Ann, Inst. Statist. Math. Vol. 53, No. 4, 708-729 (2001) ©2001 The Institute of Statistical Mathematics A BOOTSTRAP APPROACH TO NONPARAMETRIC REGRESSION FOR RIGHT CENSORED DATA GANG LI’ AND SOMNATH DaTTA? 1 Department of Biostatistics, School of Public Health, University of California, Los Angeles, CA 90095-1772, U.S.A. 2 Department of Statistics, Franklin College of Arts and Sciences, University of Georgia, 204 Statstres Building, Athens, GA 30602-1958, U.S.A., e-mail: datta@stat.uga.edu (Received May 13, 1999; revised April 4, 2000) Abstract. In this paper a two-stage bootstrap method is proposed for nonparamet- ric regression with right censored data. The method is applied to construct confidence intervals and bands for a conditional survival function. Its asymptotic validity is es- tablished using counting process techniques and martingale central limit theory. The performance of the bootstrap method is investigated in a Monte Carlo study. An illustration is given using a real data. Key words and phrases: Bootstrap, Beran’s estimate, censoring, confidence bands, confidence intervals, Kaplan-Meier estimate, nonparametric regression, quantile re- gression. 1. Introduction and a review In lifetime data analysis, nonparametrically estimated conditional survival curves (such as the conditional Kaplan-Meier estimate) are useful for assessing the influence of risk factors, predicting survival probabilities, and checking goodness-of-fit of vari- ous survival regression models. However, it has not been an easy task to assess the variability of the estimated conditional survival curves. Consider the right censored survival regression data consisting of n iid. triples (X41, 61, 21),---,(Xn» Snr Zn), where X; = min(T;, Ci), 6 = I(T; < Cj), and T; > 0, C; > 0, and Z; represent the survival time, the censoring time, and the covariate, respectively, for the i-th subject under study, i =1,...,n. To ensure the identifiability of the model, we assume that for each i, T; and C; are conditionally independent given Z;. Let S(t | z) = P(T; >t | Z = 2) and A(t | z) = — fj S(ds | 2)/S(s— | z) denote the conditional survival function and the conditional cumulative hazard function of T; given Z; = z, respectively. We study the problem of constructing nonparametric confidence bands and intervals for S(t | z) and A(é | z) using the optimal rate conditional Kaplan-Meier estimate of Beran (1981). Such confidence bands and intervals can be used to assess the variability of the estimated con- ditional survival probabilities and provide a useful scale against which unusual features of the estimated conditional survival curve may be evaluated. Nonparametric estimation of the conditional survival function and its related func- tions was initiated by Beran (1981) and has been further studied by Dabrowska (1987, 1989, 1992), Li (1997), Li and Doss (1995), McKeague and Utikal (1990), and others. For the convenience of discussion, let us consider a simple version of Beran’s (1981) esti- 708BOOTSTRAPPING CENSORED REGRESSION 709 mate ),(t | z) of S(t | z) that is defined as the Kaplan-Meier estimate constructed using only those observations whose covariate Z fall inside a neighborhood (z — hn, z+ hn) of z, where hy is called the bandwidth (the general definition of Beran’s estimate is given in Section 2). Because only a portion of the data are used, the rate of convergence of Sp(t | z) to S(¢ | z) is typically slower than the n~'/? rate for the ordinary Kaplan-Meier estimate. The actual rate of convergence depends on how fast hn goes to 0 as well as on the smoothness of S. A decrease in the bandwidth would reduce the bias but, on the other hand, increase the variance of the estimate, and vice versa. Under certain smoothness conditions, it has been shown that (cf. Dabrowska (1987) and Li (1997) if hy is of order n~/5, then S)(t | z) converges to S(t | z) at an optimal rate. Moreover, _ (nhn)¥/?