 2 Partially Linear Accelerated Failure Time Model 2.1 Motivation Censored data can be analyzed under the accelerated failure time model, if the relationship between the logarithm of failure time and the covariates is assumed to be linear a priori While this has the advantage of producing good model estimates when the true relationship is consistent with the linear assumption, the resulting estimates... alternative to the proportional hazard model, the accelerated failure time model (Kalbfleisch and Prentice, 1980, p3234; Cox and Oakes, 1983, p.64-65), has become more appealing in handling the censored failure time data See also, Wei (1992) 1.1 Introduction to Survival Analysis 1.1.3 4 The Censored Accelerated Failure Time Model In parallel with the parametric proportional hazards model, the censored accelerated. .. nonlinear effect on the time to prostate cancer recurrence (Qi et al., 2011) In these studies, the uncertain effect is treated as the nonparametric component whenever the accelerated failure time partially linear model is employed If the predictor with the uncertain effect is independent of other predictors, the regression coefficients can still be estimated in a pseudo linear model When the independence... (1.3) Thus, the failure time variable T admits the linear regression form log T = β T X + ε, (1.5) with ε = log T0 It is called as accelerated failure time model or AFT model for short This ordinary regression model form (1.5) is easier to interpret the estimates of regression coefficients and requires no proportional hazards assumption as compared to model (1.2) These are important reasons for its increasing... resulting estimates may not be good when the dependence of the response on one of the covariates is uncertain To solve this, the accelerated failure time partially linear model is often used, incorporating 2.1 Motivation a nonparametric component into the accelerated failure time model for more flexibility This model can also be viewed as partially linear model for censored data Deviations from an assumed... popularity The estimation methods and their theoretical properties for the censored accelerated failure time model have been studied extensively, for example, the least square based approach ( Ritov, 1990; Lai and Ying, 1991a) as in Buckley and James (1979) and the rank based approach (Tsiatis, 1990; Lai and Ying, 1991b; Ying, 1993; Jin et al., 2003) proposed initially by Prentice (1978) Since the latter... In the absence of local smoothing, a global Gehan estimator based on piecewise linear approximation of the nonparametric term is also proposed In Chapter 3, we extended the local rank procedure along with kernel smoothing to the varying coefficients model for failure time data This is a pioneer work which allows one to capture the nonlinear interaction effects of covariates on the logarithm of failure time . RANK INFERENCES FOR THE ACCELERATED FAILURE TIME MODELS ZHOU FANG NATIONAL UNIVERSITY OF SINGAPORE 2014 RANK INFERENCES FOR THE ACCELERATED FAILURE TIME MODELS ZHOU FANG (B.Sc assumptions for censored regression than the former, it provides a more powerful tool for the study of the model (1.5) in practice. Therefore, our research mainly focuses on the rank- based inferences for. the rank- based inferences for the censored accelerated failure time model. On the basis of Jin et al. (2003), for the accelerated failure time model (1.5) with the observed censored data (1.1),