HHS Public Access

Author manuscript
J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.
Author Manuscript

Published in final edited form as:

J Biopharm Stat. 2018 ; 28(5): 927–938. doi:10.1080/10543406.2017.1405012.

Statistical Methods for Conditional Survival Analysis

Sin-Ho Junga, Ho Yun Leeb, and Shein-Chung Chowa
aDepartment of Biostatistics and Bioinformatics, Duke University, Durham, NC, USA;
bDepartment of Radiology, Samsung Medical Center, Sungkyunkwan University School of
Medicine, Seoul, Korea
Author Manuscript

SUMMARY
We investigate the survival distribution of the patients who have survived over a certain time
period. This is called a conditional survival distribution. In this paper, we show that one-sample
estimation, two-sample comparison and regression analysis of conditional survival distributions
can be conducted using the regular methods for unconditional survival distributions that are
provided by the standard statistical software, such as SAS and SPSS. We conduct extensive
simulations to evaluate the finite sample property of these conditional survival analysis methods.
We illustrate these methods with real clinical data.

Keywords
Delta method; Fieller method; Kaplan-Meier estimator; Log-rank test; Martingale central limit
Author Manuscript

theorem; Proportional hazards model

1 Introduction
Traditionally, survival analysis in clinical researches has been to investigate the distribution
of patients’ survival times measured from the diagnosis of a disease or the start of a
treatment (i.e., baseline). This type of analysis provides the survival probability of patients
expected at the start of treatment that will be useful to predict their prognosis before starting
the treatment. However, the survival probability evolves over time and usually decrease with
increased survivorship, so that both patients and clinicians are interested in the change in
survival probability over the progress of treatment and disease. Residual lifetime of
individuals when they have survived over a relevant landmark of time can serve towards this
Author Manuscript

end. For example, when comparing the efficacy of an intensive treatment with a standard
treatment, patients receiving the prior may have a higher risk due to treatment-related
mortality during treatment period, but may have a much lower risk once they survive over
the treatment period with the disease cured.

While researches on residual lifetime theory have been very active in reliability area (e.g.
Bryson and Siddiqui, 1969; Hollander and Proschan, 1975; Muth, 1977; Ruiz and Navarro,

Correspondence: Sin-Ho Jung, Department of Biostatistics and Bioinformatics, Duke University, 2424 Erwin Road, Hock Plaza, Suite
1102, DUMC Box 2721, Durham, NC 27710, USA, [email protected].
 Jung et al. Page 2

1994), those in biostatistics field have been sparse. Among some biostatistical examples are
Author Manuscript

Jeong and Jung (2008) on two-sample comparison of median residual lifetime, and Jung,
Jeong and Bandos (2009) extending the two-sample problem to regression analysis. More
recently, residual lifetime analysis has been very popularly used in clinical trials to analyze
the change in survival distribution of patients as a progress of disease (e.g. Zamboni et al.
2010; Zabor et al. 2013; Bischof et al. 2015; Mertens et al. 2015) under the name of
conditional survival analysis. Let T be the survival variable of a population with survivor
function S(t) = P (T ≥ t). The t-year conditional survival distribution for patients who have
survived for t0 years, P (T ≥ t + t0|T ≥ t0), is denoted as S(t|t0) = S(t + t0)/S(t0) for t ≥ 0. In
the clinical literatures (e.g. Zabor et al. 2013; Mertens et al. 2015), investigators estimate the
conditional survival distributions by replacing the survivor functions with their Kaplan-
Meier (1958) estimates, but they do not provide a formal statistical testing to compare them
between patient groups.
Author Manuscript

This paper can be regarded as a review paper supporting the analysis methods that are
popularly used in medical field without theoretical justification. In this paper, we present
analysis methods for conditional survival distributions including confidence interval of
conditional survivor function for 1-sample problem, the conditional log-rank test for 2-
sample test, and the conditional Cox (1972) proportional hazards model for regression
analysis. We present simulation results to evaluate the performance of these methods. The
proposed methods are demonstrated with real clinical data.

2 Analysis of Conditional Survival Distributions

2.1 One-Sample Problem
Suppose the lifetimes from n patients T1, …,Tn are independent and identically distributed
Author Manuscript

with survivor function S(t) = P(Ti ≥ t) and cumulative hazard function ⋀(t) = — log S(t).
From patient i, we observe (Xi,δi), where Xi is the minimum of Ti and censoring time Ci,
and δi is an event indicator taking 1 if patient i had an event and 0 otherwise. We assume
that censoring times are independent of the survival times.

For patients who have survived for at least t0(≥ 0) years, the probability that they live
additional t years, S(t|t0) = P(Ti ≥ t + t0|Ti ≥ t0), is given as

S t t0 = S t + t0 /S t0 for t ≥ 0 (1)

by the definition of conditional probabilities. S(t|t0) is called the conditional survivor

Author Manuscript

function for patients who have survived for t0 years.

