MAXIMUM UTILIZATION ANALYZING SURVIVAL

SIDNEY J CUTLER, M.A., FRED EDERER, B.S. BETHESD~L, MD. AND

OF THE LIFE TABLE METHOD

IN

(Receivedfor publication Aug. 18, 1958)

of patient survival is necessary for the evaluation of 44EASUREMENT treatment of usually fatal chronic diseases. This is particularly true for

cancer. The American College of Surgeons, recognizing this, requires the maintenance of a cancer case registration and follow-up program for approval of a Acceptance of survival as a criterion for measuring hospital cancer program. the effectiveness of cancer therapy is also attested to by the very large number of papers published every year reporting on the survival experience of cancer patients. Although the proportion of patients alive 5 years after diagnosis (S-year survival rate) is the most frequently used index for measuring the efficacy of therapy in cancer, an increasing number of investigators are reporting on the manner in which patient populations are depleted during a period of time, e.g., survival curves. A popular and relatively simple technique for describing survival experience over time is known as the actuarial or life table method. Whereas the method and its uses have been admirably described by a number of authors,2-6 A principal adone important aspect has received relatively little attention. vantage of the life table method is that it makes possible the use of all survival information accumulated up to the closing date of the study. Thus, in computing a S-year survival rate one need not restrict the material to only those patients who entered observation 5 or more years prior to the closing date. We will show that patients who entered observation 4, 3, 2, and even one year prior to the closing date contribute much useful information to the evaluation of 5-year survival. Let us consider a group of patients entering observation continuously beginning with Jan. 1, 1946. Sometime early in 1952, we decide to analyze the survival experience of these patients to obtain a 5-year survival rate. We choose Dec. 31, 1.951, as the closing date, i.e., the follow-up status and survival time of each patient is recorded as of that date.
699

CliTLEK

.\ND EDEKEIZ

J. Chron. Dis. December, 1958

Of the

patients

entering

the study

during entering
I

the

6 >.ears ended

on Dec.

31

1951, only those diagnosed 5 years.* The exposure is shown in Table I.

in 1946 were exposed

to the risk of dl-ing for at least in each of the calendar >.ears

time for patients

IADLE

CALENDAR

YEAR

OF DIAGNOSIS

YlT4RS

OF I<SPOSURE

TO RISK

OF DYING

1946 1947 1948 1949 1950 1951

5 4 3 2 1 Less

to 6 to 5 to I to 3 to 2 than 1

It might

be supposed,

intuitively,

that

the patients a S-year

who entered survival

observation

from 1947 to 1951 are of no value in computing 31, 1951, since each of these patients This, should through reliable expcsure however, not is not true. Merrell from rates the for whom less than the required be discarded survival is just maximum S-year time utilization short survival and Shulman5 of years analysis. series The Data

rate as of Dec. out that patients is available that, possible is and to Registo compute

was under observation survival Wilder method, even primary

for less than 5 years. information

have pointed has when

number

demonstrated the longest

of the life table for a large information purposes.$ of 5 years.t

it is possible objective

of this paper Cancer

to show how partial

can be included

in the life table

show how much is gained by doing so. ter are used for illustrative

from the Connecticut

THlS ANA\TOMY Oli THE I,IFE T.W,E$

Table patients through

II provides with localized

the basic cancer

facts,

as of Dec.

31, 1951,

concerning

126 male 1946

of the kidney,

diagnosed

during

the period

1951.

The cases are divided

into 6 cohorts,

one for each year of diagnosis. column gives the time F-or

The columns of Table II are described Column 1. EVeears After Diagnosis elapsed example, of patients column from the date of diagnosis a patient the third )ear after who was diagnosed

here. (x to x+2).-This of one year, 20, 1946, i.e., during for removal Jan.

in intervals diagnosis,

i.e., O-1, 1-2, etc.

died during

and died on Oct. 5, 1948, interval 2-3. The number in the appropriate from observation.

that left observation

during each interval to the reason

is entered

(3, 4, or S), according

Columns 2. Alive at Beginning of In,terval (1,).-The entry on the first line of this column indicates the number of cases alive at diagnosis, i.e., the initial number of patients 3. in the cohort. During Interval (d,).Column Died

Tn practice, other *In this example, date of entry into the study is dnfirxxl as data of diagnosis. refnrrnce dates, such as date of initiation of a particular course) of therapy, may be used. tin the snrirs reported by Wilder, the range of exposure time was from one day to, but not including. 5 years. $We wish to thank Dr. Matthew H. Griswold, Director, Division of Canrrr and Other Chronic Diseasw. Connecticut State Department of Health, for his courtesy in making these data available. BWe borrowed the phrase anatomy of the life table from Pearls2 excellent textbook Biometry and
Medicnl Stntistics.

