Documente Academic
Documente Profesional
Documente Cultură
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
International Biometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Biometrics.
http://www.jstor.org
SUMMARY
A nonparametricasymptoticconfidenceintervalfor the median survivaltime is developed for the
case where data are subject to arbitraryright censoring. This is accomplishedby inverting a
generalizationof the sign test for censoreddata. A simulationstudy shows that this nonparametric
confidenceintervalperformswell for a variety of underlyingsurvivalfunctions.The procedureis
appliedto data from a clinicaltrial that comparedf'ourdosage regimensof 5-uorouracil.
1. Introduction
It has become a commonpracticein the medicalliteratureto give a point estimatefor the
mediansurvivaltime. Peto et al. (1977) observed that this can be very misleading:if the
survivalcurve is relativelyflat in the neighborhoodof 50/Osurvival,there can be a great
deal of variability in the estimated median. It would be more appropriateto cite a
confidenceintervalfor the median.
Confidenceintervalscan be derived with the assumptionof a parametricform for the
underlyingsurvivaldistribution.The results of Bartholomew(1957), for example, lead to
an interval estimate for the median when the survivalcurve is exponential. However, a
distribution-freeconfidenceintervalis often desirable.
In the one-sample problem without censoring,a nonparametricconfidenceintervalfor
the median is developed by inversion of the sign test. Similarly,for the problem with
censoring, a nonparametricconfidence interval for the median can be obtained by
generalizationof the sign test to enable censoreddata to be handled.This generalizedsign
test is given in 3, and the resultingconfidenceintervalis presented in 4. Some Monte
Carloresultsof the proposed confidenceintervalare given in 5, followed by an example
using data from a clinical trial of treatmentsfor colorectal cancer.
2. The SurvivalFunction and its Quanfiles
Let X1, X2, . . ., X be the true survivaltimes of a sample of size n. They are assumedto
be irldependent,identically distributedrandom variables with survival function S(t)=
pr(X:>t) and cumulativedistributionfunction F(t)=1-S(t). When observationsare
subject to arbitraryright censoring, the period of followup for the ith individual is
restricted to an amount Ti. Then the observed survival time for the ith individual is
Xi = min(X, Ti). One also observes bi, which will indicate if Xi is censored or not. Thus,
if Xi <X, the observationis said to be censored and we set bi= 0. On the other hand, if
Xi = X, the death is observed and we set bi = 1.
The product is taken over all distinct observed survival times {Xi};ni is the numberof
observationsgreaterthan or equal to Xi (that is, ni is the numberof individualsat risk at
time Xi), and di is the numberof observeddeaths that occurredat time Xi In case of a tie
between a censored observationand an observed deathSthe convention is adopted that
the observed death is ranked before the censored observation.If the largest observation,
say X(n),is censored, then S(t) is defined to be O for tX(n).
The estimator S(t) is a right-continuousstep function which jumps at the observed
death points. It has some desirable properties. Kaplan and Meier showed that it is the
distributionfunction which maximizesthe likelihood of the observationswithin the class
of all cumulative distributionfunctions which have probabilitymass at each observed
death time. Efron (1967) showed that the Kaplan-Meier estimate S(t) of a survival
distributionS(t) satisfies the self-consistencyrelation
where Nx(t) is the numberof Xt greaterthan t, it is assumedthat there are no ties among
the observed survival times; note that Efron originally worked with left-continuous
survivalfuIlctions.
The pth quantile of the distributionF is defined to be (F)-l(p)=inf{t:F(t)p}.
J. Sander,in TechnicalReport No. 5, Biostatistics,StanfordUniversity(1975), studiedthe
large sample propertiesof (F0)1(p) under the model of randomcensorship,and showed
under suitable regularityconditions that n2{(Ft')-l(p)-(F)-l(p)} converges weakly to a
zero-mean Gaussianprocess. As in the uncensoredcase, the expressionfor the varianceof
the pth sample quantile depends explicitly on f0{(F0)-1(p)},where f is the continuous
density correspondingto F. The problem with the use of this variance to derive a
confidenceinterval is the difficultyin estimatingthe density at the median.
rl Xi>M,
Q(Xi, bi)- t S(M)/S(Xi), Xi < M, bi = ,
tO, Xi<M, bi-1.
Define the test statistic U= (l/n) 1 Q(Xi, bi). Then we have the followingresult:
Theorem.The test statistic U (used to test Ho median=M) is equal to S(M), the
Kaplan-Meierestimate evaluated at M.
The proof of the theorem, which uses the propertyof self-consistency,is contained in
Appendix 1.
The asymptoticdistributionof the Kaplan-Meierestimate is well known. Breslow and
Crowley (1974) showed that
S(M)-S(M)
r
is approximatelyX2(l)
Under the null hypothesis S(M)=2, and an approximateoe-leveltest is not to reject
Ho when
To find the region Rs,e,it is necessary only to check if observed death times are in the
region.This is because the Kaplan-Meierestimate and its estimatedvariancejump only at
observed death times. Thus, if tl and t2 are two consecutive observed death times with
tl<t2 and tleR,X, then the interval [tl, t2) is contained in R,,. One would expect the
confidence region Rs, to be an interval which includes the estimated median. This is
generallytrue, as the next results show.