The conditional cumulative hazard function ⋀(t|t0) = – log S(t|t0) is given as ⋀(t|t0) = ⋀(t +
t0) – ⋀(t0) from (1). Hence, for t ≥ 0, the conditional hazard function λ(t|t0) = ∂⋀(t|t0)/∂t is
identical to the unconditional (or marginal) hazard function λ(t + t0) = ∂⋀(t + t0)/∂t. Jeong,
Jung and Costantino (2008) propose a nonparametric inference method on conditional
median residual lifetime θ that satisfies S(θ|t0) = 1/2.

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 3

Let S t denote the Kaplan-Meier estimator of S(t). Then, S(t|t0) is consistently estimated by
Author Manuscript

S t t0 = S t + t0 /S t0 for t ≥ 0.

From Corollary 3.2.1 of Fleming and Harrington (1991), we have

t + t0

S t + t0 − S t + t0
S t + t0 ∫ 0
Y s −1
dM s
n = − n + op 1 (2)
S t0 − S t0 t0
S t0 ∫
0
Y s −1
dM s

t
∫0
Author Manuscript

where M t = dN s − Y s d Λ s is a 0-mean martingale, and N t = ∑ni = 1 δiI X i ≤ t

and Y t = ∑ni = 1 I X i ≥ t denote the event and the at-risk processes, respectively. Let y(t)
denote the uniform limit of n−1Y (t) for t ≤ τ, where τ denotes the minimum of the upper
limits of the supports of censoring and survival distributions. By the martingale central limit
theorem, n S t + t0 − S t + t0 , S t0 − S t0 converges to N (0, Σ) in distribution, where

t + t0 t0
S t + t0
2
∫0 y s −1d Λ s S t0 S t + t0 ∫0 y s −1d Λ s
Σ = ,
t0 t0
S t0 S t + t0 ∫0 y s −1d Λ s S t0
2
∫0 y s −1d Λ s
Author Manuscript

which can be consistently estimated by replacing y(t), S(t) and d⋀(t) with their consistent
estimators Y(t)/n, S t and Y(t)−1dN(t), respectively. That is, a consistent estimator of Σ is
given as

2
S t + t0 a t + t0 S t0 S t + t0 a t0
Σ = n ,
2
S t0 S t + t0 a t0 S t0 a t0

t
where na t is a consistent estimator of ∫0 y s −1d Λ s that is given as

n
Author Manuscript

t δiI 0 ≤ X i ≤ t
at = ∫0 Y s −1d Λ s = ∑
i = 1 ∑n
2
j = 1 I X j ≥ Xi

t
by using the Nelson-Aalen (Nelson, 1969; Aalen, 1978) estimator Λ s = ∫0 Y s −1dN s
The partial differentiation of θ1/θ2 with θ1 = S(t + t0) and θ2 = S(t0) is given as

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 4

T
S t + t0
Author Manuscript

1
∇ = , − .
St 2
0 St
0

Hence, by the delta-method, n S t t0 − S t t0 converges to N (0,σ2) in distribution, where

t + t0
σ 2 = ∇T Σ ∇ = S t t0
2
t0
∫y s −1dΛ s .

A consistent estimator σ 2 can be estimated by replacing S(t), y(t) and d⋀(t) with their
consistent estimators S t , n−1Y(t) and Y(t)−1dN(t), respectively. That is,
Author Manuscript

n δiI t0 ≤ X i ≤ t + t0
2
σ 2 = nS t t0 ∑ .
2
i = 1 ∑n I X ≥ X
j=1 j i

A 100(1 — α)% confidence interval for the conditional survival probability S(t|t0) can be
calculated using this asymptotic result, i.e.

S t + t0 /S t0 ± z1 − α/2σ,

where z1-α denote the 100(1 — α) percentile of the standard normal distribution.

These inferences are available only when there are patients who are at risk at t0 in the data
Author Manuscript

set. In fact, these inferences can be much simplified. Suppose that P(Ti ≥ t0)P (C ≥ t0) > 0,
i.e. the maximum survival time is longer than t0 and some patients are followed for longer
than t0. Then, for t > 0, we have

P T i ≥ t + t0, X i ≥ t0
P T i ≥ t + t0 X i ≥ t0 = . (3)
P X i ≥ t0

The right hand side of (3) equals

P T i ≥ t + t0, T i ≥ t0, Ci ≥ t0 P T i ≥ t + t0, Ci ≥ t0 P T i ≥ t + t0

= =
P T i ≥ t0, Ci ≥ t0 P T i ≥ t0, Ci ≥ t0 P T i ≥ t0
Author Manuscript

since Xi = min(Ti,Ci), and Ti and Ci are independent. Hence, we have S(t|t0) = P(Ti ≥ t + t0|
Xi ≥ t0), denoting the survival probability at t + t0 for patients who are at risk at t0. This
implies that the conditional survival probability S(t|t0) can be estimated by calculating the
Kaplan-Meier estimator at t + t0 from the patients who are at risk at t0.