Volume Number

8 6

LIFE TABLE
~A1~1.11 II.

METHOD IN ANALYZING
1 jA7.A FOR SINGLIi

SURVIVAL
COHORTS Ikz~Iosetl ,1

701
19-I-h-1951 aml

SI~R\TVAT.

EKAR

(126 hlalc

Conwcticrlt

Resitlents \Iith Localizetl Kidnev Ca~~rcr: Follo\rwl lhrough I kc. 31,1951)

YEARS DIAGb

AITIR OSIS 1

ALI\Ii NING

AT

HI<(;INDCrRINC;

OLLOW-UP INTERVAL (4) ux

WITHDRAWN (5) w x

ALIVE

OF INTIiR\AI.

DITRINC; INTERVAL*

(1) X TO >: +

Patients diagnosed in 1946 (1946 cohort) o-1 1-2 2-3 3-4 4-5 S-6 9 4

4 t :

._____Patients diagnosed in 1947 (1947 cohort) 0-l l-2 2-1 3-2 4-5

18

1: 10
6

f,

Patients diagnosed in 1948 (1948 cohort) O-I l-2 2-.3 3-4 21 10 7

11 1

7 -

I
2; 16 12 3 1

Patients diagnosed in 1949 (1949 cohort)

O-l l-2 2-3

15

Patients diagnosed in 1950 (1950 cohort) _______ o-:! l-2! 19 13 5 1 : 11

Patients diagnosed in 1951 (19.51 cohort) O-l

25

~ I

1.5

*Alive at closing date of study.

TABLE

(126 Male Connecticut

III. COMBINED LIFE TABLE AND COMPUTATION OF S-YEAR SURVIVAL RATE Residents With Localized Kidney Cancer Diagnosed 1946-1951 and Followed Through

Dec. 31, 1951)

E.FFECTIVE NUMBEI TO THE ROPORTION PROPORTION SURVIVING DYING M COL. 5) COL. LOST TO FOLWITHDRAWN EXPOSED RISK -( COL. 2 OF DYING ALIVE VG INTERVAI (COL. DURLOW-UP NG INTERVAL DUR-

C CMULATIVE SURVIVING NOSIS

PROPORTIOS FROM THROUGH OF INTERVAL DIAGEND 0

ALIVE

AT

DIED

YEARS

AFTER

BEGINNING

DURING

DIAGNOSIS

OF INTERVAL

INTERVAL

- w

(1 -

COL. 7) (P,X (8) PK

P?X

. .x
. (9) pX

Px)

(1) XTOXf w* W,

3f COL. 6) (7) qx

126 4 6 1.5 2 E 116.5 51.5 30.5 0.40 0.10 0.07 0.12 0.00

O-l i-2 2-3 3-4 4-5 5-q

obtained survival rate; it is included here by summing, cell by cell, the survival merely data of the to complete

z 21 10 4

0.60 0.90 0.93 0.88 1.00

0.60 0.54 0.50 0.44 0.441

*Columns

2 through

5 of this

table

were

6 yearly

cohorts the account

of Table of the initial

II. 126 patients. Bow were

iThis

line is not

needed

for

computing

the 5-year

alive

at the closing

date

of the study.

:Five-year

survival

rate.

Volume 8 Number 6

LIFE TABLE

METHOD

IN ANALYZING

SURVIVAL

703 this column patient we lost to be

Column 4.
enter Dec. alive. the 31, 1951,

Lost to Follow-up During Interval (u,). -In

of patients whose The survival status was unknown. observed length of observation

number

as of the for each is entered that similar

closing

date,

to follow-up Thus, line, i.e.,

is the time elapsed a patient 3-4. the life table under

from date of diagnosis for 3 years and 4 months method it is usually of lost that

to date last known

on the fourth

interval

In applying cases remaining from the analysis vival experience information. Column enter Dec. vation during the 31, 1951. depends the third II, but 5.

