Lecture 2

ESTIMATING THE SURVIVAL

FUNCTION

— One-sample nonparametric methods

There are commonly three methods for estimating a sur-

vivorship function
S(t) = P (T > t)

without resorting to parametric models:

(1) Kaplan-Meier

(2) Nelson-Aalen or Fleming-Harrington (via esti-

mating the cumulative hazard)

(3) Life-table (Actuarial Estimator)

We will mainly consider the first two.

1
 (1) The Kaplan-Meier Estimator

The Kaplan-Meier (or KM) estimator is probably the most

popular approach.

Motivation (no censoring):

Remission times (weeks) for 21 leukemia patients receiving
control treatment (Table 1.1 of Cox & Oakes; Freireich et
al., 1963):
1, 1, 2, 2, 3, 4, 4, 5, 5, 8, 8, 8, 8, 11, 11, 12, 12, 15, 17, 22, 23

We estimate S(10), the probability that an individual sur-

8
vives to week 10 or later, by 21 .

How would you calculate the standard error of the estimated

survival?
8
Ŝ(10) = P̂ (T > 10) = = 0.381
21
(Answer: se[Ŝ(10)] = 0.106)

12 8
What about Ŝ(8)? Is it 21 or 21 ?

2
A table of Ŝ(t):
Values of t Ŝ(t)
t < 1 21/21=1.000
1 ≤ t < 2 19/21=0.905
2 ≤ t < 3 17/21=0.809
3≤t<4
4≤t<5
5≤t<8
8 ≤ t < 11
11 ≤ t < 12
12 ≤ t < 15
15 ≤ t < 17
17 ≤ t < 22
22 ≤ t < 23

In most software packages, the survival function is evaluated

just after time t, i.e., at t+. In this case, we only count the
individuals with T > t.

3
 1.0
0.8
0.6
Survival
0.4
0.2
0.0

0 5 10 15 20 25
Time

Figure 1: Example for leukemia data (control arm)

4
Empirical Survival Function:

When there is no censoring, the general formula is:

Pn
# individuals with T > t I(Ti > t)
Sn(t) = = i=1
total sample size n

Note that Fn(t) = 1 − Sn(t) is the empirical CDF.

Also I(Ti > t) ∼ Bernoulli(S(t)), so that

1. Sn(t) converges in probability to S(t) (consistency);

√
2. n{Sn(t) − S(t)} → N (0, S(t)[1 − S(t)]) in distribu-
tion.

[Make sure that you know these.]

5
What if there is censoring?

Consider the treated group from Table 1.1 of Cox and Oakes:

6, 6, 6, 6+, 7, 9+, 10, 10+, 11+, 13, 16, 17+

19+, 20+, 22, 23, 25+, 32+, 32+, 34+, 35+

+
[Note: times with are right censored]

We know S(5)= 21/21, because everyone survived at least

until week 5 or greater. But, we can’t say S(7) = 17/21,
because we don’t know the status of the person who was
censored at time 6.

In a 1958 paper in the Journal of the American Statistical

Association, Kaplan and Meier proposed a way to estimate
S(t) nonparametrically, even in the presence of censoring.
The method is based on the ideas of conditional proba-
bility.

6
[Reading:]
A quick review of conditional probability

Conditional Probability: Suppose A and B are two

events. Then,
P (A ∩ B)
P (A|B) =
P (B)

Multiplication law of probability: can be obtained

from the above relationship, by multiplying both sides by
P (B):
P (A ∩ B) = P (A|B) P (B)

Extension to more than 2 events:

Suppose A1, A2...Ak are k different events. Then, the prob-
ability of all k events happening together can be written as
a product of conditional probabilities:

P (A1 ∩ A2... ∩ Ak ) = P (Ak |Ak−1 ∩ ... ∩ A1) ×

×P (Ak−1|Ak−2 ∩ ... ∩ A1)
...
×P (A2|A1)
×P (A1)

7
Now, let’s apply these ideas to estimate S(t):

– Intuition behind the Kaplan-Meier Estimator

Think of dividing the observed timespan of the study into a

series of fine intervals so that there is a separate interval for
each time of death or censoring:

D C C D D D

Using the law of conditional probability,

Y
P (T > t) = P (survive j-th interval Ij | survived to start of Ij )
j
Y
= λj
j

where the product is taken over all the intervals preceding

time t.