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 5

This relationship becomes clearer by the definition of Kaplan-Meier estimator. For

Author Manuscript

simplicity of notation, suppose that there are no ties among Xi, …,Xn. Then, by the
definition of Kaplan-Meier estimator, (2) is expressed as

∏i: X 1 − δi /Y X i
i ≤ t + t0
S t t0 =
∏i: X ≤ t 1 − δi /Y X i
i 0
= ∏ 1 − δi /Y X i ,
i: t0 < X i ≤ t + t0

which is the Kaplan-Meier estimator at t + t0 calculated from the data set consisting of
patients who are at risk at time t0, and its variance is consistently estimated by n−1σ 2. Hence,
one-sample inference of conditional survival distribution using the delta-method will be
Author Manuscript

identical to that based on the standard (or, unconditional) survival distribution using Kaplan-
Meier estimator to the subset of data consisting of the patients who are at risk at t0. These
results hold with tied survival data too.

An alternative confidence interval for S(t|t0) can be obtained using the Fieller’s (1954)
method. Let ρ = S t0 /S t + t0 . By using the asymptotic result for Kaplan-Meier estimator,
n ρS t + t0 − S t0 is asymptotically normal with mean 0, and its variance can be

consistently estimated by n ρ2σ 22 − 2ρσ 12 + σ11 , where σ i j is the (i, j)-component of n−1 Σ.
Hence, we have

ρS t + t0 − S t0
P −z1 − α/2 < < z1 − α/2 = 1 − α . (4)
ρσ 11 − 2ρσ 12 + σ 22
Author Manuscript

We can obtain a 100(1 – α)% confidence interval of S(t|t0) by solving the equation within
the probability of (4) with respect to p, i.e.

f1 ± f 21 − f 0 f 2
,
f2

where f 0 = S t + t0 2 − σ 11z12 − α/2, f 1 = S t + t0 S t0 − σ 12z21 − α/2 and

f 2 = S t0 2 − σ 22z21 − α/2. This formula gives an appropriate confidence interval when

Author Manuscript

f 21 > f 0 f 2

2.2 Two-Sample Log-Rank Test

Suppose that nk patients are randomized to arm k(= 1, 2), and the survival time T k1, …, T kn
k
from the nk patients of arm k are independent and identically distributed with survivor
function Sk(t) = P(Tki ≥ t) and cumulative hazard function Λk(t) = – log Sk(t). From patient

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 6

i(= 1,–, nk) in arm k(= 1, 2), we observe (Xki, δki), where Xki is the minimum of Tki and
censoring time Cki, and δki is an event indicator taking 1 if the patient had an event and 0
Author Manuscript

otherwise. We assume that the censoring times are independent of the survival times within
n n
each arm. Let N k t = ∑i k= 1 δkiI X ki ≤ t and Y k t = ∑i k= 1 I X ki ≥ t denote the event and
the at-risk processes for arm k, respectively. Also, let n = n1 + n2, N(t) = N1(t) + N2(t) and
Y(t) = Y1 (t) + Y2 (t).

For conditional survivor function Sk(t|t0) = Sk (t + t0)/Sk (t0), we want to derive a log-rank
test to test H0 : S1(t|t0) = S2(t|t0) for all t ≥ 0 against H1 : S1(t|t0) ≠ S2(t|t0) for some t ≥ 0.
From the previous section, d⋀(t|t0) = d⋀(t + t0) for t ≥ 0, so that we can consider a log-rank
test for comparing conditional survival distributions for patients who have survived over t0,

∞ ∞
Wt = n ∫0 H t + t0 d Λ1 t t0 − d Λ2 t t0 = n ∫0 H t + t0 d Λ1 t + t0 − d Λ2 t + t0 ,
Author Manuscript

which is identical to

∞
Wt =
0
n ∫t H t d Λ1 t − d Λ2 t
0

where H(t) is a predictable function that is uniformly convergent to h(t) over [t0,τ] and τ is
the minimum of the supports of the censoring and survival distributions. The logrank
statistic (Peto and Peto, 1972) uses H(t) = n−1Y1(t)Y2(t)/Y(t), the Gehan-Wilcoxon test
(Gehan, 1965) uses H(t) = n−2Y1(t)Y2(t), and the Prentice-Wilcoxon test (Prentice, 1978)
uses H t = n−1S − t Y 1 t Y 2 t /Y t , where S − is the left-continuous version of the Kaplan-
Author Manuscript

Meier (1958) estimate from the pooled data.

Note that W t has the same expression as the standard rank tests W0 except that the range of
0
the integration is restricted to [t0, ∞). Using the same arguments as those used for the
standard rank tests (e.g. Gill, 1980; Fleming and Harrington, 1991), we can show that
W t /σ t is asymptotically standard normal with
0 0

∞ H t2
σ2 = n
t0 ∫
t0 Y 1 t Y2 t
dN t
Author Manuscript

under H0. Hence, we reject H0 in favor of H1, if W t /σ t > z1 − α/2 with two-sided type I
0 0
error rate α.