(Sat | 2) = SE 2) SUE 2), where for fixed z, U(t | z) is a continuous Gaussian martingale process with a nonzero mean (see (2.6) and (2.7) below for explicit expressions of the mean and variance function of U). In particular, the optimal rate of convergence for §,(t | z) is (nhy)~/? = O(n=2/5), It, however, remains an open problem as to how to construct confidence bands and intervals for S(t | z) using the optimal rate weak convergence result (1.1). The major hurdles are that U has an unknown nonzero mean and that the distribution of sup, |U(E | 2)| is intractable. One possible solution is to “undersmooth” the Beran estimate: if nh5, — 0, then the limiting process U will have a zero mean and a Hall- Wellner (1980) type confidence band for S(t | z) can be constructed (cf. Li and Doss (1995)). The drawback of this approach is that undersmoothing slows down the rate of convergence, which is not desired. ‘The main purpose of this paper is to study a bootstrap method for censored non- parametric regression and use it to construct confidence bands and intervals for S(t | 2) from the optimal rate Beran estimate. The basic idea of bootstrap is to first resam- ple from the observed data, then reconstruct the estimate of interest, say Beran’s es- timate $f(t | z), from the resampled data, and finally approximate the distribution of (nhn)/2(Sa(t | z) — S(t | z)) by that of its bootstrap version which can be obtained via computer simulation. An advantage of the bootstrap approach is that with an appro- priately designed resampling method, the bootstrap will correctly account for the bias of the estimated survival function §),(t | z). Therefore it does not require additional bias estimation or the use of suboptimal rate estimate (undersmoothing) for construct- ing confidence bands or intervals. It also automatically adapts to different variances of the estimated survival function at different covariate locations. However, it is not obvious what resampling scheme should be used for censored nonparametric regression as discussed below. ‘The idea of bootstrap was first introduced by Efron (1979) for homogeneous i.i.d. complete data setup in which the bootstrap is carried out by “resampling with replace- ment” from the sample data. This approach has since become a powerful tool in many statistical applications. See, for example, Efron and Tibshirani (1993) for the boot- strap method and its applications. Bootstrap for right censored data with no covariate was first studied by Efron (1981) who proposed two equivalent versions of bootstrap: a “simple” version and an “obvious” version. The “simple” bootstrap of Efron (1981) “resamples with replacement” from the observed pairs {(X;,6:),i = 1,...,n}. The va- lidity of this method was established by Akritas (1986) and Lo and Singh (1986). For710 GANG LI AND SOMNATH DATTA the nonparametric right censored regression setup considered in this paper, it seemed at first sight that a natural extension of Efron’s (1981) “simple” bootstrap would be to resample with replacement from the observed triples {(X;,6;, Z:),i = 1,...,n}. This resampling scheme, however, has a serious flaw that it fails to adequately take account of the type of bias of the optimal rate Beran estimate $,. Consequently, it will not give asymptotically correct results. This is very similar to the inappropriateness of the “usual” bootstrap in nonparametric density estimation and in ordinary nonparametric regression with complete data (cf. Hall (1992), Sections 4.4.2 and 4.5). Bootstrap in nonparametric regression with no censoring has been investigated by Hall ((1992), Sec- tion 4.5), Hardle and Bowman (1988) and Hardle and Marron (1991), among others. An ordinary nonparametric regression model for an observed data set (T;, Zi), i=1,...,7n, is Ty = m(Z) +6, 4 ym, where m(z) = E(T | Z = z) is the regression mean and the errors ¢; are independent and identically distributed with zero mean. In this case, the essential idea used in the aforementioned works is to resample from the estimated resid- uals é; = T;—thaa(Z;), 1 < i < n where 1h, (z) is a pilot nonparametric estimate (such as the kernel estimate) of the conditional mean m(z). Some of the resampling techniques for ef proposed in the literature include resampling from a set residuals determined by a window function (cf. Hardle and Bowman (1988)) and resampling each residual from a two-point distribution (wild bootstrapping) (cf. Hardle and Marron (1991)). After resam- pling, the bootstrap sample are formed as {(Tj7, Z;)}_, where Tj = ritg(Z;) +f, and g is taken to be larger than h. Obviously, it is not easy to extend