Censoring Issues in Survival Analysis: Kwan-Moon Leung, Robert M. Elashoff, and Abdelmonem A. Afifi

P1: EAK/vks P2: MBL/rsk QC: MBL/vmw T1: MBL
March 19, 1997 10:28 Annual Reviews AR028-04 AR28-04
Annu. Rev. Public Health 1997. 18:83–104

Copyright c 1997 by Annual Reviews Inc. All rights reserved
CENSORING ISSUES IN SURVIVAL

ANALYSIS
Kwan-Moon Leung1,2 , Robert M. Elashoff 1 ,
Annu. Rev. Public. Health. 1997.18:83-104. Downloaded from www.annualreviews.org
and Abdelmonem A. Afifi1

by Universidad Nacional de Colombia on 08/15/13. For personal use only.
1 Department of Biostatistics, School of Public Health, University of California,

Los Angeles, California 90095: e-mail, [email protected]; 2 Quality Initiatives
Division, Health Net, Woodland Hills, California 91367
KEY WORDS: survival analysis, right censoring, interval censoring, informative censoring,
ignorability
ABSTRACT
A key characteristic that distinguishes survival analysis from other areas in statis-
tics is that survival data are usually censored. Censoring occurs when incomplete
information is available about the survival time of some individuals. We define
censoring through some practical examples extracted from the literature in various
fields of public health. With few exceptions, the censoring mechanisms in most
observational studies are unknown and hence it is necessary to make assumptions
about censoring when the common statistical methods are used to analyze cen-
sored data. In addition, we present situations in which censoring mechanisms can
be ignored. The effects of the censoring assumptions are demonstrated through
actual studies.
INTRODUCTION
Survival analysis is used in various fields for analyzing data involving the
duration between two events, or more generally the times of transition among
several states or conditions. It is also known as lifetime data analysis, reliability
analysis, time to event analysis, and event history analysis depending on the
type of application. In this paper, the term survival time is used interchangeably
with the terms risk period, lifetime, failure time, and time to a certain event.
To determine the survival time, we need to define two time points: the time of
origin, i.e. the time at which an original event, such as birth, occurs and the
time of failure, i.e. the time at which the final event, such as death, occurs. A
83
0163-7525/97/0510-0083$08.00
84 LEUNG, ELASHOFF & AFIFI
subject is said to be at risk if the original event has occurred, but the final event
has not.
A key characteristic that distinguishes survival analysis from other areas
in statistics is that survival data are usually censored or incomplete in some
way. We define censoring through some practical examples, then describe
the common statistical methods used to analyze censored data and discuss the
necessity of making assumptions about censoring when those methods are used.
We also discuss situations in which the censoring mechanism can be ignored,
and investigate the effects of different censoring assumptions in actual studies.
Finally, we indicate some trends in the future of research on censoring.
ORIGINS OF CENSORING
Censoring occurs when incomplete information is available about the survival
time of some individuals. In this section, we present a number of concrete
examples extracted from the literature in various fields of public health. The
purpose is to use real-life situations to illustrate types of censoring and to
motivate the discussion presented in the later sections.
Examples
Example 1. Health insurance and mortality To examine the relationship be-
tween the status of insurance and the risk of subsequent mortality, adults older
than 25 years who reported that they were uninsured or privately insured in the
first National Health and Nutrition Examination Survey (NHANES I) (9) were
followed prospectively from initial interviews between 1971 and 1975 until
1987 (end of the Epidemiologic Follow-up Study, NHEFS). There were a total
of 4882 eligible subjects, of whom 669 subjects were uninsured.
The time period of interest was the time from the start of follow-up to death,
and the research question is whether this variable is affected by whether or not
the individual is insured. In general, an observation is said to be right censored
if the person was alive at study termination or was lost to follow-up at any
time during the study. By right censoring, it is meant that the survival time is
only known to exceed a certain value. In this study, the analysis was adjusted
for other factors such as baseline age, gender, race, smoking status, alcohol
consumption, obesity, self-rated health, employment status, and so forth.
Example 2. Do men and women relapse into alcoholism for different reasons?
A sample of 44 women and 50 men attending an alcohol treatment facility oper-
ated by the Western Australian Alcohol and Drug Authority were studied (29).
A range of demographic, social, and psychological measures were observed to
determine whether women and men relapse for different reasons. The length
of follow-up was three months and the variable of interest was the time from
CENSORING ISSUES 85
start of follow-up to relapse. After 3 months, 25 subjects were lost to follow-up

(6 women and 19 men), and 42 subjects (22 women and 20 men) had relapsed.
Here again, right-censoring occurred.
Note that the numbers of loss to follow-up are quite different between women
and men. These unequal censoring rates can cause the analysis to lose power
when assessing gender effect (see, for example, Reference 15).
Example 3. Survival with inoperable lung cancer A sample of 61 patients with

inoperable lung cancer who were treated with the drug cyclophosphamide at the
Eastern Cooperative Oncology Group were studied (22). The primary outcome
variable was the time from diagnosis to death. In this sample, 28 right-censored
observations represented patients whose treatment was terminated because of
the appearance of metastatic disease or because of a significant increase in the