For example, for the conditional log-rank test, we have

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 7

∞ Y1 t Y2 t
1
∫t
Author Manuscript

Wt = d Λ1 t − d Λ2 t
0 n Y t
0
n1 n n2 n
1 ∑i′2= 1 I X 2i′ ≥ X 1i ∑i′1= 1 I X 1i′ ≥
X 2i
=
n
∑ δ1i I X 1i ≥ t0
n
− ∑ δ2iI X 2i ≥ t0
n
i=1 ∑k2 = 1 ∑i′k= 1 I X ki′ ≥ X 1i i=1 ∑2k = 1 ∑i′k= 1 I X ki′ ≥ X 2i

and

∞ Y1 t Y2 t
σ2 =
t0
1
n ∫t Y t2
dN t
0
n n
1
2 nk
∑i′1= 1 I X 1i′ ≥ X ki ∑i′2= 1 I X 2i′ ≥ X ki
=
n ∑ ∑ δkiI X ki ≥ t0
2
.
Author Manuscript

k = 1i = 1 n
∑2k′ = 1 ∑i′k′= 1 I X k′i′ ≥ X ki

From the expression of W t and σ 2t , it is obvious that the conditional log-rank test at t0 can
0 0
be carried out by applying the standard log-rank test to the data set consisting of patients
who are at risk at t0, 𝒟 t0 = X ki, δki : X ki ≥ t0, k = 1, 2, i = , …, nk . The 2-sample
conditional log-rank test can be easily extended to the log-rank test for K-sample cases with
K > 2. These results holds for other types of conditional rank tests.

2.3 Regression Method

From patient i = 1,…, n, we observe covariates zi = (z1i, …, zmi)T together with the
Author Manuscript

minimum of the survival and censoring times Xi and event indicator δi. We assume that,
given zi, the survival and censoring times are independent. Suppose that the conditional
survival distribution for patients who have survived over t0 has a proportional hazards model

λi t t0 = λ0 t t0 exp βT zi (5)

for t ≥ 0, where λi(t|t0) denotes the baseline conditional hazard function. As was shown in
the previous sections, the conditional hazard function λi(t|t0) is identical to the unconditional
hazard function λi(t + t0), so that (5) can be expressed as the regular proportional hazards
model
Author Manuscript

λi t = λ0 t exp βT zi

for t ≥ t0. Hence, if the (unconditional) survival distribution has a proportional hazards
model with constant covariate effect over the whole time span, then the conditional survival
distribution for any t0(> 0) has the same proportional hazards model. However, if the
covariate effect changes over time, then the regression model for conditional survival
distribution changes in t0.

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 8

The partial score and information matrix (Cox 1972) are given as
Author Manuscript

βT Z j
n ∞ ∑nj = 1 Z jY j t e
Ut β = ∑
0 i = 1 t0
∫ Zi −
βT Z
Y i t dN i t

∑nj = 1 Y j t e j

and

T ⊗2
βT Z j
n ⊗ 2Y t e β Z j n
n∞ ∑j = 1 j
Z ∑j = 1 j j
Z Y t e
At β = ∑
0 t
i=1 0
∫ j
β TZ
−
βT Z
Y i t dN i t

∑nj = 1 Y j t e j ∑nj = 1 Y j t e j
Author Manuscript

respectively, where z⊗2 = zzT for a column vector z. Using the same asymptotic theory for
the Cox regression method, we can show that n β − β is approximately normal with mean
0 and its variance-covariance matrix can be consistently approximated by nAt−1 β .
0

For a univariate proportional hazards model with a dichotomous covariate, it is easy to show
that the partial score test under β = 0, U t 0 / At 0 , is identical to the conditional log-rank
0 0
test, W t /σ t , that was discussed in the previous section. From the expression of the partial
0 0
score and information, it is obvious that the conditional Cox regression model (5) can be
fitted by applying the standard Cox regression method to the data set consisting of the
patients who are at risk at t0, i.e. D (t0) = {(Xi,δi,zi) : Xi ≥ t0,i = 1, …,n}. Kurta et al. (2014)
Author Manuscript

proposed this analysis method without any theoretical justification.

3 Numerical Studies
3.1 Simulations
We want to show that the standard inference methods using the subset of data appropriately
reflect the conditional survival distribution concept. At first, we conduct simulations on one-
sample problems using a piecewise exponential distribution with survivor function

exp – λ1 t if 0 ≤ t < 2
St =
exp – λ2 t − 2 λ1 − λ2 if t ≥ 2
Author Manuscript

This distribution has a hazard function of λ1 for 0 ≤ t ≤ 2 and λ2 for t ≥ 2. By choosing λ1 =

0.3466 and λ2 = λ1/2, the median survival for the whole patients is 2 years at the baseline,
while that for those who have survived the first t0 = 2 years is 4 years starting from the 2-
year time point. Survival times are generated from S(t) and censoring times from U(0,a),
where a is chosen for 30% of censoring rate. With a fixed at this value, we consider
censoring distribution U(b, a + b) with b chosen for 15% of censoring. We generate 10,000

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 9

simulation samples of size n = 200. From each sample, we estimate S(t|t0) and its 95%
Author Manuscript

confidence interval by the delta-method and Fieller’s method for t0 = 1, 2, 3, and 4, and t =
1, 2,…, 5 – t0.