these methods to the nonparametric censored regression setup because the estimated residual is not available when the survival time 7’ is censored. In this article, we study a different resampling approach for the nonparametric censored regression setup. The proposed resampling is carried out in two steps. In the first stage, we resample with replacement from the set {Z1,...,Zn} to obtain the bootstrap sample for the covariate {Z{,..., 24}. Then, in the second stage, we generate a pair (X}, 67) for each Z} using ideas similar to that of Efron (1981). Let us use a special case to explain how it works. For each Z} (1 < i 0 determines the size of the neighborhood. It can be shown (cf. Efron (1981)) that the resampling methods used in the second stage is equivalent to generating Tr ~ Sp(t | Zf), Cf ~ Gait | 22) and letting X} = min{T},Cy}, 6 = I(Tf < C3), where S,(t | Zt) and G,(t | 23) are the Kaplan-Meier estimates of the survival distributions of T’ and C, respectively, constructed from the data A;. It is interesting to mention that if g, = 0, then this two-step procedure is equivalent to resampling with replacement from the set of triples {(Xi,6i,%), i = 1,...n}. This provides an intuitive explanation why “resample with replacement” from the sample triples would not be appropriate for bootstrapping the estimated conditional survival function $),(¢ | z). In order to properly account for the bias of 5,(t | z), gn has to be larger than hn. Recall that hy is the bandwidth used in the construction of Beran’s estimate whose distribution needs to be bootstrapped. Although the technical reasons will be discussed in later sections, we point out that intuitively, a larger gn enables one to catch more bias. The conditions imposed on gn will be given in the next section. More discussion on this point can be found in Remarks 2.1 and 4.1. Earlier, Van Keilegom and Veraverbeke (1997) studied bootstrap for censored non- parametric regression under fixed design where one only needs to resample the survival times from the conditional distribution. For random designs, one has to decide how toBOOTSTRAPPING CENSORED REGRESSION m deal with the random covariate in the resampling process. Some discussion of the dif- ference in bootstrapping fixed and random design linear regression models for complete data can be found in Freedman (1981). In this paper we propose a two-step bootstrap for random design that involves resampling the covariates in the first stage and the sur- vival times in the second stage as discussed earlier. This does not resemble resampling residuals as what has been done for nonparametric regression with complete data. To the best of our knowledge, the two-step procedure is new in the literature even for com- plete data. The additional phase of resampling of the covariates also introduces more technical challenge for establishing the asymptotic validity of the bootstrap. Our jus- tification of the two-stage bootstrap for random design is different from that of Van Kellegom and Veraverbeke (1997) for fixed design. Our approuch may also be adopted to study bootstrap in nonparametric regression for other important incomplete survival data such as left-truncated data and both left-truncated and right-censored data, which are well known special cases of the counting process model. We finally note that boot- strap methods have been studied for some semiparametric regression models such as the Cox proportional hazards model; see, e.g., Hjort (1985) and Burr (1994), among others. As mentioned earlier, a major use of bootstrap in this paper is to solve the problem of bandwidth selection in censored regression. We will present some numerical results on the performance of the proposed bootstrap method and illustrate its use through a real data example. It is worth noting that Li and Doss (1995) proposed a different class of estimators for the conditional survival function using local linear hazard smoothing. ‘Their estimators have less bias near the boundary of the covariate space than Beran’s estimators. Although nontrivial modifications may be needed, bootstrapping the esti- mators of Li and Doss (1995) could be studied along similar lines. Finally, we note that random design is more common than fixed design in clinical