Abstract
We study nonparametric estimation of the illness-death model using left-truncated and right-censored data. The general aim is to estimate the multivariate distribution of a progressive multi-state process. Maximum likelihood estimation under censoring suffers from problems of uniqueness and consistency, so instead we review and extend methods that are based on inverse probability weighting. For univariate left-truncated and right-censored data, nonparametric maximum likelihood estimation can be considerably improved when exploiting knowledge on the truncation distribution. We aim to examine the gain in using such knowledge for inverse probability weighting estimators in the illness-death framework. Additionally, we compare the weights that use truncation variables with the weights that integrate them out, showing, by simulation, that the latter performs more stably and efficiently. We apply the methods to intensive care units data collected in a cross-sectional design, and discuss how the estimators can be easily modified to more general multi-state models.
Similar content being viewed by others
References
Andersen P, Borgan O, Gill RD, Keiding N (1993) Statistical models based on counting processes. Springer-Verlag, New York
Asgharian M, M’Lan C, Wolfson D (2002) Length-biased sampling with right censoring: an unconditional approach. J Am Stat Assoc 97:201–209
Chang S, Tzeng S (2006) Nonparametric estimation of sojourn time distributions for truncated serial event data—a weight-adjusted approach. Lifetime Data Anal 12:53–67
Datta S, Satten G (2001) Validity of the Aalen-Johansen estimators of stage occupation probabilities and Nelson-Aalen estimators of integrated transition hazards for non-Markov models. Stat Probab Lett 55:403–411
Gill R (1992) Multivariate survival analysis. Theory Probab Appl 37(1):18–31
Gill R, van der Laan M, Wellner J (1995) Inefficient esttimators of the bivariate survival function for three models. Annales de l’Institut Henri Poincaré - Probabilités et Statistiques 31(3):545–597
Hougaard P (2000) Analysis of multivariate survival data. Springer, New York
Huang Y, Wang M-C (1995) Estimating the occurrence rate for prevalent survival data in competing risks model. J Am Stat Assoc 90(432):1406–1415
Kalbfleisch J, Prentice R (2002) The statistical analysis of failure time data. Wiley, Hoboken
Keiding N (1991) Age-specific incidence and prevalence: a statistical perspective. J R Stat Soc Ser A 154(3):371–412
Kosorok M (2008) Introduction to empirical processes and semiparametric inference. Springer, New York
Lin D, Sun W, Ying Z (1999) Nonparametric estimation of the gap time distributions for serial events with censored data. Biometrika 86(1):59–70
Mandel M (2010) The competing risks illness-death model under cross-sectional sampling. Biostatistics 11(2):290–303
Mandel M, Betensky R (2007) Testing goodness of fit of a uniform truncation model. Biometrics 63(2):405–412
Mnatzaganian G, Galai N, Sprung CD, Zitser-Gurevich Y, Mandel M, Ben-Hur D, Gurman G, Klein M, Lev A, Levi L et al (2005) Increased risk of bloodstream and urinary infections in intensive care unit (ICU) patients compared with patients fitting ICU admission criteria treated in regular wards. J Hosp Infect 59:331–342
Neuhaus G (1971) On weak convergence of stochastic processes with multidimensional time parameter. Ann Math Stat 42(4):1285–1295