Table 1 reports mean bias and the sample standard deviation over the simulation samples
(SSD) of the estimator, S t t0 , and empirical coverage probabilities of the two confidence
interval methods. Table 1 also reports the mean of the standard deviation (MSD) of S t t0
estimated by the delta method over the simulation samples. We observe that the estimated
conditional probabilities have very small bias and the bias tends to increase in censoring
proportion. As expected, SSD and MSD tend to increase in censoring proportion. They are
very close for each simulation setting, but MSD is slightly smaller than SSD. This may
result in slightly anti-conservative empirical coverage probability of the confidence intervals
by the delta method. As t0 and t increase, the number of subjects at risk decreases and the
Author Manuscript

bias tends to be negative. The two confidence interval methods have empirical coverage
probabilities close to the nominal 95% overall, but Fieller’s method always have slightly
larger average length and more accurate coverage probability than the delta method. It is
known that Fieller’s method usually provides better large sample approximation for a ratio
of parameters than the delta method, e.g. Herson (1975). For each method, the average
length increases in censoring proportion.

Now, we investigate the finite sample properties of the two-sample log-rank test on
conditional distributions. Suppose that arm 1 has an exponential distribution with Si (t) =
exp(–λ1t) for t ≥ 0, and arm 2 has a piecewise exponential distribution with survivor
function

exp – λ2 t if 0 ≤ t < 2
Author Manuscript

S2 t =
exp – λ1 t − 2 λ2 − λ1 if t ≥ 2

Note that, if λ1 ≠ λ2, then the two arms have different survival distributions, but with t0 ≥ 2,
their conditional distributions are identical with Sk(t|t0) = exp(–λ1t) and λk(t|t0) = λk(t + t0)
= λ1 for t ≥ 0. So, the log-rank test will have some power for t0 < 2, but not for t0 ≥ 2. We set
λ1 = 0.3466, λ2 = λ1/2, and n1 = n2 = 100, 150 or 200. We consider 15% and 30%
censoring by uniform censoring variables as in the previous simulations. We generate 10,000
samples, apply the 2-sample log-rank test with 2-sided α = 0.05 for t0 = 0,1, 2, 3, 4 to each
sample, and estimate the empirical power as the proportion of samples that the log-rank test
rejects the null hypothesis that two arms have the same conditional survival distributions.
Note that the test with t0 = 0 corresponds to the standard log-rank test to compare two
Author Manuscript

unconditional distributions.

Table 2 reports the empirical power of the log-rank tests. As expected, the empirical power
of the conditional log-rank test is close to the nominal level α = 0.05 with t0 ≥ 2, for which
two conditional distributions are identical. However, with t0 = 0 and 1, it has some power,
and the power becomes higher with a smaller t0(= 0) since the time interval over which the
two conditional distributions are different is wider in this case. The empirical power for t0 =

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 10

0 and 1 also increases in n(= n1 + n2), while that for t0 ≥ 2 is close to the nominal α = 5%
Author Manuscript

regardless of the sample size. We observe that the power with t0 < 2 does not much depend
on the censoring proportion under the simulation setting.

We consider two regression models for simulations on Cox regression analysis of

conditional survival distributions. In Model 1, given covariate value zi, the hazard function is
given as

λ0 exp βzi if 0 ≤ t < 2

λi t =
λ0 if t ≥ 2

Since the hazard function for t ≥ 2 does not depend on zi, the conditional Cox regression
with t0 > 2 will be free of the covariate. The cumulative hazard function is given as
Author Manuscript

λ0 t exp βzi if 0 ≤ t < 2

Λi t =
2λ0 exp βzi + λ0 t − 2 if t ≥ 2

and the survivor function is given as

−Λi t exp − λ0 t exp βzi if 0 ≤ t < 2

Si t = e =
exp − 2λ0 exp βzi − λ0 t − 2 if t ≥ 2

For a U(0,1) random variable Ui, we generate Ti by solving Si(Ti) = Ui. We set β= 0.3 and
λ0 = 0.3 for Model 1.
Author Manuscript

In Model 2, we consider a piecewise exponential distribution with a time-dependent

covariate effect: for patient i with covariate value zi, λi(0) = 0,

λi t = λ0 exp β/ jzi if j − 1 < t ≤ j

for j = 1, 2,.… For this model, the covariate effect β/j decreases in t. The cumulative hazard
function given zi is,

k
Λi t = λ0 ∑ exp β/ jzi − k − t λ0 exp β/kzi if k − 1 < t ≤ k .
j=1
Author Manuscript

Since ⋀i(t) = – log Si(t) and Si(Ti) ~ U(0,1) for Ti with cumulative survivor function Si(t),
we generate Ti by solving equation ⋀i(Ti) = – log Ui for Ui ~ U(0,1), i.e.

λ0 ∑kj = 1 expβ/ jzi + log U i k−1 k

Ti = k − if λ0 ∑ exp β/ jzi < − log U i ≤ λ0 ∑ exp β/ jzi .
λ0 exp β/kz j=1 j=1
i

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 11

We set λ0 = 0.3 and β = 0.4 for Model 2.