assumed cases was

subsequent to
to that of lost of surcases

date of last contact, the survival

follow-up. is equivalent

experience to assuming

In contrast,

complete

omission

from date of diagnosis the

of lost cases was similar

to that for cases with complete

follow-up column we date,

Withdrawn Alive During

of patients interval known during to The which

Interval (w,).-In
have these been alive patients

this

number

on the closing withdrew all patients that,

from obserdiagnosed cohort for all for was For

on their year

date of diagnosis. diagnosis,

For example, 2-3. Note

in 1949 alad alive on Dec. in Table intervals

31, 1951, are recorded interval by dashes) (1946)

as withdrawals

from observation for each column

after

zeros (symbolized the last. only

are entered provided

on

in this survival

Although a full 5 years, obtained example,

one of the cohorts all the information the entries

information Table III cell by cell.

we used the available

information in Table on the first

all 6 cohorts. summing

by pooling by summing

II,

line of Column

3 for each yearly one year

cohort in Table 11, we obtained of diagnosis, shown in Table III. In practice, directly, cohorts. the cohorts after the data III, as in Table contributed (Table after rather

the total

of 47 cases cohort

who died within

for the pooled than procedure

of 126 cases would be tabulated tabulations for 6 individual each of Tables year three 6 7 were information by comparing in 1948, in 1948, some after

by summing data.

\JTe used the latter

to show how much For example, known of the each

to the pooled III, Line

II and II I, we find that diagnosis and in 1949, alive

of the 5 patients 2, Column Similarly, Thus, survival

to have died in the second diagnosed

3), one was diagnosed 18 patients diagnosed of 5 years from

one in 1950. after diagnosis. of patient First,

in 1947, information diagnosis.

were alive 4 years 3 years to our knowledge A statistical be discussed in Columns

COMIl_TATION

diagnosis;

of the 21 patients cohort during a period resulting

contributes

measure later. 1 through

of the gain however, 5 of Table

RATES

in precision III

this procedure rates.

will

we will explain

how the basic data summarized survival

are used to compute

OF SCRVIVAL

to mind as it symbol for lost cases, because 1 is a standard life table notation for alive at beginning of interval. tFor a detailed account of the mechanics of recording and tabulating survival data, see Berkson and Gage, pp. 4-5.

The first step in preparing a life table is to distribute the deaths, losses, and This withdrawals with respect to the interval in which they left 0bservation.t __~__. *WF!ale using the letter u to represent untraced cases, rather than the letter 1 which comes

704
iuformatiou entries which is summarized 3, 4,
on

CUTLER

AND EDERER

J. Chron. Dis. December, 1958

iu (~oiumns
wd 5

3, 4, and 5 of Table the total uumber

11 I. of cases

The sum of the in the study,,

iu (olumus is entered

equals

the first liue of (~01um11 2 (126 (~;\sw). according alive from to the formula: (d, + number losses,
11, + W,).

Sue-c-cssive clltries

it1

this column
For year (47 + example, by

are obtained
the number

Ix+1 = 1, obtained 4 + subtracting 15). the

at the

beginning alive

of the second at the during

year the

(60)

was first

beginning

of the

(126)) the sum of the deaths,

and withdrawals

first year

The life table up year (Columns

is completed by a series of four computations 6 through 9). from observation iu 1951,

for each follo\vis assumed were exposed For example, (withdrawn 15 patients
on

Column 6.
that patients

E$ective Number Exposed to Risk of Dying (I,).-It

during an interval the interval.* 31, 1951 for these on the average, to assume was observed exposed for one-half

lost or withdrawn diagnosed

to the risk of dying, of the 25 patients alive). average, The number was roughly It is reasonable equally each patient effective

15 were alive on Dec. the calendar year. year

that the date of diagnosis during for one-half to risk Thus, (UX + w,) /2.

distributed

1951 and that,

the the lost

number

is obtained one-half

by subtracting

from

alive at the beginning during the year.

of the year,

the sum of the number

and withdrawn Column 7.

