Product Limit Estimator
Product Limit Estimator
Product Limit Estimator
Nelson-Åalen Estimators
Abraham Weishaus
Outline
Example y r s Ŝ(y )
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 5/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 6/71
Example
You are given the results of a study as shown to
the right. d
Calculate Ŝ(7) using the Kaplan-Meier product 0 — 5
limit estimator. 0 2 —
0 — 2
y r s Ŝ(y ) 0 — 4
5 1 5 —
2 6 1 6 1 — 6
5 3 7 —
5 4 1 8 5 — 7
7 2 1 5 7 — 8
16
5
16
= 0.3125
Example
An 11-year mortality study on a group of 10 men age 80 re-
sults in the following survival times:
3 6 6 9 10
0.40515/ 11 = 0.2915
Example
An 11-year mortality study on a group of 10 men age 80 re-
sults in the following survival times:
3 6 6 9 10
Example
In a mortality study: 0.9204 1
=1−
0.9330 r13
(i) There were 80 individuals in the study at time 12.
1
(ii) There was 1 death at time 13, and no other deaths in = 0.013514
the interval [12, 13). r13
(iii) Ŝ(12) = 0.9330, Ŝ(13) = 0.9204 using the Kaplan- r13 = 74
Meier product limit estimator. Censored = 80 − 74 = 6
Determine the number of lives censored in the time inter-
val [12, 13).
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 10/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 11/71
Product-Limit Example
Version of 4-S01:4
The following are observation times:
1 2 3 4 5
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 13/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 14/71
Example
An 11-year mortality study on a group of 10 men age 80 re- y r s Ĥ(y )
1
sults in the following survival times: 3 10 1 10
= 0.1
1 2
3 6 6 9 10 6 8 2 10
+ 8
= 0.35
1
In addition, one man apiece leaves the study at times 4 and 9 5 1 0.35 + 5
= 0.55
8. The other 3 survive to the end. 1
10 4 1 0.55 + 4
= 0.80
Calculate the Nelson-Åalen estimator of the cumulative
hazard rate function. In particular, calculate Ĥ(6) = 0.35
1. Ĥ(6) Ŝ(8) = e−0.35 = 0.7047
Pr(X < 9) = 1 − e−0.35 = 0.2953
c
2. Ŝ(8)
3. Pr(X
c < 9)
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 16/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 17/71
1 2 3 4
Team 1 31 27 28 32
Team 2 29 38 27 30
Assume the two teams have the same distribution of scores.
Using the Nelson-Åalen estimator, calculate the probability
that a team will score more than 30 in a quarter.
2 1 1 1
Ĥ(30) = + + + = 0.86667
8 6 5 4
−0.86667
Ŝ(30) = e = 0.4204
1 2 3 4
Team 1 31 27 28 ??
Team 2 29 38 27 ??
It is also known that the first team scored more than 30 in
the fourth quarter.
Assume the two teams have the same distribution of scores.
Using the Nelson-Åalen estimator, calculate the probability
that a team will score more than 31 in a quarter.
2 1 1 1
Ĥ(31) = + + + = 1.23571
7 5 4 2
−1.23571
Ŝ(31) = e = 0.2906
Example
You are given the following data for a study:
180
145
110
S1 (3) = = 0.6032
y r s 200 170 140
1 200 20 20 25 30
2 170 25 H2 (3) = + + = 0.461345
3 140 30 200 170 140
−0.461345
S2 (3) = e = 0.6304
S1 (3) is the Kaplan-Meier estimate. S2 (3) is the Nelson-
Åalen estimate. 0.6032 − 0.6304 = −0.0272
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 20/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 21/71
Example 3
≈ 0.1582
In a mortality study, you are given that deaths occur one at n−1
a time and there are no withdrawals. Using the Nelson-Åalen 3
estimator, Ĥ(y3 ) = 0.1582. n−1≈ = 18.96 ⇒ n = 20
0.1582
Calculate Ĥ(y5 ). 1 1
Ĥ(y5 ) = 0.1582 + + = 0.2795
17 16
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 22/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 23/71
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 24/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 25/71
New Section: Greenwood and Åalen Variance
Greenwood Approximation
j
X s
Ô Ŝ(yj ) = Ŝ(yj )2
Vr
r (r − s )
=1
j
X s
Ô Ĥ(yj ) =
Vr 2
r
=1
Example
An 11-year mortality study on a group of 10 men age 80 re-
sults in the following survival times:
1 2
Vr Ŝ(6) = 0.6752 + = 0.024047
3 6 6 9 10 (10)(9) (8)(6)
1 2
Vr Ĥ(6) = + 2 = 0.04125
In addition, one man apiece leaves the study at times 4 and
8. The other 3 survive to the end. 102 8
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 29/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 30/71
y s r
1 15 100
8 20 65
17 13 40
25 31 31
v
t 15 20 13
u
+ + = 0.119828
1002 652 402
Example
yj rj sj
For a mortality study on 10 individuals, death times were
9 8 1
16 5 1
5 9 16 22 25 27
and censored observations occurred at times 7, 10, 14, and Ŝ(20 | X > 5) = 0.7
16. 1 1
Ô Ŝ(20 | X > 5) = 0.72 +
Vr
The probability that an individual dies no later than time (8)(7) (5)(4)
20, given survival to time 5, is estimated using the Kaplan- = 0.03325
Meier estimator to be 0.3.
