You are on page 1of 23

Survival Function

Jerry Dwi Trijoyo Purnomo

INSTITUT TEKNOLOGI SEPULUH NOPEMBER (ITS)


Surabaya - Indonesia

www.its.ac.id
Review
Start Event
Event: Death, disease, relapse
Time = survival time
Event = failure
Censoring = don’t know survival time exactly

2
Notation

3
The Estimated Median

4
Data
• Suppose we have the observed data under right censored:
1.3 1.5+ 1.6 1.8+ 3.2
• We first treat this data as if there were NO censored observations. Let ti denote
an ordered observed value. The empirical survivor function (esf), denoted by
S n (t ) is defined to be

# of observations > t # {ti > t}


S n (t ) = =
n n

5
The Mean Survival Time (1/2)
• The mean in survival is defined as

E (t ) = ∫ S n (t ) dt = area under S n (t )
0

• The esf is a consistent estimator of the true survivor function S(t). The exact
distribution of nSn(t), for each fixed t, is binomial (n, p), where n = the number of
observations and p = P(T > t). Further, it follows from
the central limit theorem that for each fixed t,

S n (t ) ≈ N ( p, p(1 − p ) n )

6
The Mean Survival Time (2/2)
• Mean in survival is not important property like as common r.v. since:
a. not robust (survival curve often has high skew)
b. tail information for survival data may be incomplete (we obtain mean
from incomplete data due the data are censored)

7
Kaplan-Meier Estimator of Survival (1/5)

• We now present the product-limit estimator of survival. This is commonly called


the Kaplan-Meier (K-M) estimator as it appeared in a seminal 1958 paper.
• The Product-limit (PL) estimator of S(t) = P(T > t):
K-M adjusts the esf to reflect the presence of right-censored observations.
• Recall the random right censoring model. On each of n individuals we observe the
pair (Yi, δi) where
1 if Ti ≤ Ci
Yi = min (Ti , Ci ) and δ i = 
0 if Ci < Ti

8
Kaplan-Meier Estimator of Survival (2/5)
• On a time line we have

• where y(i) denotes the ith distinct ordered censored or uncensored observation
and is the right endpoint of the interval I i , i = 1, 2, . . . , n' ≤ n.

9
Kaplan-Meier Estimator of Survival (3/5)

• The people at risk at the beginning of the interval I i are those people who
survived (not dead, lost, or withdrawn) the previous interval I i −1. Let R(t) denote
the risk set just before time t and let
ni = # in R(y(i))
= # alive (and not censored) just before y(i)
di = # died at time y(i)
pi = P(surviving through I i| alive at beginning I i )
= P(T > y(i) | T > y(i−1))
qi = 1− pi = P(die in I i| alive at beginning I i ).

10
Kaplan-Meier Estimator of Survival (4/5)

• Recall the general multiplication rule for joint events A1 and A2 :


P(A1 ∩ A2) = P(A2 | A1)P(A1)
• From repeated application of this product rule the survivor function can
be expressed as
S (t ) = P(T > t ) = ∏p i
y( i ) ≤t

11
Kaplan-Meier Estimator of Survival (5/5)

• The estimates of pi and qi are


di d i  ni − d i 
qˆi = and pˆ i = 1 − qˆ i = 1 − =  
ni ni  ni 

• The K-M estimator of the survivor function is

 −  k
 −  k
 di 
S (t ) = ∏ pˆ i = ∏ 
n d n d
ˆ i i
 = ∏  i i
 = ∏ 1 − 
y( i ) ≤t y( i ) ≤t  ni  i =1  ni  i =1  ni 

where y(k) ≤ t < y(k+1).

12
Estimate of Variance
• The estimate of Ŝ (t ) :
Greenwood’s formula (1926):

( )
k
var S (t ) = S (t ) ∑ = S (t )∑
ˆ ˆ 2 d i ˆ 2 di
y( i ) ≤t ni (ni − d i ) i =1 ni (ni − d i )

where y(k) ≤ t < y(k+1).


• The theory tells us that for each fixed value t
( ( ))
Sˆ (t ) ≈ N S (t ), var Sˆ (t )

13
(1-α) x 100% CI for S(t)
• Thus, at time t, an approximate (1 − α) × 100% confidence interval
for the probability of survival, S(t) = P(T > t), is given by

( )
Sˆ (t ) ± Zα 2 × s.e. Sˆ (t )
( )
where s.e. Sˆ (t ) is the square root of Greenwood’s formula for the
estimated variance.