8
4 possibilities for each interval:
(1) No death or censoring - conditional probability of
surviving the interval is estimated to be 1;

(2) Censoring - assume they survive to the end of the

interval (the intervals are very small), so that the condi-
tional probability of surviving the interval is again esti-
mated to be 1;

(3) Death, but no censoring - conditional probability

of not surviving the interval is estimated by # deaths
(d) divided by # ‘at risk’ (r) at the beginning of the
interval. So the estimated conditional probability of sur-
viving the interval is 1 − d/r;

(4) Tied deaths and censoring - assume censorings last

to the end of the interval, so that the estimated condi-
tional probability of surviving the interval is still 1−d/r.

General Formula for jth interval:

It turns out we can write a general formula for the estimated
conditional probability of surviving the j-th interval that
holds for all 4 cases:
dj
1−
rj

9
Here as the intervals get finer and finer, the approximations
made in estimating the probabilities of getting through each
interval become more and more accurate, at the end the
estimator converges to the true S(t) in probability (proof
not shown here).

This intuition explains why an alternative name for the KM

is the product-limit estimator.

10
 The Kaplan-Meier estimator of the survivorship
function (or survival probability) S(t) = P (T > t)
is:
Q rj −dj
Ŝ(t) = j:τj ≤t rj
 
Q dj
= j:τj ≤t 1− rj

where
• τ1, ...τK is the set of K distinct uncensored failure times
observed in the sample
• dj is the number of failures at τj
• rj is the number of individuals “at risk” right before
the j-th failure time (everyone who died or censored
at or after that time).

Furthermore, let cj be the number of censored observations

between the j-th and (j + 1)-st failure times. Any censoring
tied at τj are included in cj , but not censorings tied at τj+1.

Note: two useful formulas are:

(1) rj = rj−1 − dj−1 − cj−1
X
(2) rj = (cl + dl )
l≥j

11
Calculating the KM - leukemia treated group

Make a table with a row for every death or censoring time:

τj dj cj rj 1 − (dj /rj ) Ŝ(τj )

18
6 3 1 21 21 = 0.857
7 1 0 17
9 0 1 16
10
11
13
16
17
19
20
22
23

Note that:

• Ŝ(t) only changes at death (failure) times;

• Ŝ(t) is 1 up to the first death time;
• Ŝ(t) only goes to 0 if the last observation is uncensored;
• When there is no censoring, the KM estimator equals
the empirical survival estimate

12
 1.0
0.8
0.6
Survival
0.4
0.2
0.0

0 5 10 15 20 25 30 35
Time

Figure 2: KM plot for treated leukemia patients

Output from a software KM Estimator:

failure time: weeks

failure/censor: remiss

Beg. Net Survivor Std.

Time Total Fail Lost Function Error [95% Conf. Int.]
-------------------------------------------------------------------
6 21 3 1 0.8571 0.0764 0.6197 0.9516
7 17 1 0 0.8067 0.0869 0.5631 0.9228
9 16 0 1 0.8067 0.0869 0.5631 0.9228
10 15 1 1 0.7529 0.0963 0.5032 0.8894
11 13 0 1 0.7529 0.0963 0.5032 0.8894
13 12 1 0 0.6902 0.1068 0.4316 0.8491
16 11 1 0 0.6275 0.1141 0.3675 0.8049

13
 17 10 0 1 0.6275 0.1141 0.3675 0.8049
19 9 0 1 0.6275 0.1141 0.3675 0.8049
20 8 0 1 0.6275 0.1141 0.3675 0.8049
22 7 1 0 0.5378 0.1282 0.2678 0.7468
23 6 1 0 0.4482 0.1346 0.1881 0.6801
25 5 0 1 0.4482 0.1346 0.1881 0.6801
32 4 0 2 0.4482 0.1346 0.1881 0.6801
34 2 0 1 0.4482 0.1346 0.1881 0.6801
35 1 0 1 0.4482 0.1346 0.1881 0.6801

(Note: the above is from Stata, but most software output

similar KM tables.)