size of their primary lesion. Since such disease progressions are usually asso-
ciated with shortened residual survival time, the censoring may be informative
(that is, censoring provides more information than the fact that survival time
exceeded a certain time) and may cause serious problems if analysis of such
data does not take this information into account.
Example 4. Long-term life-style intervention with high-normal blood pressure
levels The study sample comprised 2382 participants (1566 men and 816
women), who were 110% to 165% of desirable body weight (11). Individuals
were allocated at random to four treatment arms using a 2 × 2 factorial design
(weight loss × Sodium reduction). This was a phase II study of the Trials of
Hypertension Prevention (TOHP).
The primary outcome variable was the time from randomization to either
diastolic blood pressure (DBP) becoming ≥90 mm Hg, systolic blood pressure
(SBP) becoming >140 mm Hg, or taking antihypertensive medication. An
observation was considered to be right censored if the subject’s blood pressure
was within normal limits, that is, had DBP ≤90 mm Hg and SBP ≤140 mm
Hg, and did not take any antihypertensive medication by the end of the study,
or if the subject was lost to follow-up. Note that had we not considered taking
antihypertensive medication as an endpoint, then the observation would have
been right censored at the time of taking antihypertensive medication. However,
this right-censored mechanism would be informative since subjects are likely to
start taking antihypertensive medication because their blood pressure exceeded
the normal limits.
Example 5. Longitudinal study of respiratory symptoms in aluminum pot-
room workers A prospective study of respiratory health in aluminum pot-
room workers was initiated in the Nordic countries on January 1, 1986 (20, 28).
The workers were followed until December 31, 1989, or until leaving potroom
work, whichever came first. The workers were supposed to have health exami-
nations and to fill out questionnaires regarding respiratory symptoms at the start
of employment, then yearly at a routine examination, if attending the plant’s
health clinic because of respiratory symptoms, or when leaving employment.
The study sample consisted of 1301 subjects who were employed during
the study period and had at least two examinations. If the subjects reported
wheezing and dyspnea, they were considered symptomatic. The time variable
of interest was the time from employment to development of symptoms. Besides
fluoride exposure, other potential covariates included age, smoking habits, and
previous exposure to dust/gases.
In addition to right censoring, that is, leaving the potroom or ending the
survey without respiratory symptoms, some observations were singly interval
censored because for them the study endpoint was established only by periodical
examinations. By singly interval censoring it is meant that the outcome variable
is not known exactly, rather it is known only up to a time interval.
Example 6. Multicenter AIDS Cohort Study (MACS) From April 1984 to
September 1993, there were 4954 men between the ages of 18 and 70 who were
recruited for the Multicenter AIDS Cohort Study (MACS) (5). The MACS is
a longitudinal study of the natural history of human immunodeficiency virus
type 1 (HIV-1) among homosexual and bisexual men. Subjects were recruited
at four centers: Los Angeles, Chicago, Pittsburgh, and Baltimore.
An important time variable was the incubation period of AIDS (time from
HIV-1 seroconversion to an AIDS-defining illness). An observation was right
censored if the subject was AIDS free on September 1, 1993, or was lost to
follow-up. The censoring issue becomes more complicated when we realize
that both the time of HIV seroconversion and the time of AIDS onset are
known only up to a time interval since those times are determined by periodical
examinations. Such observations are called doubly interval censored, i.e. the
survival time (incubation period of AIDS) is subject to interval censoring on
the left and on the right.
Example 7. Survival with malignant melanoma In the period between 1971
and 1993, approximately 6000 patients with malignant melanoma were treated
by the staffs of the John Wayne Cancer Institute (JWCI) (24). The primary
objective of the study reported here was to examine the efficacy of a new
polyvalent melanoma cell vaccine (MCV) in treating patients with metastatic
disease. Such treated patients represent a subset of the JWCI patients. To provide
appropriate treatments, patients were followed periodically after admission to
detect any change in disease stages. Excluding patients with stage III disease
when first seen at the JWCI, we had 1548 patients in the data set. By the time of
analysis, 890 had advanced to stage III, 788 of whom died. Beside the indicator
CENSORING ISSUES 87
of treatment (MCV or no MCV), other covariates include gender, distant site,

Breslow’s depth of patient’s primary tumor, and the time interval between the
first diagnosed stage II disease and disease metastasis.
Here we note that, by the time a stage III was diagnosed (regional bone
recurrence or metastatic disease), metastasis had already occurred. Thus, the
date of the first metastasis is only known to lie between the dates of the last
diagnosed stage II disease (regional recurrence without involving bone) and
the first diagnosed stage III disease. Such observations were thus left interval
censored.
Types of Censoring Mechanisms

POINT CENSORING We now summarize the different types of censoring mech-
anisms illustrated in the above examples. We begin with observations that

may be point-censored as depicted in Figure 1, where the subjects are under
observation from time T0 to time T1 , and the survival time or censoring time
are known exactly. Such situations arise when subjects are continuously mon-
itored or when the time of event is well documented, for example, when the
subjects are hospitalized or when the event is death. We call censoring arising
from these situations point censoring to distinguish it from interval censoring,
to be discussed shortly, in which the survival time is known only up to a time
interval. The solid line represents the risk period for each subject. The line
ending with an asterisk (∗) indicates an occurrence of the event of interest, and
the line ending with an open point (o) indicates an occurrence of an event other
than the event of interest, e.g. loss to follow-up or death due to causes other
than the one under study.
In Figure 1, the entire risk period for Subject A falls within the observation
period and the time of occurrence of the event is known; hence there is no
censoring for this observation. For subject B, the risk period starts during the
observation period and the event occurs after follow-up is terminated at T1 .
Figure 1 Types of point-censored observations.

The observation of Subject B is therefore right censored at T1 . This type of

censoring occurred in all the above examples. For Subject C, the observation
is also right censored but it is so because an event other than the one of interest
occurs during the observation period and takes the subject out of the risk set
(the set of subjects who are at risk). To distinguish between these two types
of right-censored data, we call censoring due to study termination (Subject B)
end-of-study censoring and censoring due to other reasons (Subject C) loss-
to-follow-up censoring. Under the assumption that the time of entry to the
study is independent of the risk period, it can be easily shown that end-of-study
censoring is independent of survival time, and hence it poses no problem to the
analysis. However, it is not always reasonable to make such an independence

assumption about loss-to-follow-up censoring, and hence we need to make
further assumptions about the censoring mechanism when analyzing the data.
Statistical methods for handling censoring mechanisms are discussed in detail
in the next section.
While cases B and C in Figure 1 represent right censoring, Subject D repre-
sents a case with left truncation. Such a problem could happen, for example,
when a subject in the AIDS study was already HIV-1 seropositive prior to enroll-
ment and the time variable of interest is the incubation period of AIDS. There
are other situations, such as for Subject E, in which the observation is both left
and right censored; we call such observations doubly censored. An example of
this sort exists in the AIDS study example where a subject was already HIV-1
seropositive when enrolled but was still AIDS-free at the end of the study. A
crucial question here is whether the time (DL ) from the beginning of the risk
period to the beginning of the observation period is known. If DL is observed,
then one could apply the methodologies developed for no censoring (or right
censoring) for analysis with proper adjustment of the risk set. However, if DL
is not observed, then we cannot specify the origin of the survival time. In this
case external information such as the distribution, over chronological time, of
the original event is needed (14, 17, 32).
Finally, in most applications there are cases where the origin and the event
both occur prior to the start of follow-up or after follow-up ends. Such cases
generally do not affect the analysis but they can affect the generalizablility
of the findings. Subjects F and G in Figure 1 represent such cases known as
completely right censored and completely left censored, respectively.
INTERVAL CENSORING In many applications, for instance, in Examples 5–7,

the time of the event may be known only up to a time interval, especially when
the time is established by periodical examinations. Figure 2 defines the relevant
time points and illustrates how interval-censored data arise.
CENSORING ISSUES 89
Figure 2 An example of singly-interval censored observations.
For subject A0 in Figure 2 (Subject A in Figure 1 but the time of event is

interval censored), instead of knowing that the event of interest occurred at

time s, we only know that the event occurred in the time interval (sL , sR ], i.e.
after SL but up to and possibly at SR . An example of this kind exists in the