14
Compute The K-M Estimator (1/3)

Product-Limit
Number Number Estimator
Time di
Sˆ 2 (t )∑
di
t
of Event at Risk  d  ∑
ti ≤t ni (ni − d i ) ni (ni − d i )
di ni Sˆ (t ) = ∏ 1 − i  ti ≤t

ti ≤t  ni 
0 1
 1 1
1.3 1 5 1 −  = 0.80 = 0.05 0.82 x 0.05 = 0.032
 5 5× 4
 1 1
1.6 1 3 (0.8) 1 −  = 0.53 0.05 + = 0.22 0.532 x 0.22 = 0.06
 3 3× 2

3.2 1 1 (0.53)1 − 1  = 0
 1

15
Compute The K-M Estimator (2/3)
Based on the table:
a. K-M estimator of P(T>t) = S(t) for t = 1.3, 1.6, 3.2:
Sˆ (1.3) = 0.8 ; t = 1.3
Sˆ (1.6 ) = 0.53 ; t = 1.6
Sˆ (3.2 ) = 0 ; t = 3.2

( )
b. var Sˆ KM (t ) for t = 1.3, 1.6 using Greenwood’s formula:
( )
var Sˆ (t ) = 0.032 ; for t = 1.3
( )
var Sˆ (t ) = 0.063 ; for t = 1.6

16
Compute The K-M Estimator (3/3)

c. The estimation of E(T) using the result in (a): µˆ = ∫ Sˆ (t )dt
µ̂ is equal to the area under the estimated survivorship function. Thus, if the times
0

to death are order as t (1) ≤ t ( 2 ) ≤  ≤ t ( m ) (if there are m uncensored observation


and t is the largest observation, then:
(m )

µˆ = 1.00t (1) + Sˆ (t (1) )(t ( 2) − t (1) ) +  + Sˆ (t ( m −1) )(t ( m ) − t ( m −1) )


= 1.00 (1.3) + 0.8 (1.6 − 1.3) + 0.53 (3.2 − 1.6)
= 1.30 + 0.24 + 0.85
= 2.39

(Lee, 2003)

17
Leukemia Data
Compare 2 treatments: 6-MP and placebo

6-MP: 6, 6, 6, 7, 10, 13, 16, 22, 23, 6+, 9+, 10+, 11+, 17+, 19+, 20+, 25+, 32+, 32+,
34+, 35+
Placebo: 1,1, 2, 2, 3, 4, 4, 5, 5, 8, 8, 8, 8, 11, 11, 12, 12, 15, 17, 22, 23

18
Compute 6-MP

6-MP
 d  di di
t di ni Sˆ (t ) = ∏ 1 − i  ∑ n (n Sˆ 2 (t )∑
ti ≤t  ni  ti ≤t i i − di ) ti ≤t ni (ni − d i )
0 1
6 3 21 0.857 0.018 0.013
7 1 17 0.807 0.022 0.014
10 1 15 0.753 0.027 0.015
13 1 12 0.690 0.035 0.017
16 1 11 0.627 0.044 0.017
22 1 7 0.538 0.068 0.020
23 1 6 0.448 0.101 0.020

19
K-M Curve for 6-MP

1.0
survival probability

0.8
0.6
0.4
0.2
0.0

0 5 10 15 20 25 30 35

survival time (in days)

Figure 1. Kaplan-Meier Curve for 6-MP


20
Compute Placebo
Placebo
 d  di di
t di ni Sˆ (t ) = ∏ 1 − i  ∑ Sˆ 2 (t )∑
ti ≤t  ni  ti ≤t ni (ni − d i ) ti ≤t ni (ni − d i )
0 1
1 2 21 0.905 0.005 0.004
2 2 19 0.810 0.011 0.007
3 1 17 0.762 0.015 0.009
4 2 16 0.667 0.024 0.011
5 2 14 0.571 0.036 0.012
8 4 12 0.381 0.078 0.011
11 2 8 0.286 0.200 0.016
12 2 6 0.190 0.283 0.010
15 1 4 0.143 0.366 0.007
17 1 3 0.095 0.533 0.005
22 1 2 0.048 1.033 0.002
23 1 1 0
21
K-M Curve for Placebo

1.0
survival probability

0.8
0.6
0.4
0.2
0.0

0 5 10 15 20

survival time (in days)

Figure 2. Kaplan-Meier Curve for Placebo


22
K-M Curve for Leukemia Data

Figure 3. Kaplan-Meier Curve for both 6-MP and Placebo


23

You might also like