1, = 1, as the probability ----of dying during

Proportion Dying During Interval (q,).-This

the interval. It is obtained

is also referred to by dividing the

*The computing procedure given here is based on the assumption that, for cases wit,hdrawn alive and cases lost to follow-up, survival subsequent to date of last contact is similar to that for cases with complete follow-up information. For cases withdrawn alive, this assumption introduces no bias, brcause there is no reason to believe that patients alive on the closing date are different from patients observed for a longer period. However, for cases lost to follow-up, this assumption may introduce a bias. Patients lost to follow-up were alive when last observed, and whether their survival experience is better than, worse than, or equal to the survival of patients remaining under follow-up is highly Farspeculative. For example, cancer patients may be lost to follow-up for a variety of reasons. advanced cases may leave their usual place of residence to enter the household of a relative; successfully treated patients may stop reporting to t.he tumor clinic, because they feel that no further medical care is required. It is therefore important to keep the proportion of cases lost to foliow-up at a minimum. Survival rates based on a series in which a substant,ial proportion of paGent have been lost to follow-up are of highly questionable value, because it is impossible to determine the extent to which they are biased. Some investigators, such as Paterson and Tod8 recommend that lost cases be counted as dead to avoid undesirable uncertainty. although (it) may result in a slight bias against t,he ellicacy of treatment. Other investigators, such as Ryan and his colleagues9 omit lost cases from the analysis of survival. The latter procedure involves the assumption that from date of diagnosis the survival experience of lost cases is similar to that of cases with complete follow-up. We prefer the first of the several possible assumptions regarding lost cases, namely that subsequent The complete to date of last contact their survival is similar to that for cases with complete follow-up. The omission of lost cases from the computation of survival rates discards available information. Registry assumption that lost cases died immediately after the date of last contact is cont.rary to fact. experience with intensive field investigation of lost cases, which resulted in recovery of some, indicates that such patients often live for several years beyond the initial date of last contact.1D Although cases withdrawn alive and cases lost to follow-up are treated alike in the computations (1) it is import,ant described here, we distinguish between the two in the life table for reasons mentioned: to be aware of the number of cases lost to follow-up because of their potential bias, and (2) other computational methods may treat the two groups differently.

Volume Number

8 6

LIFE

TABLE

METHOD

IN ANALYZING

SURVIVAL

70.5

number

of deaths

by the effective

number

qx= 1, To express as a percentage, Column ternately is obtained

lo

exposed d,._

to risk:

multiply

by 100. the In,terval &,).-This the interval, dying

qx.

8.

Proportion

Surviving

is referred

to alIt

as the probability by subtracting as a percentage, 9. Cumulative

of surviving the proportion

px = 1 -

or the survival the interval

rate.

during

from unity:

express Column

multiply

by 100. Surviving From Diagnosis survival to End rate. of It to as the cumulative the proportion
ps x . . . px.*

Proportion referred
pz x

Interval

(I;,).-This

is generally-

is obtained

by cumulatively entries

multiplying
p1 x

surviving 2-year,

each interval 3-year, 4-year,

px =
Note that successive and S-year survival cumulative rates survival

in this column rates in drawing

give the l-year, II I). curve. The

(Table a survival

successive

cumulative out in inter-

are plotted

Although of days, intervals the first intervals. of deaths described

the computations

illustrated

in Table

III werecarried

vals of one year after

the date of diagnosis,

the life table

may be set up in terms

weeks, months, years, etc. In fact, the life table may be organized in of varying length. For example, one might record experience during ::ear This occur in monthly type during intervals, and The the experience thereafter a large in annual proportion rates of presentation the first year. ma)- be desirable method when

of computing

survival

here may be used whatever

EXPERIENCE

the size of the intervals.

WITH PARTIAL FOLLOW-UP

Gz\IN IN I.TILI%ING

OF (OHORTS

The

Istandard

error

provides

a measure

of the

confidence

with

which

one

may interpret a statistical result. Thus, the standard error of the survival rate indicates the extent to which the computed rate may have been influenced by sampling error per cent errors variati0n.t confidence For example, the computed interval. by adding survival and subtracting one obtains twice the standard an approximate 9.5 to and from rate,

the same conditions on either

This means that in repeated observations under the true survival rate will lie within a range of two standard rate, an average of 95 times in 100.

side of the computed

Thus, the computed 5-year survival rate for male patients with localized cancer of the kidney is 44 per cent. The standard error, computed according to the method explained iu the Appendix, is 6 per cent. It is therefore likely that
*This f~xmula is based on the assumption that the various interval survival probabilities are statistically independent. tThe 126 cases of localized cancer of the kidney are in effect a sample from a population of male patients with localized kidney cancer. An illus,tration of sampling variation may be drawn from baseball. A 0.250 hitter may, in four times at bat, get one hit. Frequently, though, he will ge6 no hits or two hits. And not too infrequently he will get three hits. If we watch a game and see a player get two hits in four times at bat, it is difticult for us to judge how good a hit,ter this player really is. We have to watch this player for many games before we can get a reliable estimate of his batting average. Survival rates are similar to batting averages in the sense that they are relative frequencies, i.e.. the numerator is part of the denominator. For each hit there must be at least one time at bat. and for each death there must be at least, one case exposed to the risk of dying,

CUTLER

AND

EDEKEK

the true 5-year 56 per cent. Admittedly, true survival 126 patients tion available (one-third follow-up explained The in Table larger

survival

rate is not smaller rate

than

32 per cent a veq.

and

not larger

than of the

the computed

does not yield

firm estimate
on a

rate, but we must bear in mind that it was based and only 9 of these patients were diagnosed

series of onl?

a full 5 >.ears prior to

the date of study.