14
[Reading:] Redistribution to the right algorithm
(Efron, 1967)

There is another way to compute the KM estimator.

In the absence of censoring, Ŝ(t) is just the proportion of

individuals with T > t. The idea behind Efron’s approach
is to spread the contributions of censored observations out
over all the possible times to their right.

Algorithm:
• Step (1): arrange the n observation times (failures or
censorings) in increasing order. If there are ties, put
censored after failures.
• Step (2): Assign weight (1/n) to each time.
• Step (3): Moving from left to right, each time you en-
counter a censored observation, distribute its mass to all
times to its right.
• Step (4): Calculate Ŝj by subtracting the final weight
for time j from Ŝj−1

15
Example of “redistribute to the right” algorithm

Consider the following event times:

2, 2.5+, 3, 3, 4, 4.5+, 5, 6, 7

The algorithm goes as follows:

(Step 1) (Step 4)
Times Step 2 Step 3a Step 3b Ŝ(τj )
2 1/9=0.11 0.889
2.5+ 1/9=0.11 0 0.889
3 2/9=0.22 0.25 0.635
4 1/9=0.11 0.13 0.508
4.5+ 1/9=0.11 0.13 0 0.508
5 1/9=0.11 0.13 0.17 0.339
6 1/9=0.11 0.13 0.17 0.169
7 1/9=0.11 0.13 0.17 0.000

This comes out the same as the product-limit approach.

16
Properties of the KM estimator

When there is no censoring, KM estimator is the

same as the empirical estimator:
# deaths after time t
Ŝ(t) =
n
where n is the number of individuals in the study.

As said before
asymp.
Ŝ(t) ∼ N (S(t), S(t)[1 − S(t)]/n)

How does censoring affect this?

• Ŝ(t) is still consistent for the true S(t);
• Ŝ(t) is still asymptotically normal;
• The variance is more complicated.

The proofs can be done using the usual method (by writing
as sum of i.i.d. terms plus op(1)) but it is laborious, or it
can be done using counting processes which was considered
more elegant, or by empirical processes method which is
very powerful for semiparametric inferences.

17
 The KM estimator is also an MLE

You can read in Cox and Oakes book Section 4.2.

Here we need to think of the distribution function F (t) as an

(infinite dimensional) parameter, and we try to find the F̂ (or
Ŝ = 1 − F̂ ) that maximizes a nonparametric likelihood.
Such a MLE is called a NPMLE.

As it turns out, such a F̂ (t) has to be discrete in order

to for the likelihood to be bounded (otherwise MLE does
not exist), with masses only at the uncensored failure times
{aj }j . (aj = τj in our previous notation)

Ex. Show the following for an absolutely continuous distri-

bution function F(t): a) for a random sample of size n from
F(t), if we allow an estimator of F(t) to be absolutely con-
tinuous (i.e. having a positive density) on any open inter-
val covering the observed data point(s), then the likelihood
function can be made arbitrarily large; b) if we restrict to
those estimates of F(t) such that the likelihood function is
bounded, the maximum of the likelihood is achieved when
we assign point mass of 1/n to each observed data point.
This shows that the empirical CDF is the NPMLE.

18
Cox and Oakes book Section 4.2 (please read if you have
time) shows that the right-censored data likelihood for such
a discrete distribution can be written as
g
d
Y
L(λ) = λj j (1 − λj )rj −dj
j=1

where λj is the discrete hazard (i.e. conditional probability)

at aj . Note that the likelihood is the same as that of g
independent binomials.