5. SimulationResults
In order to investigate the coverage probabilityand length of the nonparametricconfi-
dence interval I,x,a computer simulationwas performed.The nonparametricinterval I,x
was comparedto two parametricintervalswhich assume an exponentialsurvivaldistribu-
tion.
The results of Bartholomew (1957) can be used to derive the maximum likelihood
estimate of the median and its asymptotic variance. For an exponential(A)survival
distributiongiven by S(t)=exp(-At),A>0, the maximum likelihood estimate of the
median M is
^ ln 2 E (observedsurvivaltimes)
M= d '
i P( M )'
Ti being the censoringtime of the ith patient. Here, we have used the expected Fisher
informationto obtain the estimate of the variance. Thus an approximate(1-oe)-level
confidenceinterval,called here the 'Bartholomewconfidence interval',is given by
{ (Pi)2'M+(Pi)24' (3)
where z2cy the 100(1-2oe) percentage point of an N(0, 1) distribution.In order to
iS
variance of logA. Here, we have used the observed Fisher informationto obtain the
estimate of variance. The confidence limits for log A are transformed to provide a
confidenceintervalfor the median.The variance-stabilized1-oe confidenceintervalfor M
is then
[ ln2 ln2 )
lexp(ln A+ d-2z2,x)' exp(ln A-d-2Z-,x) J '
where A = d/ (observedsurvivaltimes).
Three experimentswere performed,each with a differentunderlyingsurvivaldistribu-
tion (exponential, Weibull and Rayleigh). Figure 1 gives a plot of the three hazard
functionscorrespondingto the three distributions.For each experiment,400 simulations
were completed.For every simulation,survivaland censoringtimes were generatedfor 50
patients. The censoring distributionwas assumed to be uniform on [0, T]. For each
simulation a 1- oe confidence interval for the median survival time was computed
(oe=.01,.05,.10,.20 and .25). Then the observed coverageprobabilitywas recorded,i.e.
20
1S _
A(t) Weibull/ /
(X 103) 10 * / /
/ / Exponentis
5 / /
X Rayleigh
I x I I . .
20 40 60 80 100
TIME
oe
Table 2
Average confidence interval length. Exponential (.01) survival
distribution
Table 3
1- oe confidence intervals:observed coverage probabilities.Weibull (0.0012, 1.5)
survivaldistribution
.
0e
Table 4
Averageconfidenceintervallength.Weibull(.0012, 1.5) survivaldistribution
12% censored 16% censored 32% censored
oz: .01 .10 .01 .10 .01 .10
Table 5
l-ot confidence intervals: observed coverage probabilities. Rayleigh (.0001) survival
distribution
of
ConfidenceIntervalfor MedianSurvival
37
he obtained results which appear quite good. Efron (1981) developed a confidence
interval by the method of bootstrappingcensored samples; that is, the distributionof
the sample median is estimatedby Monte Carlo simulationof samplesgeneratedfrom S.
These techniquesare summarizedand reviewed by Reid (1981).
6. An Example
The confidenceintervalprocedurediscussedearlier was applied to data from a Phase III
colore,ctalcance,rclinicaltrial (Ansfieldet al., 1977). [The data analyzedwere an updated
versionreportedin the CentralOncologyGroupFinal Report (COG 7030), Spring1977.]
Four dosage regimensof S-fluorouracilwere compared.Table 7 summarizesthe data for
each of the four groups. The 95/Oconfidenceintervalsfor each of the four medians are
also shown. Treatment1, an intravenousloading-courseschedule, had the largest median
survival time of 61 weeks. Although the medians 'appear' to be different, all four
confidence intervals overlap. In fact, the Gehan (Breslow) test shows no significant
difference among the groups (P=.22). Since there were ties in the data, the Kaplan-
Meier variancewas estimatedby (2). Table 8 details the calculationsfor Treatment 1. A
numerical check shows that all the confidence regions are indeed proper intervals
(R,,,--I,,,). Figure 2 shows'the Kaplan-Meierestimate for Treatment 1 and several 1-oe
confidenceintervals (oc=.01,.05,.10,.20 and .25).
75 /O C43,64)
80% [4C)64)
100 90% [39 73)
90-q%, 95% [38,73)
80- \
70- tbo
> 60- 50 ^
(t) 50 _ i3'o 4t
a} 40- O
980 O
0 95 '-_
10 - I 99 . 1 -0vv
O ' I I I 1 1 1 , I I
O 25 50 75 100 125 150 175 200 225
Weeks
Table 7
Samplesizes, mediansand 95% confidenceintervalsfor four treatments
_ , . . . .