Furthermore, whereas the survival rate based on all informaon these 126 patients provides at least a rough idea of the true rate discarding would result that method applied the information in an extremely to the total portion survival follows. series of 126 cases, based illustrated of the group. rates We have thereon successively on the cohorts unreliable with partial This is estimate.

to one-half), information computing III,

in the discussion can be applied cohorts.

to any selected of S-year

fore used it to compute patient

a series

A S-year survival rate was computed for the 9 patients diagnosed in 1946, all of whom had a S-J-ear exposure time. \Ve then added the 1X patients diagnosed in 1947, who had a 4-year exposure time, and computed a S-year This was Table survival procedure utilized IV. The a standard 1946 cohort of 9 cases yielded The a 5-year survival rate of 53 per cent, tells
us

rate in

based

on

the the

available the S-year

information experience rate.

for these

27 patients.

was continued estimating standard

until

known survival

of all 126 patients

and their

corresponding

errors are shown

The successive rates in the uppermost section of with

error of 17 per cent. estimate;

large standard

error

that

this is a

very unreliable a survival addition error addition from

the true rate is probabl>-

between

19 and X7 per cent,*

a very wide range.

The combined experience of the 1946 and 1947 cohorts yielded rate of 46 per cent, with a standard error of 10 per cent. Thus, the on cases with 4 full years of exposure to 10 per cent, information a relative on cohort decrease 1948 reduced the standard The of exposure) 17 per cent of 43 per cent.

of information of the available

(3 full years

reduces the standard error to 7.5 per cent, etc. The utilization of all available information on all the cohorts results in a standard error of 6.0 per cent. Thus, the standard We series groups breast then error of the survival survival cancer breast rate rates based and
on

all available

information errors

is 6.5 for

per cent less than the standard computed of successively of patients: cancer, enlarged kidney

error based cohorts cancer

on cases with a full 5 J-ears of exposure. corresponding for each involvement, standard of four additional in women;

of patients with

with regional

in men; localized

in women;

regional

involvement,

and cancer of the lip, both sexes combined (Table illustrate the advantage of utilizing all available of varying graphically rate than --__*These are the 95 per cent confidence limits: 53 * Z(17).

IV). \Ie did this ill order to experience for patient groups The results 1951 are shown survival is smaller one-third error of the 5-year by at least

size and with in Fig. 1.

on

varying

mortality

experience.

In ever) instance,
experience standard

the standard of cohorts

based

the combined

1946 through

the corresponding

error for the 1946 cohort

\olume

Number 6

LIFE

TABLE

METHOD

IN ANALl-ZING

SURVIVAL

707

'IABLE IV. FIVE-YEAR SUR\-IVAL KATES AXD -THEIR STANDARD ERRORS FOR FIVE GROWS OF CANCER PATIENTS, SHOWING THE ~?DuCTIOK IN STANDARD ERROR \%ITHINCREASE IN COHORT SIZE

COHORT

NUMBEROF CASES DIAGNOSED

5-YEAR SURVIVAL RATE

STANDARD ERROR OF 5-YEAR SURVIVAL RATE

PER CENT REDUCTION INSTANDARD ERROROF~-YEAR SURVIVAL RATE

Kidney, localized
1946 1946m-1947 1946.-1948 1946~-1949 1946m.1950 1946-.1951 0.53 0.46 0.43 0.43 0.45 0.44 0.171 0.098 0.075 0.064 0.063 0.060

48 82 101 126

2:

h
1946 1946-.1947 1946~-1948 1946m-1949 1946~-1950 1946m.1951 11 23 30 39

idney,

regional

0.18 0.33 0.28 0.25 0.23 0.24

0.116 0.101 0.091 0.071 0.069 0.070

13 22 36 41 40

Breast, localized
1946 1946m.1947 1946- 1948 1946~-1949 1946&l 950 1946~-1951 225 454 695 963 1,227 1,490 0.64 0.64 0.64 0.64 0.65 0.65 0.033 0.025 0.023 0.022 0.022 0.021