trials and epidemiological studies where the covariate is usually observational. In Section 2, we give two equivalent resampling algorithms for bootstrapping the optimal rate Beran estimate of the conditional survival function. The algorithms actually use weighted resampling which include the local resampling discussed earlier as a special case. We state conditions on g, under which the proposed method is asymptotically valid. We describe how to construct confidence intervals and equal precision bands using the proposed bootstrap method. We also discuss the use of bootstrap for data- driven bandwidth selection in censored nonparametric regression. Further extensions to pointwise and simultaneous confidence intervals for a quantile regression function at different covariate points are also discussed. In Section 3, we illustrate our method on a real data set and present numerical results about the performance of the proposed bootstrap method for finite sample sizes. The proofs are collected in Section 4. 2. Main results 2.1 Notations and assumptions Recall that A(t | z) and S(¢| z) denote the conditional cumulative hazard function and conditional survival function, respectively, of the survival time T given Z = z. Let G(t | z) = P(C; > t | 2; = 2) denote the conditional survival function of the censoring time C given Z = z. Assume that Z has a density f(z) and let a(t | z) = GA(¢ | z)/dt denote the conditional hazard function of T given Z = 2: Let NW) =X St6=1) and Vi) = MX 2), ¢72 GANG LI AND SOMNATH DATTA In addition, define Fi(t | 2) = E(Ni(t) | 4 = 2) = P(X $6 = 1] % (2.1) F(t | 2) = E(t) | 4 = 2) = P(X 2t| 2 =2), Hy(t,2) = He(t|2)f(z), = 1,2. The following assumptions will come into play in what follows. AssumPTION A. Let w(-) be a density function vanishing outside [—1, 1]. (A.1) fwo(u) =0 and f'urw(u)du < oo. ia 2) w(-) is of bounded variation. (A.3) The derivative function w’(-) is of bounded variation. Let r > 0 be areal number such that S(r | 2)G(r | z) > 0, I= [a,b] (-oo 0) be a 6 ee of I. For any real function f(t, z) and any interval J, denote || f||; = sup{|f(¢,z)|:0 0 and inf{ f(z) : z € Ig} > 0 for some 6 > 0. (B.2) ||0'H,/d2"||f, < 00 for i =0,1,2,3 and k = 1,2. (B.3) For each s € [0,7], a(s | z) is twice differentiable with respect to z and satisfies a(s | w) = where there exists a constant Ky > 0 independent of s and z such that a(s | 2), 0,(s | 2), a,(s | 2), and 7(s,2) are all bounded by K; for all s € (0,7] and z € I. Here a’, and a, denote the first and second order partial derivatives of a with respect to 2. 8 |2) +a,(s | z)(w— 2) +a,4(s | z)(w— 2)? + 418, 2)(w— 2) 2.2 The Beran estimate For readers’ convenience, we review some results about the Beran-type estimates. ‘The Beran (1981) kernel estimates of the conditional cumulative hazard function A(¢ | z) and the conditional survival function S(¢ | z) are defined by “ co hero [BAA (=). and Sp(t | z) = [1, 0 is the bandwidth parameter. ‘The Beran estimates can be considered as weighted average estimates. For example, the jump size AA,(t | z) of An(¢ | 2) at time t is a weighted average of the jumps of the N;’s at time t for those subjects who are at risk. It is easy to see that if Wy; = 1 for all ¢, then A), and $), reduce to the ordinary Nelson-Aalen estimate and Kaplan-Meier estimate for homogeneous data. If w(-) is taken to be the uniform density on (~1,1), then Ay, is , 1si c > 0, then, under regularity conditions (24) Vahn(Ant:|2) = AC 12)) 4 012), (2.5) ay (Sn(-12)— 8-12) 4 8( 12)UC 12), in D(0,7] for any r such that $(r | z)G(r | 2) > 0, where D[0, 7] is the standard Skorohod space on [0,7] and U(- | z) is a continuous Gaussian martingale with mean (26) ie |2) = vel fPutanad] [f ac us] tvi [ freee a| [ f Pwenaa] and variance function (2.7) o(t|2)= [ [ween] if Has). 2.3 Bootstrapping the Beran estimate We first give two equivalent algorithms for bootstrapping the Beran estimate of the conditional survival function. Then we state the main results that provide a theoretical justification of the proposed bootstrap method. The Simple Weighted Bootstrap. Generate Zj,...,Z} i.i.d. from the empirical dis- tribution of {Zi,...,Zn}- For each i, generate a pair (X}, 6) from the weighted empir- ical distribution Fy(-,- | Z}) of {(X1,61)s---s(Xns6n)}, where (2.8) F(u,v | z) = = YL Wale) 2)I(X; | |= SEeite ‘ Fig. 2. 