AIDS study where subjects were periodically examined to determine the HIV-1
status. In Figure 2, “−” means an HIV-1 negative result was reported and “+”
means an HIV-1 positive result was reported. In fact, observations from most
studies with a nonlethal outcome are interval censored since we usually cannot
monitor subjects continuously. The issue is whether we should analyze the
data as interval censored or point censored. For instance, if the median survival
time is 5 years and the intervals are about 3 to 6 months wide, then we have
no reason to complicate the analysis by considering interval censoring. On the
other hand, if the intervals are about 1 year or longer, then we should account
for such uncertainty in the analysis. It is also interesting to note that interval
censoring theoretically includes both left and right censoring.
Figure 3 extends Figure 2 to represent more complicated but practical situa-
tions in which we continue to periodically monitor Subject A0 after an immediate
event (denoted by “∗” in Figure 3), for example disease metastasis or HIV-1
seroconversion up to a final event (denoted by “#” in Figure 3), for example
death or AIDS onset. The time between the event “∗” and the event “#” is
known up to an interval on both ends.
We call such an observation doubly interval censored to distinguish it from
singly interval censored observations. To analyze doubly interval censored data,
Figure 3 An example of doubly interval censored observations.

it is tempting to transform the observations to the singly interval censoring form,

that is, for Subject A00 we create the interval (tL − sR , tR − sL ], and then apply
the methods developed for singly interval censored data. However, De Gruttola
& Lagakos (4) pointed out that this approach is not valid.
COMMON STATISTICAL METHODS

FOR CENSORED DATA
In this section we discuss the common statistical methods used to analyze
censored data. As for other types of incomplete data, several approaches have
been proposed (see 25 for discussion of statistical methods for incomplete

data). As described below, there are four basic approaches for the analysis of
censored data: complete data analysis, the imputation approach, analysis with
dichotomized data, and the likelihood-based approach.
Complete-data analysis As many researchers and statistical packages do when
faced with incomplete data, one can simply ignore the censored observations
and analyze only the uncensored complete observations. The main advantage
of this approach is simplicity. However, there are some disadvantages. (a) Loss
of efficiency: The loss in sample size can be considerable since it is not unusual,
especially in medical or epidemiological studies, that 50% or more observations
are censored. (b) Estimation bias: Inferences based on analyzing the uncen-
sored observations only may be biased. It is a common misconception that
one need not make any assumptions about the censoring mechanism when per-
forming a complete-data analysis. In reality, such an analysis requires a strong
assumption regarding the censoring mechanism: As in the incomplete data situ-
ations, complete-data analysis produces unbiased estimates only if the missing
(censored observations) are missing (censored) completely at random (25).
Imputation approach Although imputation is one of the popular approaches
for handling incomplete data, it may not be appropriate for censored data. In the
context of right censoring, there are two extreme ways to impute the missing
survival times: (a) assuming all censored cases fail right after the time of
censoring, that is, left-point imputation or (b) assuming all censored cases never
fail, that is, right-point imputation. It is clear that neither of these approaches
is appropriate since the survival probabilities would be underestimated and
overestimated, respectively. Another approach is to assume that the failure
time after censoring follows a specific model and estimate the model parameters
in order to impute the residual survival time (time from censoring to failure).
However, this approach depends on the model assumptions, which are very
difficult to check without information on survival after censoring (the missing
information).
CENSORING ISSUES 91
In the context of interval censoring, the inappropriateness of imputation is

less clear. Many researchers use imputation techniques, especially right-point
or mid-point imputation, when the observations are interval censored. This may
be due to lack of statistical software packages for analysis of interval-censored
data. However, many authors (23, 24) have pointed out that both right-point
and mid-point imputations may generate seriously biased results, as will be
seen in later sections.
Analysis based on dichotomized data The problem of right censoring and
interval censoring may be avoided if one analyzes the incidence of occurrence
versus nonoccurrence of the event within a fixed period of time and disregards
the survival times. In this case, the dichotomized data can be easily analyzed
by the standard techniques for binary outcomes, such as contingency tables and
logistic regression. However, there are some disadvantages of this approach:
1. It cannot distinguish between loss-to-follow-up and end-of-study censoring.
2. Variability in the timing of the event among those who had the event within
the observation period cannot be modeled. Let us consider an extreme
example: Suppose we are studying the effect of a new drug on patients who
underwent surgery for a particular disease. Eighty percent of the subjects
who have the placebo have recurrence shortly after the surgery, while 80% of
the subjects who took the new drug remained disease free for at least 5 years
but had recurrence within 10 years. All other patients did not have recurrence
after 10 years. If we analyze these data with dichotomized outcomes, we
may find no difference between treatment the groups when the observation
period is 10 years. However, there would likely be a significant difference
between the treatment groups when the observation period is 5 years.
3. No time-dependent covariates (such as age, smoking status or alcohol con-

sumption status) can be used in modeling.
However, the approach of analyzing dichotomized data may be acceptable

when the risk of failure is low, the risk periods (survival times) are long, and the
covariates are associated with preventing the event rather than with prolonging
the survival time. Such situations are common in many epidemiologic studies.
Likelihood-based approach Perhaps the most effective approach to censoring
problems is to use methods of estimation that adjust for whether or not an in-
dividual observation is censored. Many of these approaches can be viewed as
maximizing the likelihood under certain model assumptions, including assump-
tions about the censoring mechanism. Likelihood-based approaches include,
for example, the Kaplan-Meier estimator of the survival function in a one-

sample problem, the log-rank test for testing equality of two survival functions
in a two-sample problem, and the Cox-regression and accelerated-failure-time
models for analysis of time to event data with covariates. The main advantage
of the likelihood-based approach is that it utilizes all the information available.
However, as in the other approaches, assumptions about the censoring mecha-
nism are still required. In the following sections, we present these assumptions,
examine how important they are, and discuss when they can be ignored.
NECESSITY OF MAKING ASSUMPTIONS