Prentice R, Moodie Z, Wu J (2004) Nonparametric estimation of the bivariate survivor function. In Lin D, Heagerty P (eds) Proceedings of the second Seattle symposium in Biostatistics. Lecture notes in statistics, vol. 179. Springer, New York
Putter H, Fiocco M, Geskus RB (2007) Tutorial in biostatistics: competing risks and multi-state models. Stat Med 26(11):2389–2430
Qin J, Shen Y (2010) Statistical methods for analyzing right-censored length-biased data under Cox model. Biometrics 66:382–392
Robins J, Rotnitzky A et al (1992) Recovery of information and adjustment for dependent censoring using surrogate markers. In: Jewell N, Dietz K, Farewell V (eds) AIDS epidemiology—methodological issues. Springer, Boston
Rubin D (1981) The Bayesian bootstrap. Ann Stat 9:130–134
Tsai W-Y (1990) Testing the assumption of independence of truncation time and failure time. Biometrika 77(1):169–177
Vakulenko-Lagun B, Mandel M (2016) Comparing estimation approaches for the illness-death model under left truncation and right censoring. Stat Med 35:1533–1548
van der Laan M (1996) Nonparametric estimation of the bivariate survival function with truncated data. J Multivar Anal 58(1):107–131
Wang M-C (1989) A semiparametric model for randomly truncated data. J Am Stat Assoc 84:742–748
Wang M-C (1991) Nonparametric estimation from cross-sectional survival data. J Am Stat Assoc 86:130–143
Wang M-C (1999) Gap time bias in incident and prevalent cohorts. Stat Sin 9:999–1010
Wang M-C, Jewell N, Tsai W-Y (1986) Asymptotic properties of the product limit estimate under random truncation. Ann Stat 14(4):1597–1605
Wang W, Wells M (1998) Nonparametric estimation of successive duration times under dependent censoring. Biometrika 85(3):561–572
Acknowledgments
We thank the two reviewers for their valuable comments and suggestions. The work was supported by The Israel Science Foundation (Grant No. 519/14) and by NSF grant DMS-1407732.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
1.1 Convergence to a Gaussian process
Definition 1
Let \(X_1,\ldots ,X_n\) be random variables. For every function f, define \(Pf\equiv E[f(X)]\). Define \(\mathbb {P}_n\) to be the empirical measure, such that \(\mathbb {P}_n f\equiv n^{-1} \sum _{i=1}^{n}f(X_i)\). Define \(\mathbb {P}_n^{(b)}\) to be the bootstrap empirical measure with weights \((M_{n1},M_{n2},\ldots ,M_{nn})\) which are i.i.d., positive random variables with expectation 1 and finite variance, and independent of \(X_1,\ldots ,X_n\), such that \(\mathbb {P}_n^{(b)} f\equiv n^{-1} \sum _{i=1}^{n} (M_{ni}/\bar{M}_n)f(X_i)\) where \(\bar{M}_n=n^{-1}\sum _{i=1}^n M_{ni}\). The convergence type for bootstrap \(\sqrt{n}(\mathbb {P}_n^{(b)}-\mathbb {P}_n)\underset{\text {M}}{\overset{\text {P}}{\rightsquigarrow }}\mathbb {G}\) is defined in Kosorok (2008, pp. 19–20). Finally, for a space \(\mathcal {X}\), define \(\ell ^{\infty }(\mathcal {X})\) to be the space of all uniformly bounded real functions on \(\mathcal {X}\).
Let
Let N be the expectation of \(\widehat{N}\), in other words, \(N(t,u,l)\equiv P(T\le t, U\le u,L\le l, C>T+U-L \mid L\le T+U)\). We have \(N(t,u,l)=G^*_{\widetilde{T}^*,\widetilde{U}^*,L^*}(t,u,l)\) as defined in Eq. (1). In particular,
Lemma 1
The process
where \((\mathbb {G}_1,\ldots ,\mathbb {G}_4)^T \in \ell ^\infty ([0,\tau ]^3)\times (\ell ^\infty [0,\tau ])^3\) is a tight zero-mean Gaussian process with covariance structure that appears in the proof. Moreover, its corresponding bootstrap process
Proof