24 30 33 33 36

Bread,

regional
0.42 0.38 0.39 0.035 0.025 0.021 0.020 0 020 0.020

1946 1946m.1947 1946- 1948 1946- 1949 1946--1950 1946m.195 1

208 443 708

29 40 43 43 43

Lip
1946 1946.-1947 1946~-1948 1946--l 949 1946--1950 1946.-1951 61 109 i69 224 283 332 0.71 0.65 0.68 0.68 0.68 0.67 0.060 0.048 0.042 0.040 0.040 0.039

20 30

33
33 35

708 The advantage of utilizing was greater informatiol1 for localized on patient kidney cohorts

J. Chron. Dis. December, 1958

lvith less than for the

>rears of exposure

cancer with
an

than average

other

This is because: (1) particularly groups. were diagnosed in the first >.ear (1946), in each of the subsequent of follow-up (0.40) years; was much larger than

few cases (9) of localized compared rate during lears

kitlne~~ ca1lcer of 23 cases the first >.ear average of

and (2) the mortalit>, in succeeding

(annual

Thus, because of this mortality pattern, the information on surviv:il 0.07). during the first year after diagnosis contributed ver). substantially to the information on survival over a S-year period. Five of the 6 annual cohorts (1946-1950) contributed agnosis. cohorts relative of the mortality after that complete information In general, the relative partial survival during and follow-up in the initial regarding survival during the first year after digain in utiliziug survival information on patient information and will vary (3) the directi), with: (2) the relative relative (1) the completeness of the

with added rates

increase

size of the cohort*;

information; diseases,

magnitude

the first few follow-up mortality group to taper the patient

intervals. is often For before relatively some high shortI> such as one year diseases, within

In cancer, diagnosis little lung or stomach results.

and in other tends cancer,

off thereafter. may

be so depleted

is gained Therefore,

by waiting

more than one year be unnecessary

evaluating

therapeutic 2-year, survival in the

it may frequently

to wait until a .5-year surA I-year, instances, changes In other

vival rate can be computed or 3-year rate may provide data for only 5 years pattern may mortalit)DISCUSSION

to evaluate the effects important information. be inadequate, time. because

of therapy. of significant

at a later

A category tionally chosen all available of localized cause small groups

of patients

with

relatively example

few new cases of survival the survival breast

per year rates-only This

was intenof utilizing 126 cases

as the principal cancer

to illustrate

the advantage

information kidney

for the computation to describe with

in men were diagnosed For example, we would therapy

in 6 years.

was done beof relative11 the by surgery the number onlv 25 with

it is frequently experience receiving Similarly,

desirable of patients

experience treated

of patients.

if we were interested cancer find that is small.

in evaluating

survival of patients therapy.

localized

in combination

with radiation,

in any one year

the combined if survival

As an illustration, for a specific

of the 225 cases diagnosed

in Connecticut

in 1946 were treated

b,. the combined subgroup

is to be evaluated

respect to age, the number of cases per year would usually be small. Therefore, in order to increase the reliability of survival rates computed for various patient groups of clinical computing trial. in a clinical ___-interest, rates it is important importance if the rates survival to utilize all available survival information. information in in a clinical as a criterion treatments It is of paramount survival trial. For example, to use all available rate may to determine
sample size.

are going to be used as criteria have been selected which of the several

a 3-year

It may be possible
for a discussion

*See the Appendix

of effective

VoIume

Kumber

1.1~~ TABLE METHOD IN ANALYZING SURVIVAL

KIDNEY,LOCALIZED 1.00 [ KIDNEY ,REGIONAL

709

.O( 1 .90 -

90

.eo ,712 -

r2 2 1 &

80
70 .60 .50 40-T 3020-(, IO 1946 l9461947 l9461946 19461949 l9461950 lS461951 t -

2
-I

2
2 2 : w > A

&I3 ,,
.5l 40.SD 20 -

2 5 w > In

J
1946 l9461947 19461946 ,9461949 ,9461950 ,9461951

BREAST,

LOCALIZED 1.00 I
.90

BREAST

,REGIONAL

-00 .70 i
.60 5040-

5 lx
4 > 2 z oz g

3o -

1;1 .20

01946 l9461947 l9461948 19461949 l9461950 l9461951

0 l94619461947 l9461946 l9461949 l9461950 19461951

LIP

.J

I I I:
i
0'
1946 19461947 lS461948 19461949 ,S461950 ,946195,

Pig. I.-Decrease in the 9-5 per cent confidence interval for the B-year survival rate as cases with less than 5 years exposure to the risk of dying we added. (The 95 per cent confidence interval is ohtained by adding -2 standard errors to the survival rate.) Source: Table Iv.