Therefore, the maximum likelihood estimator of λj is (why):

λ̂j = dj /rj

For a discrete survival distribution

Y
S(t) = (1 − λj )
j:aj ≤t

Now we plug in the MLE’s of λ to estimate S(t) (why):

Y
Ŝ(t) = (1 − λ̂j )
j:aj ≤t

Y  dj

= 1−
j:aj ≤t
rj

This is the NPMLE of S.

19
One can often show that an NPMLE behaves like a classic
MLE:
• consistent for the true parameter (function);
• asymptotically normal (converges in distribution to a
Gaussian process).

For a semparametric model (that we’ll talk about later),

it is often semiparametrically efficient.

20
 Greenwood’s formula for variance

Note that the KM estimator is

Y
Ŝ(t) = (1 − λ̂j )
j:τj ≤t

where λ̂j = dj /rj .

Since the λ̂j ’s are just binomial proportions given rj ’s, then
λ̂j (1 − λ̂j )
Var(λ̂j ) ≈
rj
Also, the λ̂j ’s are asymptotically independent.

Since Ŝ(t) is a function of the λj ’s, we can estimate its vari-

ance using the Delta method.

Delta method: If Yn is (asymptotically) normal

with mean µ and variance σ 2, g is differentiable and
g 0(µ) 6= 0, then g(Yn) is approximately normally
distributed with mean g(µ) and variance [g 0(µ)]2σ 2.

21
Greenwood’s formula (continued)

Instead of dealing with Ŝ(t) directly, we will look at its log

(why?): X
log[Ŝ(t)] = log(1 − λ̂j )
j:τj ≤t

Thus, by approximate independence of the λ̂j ’s,

X
Var(log[Ŝ(t)]) =
c Var[log(1
c − λ̂j )]
j:τj ≤t
!2
X 1
= Var(
c λ̂j )
j:τj ≤t 1 − λ̂j
!2
X 1 λ̂j (1 − λ̂j )
=
j:τj ≤t 1 − λ̂j rj
X λ̂j
=
j:τj ≤t (1 − λ̂j )rj
X dj
=
j:τj
(r − dj )rj
≤t j

Now, Ŝ(t) = exp[log[Ŝ(t)]]. Using Delta method once again,

h i
2
c Ŝ(t)) = [Ŝ(t)] · Var
Var( c log[Ŝ(t)]

22
Greenwood’s Formula:
2
P dj
Var(Ŝ(t)) = [Ŝ(t)]
c
j:τj ≤t (rj −dj )rj

23
Confidence intervals

For a 95% confidence interval, we could use

Ŝ(t) ± z1−α/2 se[Ŝ(t)]
where se[Ŝ(t)] is calculated using Greenwood’s formula. Here
t is fixed, and this is refered to as pointwise confidence in-
terval.

Problem: This approach can yield values > 1 or < 0.

Better approach: Get a 95% confidence interval for

L(t) = log(− log(S(t)))
Since this quantity is unrestricted, the confidence interval
will be in the right range when we transform back:
S(t) = exp(− exp(L(t)).

[ To see why this works:

0 ≤ Ŝ(t) ≤ 1
−∞ ≤ log[Ŝ(t)] ≤ 0
0 ≤ − log[Ŝ(t)] ≤ ∞
h i
−∞ ≤ log − log[Ŝ(t)] ≤ ∞
]

24
[Read] Log-log Approach for Confidence Intervals:

(1) Define L(t) = log(− log(S(t)))

(2) Form a 95% confidence interval for L(t) based on L̂(t),

yielding [L̂(t) − A, L̂(t) + A]

(3) Since S(t) = exp(− exp(L(t)), the confidence bounds

for the 95% CI of S(t) are:
h i
L̂(t)+A L̂(t)−A
exp{−e }, exp{−e }

(note that the upper and lower bounds switch)

(4) Substituting L̂(t) = log(− log(Ŝ(t))) back into the above

bounds, we get confidence bounds of
 
eA e−A
[Ŝ(t)] , [Ŝ(t)]

25
What is A?

• A is 1.96 · se(L̂(t))
• To calculate this, we need to calculate
h i
Var(L̂(t)) = Var log(− log(Ŝ(t)))