Write \(\widehat{N}(t,u,l) \equiv \mathbb {P}_n f_{t,u,l}(F^*,\widetilde{T}^*,\widetilde{U}^*,L^*)\), where \(f_{t,u,l}(F^*,\widetilde{T}^*,\widetilde{U}^*,L^*)=F^* I(\widetilde{T}^*\le t,\widetilde{U}^*\le u,L^*\le l)\). Note that the class
is a P-Donsker class. Hence, \(\sqrt{n}(\widehat{N}-N)\rightsquigarrow \mathbb {G}_1\), where \(\mathbb {G}_1\) is a Brownian bridge on \(\ell ^{\infty }([0,\tau ]^3)\) with covariance
Let \(\Lambda _C(s)\) be the cumulative hazard of C, and recall that C is randomly censored by \(T^*+U^*-L^*\). Let \(\pi (s)=P(\widetilde{T}^*+\widetilde{U}^*\ge s)\). Write
By Kosorok (2008, Chap. 4.3), \(\sqrt{n}(\widehat{S}_C-S_C)=\sqrt{n}(\mathbb {P}_n-P)\nu +o_p(1)\). Note that the random process \(\nu \), as a process in \(s\in [0,\tau ]\), is P-Donsker by Corollary 9.32 combined with Lemma 4.1 of Kosorok (2008). Hence, \(\sqrt{n}(\widehat{S}_C-S_C)\rightsquigarrow \mathbb {G}_2\) where \(\mathbb {G}_2\) is a Brownian bridge on \(\ell ^{\infty }([0,\tau ])\) with covariance
For \(\mathbb {G}_3\) and \(\mathbb {G}_4\) we use results from Wang (1991). Let
where \(\beta \equiv P(T+U\ge L)\). By Wang (1991, Sect. 4), \(\sqrt{n}(\widehat{S}_{T+U}-S_{T+U})=\sqrt{n}(\mathbb {P}_n-P)\xi +o_p(1)\) and \(\sqrt{n}(\widehat{F}_L-F_L)=\sqrt{n}(\mathbb {P}_n-P)\vartheta +o_p(1) \). Note that the random processes \(\xi \) and \(\psi \), as processes in \(s\in [0,\tau ]\), are P-Donsker by Lemma 4.1 of Kosorok (2008). The process \(\zeta \) is also P-Donsker by the boundedness of \(S_{T+U}\) on \([0,\tau ]\) and therefore also their sum. Hence, \(\sqrt{n}(\widehat{S}_{T+U}-S_{T+U})\rightsquigarrow \mathbb {G}_3\) where \(\mathbb {G}_3\) is a Brownian bridge on \(\ell ^{\infty }([0,\tau ])\) with covariance given in Wang (1991, Lemma 4.1). Similarly, \(\sqrt{n}(\widehat{F}_L-F_L)\rightsquigarrow \mathbb {G}_4\) where \(\mathbb {G}_4\) is a Brownian bridge on \(\ell ^{\infty }([0,\tau ])\) with covariance given in Wang (1991, Theorem 4.1).
Since all four Gaussian processes are tight, by Lemmas 7.12 and 7.14 of Kosorok (2008), the joint process \((\mathbb {G}_1,\ldots ,\mathbb {G}_4)^T\) is also tight in \(\ell ^{\infty }([0,\tau ]^3)\times (\ell ^{\infty }[0,\tau ])^3\). By the Cramer-Wald device it is also zero-mean Gaussian with covariance
where the expectation is taken with respect to the random variables \(\widetilde{T}^*,\widetilde{U}^*,L^*,F^*\). Since all the functions’ classes are Donsker, by Theorem 2.6 of Kosorok (2008), the bootstrap version in (11) also holds.
1.2 Hadamard differentiability
Definition 2
Let \(\mathbb {D}\) and \(\mathbb {E}\) be normed spaces. Then \(\phi :\mathbb {D}\mapsto \mathbb {E}\) is Hadamard differentiable at \(A\in \mathbb {D}\) if there exists a linear and continuous function \(\phi _A':\mathbb {D}\mapsto \mathbb {E}\) such that
for all converging sequences \(h_n\rightarrow 0\) and \(a_n\rightarrow a\) with \(h_n\in \mathbb {R}\), \(a_n\in \mathbb {D}\) and \(A+h_na_n\in \mathbb {D}\) (Kosorok 2008, Sect. 2.2.4).
Definition 3
The space \(D[0,\tau ]\) is the space of all cadlag functions (right continuous functions with left limits) from \([0,\tau ]\) to \(\mathbb {R}\) equipped with the sup-norm. Denote by \(BV_M[0,\tau ]\) the space of all functions with bounded variation, that is, all the functions \(A\in D[0,\tau ]\) such that \(\int _0^\tau |dA(t)|\equiv |A(0)|+\int _{(0,\tau )}|dA(t)| <M\) (see Kosorok 2008 Sect. 12.2.2). Finally, the space \(D([0,\tau ]^p)\) is the space of all cadlag p-variate functions equipped with the sup-norm (see Neuhaus 1971, for details).