Author Manuscript

For each of the survival models, we generate 10,000 simulation samples of size n = 500 and
generate 15% and 30% censoring from uniform distributions U(b, a + b) as in the previous
simulations. For each subject, covariate zi is generated from the standard normal
distribution. From each sample, conditioning on (Ti ≥ t0) with t0 = 0, 1, 2, 3, or 4, we fit a
proportional hazards model with a time-independent regression coefficient, estimate the
regression coefficient, and test on H0 : β= 0 with 2-sided α = 0.05.

Table 3 report the mean regression estimate and empirical power under the two models. For
Model 1, with t0 ≥ 2, the mean regression estimate is close to 0 and the empirical power is
close to the nominal 0.05 level as expected. But, for t0 < 2, the regression estimate is smaller
than β= 0.3 since the covariate effect is diluted over the time interval t ≥ 2 which has no
covariate effect. The regression estimate is smaller with t0 = 1 than with t0 = 0 since the
Author Manuscript

former case has a narrower time interval with non-zero regression coefficient. With 30% of
censoring, the regression estimate is larger since the additional censoring over 15%
censoring occurs over t ≥ 2 (b = 2.2 for Model 1) for which the covariate has no effect. For
Model 2, we observe that the regression estimate decays in t0. And the decaying trend is
more prominent with 30% censoring since the additional censoring occurs after b = 2.1
where the covariate effect is smaller than that over the earlier time interval. The empirical
power quickly decreases in t0 since both the mean covariate effect and the number of
observations used in analysis decrease. However, the decrease of power in censoring
proportion is smaller since the additional censoring occurs over the time interval with
smaller covariate effect.

3.2 Real Data Analysis

Author Manuscript

Kim et al. (2016) report analysis results of a retrospective record study on 723 lung adeno‐
carcinoma patients. All the patients underwent complete resection and mediastinal lymph
node dissection with or without postsurgical adjuvant therapy. From each patient, overall
survival (OS), time to death of any cause from surgery for tumor resection, and progression-
free survival (PFS), time to tumor progression, were observed as outcomes together with risk
factors including ECOG performance score (PS), with or without adjuvant chemotherapy
(adj), tumor-shadow disappearance ratio (TDR) on CT value, and maximum standardized
uptake value (SUV) on 18F-uoro-2-deoxyglucose (FDG)-PET/CT. SUV is log-transformed
to lower the effect of outliers. The objectives of the study is to associate OS and PFS with
the latter four clinical and image predictors using conditional survival analysis. We report
the analysis results on OS to illustrate the conditional survival analysis methods. At first,
patients are partitioned into two groups by PS = 0 and PS ≥ 1. Figure 1 displays the
Author Manuscript

conditional survivor functions of the two PS groups and conditional log-rank p-value for t0
= 0,1,…, 6. We observe that the effect of PS decays as time passes from surgery and
becomes insignificant for patients who have survived for t0 = 4 years or longer. For each t0
value, we regress the conditional survival at t0 = 0, 1, …, 6 on these four covariates using a
multivariate proportional hazards model. Figure 2 displays the regression estimate of each
covariate and its 95% confidence interval against t0. The covariate effect diminishes for
longer survivors except history of adjuvant therapy which has a strong and consistent

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 12

negative effect on OS. A high log-SUV tends to be associated with shorter OS, but it is not
Author Manuscript

so significant for the t0 values considered. Poor PS is significantly associated with poor OS
until t0 = 2 years and its effect becomes weak after t0 = 4 years. High TDR is associated with
longer until about t0 = 4 years, but its effect diminishes among survivors over t0 = 5 years.

4 Conclusions
Conditional survival analysis has been popularly used to investigate the long term effect of
treatment and baseline characteristics on the prognosis in clinical researches. As a reviewer
points out, this analysis provides some new insight on the difference and effect of non-
proportional hazards, early and late treatment and baseline patient characteristics. One
meaningful scenario is that an aggressive surgical treatment may have a high early mortality,
but leads to much higher survival or even cure after the treatment period, while a
chemotherapy does not have a severe treatment-related mortality, but leads to a moderate
Author Manuscript

treatment effect over a long time span. The conditional survival analysis would be
particularly useful to compare this kind of early and late survival benefit between treatments.

Without any theoretical justification, investigators have applied the standard survival
analysis methods, such as log-rank test and Cox regression, to the data removing the patients
whose censoring or event times are shorter than t0 claiming that this results in the survival
distributions of the survivors over t0. This paper is to theoretically justify this claim. We
have reviewed inference methods of one-sample, two-sample and regression analysis for
conditional survival distributions. For a reliable estimation of S(t|t0), we need enough
number of patients who are at risk at t0 and enough number of patients followed for at least t
+ t0 unless all patients have events before this time point. Hence, a conditional survival
analysis will not be available unless the follow-up period is long enough. Since the
Author Manuscript

conditional survival analyses among the patients who survived over t0 are identical to the
standard survival analyses using the data set consisting of patients who are at risk at time t0.
Hence, we can conduct any conditional survival analysis using existing statistical softwares,
such as SAS or SPSS, with the standard survival analysis procedures. These methods are
based on large sample theory. Through simulations, we find that these methods accurately
reflect the change in risk function over time and have good finite sample properties.