90% simultaneous confidence bands for S(t | 0.5) and S(t | 2.35) over 5 years. In Table 1 we report the results of simulations to estimate the coverage probabilities of the bootstrap confidence band for the conditional survival function S(t | 0.5) on the interval [0,2]. Here the covariate Z has a uniform distribution on (0,1) and the conditional survival and censoring distributions (given Z = z) are F(t | z) = 1- exp(—2"t) and G(t | z) = 1—exp(—bzt). The parameter 6 is adjusted to give the prescribed censoring rate. Note that for this example, P(T > C | Z = z) = b/(1 +) which does not involve z. In the simulation, we used hy = cn~"/5 and gy = en~°! with718 GANG LI AND SOMNATH DATTA ‘Table 1. Observed coverage probabilities of bootstrap bands for S(¢ | z) over the interval [0, 2]. Nominal Level=.90 Nominal Level: Sample size Censoring rate n 5% 20% 40% 70% 5% 20% 40% 70% 50 ol 92 90 90 95 95 93 95 100 94 =92 92 81 96 96 95 88 200 88 88 87 80 91 91 89 87 500 91 90 90 84° 97) 98 97 95 1000 90.90 91 85 95 95 95 94 ‘Table 2. Observed coverage probabilities of bootstrap confidence intervals for S(¢ | 2) at t = 2. Nominal Level=. Sample size Censoring rate Censoring rate n 5% 20% 40% 70% 5% 20% 40% 70% 50 91.900 8780S 100 92 90 88 71 97 97958. 200 91 93 94 73979958 500 97 95 92 83) 9897998 1000 $60 00) OO re Onl Odi) 20s) 2 00 ¢ = 0.70 which is close to the optimal value of c for a global bandwidth. For a given sample, 1,000 bootstrap samples were used to construct confidence bands and intervals. Each entry in the table was based on 100 Monte Carlo samples. (It would be ideal to use a larger number of Monte Carlo samples. However, this would require enormous computation time for the simulation when the sample size n exceeds 500. We did run the simulation for n = 50 and 100 using 1,000 Monte Carlo samples and the results were consistent with those reported here.) Table 2 was similarly constructed except that it reports estimated coverage proba- bilities of the bootstrap confidence intervals for a single conditional survival probability S(2| 0.5). It is seen from both tables that with the exception of heavy censoring, the coverage probabilities are observed to be close to their nominal values. In the case of heavy censoring, a much larger sample size would be needed. 4. Proofs PROOF OF PROPOSITION 2.1. The proof is similar to that of Efron (1981) and is omitted. A detailed proof can be obtained from the author upon request. (1 Next we prove Theorem 1. Because (2.12) is a consequence of (2.11) and the functional delta method, we only prove (2.11). Our proof of (2.11) is carried out by verifying the conditions of the martingale central limit theorem for Vnhn(Aj(t | z) — Ag(t | 2)), conditional onBOOTSTRAPPING CENSORED REGRESSION n9 the data. It is important to note that the techniques involved in the proof are sub- stantially different from those used in Li (1997) for establishing weak convergence of Vink (An(t | 2) — A(t | 2)). In particular, we need to derive the rate of uniform (in z) almost sure convergence for the Beran estimate A and its partial derivative 0A/dz and for some nonparametric mean regression estimators. In contrast, only weak consistency results with fixed z were needed in Li (1997). The following lemma gives the rates of uniform strong convergence for A and 04/dz, which will play a crucial role in the proof of Theorem 1. LEMMA 1. (Rate of uniform strong convergence for A, and its first order derivative) Assume that (A.1), (B.1), and (B.2) hold. (a) Fon = = (Seema 0 and Y*, gh < co for some p > 0, then, with proba- bility 1, (4.1) \|Ag — Allf = O(max(an, gn) a5 > 00, where ||f|l7 =sup{|f(t,2)| 10 0, then, with proba- bility 1, aA, aal|” = F . ve 24 = O(max(ta,9)) a3 n+00 (4.2) | PRoor. (a) Let Hgi(t,z) = yu (42 *) Ni(t) and Aga(t,z) = ia” (42 2) x N9n It follows from Dabrowska ((1989), p. 1165, lines 5 and 7) that (4.3) P(\|Hgn — EAgelli > ©) < conga’ exp(—cixe’ngn), k=1,2, forall €>0, where coi, Coz, C11, and ciz are some universal positive constants. Letting ¢ = an {(1+ p)/ciz}/?, the above inequalities reduces to P(\\Hox — EA gell7 > en{(1 + p)/crn}/) < congh, k= 1,2. This, combined with the Borel-Cantelli lemma, implies that