ABOUT CENSORING
With some exceptions, the censoring mechanisms in most observational studies

are unknown. Two such exceptions are the so-called type I and type II censoring
designs. Type II designs (often used in engineering) are studies in which a total
of n subjects are under observation but, instead of continuing until all subjects
fail, the study is terminated when the r th subject fails. In principle, the analysis
of type II censoring data poses no problem to the investigator. However, type
II censoring designs are rarely used by biomedical or public health researchers
and, therefore, we do not discuss them further in this paper.
By a type I censoring design we mean a study in which every subject is
under observation for a specified period C0 or until failure. A slightly more
complicated type I censoring design is one in which each subject has his/her
own fixed censoring time Ci , instead of a common censoring time C0 . In this
study design, the likelihood function for each subject can be represented by one
of the following two probabilities: the probability that the event occurred in
a small interval including time ti [denoted by f i (ti )] or the probability that the
subject did not have the event at Ci [denoted by Si (Ci )]. Mathematically, the
likelihood function of the ith subject can be written as
L i = f i (ti )δi Si (Ci )1−δi , 1.
where δi is an indicator of whether the actual survival time of the ith subject is
observed or not, that is, δi = 1 if observed and = 0 if not. Having δi = 0 means
that the subject survived the observation period Ci without the event occurring,
and hence the observation was censored. Note also that Si is the survival
function of interest and f i is the corresponding density function. The likelihood
function of the whole sample is the product of the individual likelihood function
of all the subjects. Many statistical procedures dealing with right censoring are
based on the likelihood function (Eq. 1).
Unfortunately, in most applications the censoring mechanisms are more
complicated than in the type I design, especially when both end-of-study and
CENSORING ISSUES 93
loss-to-follow-up censoring are involved. Thus, we usually cannot apply the

likelihood function (Eq. 1) directly without making further assumptions about
the censoring mechanism. In medial and epidemiological studies, censoring
times Ci are often random rather than fixed. For instance, in example 3, pa-
tients may enter the study in more or less a random fashion according to their
times of diagnosis. If the study is terminated at a preassigned date, then the
end-of-study censoring times (times from subjects’ entry to study termination)
are effectively random. Similarly, the loss-to-follow-up censoring times are
also effectively random since the times at which loss to follow-up occurs are
not known in advance.
To analyze data of this kind, we may proceed by considering the joint dis-
tribution of T (survival time) and C (censoring time), that is, the likelihood
function
Yn
L= f i (ti , ci ), 2.
i=1
where f i (ti , ci ) is the density function of survival time T = ti and censoring time
C = ci . However, we are often not interested in modeling both the survival
time and censoring time. Instead, we are only interested in the distribution
of the survival time or the effects of certain covariates on the survival time.
Furthermore, for the ith subject, we only observe yi = min{ti , ci } and the
censoring indicator δi instead of (ti , ci ). Under these conditions, the distribution
function of T is described by statisticians as being nonidentifiable (33) unless
we make further assumptions. Here nonidentifiability means that there exist
more than one distribution function of T that are compatible with the data.
A simple but commonly used assumption to resolve this problem is indepen-
dent censoring, that is, we assume that the survival time T and the censoring time
C are independent. Note that some authors use the term random censoring when
they actually mean independent censoring. Under the independent censoring
assumption, analyses can be simply based on the likelihood function (Eq. 1).
For left truncation, data can be handled in a similar manner to right censoring
(for example, 14, 32). For interval censoring, the situation is slightly different
in that it requires the knowledge of the examination scheme (prospective study)
or sampling plan (retrospective study) as explained below.
Suppose the disease process is denoted by X (t) indicating the disease state at
time t. We assume that a particular subject is observed at times t0 < t1 < · · · <
tm to be in states s0 , s1 , . . . , sm , respectively. Here s j , j = 1, . . . , m could
represent the disease stages. For these observations, the likelihood is given by
L = Pr{X (t0 ) = s0 , . . . , X (tm ) = sm ; T0 = t0 , . . . , Tm = tm ; M = m} 3.
Here, not only are the examination times T0 , T1 , . . . , Tm assumed random, the
number of examinations M is also assumed to be a random variable. However,

we generally are not interested in modeling the examination times or their num-
ber. Rather, we are usually interested only in the disease states and therefore,
for interval-censored data, inference is usually based on
L = Pr{X (t0 ) = s0 , . . . , X (tm ) = sm }. 4.
Authors often assume that the examination scheme (or the sampling plan) has
been prespecified or has at least been chosen completely independently of the
disease process. In this case, analysis can be based on the likelihood function
(Eq. 4).
IGNORABILITY OF THE CENSORING MECHANISM

RIGHT CENSORING In the statistical inference procedures included in most
statistical packages, the assumptions used in the likelihood function (∗) or its
equivalent are often either not stated explicitly or are stated in vague terms.
As a result, it is likely that many users of survival analysis are either unaware
that they are making certain assumptions about the censoring mechanism or are
unclear about precisely what these assumptions are. We think it is important
that such censoring assumptions be clearly understood since only then can we
identify those situations in which these assumptions can be ignored.
As Kaplan & Meier (17) noted: “In practice this assumption (independent
censoring) deserves special scrutiny.” However, the Kaplan-Meier estimator
may overestimate the survival function of T if the survival time and the cen-
soring time are positively correlated, and underestimate the survival function if
the times are negatively correlated.
Independent censoring may not hold for all situations but under some depen-
dence conditions we can still use the likelihood function (Eq. 1). Lagakos (21)
presented two such situations:
(a) Nonprognostic censoring A censored observation at time Ci indicates
that the survival time exceeds Ci and carries no prognostic information about
subsequent survival times for either the same individual or other individuals.
(b) Noninformative censoring (also known as constant sum condition) The
instantaneous probability of failure in a small interval about y = min{t, c} given
survival to y is unchanged by the additional information that the subject was
uncensored up to time y (see also 16).
Lagakos (21) proved that nonprognostic censoring models and independent
censoring models are special cases of the noninformative censoring model.
In many situations, censoring can be recognized to be noninformative (for
example, end-of-study censoring), and hence standard procedures assuming the
CENSORING ISSUES 95
likelihood function (Eq. 1) can be used. However, in other situations, it is not

clear whether censoring is noninformative; in fact, it is sometimes clear that
censoring is related to the survival times as, for instance, in examples 3 and 4.
In spite of its crucial importance, the noninformative censoring assumption is
not possible to test without making additional restrictions, for example, restric-
tions on the joint distribution of T and C. As mentioned above, the problem
is that given only the sample data (yi , δi ), the survival function of T is not
identifiable.
Several authors have proposed models for informative censoring and tests for
noninformative censoring under various conditions. For instance, Moeschberger
(26) suggested modeling the joint distribution of (T , C) as in (Eq. 2). Lagakos