Lemma 2
Let \(D_1\equiv \{f\in D[0,\tau ]\, :\, \inf _t|f(t)|>0\} \), \(D_2\equiv BV_M[0,\tau ]\times BV_M[0,\tau ]\), \(D_3\equiv D[0,\tau ]^2\times D[0,\tau ]^2\). Then
-
(i)
The function
$$\begin{aligned}&H_1: D[0,\tau ]\times D[0,\tau ]\mapsto D([0,\tau ]^3)\quad ;\\&\quad H_1(\, (A,B)\,)(t,u,l)=A(t+u)B(t+u-l) \end{aligned}$$is Hadamard differentiable with derivative
$$\begin{aligned} H_{1,(A,B)}'(a,b)(t,u,l)= & {} A(t+u)b(t+u-l)+a(t+u)B(t+u-l)\\= & {} H_1(A,b)(t,u,l)+H_1(a,B)(t,u,l)\,. \end{aligned}$$ -
(ii)
The function
$$\begin{aligned} H_2: D_1\mapsto D[0,\tau ] \quad ; \quad H_2(A)(t)=\frac{1}{A(t)} \end{aligned}$$is Hadamard differentiable with derivative \(H_{2,(A)}'(a)(t)=-a(t)/A(t)^2\).
-
(iii)
Let \(C(t_0,u_0)=\{(t,u,l)\in [0,\tau ]^3\,:\, l\le t+u\le \tau ,t\le t_0,u\le u_0\}\). The function
$$\begin{aligned} H_3: D_3\mapsto D[0,\tau ]^2\quad ;\quad H_3(\, (A,B)\, )(t_0,u_0)=\int _{C(t_0,u_0)} A(t,u,l)dB(t,u,l) \end{aligned}$$is Hadamard differentiable with derivative
$$\begin{aligned} H_{3,(A,B)}'(a,b)(t_0,u_0)= & {} \int _{C(t_0,u_0)} A(t,u,l)db(t,u,l)\\&+\int _{C(t_0,u_0)} a(t,u,l)dB(t,u,l)\,. \end{aligned}$$ -
(iv)
The function
$$\begin{aligned} H_4: D_2\mapsto D([0,\tau ]^3)\quad ;\quad H_4(\, (A,B)\, )(t,u,l)=\int _{(0,t+u]} A(t+u-s)dB(s) \end{aligned}$$is Hadamard differentiable with derivative
$$\begin{aligned} H_{4,(A,B)}'(a,b)(t,u,l)&=\int _{(0,t+u])} A(t+u-s)db(s)+\int _{(0,t+u]} a(t+u-s)dB(s)\\&= H_4(A,b)(t,u,l)+H_4(a,B)(t,u,l)\,. \end{aligned}$$
Proof
For the proofs of i and iv, let \(h_n\rightarrow 0\) and \((a_n,b_n)\rightarrow (a,b)\) in the appropriate space.
-
(i)
Write
$$\begin{aligned}&\frac{H_1(A+h_na_n,B+h_nb_n)(t,u,l)-H_{1}(t,u,l)}{h_n}-H_1'{(A,B)} (a,b)(t,u,l)\\&\quad =A(t+u)b_n(t+u-l)+a_n(t+u)B(t+u-l)+h_na_n(t+u)b_n(t+u-l)\\&\quad \quad -\left\{ A(t+u)b(t+u-l)+a(t+u)B(t+u-l)\right\} \rightarrow 0\,. \end{aligned}$$ -
(ii)
The proof appears in kosorok (2008, Sect. 2.2.4).
-
(iii)
The proof appears in Gill et al. (1995, Sect. 2, Illustration 1).
-
(iv)
Write
$$\begin{aligned}&\frac{H_4( A+h_na_n,B+h_nb_n)(t,u,l)-H_4(A,B)(t,u,l)}{h_n}-H_{4,(A,B)}'(a,b)(t,u,l)\\&=h_n^{-1}\left( \int _0^{t+u} (A+h_na_n)(t+u-s)d(B+h_nb_n)(s)\right. \\&\quad \quad \left. -\int _0^{t+u} A(t+u-s)dB(s)\right) \\&\quad -\left( \int _0^{t+u} A(t+u-s)db(s)+\int _0^{t+u} a(t+u-s)dB(s)\right) \\&=\int _0^{t+u}A(t+u-s)d(b_n-b)(s)+\int _0^{t+u} (a_n-a)(t+u-s)dB(s)\\&\quad +h_n\int _0^{t+u}a_n(t+u-s)db_n(s)\,. \end{aligned}$$The first term in the above equation goes to zero using the same arguments as in proof 12.3 of Kosorok (2008, p. 242). The second term goes to zero since \(a_n\rightarrow a\) and B is bounded in total variation. The third term goes to zero since \(h_n\rightarrow 0\) and \(a_n\) and \(b_n\) are bounded, which completes the proof.