Calculate the variance of this estimate using the Green-
wood formula.
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 32/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 33/71
Recursions Solution to Recursions Example
Example
(i) Ŝ(yj ) is the Kaplan-Meier prod-
s4 0.63
uct limit estimate of the survival =1− = 0.1
function at time yj . r4 0.70
Ô Ĥ(yj ) is the estimated vari- r4 = 10s4
(ii) Vr
Ŝ(yj ) Ô Ĥ(yj )
j Vr s4
ance of the Nelson-Åalen esti-
= 0.0105 − 0.0080 = 0.0025
mate of the cumulative hazard 3 0.70 0.0080 r42
rate function at time yj . 4 0.63 0.0105
1
(iii) Some values for Ŝ(yj ) and = 0.0025
100s4
Ô Ĥ(yj ) are shown to the
Vr 1
right. s4 = = 4
0.25
Determine the number of events at
time y4 .
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 34/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 35/71
Example
A mortality study is conducted 2 3
on 50 lives observed from time 0. Time Deaths Censored + = 0.0027589
You are given the information to 10 2 0 (50)(48) (48 − c18 )(45 − c18 )
the right. You are also given that 18 0 c18 3
using the Greenwood formula, = 0.0019256
24 3 0 (48 − c18 )(45 − c18 )
30 0 4 3
Ô Ŝ(35)
Vr 37 1 0 (48 − c18 )(45 − c18 ) = = 1558
= 0.0027589 0.0019256
Ŝ(35)2
c18 = 7
Determine c18 .
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 36/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 37/71
C-S05:15 We have
Twelve policyholders were monitored from the starting date s s
Ĥ(y ) Ô Ĥ(y )
of the policy to the time of first claim. The observed data are y r s Vr
as follows:
r r2
1 12 2 0.166667 0.166667 0.013889 0.013889
Time of First Claim 1 2 3 4 5 6 7 2 10 1 0.100000 0.266667 0.010000 0.023889
Number of Claims 2 1 2 2 1 2 2 3 9 2 0.222222 0.488889 0.024691 0.048580
Using the Nelson-Åalen estimator, calculate the 95% linear 4 7 2 0.285714 0.774603 0.040816 0.089397
confidence interval for the cumulative hazard rate function Then
H(4.5). p
1.96 0.089397 = 0.5860
(A) (0.189, 1.361) (D) (0.283, 1.266) 0.7746 − 0.5860 = 0.1886
(B) (0.206, 1.545) (E) (0.314, 1.437) 0.7746 + 0.5860 = 1.3606 (A)
(C) (0.248, 1.402)
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 39/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 40/71
S()
Confidence interval is (S1/ U , SU ), where
p
Z Vr(S)
U = exp
S ln S
H()
Confidence interval is (H/ U, HU) where
p
Z Vr(H)
U = exp
H
0.8108 + 0.8892
S() = = 0.85
2
Æ
Z Vr(S) = 0.8892 − 0.85 = 0.0392
0.0392
U = exp = exp(−0.28377) = 0.75294
0.85 ln 0.85
(S1/ U , SU ) = (0.851/ 0.75294 , 0.850.75294 ) = (0.8059, 0.8848)
Example
A 95% linear confidence interval for H() is (0.02, 0.06).
Construct a 95% log-transformed confidence interval for
H().