710

CGTLER

AND EDERER

J. Chron. Uis. December, 1958

being tested yields the best survival

death or for a full 3 years. possible point. Thus, the earliest

before all patients

treatments

have

been

followed

to at

inferior

would be discontinued

We have illustrated S-year survival data including terms patients two-thirds.

REFERENCES

the life table

method

for computing entered

survival

rates gained

with b)

for cancer to survive

patients,

emphasizing

the advantage

survival

information in standard

on cases which a full 5 years. error

the series too late to have

had the opportunity of reduction in this paper,

The advantage is measured in of the survival rate. For the five series of in standard error ranged from one-third to

the reduction

1. 2. 1 j: 5. 6. 7. 8. 9. ::: 12.

Manual for Cancer Programs, Bull. &4m. Coil. Surgeons 38:149, 1953. Pearl, I~.: Introduction to Medical Biometry and Statistics, Philadelphia and London, New York, 1956, Oxford liniversity Press. Berkson, J., and Gage, R. P.: Calculation of Survival Rates for Cancer, Proc. Staff Meet., Mayo Clin. 25:270, 1950. Merrell, M., and Shulman, L. E.: Determination of Prognosis in Chronic Disease, Illustrated by Systemic Lupus Erythematosus, J. CHRON. Drs. 1:12, 1955. Griswold, M. H., Wilder, C. S., Cutler, S. J., and Pollack, E. S.: Cancer in Connecticut, 1935-1951, Hartford, 1955, Connecticut State Department of Health, pp. 112-113. Wilder, C. S.: Estimated Cancer Survival Rates Confirmed, Connecticut Health Bulletin 70:217. 1956. Paterson, 1C.l and Tod, hl. C.: Presentation of the Results of Cancer Treatment, Brit. J. Radlol. 23:146, 1950. Ryan, A. J., et al.: Breast Cancer in Connecticut, 1935-1953, J..4.M.ii. 167:298, 1958. Griswold, M. H.: Personal communication. Cutler, S. J., Griswold, M. H., and Eisenberg, H.: An Interpretation of Sure-it-al Rates: Cancer of the Breast, J. Nat. Cancer Inst. 19:1107, 1957. Greenwood, M.: Reports on Public Health and Medical Subjects, No. 33, .L\ppcndis 1, The Errors of Sampling of the Survivorship Tables, London, 1926, H. M. Stationer) Office.

1923, W. B. Saunders Company. Hill, A. B.: Principles of Medical Statistics.

APPENDIX Contputing the Standnud Error of the 5-Ymv Snrvhd Kate.-The method for con,puting the standard error of the S-year survival rate was developed by Greellwood (see ref. 12) and is also described by Merrell and Shulman (see ref. 5). The formula is

Sa =

rq J

5 9

qx d, = i

-__

(II dl + rate. ;-

(12 __ d?

. + A__,
1; the standard d; error of the

-__ x = 1 l,

\ If1 -

where sg is the standard k-year survival rate is

error of the S-year survival

In general,

Columns 10 and 11 of Appendix Table I show how the calculation of the standard error of the S-year survival rate is carried out as a continuation of the computation of the sur\-ival rate.* The first 9 columns are a replica of Table III. (1) Subtract d, from I, for each line (Column 10); (2) divide g, by I, - d, for each line (Column 11); (3) total the entries in the first 5 lines of Col__--_
*The standard error computed in this illustratkxl is, itself, only ai11 estimate of the true standard wror. For PXAnd, since it is based on relatively small numbers of cases, it is not a very reliable estimate. ample, had there been, due to sampling variation, one death in the last interval, rather than rmne, the computed standard error would be 0.0216 rather than 0.0187.

;~PPEKDIX TABLE I.