If the marginal survival distribution satisfies a proportional hazards model (PHM)

assumption with time-fixed covariate effect, then the regression estimates from conditional
survival analysis will give similar regression estimates for various t0 values. We may be able
to develop a goodness of fit test for PHM assumption of a marginal survival distribution
using this concept. Jung and Wieand (1999) propose a goodness of fit test for PHM using a
similar approach. By plotting the trend of regression estimates of conditional survival
Author Manuscript

analysis over t0, we can also model the time trend of covariates with time-varying effect,
refer to Therneau and Grambsch (2000).

Acknowledgments
Funding

This research was supported by a grant from the National Cancer Institute (CA142538–01).

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 13

REFERENCES
Author Manuscript

Aalen OO (1978). Nonparametric estimation of partial transition probabilities in multiple decrement

models. Annals of Statistics 6:534–545.
Bischof DA, Kim Y, Dodson R, Jimenez MC, Behman R, Cocieru A, Fisher SB,Groeschl RT, Squires
MH, Maithel SK, Blazer DG, Kooby DA, Gamblin TC, Bauer TW, Quereshy FA, Karanicolas PJ,
Law CH, Pawlik TM (2015). Conditional disease-free survival after surgical resection of
gastrointestinal stromal tumors: A multi-institutional analysis of 502 patients. JAMA Surgery
150:299–306. [PubMed: 25671681]
Bryson C, Siddiqui MM (1969). Some criteria for aging. Journal of the American Statistical
Association 64:1472–1483.
Cox DR (1972). Regression models and life tables. Journal of the Royal Statistical Society, Ser. B
34:187–220.
Fieller EC (1954). Some problems in interval estimation. Journal of the Royal Statistical Society, Ser.
B bf 16:175185.
Fleming TR, Harrington DP (1991). Counting Processes and Survival Analysis. New York: Wiley.
Author Manuscript

Gehan EA (1965). A generalized Wilcoxon test for comparing arbitrarily single censored samples.
Biometrika 52:203–223. [PubMed: 14341275]
Gill RD (1980). Censoring and Stochastic Integrals Mathematical Centre Tracts 124, Mathematisch
Centrum, Amsterdam.
Herson J (1975). Fieller’s theorem versus the delta method for significance intervals for ratios. Journal
of Statistical Computing and Simulation 3:265–274.
Hollander E, Proschan F (1975). Tests for mean residual life. Biom etrika 62:585–593.
Jeong JH, Jung SH, Costantino JP (2008). Nonparametric inference on median residual life function.
Biometrics 64:157–163. [PubMed: 17501936]
Jung SH, Jeong JH, Bandos H (2009). Regression on median residual life. Biometrics 65:1203–1212.
[PubMed: 19432781]
Jung SH, Wieand S (1999). Analysis of Goodness-of-Fit for Cox Regression Model. Statistics and
Probability Letters 41:379–82.
Kaplan EL, Meier P (1958). Nonparametric estimation from incomplete observations. Journal of
Author Manuscript

American Statistical Association 53:457481.

Kim W, Lee HY, Jung SH, Kim HK, Choi YS, Kim J, Zo J, Shim YM, Han J, Jeong JY, Choi JY, Lee
KS (2016). Dynamic prognostication using conditional survival analysis for patients with operable
lung adenocarcinoma. Oncotarget DOI: 10.18632/oncotarget.12920.
Kurta ML, Edwards RP, Moysich KB, McDonough K, Bertolet M, Weissfeld JL, Catov JM, Modugno
F, Bunker CH, Ness RB, Diergaarde B (2014). Prognosis and conditional disease-free survival
among patients with ovarian cancer. Journal of Clinical Oncology 32:4102–4112. [PubMed:
25403208]
Mertens AC, Yong J, Dietz AC, Kreiter E, Yasui Y, Bleyer A, Armstrong GT, Robison LL, Wasilewski-
Masker K (2015). Conditional survival in pediatric malignancies: Analysis of data from the
childhood cancer survivor study and the surveillance, epidemiology, and end results program.
Cancer 121:1108–1117. [PubMed: 25557134]
Muth EJ (1977). Reliability models with positive memory derived from the mean residual life function,
In Theory and Applications of Reliability, (Edited by Tsokos CP and Shimi IN), PP 401–434,
Academic Press.
Author Manuscript

Nelson W (1969). Hazard plotting for incomplete failure data. Journal of Quality Technology 1:27–52.
Peto R, Peto J (1972). Asymptotically efficient rank invariant test procedures (with discussion). Journal
of the Royal Statistical Society, Ser. A 135:185–206.
Prentice RL (1978). Linear rank tests with right censored data. Biometrika 65:167–179.
Ruiz JM, Navarro J (1994). Characterization of distributions by relationships between failure rate and
mean residual life. IEEE Transactions on Reliability 43:640–644.
Therneau TM, Grambsch PM (2000). Modeling Survival Data: Extending the Cox Model. Springer,
New York, NY, USA.