& Williams (22) introduced a semiparametric model, called the cone model,
which includes an exponential survival function of T (with parameter λ), an un-

specified function c(y) that measures the relative odds of observing a failure at
y = min(t, c) and a scalar parameter θ ∈ [0, 1]. Here the parameter θ reflects
the degree to which censoring affects survival with θ = 1 indicating nonin-
formative. Moeschberger (26) and Lagakos & Williams (22) presented the
maximum likelihood estimates and large-sample test for noninformative cen-
soring. In addition, several nonparametric estimates of the survival function
were presented by various authors. Peterson (27) suggested an estimate of the
bounds for the survival function without assuming any censoring mechanism
or parametric distribution of T . Fisher & Kanarek (8) proposed a nonparamet-
ric estimate of the survival function under a “stretch/contract” model in which
censoring coincides with an event that alters subsequent survival times by a
known scale parameter α, that is, they assume Pr{T > t | C = c < t) =
Pr{T > c + α(t − c) | C > c + α(t − c)}. Here α < 1 means censoring is
favorable for survival whereas α > 1 means censoring is unfavorable for sur-
vival. In particular, α = 1 corresponds to noninformative censoring. Slud &
Rubinstein (31) and Klein & Moeschberger (19) generalized the Kaplan-Meier
estimator with informative censoring situations where certain measures of the
dependence of survival and censoring times are known. In (31) they assume
knowledge of ρ(t) = h(t | C < t)/ h(t | C ≥ t) which measures the relative
risk of failure at time T among previously censored subjects as compared with
subjects not yet censored, where h(t | ·) is the conditional hazard function of
T . In (19) they assume knowledge of the parameter θ that relates to Kendall’s
coefficient of concordance τ of (T , C) by τ = (θ − 1)/(θ + 1).
Interval Censoring
Compared with right censoring situations, relatively few articles were devoted
to the problem of ignorability of interval censoring. Similar to the defini-
tion of noninformative censoring for right-censored data, Gruger et al (10)
defined noninformative examination schemes (or sampling plans) for interval

censoring problems. Basically, an examination scheme (including the num-
ber of examinations and their times) is called noninformative if the likeli-
hood function given the examination scheme is proportional to the likelihood
function obtained when the examination scheme is fixed in advance. They
considered the following examination schemes that are often seen in prac-
tice and concluded that they satisfy the noninformative examination scheme
assumption.
1. Examination at regular intervals Under this scheme, all subjects are ex-
amined or observed at preassigned intervals. This situation occurs frequently

in medical studies since examination times are often preassigned.
2. Random sampling Under this scheme, all subjects are examined or ob-
served in a more or less random fashion and the examination times are inde-
pendent of the subjects’ disease history.
3. Doctor’s care Under this scheme, the next examination time is chosen on
the basis of the subject’s current observed status. For example, if a patient is
in a critical stage, then the time of the next examination will be chosen to be in
the very near future.
Under the above examination schemes, Gruger et al (10) showed that the
likelihood function (Eq. 4) can be used to obtain an estimate of the survival
function (4, 23) or estimates of the regression coefficients of survival times on
the covariates (7, 18, 24). Gruger et al (10) considered another situation called
patient self-selection. In this scheme a patient’s examination is initialized when
the patient feels unwell and/or when symptoms suggest that the disease is
advancing. Alternatively, a patient who feels unwell might refuse to appear
for examination because of loss of confidence in the efficacy of the treatment.
Under this scheme, the examination times are no longer noninformative and
must be taken into account when analyzing the data. That task, however,
requires modeling the joint distribution of the disease state and the examination
times [that is, the likelihood function (Eq. 3)]. Gruger et al (10) suggested that
investigators should plan in advance in order to avoid this difficulty.
Unlike the case of right censoring, there are only a very few articles on testing
the assumption of noninformative examination schemes for interval-censored
data. Heitjan (12) and Heitjan & Rubin (13) introduced the concept of “coarse”
data that have right censoring and interval censoring as special cases. They
provided the conditions under which it is appropriate to ignore the stochastic
nature of the coarsening (censoring), and called such conditions coarsening at
random. However, parametric assumptions are generally required in order to
test the coarsening at random conditions.
CENSORING ISSUES 97
EFFECTS OF THE CENSORING ASSUMPTIONS

In this section, we analyze the data from Examples 3, 5, and 7 (discussed in
the section on Origins of Censoring) to illustrate the importance of censoring
assumptions for right censoring, singly interval censoring, and doubly interval
censoring, respectively.
Right censoring: Survival with lung cancer (Example 3 continued) The prob-
lem and the data were described in the first section. Recall that the observa-
tions were censored because the patients experienced metastasic disease or a
significant increase in the size of their primary lesion. Such disease progres-
sion usually indicates a shortened residual survival time. Thus, we cannot

apply the standard procedure (assuming noninformative censoring) to analyze

the data. Table 1 lists the survival times and the censoring times of those
61 patients (33 uncensored observations and 28 censored observations). For
these 28 patients with originally censored observations, the ultimate survival
times were obtained later. Notice that those ultimate survival times, in gen-
eral, cannot be treated as the actual survival times since those patients were
removed from the study at the time of censoring. However, for the purpose
of illustration we will treat those times as if they were the actual survival
times.
Table 1 Survival times in weeks of 61 patients with inoperable

adenocarcinoma of the lung
33 Uncensored observations 28 Censored observations
0.43 15.71 0.14 (3.00)∗ 8.71 (20.43)