with probability 1, (4.4) Wg — EAgell7 =O(an), k= 1,2. Note that EAy,(t,z) = f w(u)Hy(t,z + gnu)du for k = 1,2. An application of the Taylor expansion and (A.1) leads to oH =O), k=1,2, a2 —rrrr—S~—SOO [fv mr waa] 5 Is for sufficiently large n. It follows from (4.4) and (4.5) that (4.6) Hoe — Hell?720 GANG LI AND SOMNATH DATTA Write : A(t 12)- at 12) = [ oe fe jttlla(s.2) —Blin(s,2) vale, a 4 : +f gal, 2 ar dHa(s,2) a fae PAals) - 62) , +f atc - nec dEHy,(s,2) lo + “fo |BHizale.) Hals.z) and denote by J; the /-th term on the right hand side of the equality, 1 = 1,...,4. It can be shown that there exist some positive constants Ki, Ke, Kg and K’, such that for sufficiently large n Wllf < Kalix ~ Bffgi|\7, Mall < Kall ya ~ BAgall?, MalIf < Kelli — Half, and all < Kall Hye — Holly. Therefore, for sufficiently large n (4.7) Ag ~All < (all + Wall + Wilh + Malle 2 SKY (\\Aoe — EAlgullh + BAe — Hell?) at with probability 1, for some positive constant K. This, combined with (4.4) and (4.5), implies (4.1). (b) Now we prove (4.2). Similar to (4.3), it can be shown that P Og _ poten Oz Oz > ‘) S bowgn' exp(—bine’ngn), k= 1,2, T for all ¢>0, for some positive constants bo1, bo2, 611 and bi. This, together with the argument leading to (4.4), implies that with probability 1, align _ Hye Oz Oz (4.8) =O(f,), k=1,2. fi Similar to (4.5), the Taylor expansion and (A.1) leads to ; : oer (4.9) |e | - a <@ [f stotwran] a =0(92), k=1,2 i Ie Finally, : OF (8,2) oA,(t | z) -f{ oz 0 oz gals, 2) andBOOTSTRAPPING CENSORED REGRESSION 721 i 4 [ea [Ae dACt | 2) 32 @z Rf ey eae ae) Similar to (4.7), it can be shown that there exists a positive constant K such that for sufficiently large n, 1A ||" 2 : : - Z| S ie (it ~ Bg || + |B gx — Helly re | ) with probability 1. This, combined with (4.4), (4.5), (4.8), and (4.9), proves (4.2). OF, — OH | 32 ae | + Oz Oz le olla OH| LEMMA 2. Assume that (£1, Z1),.--,(6n;%n) are tid random vectors taking values in {0,1] x R and that the density function f(z) of Z satisfies infacz<» f(z) > 0 for some —00 Wasl2)éi, ‘where Wa;(z) is defined by (2.3). In addition, for any function g(z), define (z) = APM f wro(uydu and n(2)=—y Assume that (A.1) and (A.2) hold and that g(z) is continuously differentiable and m(z) and f(z) are twice continuously differentiable on (a ~ €,b + €) for some € > 0. (a) If hn 0 and for some fired 0 0, then, as n— 00, (4.11) sup hz" |m,(z) ~ m(z)| £5 0. asz 0) < exexp(—cap?ai*”) aS2 0, for some universal positive constants c1, ¢2, a1, and ag. This, combined with the Borel- Cantelli lemma and (4.16), implies that (4.14) holds. We can prove (4.15) along the same lines. Therefore (4.11) is proved. (b) From the Taylor expansion and (A.1), we have Evbal(2) = f Foy) LOO) di, <1. v ysztAhnu ‘Thus, similar to (4.16), we have (4.17) sup |EYn(z) — ¥(z)| > 0. ones Let F%(z) denote the cumulative distribution function of Z; and let F(z) be the em- pirical distribution function of Z;,...,Zn- By integrating by parts, it is shown that for 2 € (a,}] and sufficiently large n, Ivn(2) - Evbn(z)| < Koh,” sup [Fy (u) ~ F7(u)], for some constant Ko > 0, where the second equality is obtained by integrating by parts. This, combined with the following inequality (cf. Nadaraya (1970), (6)) a P (sup ieF(u) — Fal > ) 0 is a universal constant) implies that P( sup |Yn(z) — Etn(z)| > ’) S eo exp(—aop?nh), oseBOOTSTRAPPING CENSORED REGRESSION 723 where ao = 2/K3. Applying the Borel-Cantelli lemma, we have (4.18) sup |Pn(z)— Evn(z)| 30 as noo. aszsb ‘Therefore, (4.12) follows from (4.17) and (4.18). 0 LeMMA 3. Let Zj,...,Z% be defined as in Subsection 2.3. Define fi(z) = xf- DL, w(7E=*) to be the bootstrap counterpart of f(z) in (4.13). Assume that hy — 0 and S~°, exp(—pnhn) < co for all p > 0. Then, for almost all sequences Z;,Zs,..-, conditional on (Z1,-..;Zn), f(z)? f(z) as n — 00 at every continuity point z of f. Proor. Let z be a continuity point of f. By Rao ((1983), Theorem 3.1.5), fa(z) —*** f(z) as n — oo. The conclusion follows from the facts that for al- most all sequences Z, Z2,...,E(f2(z) | Zi,..