Treatment
1 2 3 4
Censored 16 8 14 7
Observed deaths 37 48 44 45
Sample size 53 56 58 52
Median survival time (weeks) 61 41 47 29
95% nonparametric confidence
interval [38,73) [31,51) [28,60) [25,46)
95% variance-stabilized
confidence interval (38,73) (28,50) (30,55) (27,48)
Table 8
Calculations for 95% confidence interval for median of
Treatment1
Observed x
death time if
(weeks) S0 SE(S) (S _ 1)2 < 3 84{SE(S0)2}
6 .962 .027
11 .943 .032
13 .925 .037
14 .906 .040
15 .887 .044
16 .868 .047
21 .829 .052
22 .810 .055
23 .791 .057
25 .771 .058
26 .752 .060
27 .732 .062
28 .713 .063
31 .692 .064
34 .671 .066
37 .649 .067
38 .627 .069 x
39 .604 ? x
40 .582 .071 x
43 .559 .072 x
44 .534 .072 x
54 .508 .074 x
61 .423 .080 x
64 .395 .079 x
73 .338 .079
83 .308 .077
88 .277 .076
90 .246 .073
91 .215 .070
106 .179 .067
128 .144 .062
144 .096 .057
ACKNOWLEDGEMENTS
This researchwas supportedin part by NIH grantCA-18332. We would like to thank Dr
John Emersonfor some helpful commentson an earliermanuscript,and Dr Fred Ansfield
for permissionto use his data.
RESUME
Un intervalle de confiance asymptotiquenon parametriquepour le temps de survie median est
developpe dans le cas ou les donnees sont arbitrairementtronquees a droite. Ceci est realise a
l'aide d'une generalisationdu test des signes pour les donnees tronquees. Une etude de simulation
montre que cet intervallenon parametriqueconvient bien pour une variete de fonctions de survie
sous jacentes. La procedureest appliquee aux donnees d'une experience clinique comparantdes
regimessuivantquatredosages de 5-fluorouracil.
REFERENCES
Ansfield, F., Klotz, J. and the Central Oncology Group (1977). A Phase III study comparingthe
clinicalutility of four dosage regimensof 5-fluorouracil.Cancer39, 34-40.
Bartholomew,D. J. (1957). A problemin life testing.Joumalof theAmericanStatisticalAssociation
52, 350-355.
Breslow, N. and Crowley, J. (1974). A large sample study of the life table and product limit
estimates under randomcensorship.Annals of Statistics2, 437-453.
Efron, B. (1967). The two-sampleproblemwith censoreddata. In Proceedingsof the Fifth Berkeley
Symposiumon MathematicalStatisticsand Probability,Vol. 4, L. LeCam and J. Neyman (eds),
831-854. Berkeleyt Universityof CaliforniaPress.
Efron, B. (1981). Censoreddata and the bootstrap.Journalof the AmericanStatisticalAssociation
76, 312-319.
Emerson, J. (1982). Nonparametricconfidence intervalsfor the median in the presence of right
censoring.Biometrics38, 17-27.
Gehan, E. A. (1965). A generalized Wilcoxon test for comparing arbitrarilysingly-censored
samples. Biometrika52, 203-223.
Greenwood, M. (1926). The Natural Durationof Cancer.Reports on Public Health and Medical
Subjects,No. 33, 1-26. London: HMSO.
Kaplan, E. L. and Meier, P. (1958). Nonparametricestimation from incomplete observations.
Journalof the AmericanStatisticalAssociation 53, 457-481.
Lehmann,E. L. (1959). TestingStatisticalHypotheses.New York: Wiley.
Peto, R., Pike, M. C., Armitage, P., Breslow, N. E., Cox, D. R., Howard, S. V., Mantel, N.,
McPherson,K., Peto, J. and Smith, P. G. (1977). Design and analysisof randomizedclinical
trials requiring prolonged observation of each patient. II. Analysis and examples. British
Joumal of Cancer35, 1-39.
Reid, N. (1981). Estimatingthe median survivaltime. Biometrika68, 601-608.
ReceivedJuly 1980; revisedJanuary1981
APPENDIX1
Proofof the Theorem
To prove that the test statistic U is the Kaplan-Meierestimate at M, write
1 lt
U - - E Q(Xi, bi)
n i=l
1 n
--E pr(Xi> M | Xi, Ai,;)
n i=l
where NX(1\4)
= numberof Xi > M with bi = 1. Also,
where NX(M)
= numberof Xi >M with Ai= O. Thus,
U=-{Nx(M)+Nx(M)+ , S(X) }
Comparisonof the term in bracketswith the definitionof selfconsistencygiven by (1) shows that
U-(l/n){nS(M)}= S(M)-
APPENDIX2
and let N be the number of patients at risk at t2. Then tleRa implies (Sl-2)2<c,,(S)2V1 and
{NS2/(N-1)-2}2 < {N/(N-1)}2(S 2)2c,,V1, since g = NS 2/(N-1). Simplifying,
Proof of PropertyB. Since tl + Ra, we have (4) with the inequalitysign reversed:
It sufficesto show that (S2-2)2 is greater than the left-hand side of the above inequality.Since
C(X < 1,