2.86 18.43 0.14 (12.43) 10.57 (25.00)
3.14 18.57 0.29 (1.14) 11.86 (17.29)
3.14 20.71 0.43 (17.14) 15.57 (21.57)
3.43 29.14 0.57 (4.43) 16.57 (45.00)
3.43 40.57 0.57 (5.43) 17.29 (24.14)
3.71 48.57 1.86 (12.14) 18.71 (29.43)
3.86 49.43 3.00 (7.86) 21.29 (26.71)
6.14 53.86 3.00 (13.86) 23.86 (29.00)
6.86 61.86 3.29 (10.57) 26.00 (53.86)
9.00 66.57 3.29 (34.43) 27.57 (49.71)
9.43 68.71 6.00 (7.86) 32.14 (63.86)
10.71 68.96 6.14 (9.29) 33.14 (99.00)
10.86 72.86 47.29 (48.71)
11.14 72.86
13.00
14.43
∗
In parentheses are the eventual failure times of the 28 censored patients.
Lagakos & Williams (22) obtained the maximum likelihood estimates (stan-
dard errors) of the cone model discussed in the last section as λ̂ = 0.0409
(0.0123) and θ̂ = 0.25 (0.36). Notice that θ̂ is significantly different from one
(recall that θ = 1 means censoring is noninformative) based on a large-sample
test of H0 : θ = 1, and hence the noninformative censoring assumption is not
satisfied. We also obtained the estimates of the upper and the lower bounds
of the survival function based on Peterson’s procedure (27), and the estimated
survival function based on the procedures proposed by Fisher & Kanarek (8),
Slud & Rubinstein (31), and Klein & Moeschberger (19) with various model
parameters. Figure 4 displays the estimates of the survival function together
with the empirical distribution function derived from all 61 complete obser-
vations. Each of the three models that account for nonignorable censoring
involves some parameters whose values must be guessed by the investigator.

In our analysis, we attempted a number of such values, but, in order to save
space, we only present the estimates of the survival function that are closest
to the empirical distribution function. In practice, however, we can only es-
timate the bounds of the survival function with a range of model parameters.
Figure 4 suggests the following conclusions. First, the Kaplan-Meier estimate
(based on the noninformative censoring assumption) overestimates the “actual”
survival functions. Second, Peterson’s bounds of the survival function are too
wide to be useful. In fact, we can use the Kaplan-Meier estimate as an up-
per bound of the survival function. Third, the estimates based on Lagakos &
Williams (22), Fisher & Kanarek (8), Slud & Rubinstein (31), and Klein &
Moeschberger (19) agree quite well overall, and they agree with the empirical
distribution function well through about 36 months. After 36 months, only the
Slud-Rubinstein’s estimate continues to agree well with the empirical distribu-
tion function. This situation may suggest that the conditions required by the
other models are not satisfied beyond 36 months. Our general conclusion from
this example is that assuming ignorable censoring can lead to a biased (in this
case an over-optimistic) estimate of the survival function. Furthermore, most
of the methods that account for noninformative censoring produce reasonable
estimates of the survival functions.
Singly interval censoring: Respiratory symptoms in aluminum potroom work-

ers (Example 5 continued) Recall from example 5 that the disease status of
the aluminum workers can only be determined at the time of the health exami-
nations, and hence the time at which a symptom first occurs is only known in
the time interval between the last examination without a symptom and the first
examination with a symptom. To illustrate the consequence of falsely assum-
ing that the occurrence of symptoms coincided with reporting, in which case
the data can be analyzed as right censored, Samuelsen & Kongerud (28) fitted
P1: EAK/vks
March 19, 1997
P2: MBL/rsk
10:28
QC: MBL/vmw
Annual Reviews
T1: MBL
AR028-04
AR28-04
CENSORING ISSUES
Figure 4 Estimates of the survival function–lung cancer example. Kaplan-Meier estimate (filled diamond 3 dashes); Lagakos-
99
Williams estimate (four dashes); Peterson estimated bounds (• two dashes); Fisher-Kanarek estimate (open triangle two dashes);
Slud-Rubinstein’s estimate (+ two dashes); Klein-Moeschberger’s estimate (∇ two dashes).
the data with techniques for right censoring and for interval censoring using
Turbull’s estimate (34). Notice that we call such an approach, i.e. replacing an
interval-censored observation by its right-endpoint, as right imputation.
Figure 5 displays the Kaplan-Meier estimate of the probability of symp-
toms based on right-imputed data and the Trunbull estimate (a generalization
of the Kaplan-Meier estimator for interval censoring). Note that the Kaplan-
Meier estimate underestimates the probability of symptoms in early follow-up
and overestimates it in late follow-up, thus resulting in an overestimate of the
length of survival probabilities. This over-optimistic estimate occurs because
the time to respiratory symptoms appears to be longer than it actually is when
the right imputation approach is used.

Doubly interval censoring: Survival with malignant melanoma (Example 7

continued) So far we have considered only the problem of survival func-
tion estimation. In this example we illustrate the effect of different censoring
assumptions on the estimates of regression coefficients for doubly interval-
censored data. We consider the data from the malignant melanoma example
described in the first section. Recall that the data structure is similar to that
in Figure 3. To be more specific, the initial event is the first diagnosed stage
Figure 5 Estimate of the probability of symptoms–respiratory symptoms example.

CENSORING ISSUES 101
II disease, the intermediate event is disease metastasis, and the final event is
death. Since the time of disease metastasis is known only to be between the
time of the last stage II disease and the time of the first stage III disease, the
time of the intermediate event is left interval-censored when one computes sur-
vival time for post-disease metastasis. In this example, the times of the final
events are either known exactly or they are right censored (i.e. there is no right
interval-censoring).
Because of the left interval-censoring, we cannot directly apply the standard
approaches for right-censored data for analysis. As mentioned in the second
section, a simple analytic approach is to impute the time of the intermediate
event (disease metastasis) by the right-point or the mid-point of the time interval
and then apply the standard techniques for right-censored data. However, this
approach may not be appropriate, as will be seen below. Table 2 presents the
estimates of the regression coefficients from the imputation approaches and
from an approach that takes the interval censoring into account as proposed by
Leung & Elashoff (24). Basically, they assumed that the time of the event within
the censored interval is governed by an unknown distribution, and proposed an
estimate of the distribution.
Comparison of these estimates suggests that large differences exist among
methods, especially for the estimates related to the interval censored time (the
time from the first diagnosed stage II disease to disease metastasis). First,
the imputation approaches led to very different estimates on the effect of site
and the effect of time between the first diagnosed stage II disease and dis-
ease metastasis (metastases) as compared to the “correct” method. Second,
the imputation approaches underestimate the standard errors of the regression
Table 2 Parameter estimates of the Weibull proportional hazards model—Melanoma Study (example 7)∗
Leung & Mid-point Right-point