+,Zn) = fa(z) > f(z), and Var(f2(z) | 2iy.) En) = aie iy Di P(A) — Fh) > 0.0 The following lemma gives a list of results that will be repeatedly used in the proof of Theorem 1. LeMMa 4, Assume that hy — 0 and gy — 0. Assume further that (4.19) Sexp(-pnhs) < 00, forall. p> 0, nat and that for some constant 00. In addition, assume that the density f(z) of Z satisfies the conditions of Lemma 2. Let Aifal(s,z) = an” (A and, for 10, 8€(0,7] f= (425) sup is? Dele, 2) 2 = 2) — Hz1(s,2) [ae] [eww 29, (4.26) a? Det ls, 2)(Zf - 2)? - | wu(u)dul 2 0, (427) vf le!?(ds,2)| = Sref"(t,2) + Opl(Rin)!?), for any p> 0, i179 i=l for almost all sequences (X1,61, Z1),---, where Ho(s, 2) is defined in (2.1). PRooF. Throughout the proof expectations are taken conditional on (%1, 61,21), «+ (Xn bas Zn). (a) Let 0< s Woe(2)¥i(s). i= It can be shown that Bllia(s.2) = a Yow (A=) thats | 24) - Hale | 20 Denote by J, and Jp the first and second term on the right hand side of the last equality. Then, as n — 00, ins [zi ¥+(4=4)| [ssp Alas | 2)— Hols |2)| 25 #(2)-0=0, where the almost sure convergence follows from (4.15) and (4.11). Moreover, by (4.14), I, —** Hp(s | 2) f (2) = Ha(s,2), 08 n + 00. Thus, BHj,(s, 2) + Ho(s,2). Similarly, it can be shown that as n —» co Var(Hia(s, 2) = —* ) Beals | 2s) ~ ABA, 2))? 250. Therefore, for any s, conditional on (X1,61,Z1);---;(Xny6n, Zn), Hfa(s,z) 9? Ho(s,z) along almost all sequences (X1,61,2Z1),-... One can also show that for any 8, conditional on (X1,61,Z1),..-;(Xns6n; Zn), Hjo(s+,z) +? Ho(s+,z) along almost all sample sequences {(Xi,6i, Z;),i = 1,2,...}. Using the standard argument similar to those in the proof of Theorem 5.5.1 of Chung (1974) and the fact that Ho(s,z) is left-continuous and nonincreasing, we conclude (4.23).BOOTSTRAPPING CENSORED REGRESSION 725 Parts (c)-(d) can be proved by applying (a) and Lemmas 2 and 3. Here we omit the details. 0 PROOF OF THEOREM 1. It is not difficult to verify that the conditions of gn in Lemma 1 and the conditions of hy and g, in Lemma 4 are satisfied if nh — c > 0 and (2.10) holds. So, all results in Lemmas 1 and 4 apply here. Throughout the proof all convergence are conditional on the observed sample data (X1, 61, 21), ...,(Xns On: Zn)> Now let’s prove (2.11). Note that (Aj(t | 2) ~ Ag(t | 2)) -(f% Yetls.nany(s) — Asi) i=l = [ Leteaucin+ [Xse CSCS (4.29) = Wilt | z) + W(t | z), where c{(s,z) and Nf(s) (1 < i 0. Moreover, integrating by parts, we have [(OL = |rtin S 7 e7?(t, 2)(Ag(t | Zf) - A(t | Z3)) i=l —nlin > [aut | 23) — A(s | Z3))def?(s, 2) S VAg— Allf rtm 37 2%(t,2) + Ag — AIF f | def%(s, 2) ei ist IA, — Allz (2 St? 2) + (aha) Op (4a) raf 40 uniformly int € [0,7], with probability 1, where in the third step we have used (4.27) and the last step follows from (4.1) and (4.24). Hence, with probability 1, (4.32) (VninWi,, VrhaW7,)() = h(Q)+h(t) 4 o7(t| 2) uniformly int € [0,7]. Similarly, one can use Lemma 1 and Lemma 4 to verify the Lindeberg condition that for any €>0, [Do labaet(s.2)°% (US Aine(s 2) > OAgls | 22) 0 j=1 set [% nh, WAY aPC) 2)dA,(s | Z3) 0 (4.33) 40, with probability 1. Therefore, (4.30) follows from (4.32), (4.33), and Rebolledo’s mar- tingale central limit theorem (cf. Andersen and Gill (1982), Theorem I.2). Now, we verify (4.31). Write (434) Vnh,W2,(t | 2) = vas f dat (s,2)(a(s | 21) — a(sle))as t+ nha f So -d((Ag(s | 27) — A(s | 23} — [Ag(s | z) - Als | 2))) = Val, [ Saeatas | Zf) — a(s | 2))ds 0 G=1 +Vnhn [dee -t{ 2 (gle 2(0))— Ato] 2100] (2 ~2) = AO +A,BOOTSTRAPPING CENSORED REGRESSION 1 where %(s) is between Zy and z, and J;(t) and Jo(¢) denote the first and second term on the right hand side of the second equation. Then, ae = f * ab(s|2) [vars Yals.(Z - 5) ds + fetal) [Vike Soa 2(Z5 - =| ds 0 = +f [vars Dals.2)nl6, 2002 — >| a 0 isl (4.35) 4 Veult | 2) uniformly in t € [0,7], where the first equality is from Assumption (B.3) and the last step follows from (4.25), (4.26), the assumption that nh> — c, and the fact that lf [vate 3 aceeit, Zj)( Zp - »| ds| < Kiln i Van 5 eil0,2)(29 ~z)"ds. Et Moreover, integrating by parts, we have Lalol < [vane Se (waar o(® att Lt 24} 3t0))| dz Ag(s|2(s))_ A(s | 4i(s)) Oz Oz “(Sate + flex (al) 2 (eee tz) + ont) 1 woo a val wast) 2+ O,(1)) (436) = oft), el Sie -al -Ideg(s, 2)] “al 3/240 Ee oe where the third step uses (4.27), the fourth step follows from (4.2), and the last step follows from (2.10). Combining (4.34), (4.35), and (4.36), we prove (4.31). 