1.3 Computation of the asymptotic variance
Proof of Theorem 1
Let \(C(t_0,u_0)=\{(t,u,l)\in [0,\tau ]^3\,:\, l\le t+u\le \tau ,t\le t_0, u\le u_0\}\). Note that
Therefore we can write \(G_{T,U}=\beta \phi (N,S_C,F_L)\), where (compare to (3))
The function \(\phi (N,S_C,F_L)\) can be decomposed as a sequences of the following mappings:
where \(H_1\), \(H_2\), and \(H_3\) are defined in Lemma 2. By (3), using the same mapping for \(t_0\le \tau \), \(u_0\le \tau \),
where
The derivative of the map \(\phi \) at \((N,S_C,F_L)\), \(\phi _{(N,S_C,F_L)}'(a,b,c)\) for \((a,b,c)\) in \(D[0,\tau ]^3\times D[0,\tau ]\times D[0,\tau ] \) can be obtained using the chain rule for Hadamard differentiable functions (Kosorok 2008, Lemma 6.19).
By the functional delta method (Kosorok 2008, Theorem 2.8), together with Slutsky’s Theorem (Kosorok 2008, Theorem 7.15) applied to the multiplication by \(\beta _n\rightarrow \beta \),
By the functional delta method for bootstrap processes (Kosorok 2008, Theorem 12.1), we also have
This completes the proof for the first estimator. For the second estimator, note the relation \(F_{L^*}(dl)=F_L(dl)S_{T+U}(l)/\int _0^\infty F_L(dy)S_{T+U}(y)\) that follows from the model \(L^* \sim L\mid L\le T+U\), where L and \(T+U\) are independent. Therefore, the inverse formula is \(F_L(dl)=F_{L^*}(dl)S^{-1}_{T+U}(l)/\int _0^\infty F_{L^*}(dy)S^{-1}_{T+U}(y)\). Let \(\beta =\int _0^\infty F_L(dy)S_{T+U}(y) = P(L\le T+U)\) as before and \(\gamma =\int _0^\infty F_{L^*}(dy)S^{-1}_{T+U}(y)\). Let \(C_1(t_0,u_0)=\{(t,u):t+u\le \tau , t\le t_0,u\le u_0\}\).
Note that
where the first equation follows from the definition of the density \(G_{T,U}(dt,tu)\); the second equation follows by multiplying and dividing by \(\gamma ^{-1} \int _0^{t+u}S_C(t+u-l)F_{L^*}(dl)S_{T+U}^{-1}\); the third equation follows from the inverse formula that relates \(F_{L^*}\) and \(F_L\) above; and the last equation follows from the definition of N in (10).
Write \(G_{T,U}(t_0,u_0)=\beta \psi (N,S_C,F_{L})(t_0,u_0)\), where (compare to (6))
The function \(\psi (N,S_C,F_L)\) can be decomposed as a sequence of the following mappings
where \(H_2\), \(H_3\), and \(H_4\) are defined in Lemma 2. By (6), using the same mappings
where \(\beta _n\) is defined in (12). The derivative of the map \(\psi \) at \((N,S_C,F_{L})\), \(\psi _{(N,S_C,F_{L})}'(a,b,c)\) for \((a,b,c)\) in \(D([0,\tau ]^3)\times (D[0,\tau ])^2 \) can be obtained using the chain rule for Hadamard differentiable functions (Kosorok 2008, Lemma 6.19).
By the functional delta method (Kosorok 2008, Theorem 2.8), together with Slutsky’s Theorem for the convergence of \(\beta _n\) to \(\beta \) (Kosorok 2008, Theorem 7.14)
By the functional delta method for bootstrap processes (Kosorok 2008, Theorem 12.1) we also have
\(\square \)
Rights and permissions
About this article
Cite this article
Vakulenko-Lagun, B., Mandel, M. & Goldberg, Y. Nonparametric estimation in the illness-death model using prevalent data. Lifetime Data Anal 23, 25–56 (2017). https://doi.org/10.1007/s10985-016-9373-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10985-016-9373-0