Coefficient Elashoff† method imputation imputation
MCV treatment (1 = treated, 0 = control) −0.194 (0.097) −0.197 (0.076) −0.311 (0.076)
Breslow Depth −0.113 (0.104) −0.068 (0.075) −0.046 (0.075)
(1 = depth ≥ 1.8 mm, 0 = depth < 1.8 mm)
Gender (1 = male, 0 = female) 0.241 (0.100) 0.271 (0.076) 0.241 (0.075)
Metastasis site (1 = distant, 0 = others) 0.298 (0.126) 0.452 (0.088) 0.541 (0.088)
I{S ≥ 2 years}∗∗ −0.991 (0.111) −0.687 (0.082) −0.087 (0.097)
λ1 (scale parameter of Weibull dist.) 0.0345 (0.0069) 0.0499 (0.0073) 0.0887 (0.0131)
λ2 (shape parameter of Weibull dist.) 1.161 (0.041) 1.043 (0.026) 0.854 (0.022)
∗
The survival time is defined as the time from disease metastasis to death.
∗∗
S represents the time from the first diagnosed stage II disease to disease metastasis and I{S ≥ 2 years} = 1 if S ≥ 2 years and
= 0 otherwise.
†
See Reference 24.
coefficient estimates. In summary, this example shows that one might obtain
biased estimates and incorrect statistical inferences by falsely assuming that the
time of event is equal to the right-point or the mid-point of the time interval.
DISCUSSION AND FUTURE DIRECTIONS

As we have demonstrated in the last section, assuming incorrect censoring
mechanisms may lead to a serious bias. In practice, there are two common types
of misassumptions regarding censoring mechanisms: For right censoring, re-
searchers usually assume independent censoring (or noninformative censoring).
For interval censoring (both singly and doubly interval censoring), researchers
often assume that the occurrence of an event coincides with the reporting time
(that is, right-imputation). In writing this paper, our intention was to direct
investigators’ attention to the dangers involved in analyzing censored survival
data and point out some techniques to avoid these pitfalls.
An approach to investigating the situation was explored by Fisher & Kanarek
(8): “In some situations a subset of the loss-to-follow-up cases may in fact
be followed although at considerable expense.” With such information, the
investigator can either test the assumption of noninformative censoring (36) or
estimate the risk of informative censoring (1). However, this strategy can be
very expensive and can only be done when censoring is nonlethal.
All the methods dealing with informative censoring discussed in the literature
assume that all censored cases are either all informative or all noninformative.
In practice, there are many situations in which all three kinds of censoring
(positive dependence, negative dependence, and noninformative censoring) are
present in one sample. Thus, it would be useful to extend the existing methods
to deal with all these situations. Furthermore, the methods described here for
estimating the survival function under various conditions assume a fixed model
parameter (Fisher-Kanarek’s α, Slud-Rubinstein’s ρ and Klein-Moeschberger’s
θ; see the last two sections for details). In practice, however, these values are
rarely known. Thus, it would also be useful if there are some guidelines for
investigators to determine the value, or at least a reasonable range of the model
parameters, based on a sample data.
When covariates are available, sometimes it may be possible to recover some
of the information lost by identifying a surrogate response variable measured on
the censored subjects and using it to predict the residual survival time. Although
Cox (3) initialized this idea, no follow-up on this subject has appeared in the
literature.
Finally, the best way to handle censoring is to prevent it from happening
by a good design; no matter how effective the statistical methods are, some
information will be lost when analyzing censored data (2, 6, 35). These design
CENSORING ISSUES 103
strategies for preventing, or at least minimizing, censoring would make a very

useful contribution to the literature.
Visit the Annual Reviews home page at

http://www.annurev.org.
Literature Cited
1. Baker SG, Wax Y, Patterson BH. 1993. censoring and left truncation. Biometrika
Regression analysis of grouped survival 64:225–30
data: informative censoring and double 15. Jennrich R. 1983. A note on the behavior of
sampling. Biometrics 49:379–89 the log rank permutation test under unequal
2. Brook RJ. 1982. On the loss of information censoring. Biometrika 70:133–37
through censoring. Biometrika 69:137– 16. Kalbfleisch JD, McKay RJ. 1979. On
44 constant-sum models for censored survival

3. Cox DR. 1983. A remark on censoring and data. Biometrika 66:87–90
surrogate response variable. J. R. Statist. 17. Kaplan EL, Meier P. 1958. Nonparametric
Soc. B 45:391–93 estimation from incomplete observations.
4. De Gruttola V, Lagakos SW. 1989. Analy- J. Am. Stat. Assoc. 53:457–81
sis of doubly-censored survival data, with 18. Kim MK, De Gruttola VD, Lagakos SW.
application to AIDS. Biometrics 45:1–11 1993. Analyzing doubly censored data with
5. Dudley J, Jin S, Hoover D, Metz S, Thack- covariates, with application to AIDS. Bio-
eray R, et al. 1995. The multicenter AIDS metrics 49:13–22
cohort study: retention after 9 12 years. Am. 19. Klein JP, Moeschberger ML. 1988. Bounds
J. Epidemiol. 142:323–30 on net survival probabilities for dependent
6. Ebrahimi N, Soofi ES. 1990. Relative in- competing risks. Biometrics 44:529–38
formation loss under type II censored ex- 20. Kongerud J, Samuelsen SO. 1991. A lon-
ponential data. Biometrika 77:429–35 gitudinal study of respiratory symptoms
7. Finkelstein DM, Wolfe RA. 1985. A semi- in aluminum potroom workers. Am. Rev.
parametric model for regression analysis Respir. Dis. 144:10–16
of interval-censored failure time data. Bio- 21. Lagakos SW. 1979. General right censor-
metrics 41:933–45 ing and its impact on the analysis of sur-
8. Fisher L, Kanarek P. 1974. Presenting cen- vival data. Biometrics 35:139–56
sored survival data when censoring and sur- 22. Lagakos SW, Williams JS. 1978. Mod-
vival times may not be independent. In Re- els for censored survival analysis: a cone
liability and Biometry, ed. F Proschan, RJ class of variable-sum models. Biometrika
Serfling, pp. 303–26. Philadelphia: SIAM 65:181–89
9. Franks P, Clancy CM, Gold MR. 1993. 23. Law CG, Brookmeyer R. 1992. Effects of
Health insurance and mortality. JAMA mid-point imputation on the analysis of
270:737–41 doubly censored data. Stat. Med. 11:1569–
10. Grüger J, Kay R, Schumacher M. 1991. 78
The validity of inferences based on incom- 24. Leung KM, Elashoff RM. 1996. A three-
plete observations in disease state models. state disease model with interval-censored
Biometrics 47:595–605 data: Estimation and applications to AIDS
11. Hebert PR, Bolt RJ, Borhani NO, Cook and cancer. Lifetime Data Anal. 2:175–94
NR, Cohen JW, et al. 1994. Design of 25. Little RJA, Rubin DB. 1987. Statistical
a multicenter trial to evaluate long-term Analysis with Missing Data. New York:
life-style intervention in adults with high- Wiley. 278 pp.
normal blood pressure levels: trials of hy- 26. Moeschberger ML. 1974. Life tests under
pertension prevention (Phase II). Ann. Epi- competing causes of failure. Technometrics
demiol. 5:130–39 16:39–47
12. Heitjan DF. 1993. Ignorability and coarse 27. Peterson AV. 1976. Bounds for a joint
data: some biomedical examples. Biomet- distribution function with fixed sub-
rics 49:1099–109 distribution functions: application to com-
13. Heitjan DF, Rubin DB. 1991. Ignorability peting risks. Proc. Natl. Acad. Sci. USA
and coarse data. Ann. Stat. 19:2244–53 73:11–13
14. Hyde J. 1977. Testing survival under right 28. Samuelsen SO, Kongerud J. 1994. In-
terval censoring in longitudinal data of der right censoring and left truncation.
respiratory symptoms in aluminium pot- Biometrika 74:883–86
room workers: a comparison of methods. 33. Tsiatis A. 1975. A nonidentifiability aspect
Stat. Med. 13:1771–80 of the problem of competing risks. Proc.
29. Saunders B, Baily S, Phillips M, Allsop S. Natl. Acad. Sci. USA 72:20–22
1993. Women with alcohol problems: Do 34. Turnbull BW. 1976. The empirical distribu-
they relapse for reasons different to their tion function with arbitrarily grouped, cen-
male counterparts? Addiction 88:1413–22 sored and truncated data. J. R. Stat. Soc. B
30. Slud E, Byar D. 1988. How dependent 38:290–95
causes of death can make risk factors ap- 35. Turrero A. 1989. On the relative efficiency
pear protective. Biometrics 44:265–69 of grouped and censored survival data.
31. Slud EV, Rubinstein LV. 1983. Depen- Biometrika 76:125–31
dent competing risks and summary survival 36. Wax Y, Baker SG, Patterson BH. 1993. A
curves. Biometrika 70:643–49 score test for noninformative censoring us-
32. Tsai WY, Jewell NP, Wang MC. 1987. ing doubly sampled grouped survival data.
A note on the product-limit estimator un- Appl. Stat. 42:159–72