• From our previous calculations, we know

X dj
Var(log[Ŝ(t)]) =
c
(r − dj )rj
j:τ ≤t j
j

• Applying the delta method again, we get:

Var(
c L̂(t)) = Var(log(−
c log[Ŝ(t)]))
1 X dj
=
[log Ŝ(t)]2 j:τ ≤t (rj − dj )rj
j

• We take the square root of the above to get se(L̂(t)),

26
 1.0
0.8
0.6
Survival
0.4
0.2
0.0

0 5 10 15 20 25 30 35
Time

Figure 3: KM Survival Estimate and Confidence intervals (type ‘log-log’)

• Different software might use different approaches to cal-

culate the CI.

Software for Kaplan-Meier Curves

• Stata - stset and sts commands

• SAS - proc lifetest
• R - survfit() in ‘survival’ package

27
R allows different types of CI (some are truncated to be
between 0 and 1):

95 percent confidence interval is of type "log"

time n.risk n.event survival std.dev lower 95% CI upper 95% CI
6 21 3 0.8571429 0.07636035 0.7198171 1.0000000
7 17 1 0.8067227 0.08693529 0.6531242 0.9964437
10 15 1 0.7529412 0.09634965 0.5859190 0.9675748
13 12 1 0.6901961 0.10681471 0.5096131 0.9347692
16 11 1 0.6274510 0.11405387 0.4393939 0.8959949
22 7 1 0.5378151 0.12823375 0.3370366 0.8582008
23 6 1 0.4481793 0.13459146 0.2487882 0.8073720

95 percent confidence interval is of type "log-log"

time n.risk n.event survival std.dev lower 95% CI upper 95% CI
6 21 3 0.8571429 0.07636035 0.6197180 0.9515517
7 17 1 0.8067227 0.08693529 0.5631466 0.9228090
10 15 1 0.7529412 0.09634965 0.5031995 0.8893618
13 12 1 0.6901961 0.10681471 0.4316102 0.8490660
16 11 1 0.6274510 0.11405387 0.3675109 0.8049122
22 7 1 0.5378151 0.12823375 0.2677789 0.7467907
23 6 1 0.4481793 0.13459146 0.1880520 0.6801426

95 percent confidence interval is of type "plain"

time n.risk n.event survival std.dev lower 95% CI upper 95% CI
6 21 3 0.8571429 0.07636035 0.7074793 1.0000000
7 17 1 0.8067227 0.08693529 0.6363327 0.9771127
10 15 1 0.7529412 0.09634965 0.5640993 0.9417830
13 12 1 0.6901961 0.10681471 0.4808431 0.8995491
16 11 1 0.6274510 0.11405387 0.4039095 0.8509924
22 7 1 0.5378151 0.12823375 0.2864816 0.7891487
23 6 1 0.4481793 0.13459146 0.1843849 0.7119737

28
Mean, Median, Quantiles based on the KM

• Mean is not well estimated with censored data, since we

often don’t observe the right tail.

• Median - by definition, this is the time, τ , such that

S(τ ) = 0.5. In practice, it is often defined as the small-
est time such that Ŝ(τ ) ≤ 0.5. The median is more
appropriate for censored survival data than the mean.

For the treated leukemia patients, we find:

Ŝ(22) = 0.5378
Ŝ(23) = 0.4482
The median is thus 23.

• Lower quartile (25th percentile):

the smallest time (LQ) such that Ŝ(LQ) ≤ 0.75

• Upper quartile (75th percentile):

the smallest time (UQ) such that Ŝ(U Q) ≤ 0.25

From class discussion: 1) how do we bootstrap survival data?

2) how do we estimate the distribution of C? Recall that
the observed data is (Xi, δi), where Xi = min(Ti, Ci), δi =
I(Ti ≤ Ci), i = 1, ..., n.
29
 Left truncated KM estimate

y
= censored observation
E = event

30
When there is left truncation, the observed data is (Qi, Xi, δi),
i = 1, ..., n.

Now the ‘risk set’ at any time t consists of subjects who have
entered the study, and have not failed or been censored by
that time, i.e. {i : Qi < t ≤ Xi}.