0 Remark 4.1. It is seen from the above proof that in order for the bootstrap to pick up the correct amount of bias (i.e. for (4.31) to hold), Se has to be uniformly strong consistent at appropriate speed (see the last two steps in (4.36). This, in turn, it requires that gp satisfies condition (2.10). As mentioned in Remark 2.1, this means that gn has to go to 0 at a slower rate than hn.728 GANG LI AND SOMNATH DATTA. Acknowledgements ‘The authors are grateful to a referee to providing constructive comments. The research of Gang Li was partially supported by a National Institute of Health grant and a National Science Foundation grant. The research of Sommath Datta was partially supported by a grant from the National Security Agency. REFERENCES Akritas, M. (1986). Bootstrapping the Kaplan-Meier estimator, J. Amer. Statist. Assoc., 81, 1032-1039. Andersen, P. K. and Gill, R. D. (1982). Cox's regression model for counting processes: A large sample study, Ann. Statist, 10, 1100-1120 Beran, R. (1981). Nonparametric regression with randomly censored survival data, Tech. Report, De- partment of Statistics, University of California, Berkeley. Burr, D. (1994). A comparison of certain bootstrap confidence intervals in the Cox model, J. Amer. Statist. Assoc., 89, 1290-1302. Cao, R. and Gonzélez-Manteiga, W. (1993). Bootstrap methods in regression smoothing, J. Non- parametr. Statist., 2, 379-388. Chung, K. L. (1974). A Course in Probability Theory, 2nd ed., Academic Press, New York. Dabrowska, D. M. (1987). Nonparametric regression with censored survival time data, Scand, J. Statist., 14, 181-197. Dabrowska, D. M. (1989). Uniform consistency of the Kernel conditional Kaplan-Meier estimate, Ann. Statist., 7, 1-26. Dabrowska, D. M. (1992). Variable bandwidth conditional Kaplan-Meier estimate, Scand. J. Statist., 19, 351-361. Efron, B. (1979). Bootstrap methods: Another look at the jackknife, Ann. Statist., 7, 1-26. Efron, B. (1981). Censored data and the bootstrap, J. Amer. Statist. Assoc., 76, 312-319. Efron, B. and Tibshirani, R. J. (1993). An Introduction to the Bootstrap, Monogr. Statist. Appl. Probab., No. 57, Chapman & Hall, London. Fleming, T. R. and Harrington, D. P. (1991). Counting Processes and Survival Analysis, Wiley, New York. Freedman, D. A. (1981). Bootstrapping regression models, Ann. Statist., 6, 1218-1228. Gill, R. D. (1980). Censoring and stochastic integrals, Mathematical Centre Tracts, 124, Mathematical Centre, Amsterdam. Hall, P. (1992). The Bootstrap and Edgeworth Expansion, Springer, New York. Hall, W. J. and Wellner, J. A. (1980). Confidence bands for a survival curve from censored data, Biometrika, 67, 133-143. Hardle, W. and Bowman, A. W. (1988). Bootstrapping in nonparametric regression: Local adaptive smoothing and confidence bands, J. Amer. Statist. Assoc., 88, 102-110. Hardle, W. and Marron, J. 8. (1991). Bootstrap simultaneous error bars for nonparametric regression, Ann. Statist., 19, 78-796. Hiort, N. L. (1985). Bootstrapping Cox’s regression model, Tech. Report, No. 241, Department of Statistics, Stanford University, California. Li, G. (1997). Optimal rate local smoothing in a multiplicative intensity counting process model, Journal of Mathematical Methods of Statistics, 6, 224-244. Li, G. and Datta, 8. (1999). A bootstrap approach to nonparametric regression for right censored data, ‘Tech. Report, Sta 99-8, Department of Statistics, University of Georgia, Athens. Li, G. and Doss, H. (1995). An approach to nonparametric regression for life history data using local linear fitting, Ann. Statist., 23, 787-823. Lo, 8. H. and Singh, K. (1986). The product-limit estimator and the bootstrap: Some asymptotic representations, Probab. Theory Related Fields, 71, 455-465. McKeague, I. W. and Utikal, K. J. (1990). Inference for a nonlinear counting process regression model, Ann. Statist., 18, 1172-1187.BOOTSTRAPPING CENSORED REGRESSION 729 Nadaraya, E. A. (1970). Remarks on nonparametric estimates for density functions and regression curves, Theory Probab. Appl., 15, 134-137. Rao, B. L. S. Prakasa (1983). Nonparametric Functional Estimation, Academic Press, New York. Van Keilegom, I. and Veraverbeke, N. (1997). Weak convergence of the bootstrapped conditional Kaplan- Meier process and its quantile process, Comm. Statist. Theory Methods, 26, 853-869.