Censoring Issues in Survival Analysis: Kwan-Moon Leung, Robert M. Elashoff, and Abdelmonem A. Afifi

Uploaded by

Copyright:

Available Formats

Censoring Issues in Survival Analysis: Kwan-Moon Leung, Robert M. Elashoff, and Abdelmonem A. Afifi

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Censoring Issues in Survival Analysis: Kwan-Moon Leung, Robert M. Elashoff, and Abdelmonem A. Afifi

Uploaded by

Copyright:

Available Formats

P1: EAK/vks P2: MBL/rsk QC: MBL/vmw T1: MBL

March 19, 1997 10:28 Annual Reviews AR028-04 AR28-04

Annu. Rev. Public Health 1997. 18:83–104

CENSORING ISSUES IN SURVIVAL

and Abdelmonem A. Afifi1

1 Department of Biostatistics, School of Public Health, University of California,

84 LEUNG, ELASHOFF & AFIFI

start of follow-up to relapse. After 3 months, 25 subjects were lost to follow-up

Example 3. Survival with inoperable lung cancer A sample of 61 patients with

the appearance of metastatic disease or because of a significant increase in the

86 LEUNG, ELASHOFF & AFIFI

of treatment (MCV or no MCV), other covariates include gender, distant site,

Types of Censoring Mechanisms

anisms illustrated in the above examples. We begin with observations that

Figure 1 Types of point-censored observations.

88 LEUNG, ELASHOFF & AFIFI

The observation of Subject B is therefore right censored at T1 . This type of

analysis. However, it is not always reasonable to make such an independence

INTERVAL CENSORING In many applications, for instance, in Examples 5–7,

Figure 2 An example of singly-interval censored observations.

For subject A0 in Figure 2 (Subject A in Figure 1 but the time of event is

interval censored), instead of knowing that the event of interest occurred at

after SL but up to and possibly at SR . An example of this kind exists in the

Figure 3 An example of doubly interval censored observations.

90 LEUNG, ELASHOFF & AFIFI

it is tempting to transform the observations to the singly interval censoring form,

COMMON STATISTICAL METHODS

been proposed (see 25 for discussion of statistical methods for incomplete

In the context of interval censoring, the inappropriateness of imputation is

1. It cannot distinguish between loss-to-follow-up and end-of-study censoring.

3. No time-dependent covariates (such as age, smoking status or alcohol con-

However, the approach of analyzing dichotomized data may be acceptable

92 LEUNG, ELASHOFF & AFIFI

for example, the Kaplan-Meier estimator of the survival function in a one-

NECESSITY OF MAKING ASSUMPTIONS

With some exceptions, the censoring mechanisms in most observational studies

loss-to-follow-up censoring are involved. Thus, we usually cannot apply the

94 LEUNG, ELASHOFF & AFIFI

number of examinations M is also assumed to be a random variable. However,

IGNORABILITY OF THE CENSORING MECHANISM

likelihood function (Eq. 1) can be used. However, in other situations, it is not

(26) suggested modeling the joint distribution of (T , C) as in (Eq. 2). Lagakos

which includes an exponential survival function of T (with parameter λ), an un-

96 LEUNG, ELASHOFF & AFIFI

defined noninformative examination schemes (or sampling plans) for interval

amined or observed at preassigned intervals. This situation occurs frequently

EFFECTS OF THE CENSORING ASSUMPTIONS

sion usually indicates a shortened residual survival time. Thus, we cannot

apply the standard procedure (assuming noninformative censoring) to analyze

Table 1 Survival times in weeks of 61 patients with inoperable

33 Uncensored observations 28 Censored observations

0.43 15.71 0.14 (3.00)∗ 8.71 (20.43)

98 LEUNG, ELASHOFF & AFIFI

involves some parameters whose values must be guessed by the investigator.

Singly interval censoring: Respiratory symptoms in aluminum potroom work-

100 LEUNG, ELASHOFF & AFIFI

the right imputation approach is used.

Doubly interval censoring: Survival with malignant melanoma (Example 7

Figure 5 Estimate of the probability of symptoms–respiratory symptoms example.