Pn
So rj = i=1 I(Qi < τj ≤ Xi).

We still have
Y  dj

Ŝ(t) = 1−
j:τ ≤t
rj
j

where τ1, ...τK is the set of K distinct uncensored failure

times observed in the sample, dj is the number of failures at
τj , and rj is the number of individuals “at risk” right before
the j-th failure time (everyone who had entered and
who died or censored at or after that time).

Read [required] Tsai, Jewell and Wang (1987).

31
• When minni=1 Qi = t0 > 0, then the KM estimates
P (T > t|T > t0).
• The left truncated KM is still an NPMLE (Wang, 1991).

• The Greenwood’s formula for variance still applies.

• In R (and most other softwares) it is handled by some-

thing like ‘Surv(time=Q, time2=X, event=δ)’.

• This approach is referred to as conditional inference, i.e. con-

ditional on the Q’s. This is implemented in most survival
software.

• Another approach is the unconditional inference, where

assumptions are made about the distribution of the Q’s;
eg. uniform, in which case this is referred to as ‘length
biased’.

32
 (2) Nelson-Aalen (Fleming-Harrington)
estimator
– Estimating the cumulative hazard
Rt
If we can estimate Λ(t) = 0 λ(u)du, the cumulative hazard
at time t, then we can estimate S(t) = e−Λ(t).

Just as we did for the KM, think of dividing the observed

time span of the study into a series of fine intervals so that
there is only one event per interval:

D C C D D D

Λ(t) can then be approximated by a sum:

X
Λ(t) ≈ λ̃j · ∆j
j:τj ≤t

where the sum is over intervals up to t, λ̃j is the value of

the hazard in the j-th interval and ∆j is the width of that
interval.

Since λ̃∆ is approximately the conditional probability of dy-

ing in the interval, we can further estimate λ̃j · ∆j by dj /rj .

33
This gives the Nelson-Aalen estimator:
X
Λ̂N A(t) = dj /rj .
j:τj ≤t

It follows that Λ̂(t), like the KM, changes only at the ob-
served death (event) times.

Example:
D C C D D D
rj n n n n-1 n-1 n-2 n-2 n-3 n-4
dj 0 0 1 0 0 0 0 1 1
cj 0 0 0 0 1 0 1 0 0
1 1
λ̂(tj )∆ 0 0 1/n 0 0 0 0 n−3 n−4

Λ̂(tj ) 0 0 1/n 1/n 1/n 1/n 1/n

Once we have Λ̂N A(t), we can obtain the Fleming-Harrington

estimator of S(t):
ŜF H (t) = exp(−Λ̂N A(t)).

34
In general, the FH estimator of the survival function should
be close to the Kaplan-Meier estimator, ŜKM (t).

We can compare the Fleming-Harrington survival estimate

to the KM estimate using a subgroup of the nursing home
data:

skm sfh
1. .91666667 .9200444
2. .83333333 .8400932
3. .75 .7601478
4. .66666667 .6802101
5. .58333333 .6002833
6. .5 .5203723
7. .41666667 .4404857
8. .33333333 .3606392
9. .25 .2808661
10. .16666667 .2012493
11. .08333333 .1220639
12. 0 .0449048

In this example, it looks like the Fleming-Harrington estima-

tor is slightly higher than the KM at every time point, but
with larger datasets the two will typically be much closer.

Question: do you think that the two estimators are asymp-

totically equivalent?
35
Note: We can also go the other way: we can take the
Kaplan-Meier estimate of S(t), and use it to calculate an
alternative estimate of the cumulative hazard function:
Λ̂KM (t) = − log ŜKM (t)

36
 [Reading:]
(3) The Lifetable or Actuarial Estimator

• one of the oldest techniques around

• used by actuaries, demographers, etc.