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 14

Zabor EC, Gonen M, Chapman PB, Panageas KS (2013). Dynamic prognostication using conditional
survival estimates. Cancer 119:3589–3592. [PubMed: 23913639]
Author Manuscript

Zamboni BA, Yothers G, Choi M, Fuller CD, Dignam JJ, Raich PC, Thomas CR, OConnell MJ,
Wolmark N, Wang SJ (2010). Conditional survival and the choice of conditioning set for patients
with colon cancer: An analysis of NSABP trials C-03 through C-07. Journal of Clinical Oncology
28:2544–2548. [PubMed: 20406942]
Zabor EC, Gonen M, Chapman PB, Panageas KS (2013). Dynamic prognostication using conditional
survival estimates. Cancer 119:3589–3592. [PubMed: 23913639]
Author Manuscript
Author Manuscript
Author Manuscript

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 15
Author Manuscript
Author Manuscript

Figure 1:
Conditional survivor functions for t0 = 0,1, 2, 3, 4, 5. The p-values are from the conditional
log-rank test between PS=0 and PS> 0 groups
Author Manuscript
Author Manuscript

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 16
Author Manuscript
Author Manuscript
Author Manuscript

Figure 2:
Regression estimate of each covariate and its 95% confidence interval against t0
Author Manuscript

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 17

Table 1.

Bias, SSD, and MSD of S t t0 , and average length and empirical coverage probability of 95% confidence
Author Manuscript

intervals by Delta and Fieller methods

Delta Method Fieller Method

t0 t Bias SSD MSD ECP Length ECP Length

Under 15% Censoring

1 1 0.0004 0.0308 0.0299 0.9407 0.1096 0.9411 0.1098
2 0.0005 0.0381 0.0379 0.9471 0.1486 0.9476 0.1489
3 0.0001 0.0396 0.0390 0.9465 0.1528 0.9468 0.1530
4 0.0000 0.0379 0.0371 0.9430 0.1453 0.9434 0.1456
2 1 0.0003 0.0386 0.0384 0.9470 0.1505 0.9477 0.1511
2 −0.0001 0.0440 0.0432 0.9442 0.1694 0.9451 0.1701
3 −0.0001 0.0434 0.0425 0.9433 0.1665 0.9435 0.1672
Author Manuscript

3 1 −0.0004 0.0487 0.0475 0.9387 0.1860 0.9425 0.1879

2 −0.0004 0.0548 0.0536 0.9404 0.2099 0.9433 0.2120
4 1 −0.0002 0.0607 0.0589 0.9350 0.2307 0.9399 0.2352
Under 30% Censoring
1 1 0.0003 0.0308 0.0299 0.9457 0.1170 0.9460 0.1172
2 0.0006 0.0425 0.0416 0.9458 0.1631 0.9462 0.1634
3 0.0000 0.0438 0.0432 0.9427 0.1696 0.9428 0.1699
4 −0.0001 0.0422 0.0415 0.9398 0.1629 0.9402 0.1632
2 1 0.0004 0.0434 0.0425 0.9438 0.1668 0.9451 0.1875
2 −0.0002 0.0488 0.0482 0.9420 0.1891 0.9433 0.1899
3 −0.0002 0.0486 0.0477 0.9398 0.1872 0.9406 0.1881
3 1 −0.0006 0.0542 0.0531 0.9349 0.2085 0.9388 0.2109
Author Manuscript

2 −0.0007 0.0620 0.0604 0.9376 0.2369 0.9418 0.2397

4 1 −0.0003 0.0686 0.0666 0.9320 0.2614 0.9374 0.2677
Author Manuscript

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 18

Table 2.

Empirical power of the conditional log-rank tests between S1(t) and S2(t) for t0 = 0,1, 2, 3, 4
Author Manuscript

n Censoring t0 = 0 1 2 3 4

200 15% 0.5490 0.1971 0.0522 0.0510 0.0558

30% 0.5608 0.1853 0.0476 0.0549 0.0551
300 15% 0.7202 0.2601 0.0508 0.0480 0.0485
30% 0.7324 0.2597 0.0515 0.0509 0.0516
400 15% 0.8451 0.3345 0.0501 0.0535 0.0546
30% 0.8490 0.3206 0.0504 0.0502 0.0505
Author Manuscript
Author Manuscript
Author Manuscript

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.

 Jung et al. Page 19

Table 3.

Mean regression estimate and empirical power of the conditional regression method
Author Manuscript

Model 1 Model 2
Censoring t0 Mean Est Power Mean Est Power

15% 0 0.2022 0.8866 0.2610 0.9817

1 0.1327 0.3706 0.1456 0.3912
2 0.0036 0.0522 0.1076 0.1724
3 0.0035 0.0530 0.0862 0.1043
4 0.0026 0.0565 0.0717 0.0768
30% 0 0.2191 0.8824 0.2787 0.9793
1 0.1470 0.3538 0.1514 0.3409
2 0.0021 0.0545 0.1102 0.1475
3 0.0004 0.0560 0.0869 0.0951
Author Manuscript

4 −0.0033 0.0553 0.0692 0.0741

Author Manuscript
Author Manuscript

J Biopharm Stat. Author manuscript; available in PMC 2019 January 01.