COMPUTATION OF THE ~-YEAR SURVIVAL RATE AND ITS STANDARD ERROR. (Data from Table III)

WITHNUMBEI R TO THE PROPORTION SURtIVING NOSIS OF INTERVAL THROUGH FROM SURVIVING DYING (COL. 1 PROPORTION DIAGEND DRAWh EXPOSED RISK OF DYING ALIVE DURIN(

C UMULATIVE PROPORTIOr J

LOST

TO

F:FFECTI\E J

DIED

FOLLOW-

ALIVE AT BEGIN-

DURING

UP DUR-

(COL. 6COL.

(COL.

7f 3)
Px) I,-d,
COL. 10)

YEARS AFTER DIAGNOSIS

NING

OF

IN-

ING IN-

2 - yi COL. 1 2 (1- car.. 7) (8)

PX (PI x p2 x . r) x

INTERVAL

TER\AL

TERVAL

; ( COL. I NTERVP LL -

w
COL. 5) (6) I,

. .x

(1) XTOX+

(10) _____-

(4) Ur

(5) WX

3t COL. 6) (7) qx

q x/I .-dx

(11)

o-1

47 5 15 11 15 7 6 4

1-2 2-3 3-4 4-S 5-61

126 60 38 21 10 4 116.5 51.5 30.5 16.5 7.0 0.40 0.10 0.07 0.12 0.00 0.60 0.90 0.93 0.88 1.00

; -

it 2 -

0.60 0.54 0.50 0.44 0.44*

69.5 46.5 28.5 14.5 7.0

0.0058 0.0022 0.0024 0.0083 0.0000

0.0187t

*Five-year survival rate. tThis is the sum of the five entries in Column 11. The square root of this number, when multiplied by the 5-year survival rate, yields the standard error of the 5-year survival rate: s = (0.44) 0.0187 = (0.44) (0.37) = 0.060. $See footnote *, Table III.

712

CUTLER

AND

EDERER

J. Chron. Dis. December, 1958

Ulll 11: 0.0187; (4) take the square root of this number: 40.0187 = 0.137; (5) multiply the result by Ps: 0.137 X 0.44 = 0.060. This is the standard error of the S-year survival rate. The standard error of survival rates for end-points other than 5 years is computed similarly. For example, to compute the standard error of the 3.year survival rate, the first three entries in Column 11 must be totaled, the square root taken, and multiplied by P,. Effective Sample S&.-The concept, effectiue sample size, provides another way of assessing the benefit of including in the life table cases with partial survival information. The concept relates to the fact that the reliability of a statistical result depends on the size of the sample, i.e., the number of cases observed. For example, the standard error of a survival rate, P, when all cases have been followed until death or for the required time interval (i.e., no losses from observation or withdrawals alive prior to the cut-off date) is given by the binomial formula

In formula where 1, is the sample size, i.e., the initial number of cases. is inversely proportional to the square root of the sample size.
;IPPENDIS r.lI3LE I I

(l),

the standard

error

SAMPLE

SIZE

1946-1951 COHORT* Kidney, localized Kidney, regional Breast, localized Breast, regional Lip
*Since 5 years tActua1 the cut-off date was Dec. 31, for 1951,

EFFECTIVE SAMPLE SIZE

1946
COHORTj

126 1,4% 1,531 332

cases diagnosed

68 37 516 595 14.5

in 1947 or later were eligible

9 11 225 208 61
for less Lhan

of observation. number of cases rligihla 5 years of observation.

Let us consider the 1946-1951 localized kidney cancer cohort (.\ppendix Table I), for which the survival rate is 0.44, and its standard error, 0.060. Of the initial 126 cases in this cohort, IVe now ask how a substantial number were withdrawn alive less than 5 years after diagnosis. large a cohort, with a S-year survival rate of 0.44 and with all cases followed to death or for a full 5 years, would have a standard error equal to 0.060. To answer this question, we solve equation (1) for I,, placing a circumflex over the I, to indicate that this is a hypothetical value:

Substituting

P = 0.44 and s = 0.060, we obtain

Had we started with The result, 68, is the eyective sample size, which we interpret as follows. about 68 cases (instead of 126) and followed them all until death or survival for 5 years and found that 44 per cent survived 5 years, then the standard error would have been equal to that we actually obtained in our cohort of 126 cases. Thus, the survival rate we obtained is as reliable as one based on 68 cases. This is in sharp contrast to 9 cases which were eligible for 5 years of observation. These three values are compared for the five cancer groups discussed in the text. In each instance, the effective sample size based on the 1946-1951 cohort is substantially larger than the number of cases eligible for 5 years of observation (1946 cohort).