• applies when the data are grouped

Our goal is still to estimate the survival function, hazard,

and density function, but this is sometimes complicated by
the fact that we don’t know exactly when during a time
interval an event occurs.

37
There are several types of lifetable methods according to the
data sources:

Population Life Tables

• cohort life table - describes the mortality experience

from birth to death for a particular cohort of people born
at about the same time. People at risk at the start of the
interval are those who survived the previous interval.
• current life table - constructed from (1) census infor-
mation on the number of individuals alive at each age,
for a given year and (2) vital statistics on the number
of deaths or failures in a given year, by age. This type
of lifetable is often reported in terms of a hypothetical
cohort of 100,000 people.

Generally, censoring is not an issue for Population Life Ta-

bles.

Clinical Life tables - applies to grouped survival data

from studies in patients with specific diseases. Because pa-
tients can enter the study at different times, or be lost to
follow-up, censoring must be allowed.

38
Notation

• the j-th time interval is [tj−1, tj )

• cj - the number of censorings in the j-th interval
• dj - the number of failures in the j-th interval
• rj is the number entering the interval

Example: 2418 Males with Angina Pectoris (chest pain,

from book by Lee, p.91)

Year after
Diagnosis j d j cj rj rj0 = rj − cj /2
[0, 1) 1 456 0 2418 2418.0
[1, 2) 2 226 39 1962 1942.5 (1962 - 39 2)
[2, 3) 3 152 22 1697 1686.0
[3, 4) 4 171 23 1523 1511.5
[4, 5) 5 135 24 1329 1317.0
[5, 6) 6 125 107 1170 1116.5
[6, 7) 7 83 133 938 871.5
etc..

39
Estimating the survivorship function

If we apply the KM formula directly to the numbers in the

table on the previous page, estimating S(t) as
Y  dj

Ŝ(t) = 1− ,
j:τ <t
rj
j

this approach is unsatisfactory for grouped data because

it treats the problem as though it were in discrete time,
with events happening only at 1 yr, 2 yr, etc. In fact, we
should try to calculate the conditional probability of dying
within the interval, given survival to the beginning of it.

What should we do with the censored subjects?

Let rj0 denote the ‘effective’ number of subjects at risk. If we

assume that censorings occur:
• at the beginning of each interval: rj0 = rj − cj
• at the end of each interval: rj0 = rj
• on average halfway through the interval:
rj0 = rj − cj /2

The last assumption yields the Actuarial Estimator. It is

appropriate if censorings occur uniformly throughout the in-
terval.

40
Constructing the lifetable

First, some additional notation for the j-th interval, [tj−1, tj ):

• Midpoint (tmj ) - useful for plotting the density and

the hazard function
• Width (bj = tj −tj−1) needed for calculating the hazard
in the j-th interval

Quantities estimated:
• Conditional probability of dying (event)
q̂j = dj /rj0

• Conditional probability of surviving

p̂j = 1 − q̂j

• Cumulative probability of surviving at tj :

Y
Ŝ(tj ) = p̂`
`≤j
Y d`

= 1−
r`0
`≤j

41
Other quantities estimated at the
midpoint of the j-th interval:

• Hazard in the j-th interval (why)

dj
λ̂(tmj ) =
bj (rj0 − dj /2)
q̂j
=
bj (1 − q̂j /2)

• density at the midpoint of the j-th interval (why)

Ŝ(tj−1) − Ŝ(tj )
fˆ(tmj ) =
bj
Ŝ(tj−1) q̂j
=
bj
Note: Another way to get this is:
fˆ(tmj ) = λ̂(tmj )Ŝ(tmj )

= λ̂(tmj )[Ŝ(tj ) + Ŝ(tj−1)]/2

42
 Estimated Survival
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0 100 200 300 400 500 600 700 800 900 1000
Lower Limit of Time Interval

Figure 4: Life table estimate of survival

43
 Estimated hazard
0.010

0.008

0.006

0.004

0.002

0.000
0 100 200 300 400 500 600 700 800 900 1000
Lower Limit of Time Interval

Figure 5: Estimated discrete hazard

44