Æ
H() = 0.04 Z Vr(H) = 0.02
0.02
U = exp = exp(0.5) = 1.64872
0.04
(H/ U, HU) = 0.04/ 1.64872, 0.04(1.64872)
= (0.0243, 0.0659)
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 44/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 45/71
4-F04:12
The interval (0.357, 0.700) is a 95% log-transformed con-
fidence interval for the cumulative hazard rate function at Æ
time t, where the cumulative hazard rate function is esti- H(t) = (0.357)(0.700) = 0.5
mated using the Nelson-Åalen estimator. S(t) = e −0.5
= 0.6066 (E)
Determine the value of the Nelson-Åalen estimate of S(t).
(A) 0.50 (B) 0.53 (C) 0.56 (D) 0.59 (E) 0.61
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 46/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 47/71
Going from Log-Transformed to Linear: S Going from Log-Transformed to Linear: S
Solution
1
ln 0.82 = ln S
U
ln 0.87 = U ln S
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 48/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 49/71
Methods
É Exact exposure
É Actuarial exposure
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 53/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 54/71
q̂j = 1 − e−ndj / ej
Example
Calculate q̂66 in the previous example.
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 56/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 57/71
ndj
q̂j =
ej
Example
Calculate q̂66 in the previous example.
ndj
q̂j =
ej
Example
Calculate q̂66 in the previous example.
ndj
q̂j =
ej
Example
Calculate q̂66 in the previous example.
2
q̂66 = = 0.444444
(54/ 12)
Insuring Age
Set birth date equal to policy date, using age last birthday or
age nearest birthday.
Example
Insuring age
Example
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 61/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 62/71
Anniversary-to-Anniversary Studies
Anniversary-to-Anniversary Study
Example
Consider a 2009 anniversary to 2011 anniversary-to-anniversary
study. For simplicity, assume all events including births, pol-
icy issue, withdrawals, and deaths occur on the first day of the
month and that a month is 1/12 of a year. The ending day of the
study will be treated as if it is 1/1/2012. The following data is
recorded for 6 individuals in the study:
Birth Policy Date of Date of
Number Date Issue Withdrawal Death
1 3/1944 5/2004 5/2011 —
2 10/1944 6/2010 — —
3 12/1944 2/2009 — 4/2011
4 1/1945 12/2008 4/2010 —
5 4/1945 1/2010 — 8/2011
6 7/1945 11/2010 — —
Calculate the total actuarial exposure for age 65 using age
nearest birthday.
Answer: 52 months
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 64/71
Variance of Estimators
Exact Exposure
Exact Exposure
dj
Ô q̂j ) = (1 − q̂j )2 n2
Vr(
e2j
Example
For age 70, there are 25 years of exact exposure and 2 deaths.
Calculate the variance of the exact exposure estimate.
Variance of Estimators
Actuarial Exposure
Actuarial Exposure
q̂j (1 − q̂j )
Ô q̂j ) =
Vr(
ej / n
Example
For age 70, there are 25 years of actuarial exposure and 2
deaths.
Calculate the variance of the actuarial exposure estimate.
2
q̂70 = = 0.08
25
(0.08)(0.92)
Ô q̂70 ) =
Vr( = 0.002944
25
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 66/71
For time cj ,
Pj is population at time cj .
nj are entrants in [cj , cj+1 )
j are withdrawals in (cj , cj+1 ]
dj are deaths in (cj , cj+1 ]
Superscript b indicates beginning of period.
Superscript m indicates middle of period.
Superscript e indicates end of period.
Pj = Pj−1 + nm
j−1
m
− dj−1 − j−1 e
− j−1 + nbj
ej = Pj + 0.5(nm
j
− jm − dj ) for exact exposure
ej = Pj + 0.5(nm
j
− jm ) for actuarial exposure
Example
You perform a mortality study on 950 lives age 45. Addi- j cj Pj nm
j
jm je dj ej p̂j
tional lives enter the study at ages 45 and later. You are given 0 45 950 50 100 0 3 925 922/ 925
the following information from the mortality study:
1 46 897 30 90 0 5 867 862/ 867
j cj nj j dj 2 47 832 40 80 786 6 812 806/ 812
0 45 50 100 3 3 48 0
1 46 30 90 5
2 47 40 80 6 j cj j+1 p̂45
All entries, other than the original 950 lives, and all with- 0 45 922/ 925 = 0.996757
drawals are assumed to occur uniformly throughout each 1 46 0.996757(862/ 867) = 0.991008
year. 2 47 0.991008(806/ 812) = 0.983686
Estimate the probability of survival for an individual age 45
to each of ages 46, 47, 48 using the actuarial method.
Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 70/71 Product-Limit and Nelson-Åalen Estimators Oct. 17–26, 2016 71/71