More Statistics tutorial at www.dumblittledoctor.

com

Survival Analysis
A Brief Introduction
More Statistics tutorial at www.dumblittledoctor.com

2
 More Statistics tutorial at www.dumblittledoctor.com

1. Survival Function, Hazard

Function
y In many medical studies, the primary
endpoint is time until an event occurs
(e.g. death, remission)
y Data are typically
yp y subject
j to censoringg
(e.g. when a study ends before the event
occurs)
y Survival Function - A function
describing the proportion of individuals
surviving
i i to or beyondb d a given
i time.
i
Notation:
⎛
◦ T: survival time of a randomly selected ⎞
⎛ ⎞ ⎟
individual S (t ) = P(T ≥ t ) = exp ⎜ − ∫ λ (u )du
t

⎝ 0 ⎠
◦ t: a specific point in time. 3
 More Statistics tutorial at www.dumblittledoctor.com

Hazard Function/Rate
y Hazard Function λ(t): instantaneous
failure rate at time t given that the
subject has survived upto time t.
P(t ≤ T < t + δ | T ≥ t ) P(t ≤ T < t + δ )
λ ( t ) = limδ →0+ =
δ P(T ≥ t ) × δ
S (t ) − S (t + δ ) 1 f (t )
= limδ →0+ × =
δ S (t ) S ( t )
d
f (t ) = F (t )
dt
y Here f(t) is the probability density
function of the survival time T. That
is, F ( t ) = 1 − S ( t ) = P (T ≤ t )
y where F(t) is the cumulative 4

di ib i i
 More Statistics tutorial at www.dumblittledoctor.com

2. The Key Word is

‘Censoring’
y Because of censoring,
g, many
y common
data analysis procedures can not
p
be adopted directly.
y
y For example, one could use the
logistic regression model to model
the relationship between survival
probability and some relevant
covariates
◦ However one should use the customized
logistic regression procedures
designed to account for censoring 5
 More Statistics tutorial at www.dumblittledoctor.com

Key Assumption:
Independent Censoring
y Those still at risk at time t in
the study are a random sample of
the population at risk at time t,
for all t

y This assumption
p means that the
hazard function, λ(t), can be
estimated in a fair/unbiased/valid
/ /
way
6
 More Statistics tutorial at www.dumblittledoctor.com
3A. Kaplan-Meier (Product-
Limit) Estimator of the
Survival Curve
y The Kaplan–Meier estimator is the
nonparametric maximum likelihood
estimate of S(t). It is a product of
th fformr − d
the r −d r −d
Sˆ (t ) = 1 1
× 2 2
× ... × i i
r1 r2 ri
rk
y is
t k the number of subjects alive
djust
k
before time tk
y d
denotes
t the
th number
b whoh died
di d att time
ti
7
 More Statistics tutorial at www.dumblittledoctor.com

Kaplan Meier Curve,

Kaplan-Meier Curve
Example
Time ti # at # events Ŝ
risk
0 20 0
11.00
00
5 20 2 [1-
(2/20)]*1 00 0 90
(2/20)]*1.00=0.90
6 18 0 [1-
(0/18)]*0 90=0 90
(0/18)]*0.90=0.90
10 15 1 [1- 8
(1/15)]*0 90=0 84
 More Statistics tutorial at www.dumblittledoctor.com

Kaplan Meier Curve

ence)
1.0
% Confide
0.9
viving (95%
0.7
7 0.8
Proporrtion Surv
0.6
0

0 5 10 15 20
Survival Time

9
 More Statistics tutorial at www.dumblittledoctor.com

Figure 1. Plot of survival distribution functions for the NCI and

the SCI Groups
Groups. The Y Y-axis
axis is the probability of not declining to
GDS 3 or above. The X-axis is the time (in years) to decline.
(Barry Reisberg et al., 2010; Alzheimer & Dementia; in press.)
10
 More Statistics tutorial at www.dumblittledoctor.com

3B. Comparing Survival

Functions
1.00

0 75
0.75
High
Survival
Distribution 0.50
Function
0.25
Low
Medium

0.00
0 10 20 30 40 50 60

Time
11
 More Statistics tutorial at www.dumblittledoctor.com

Log--Rank Test
Log

The log-rank test

• tests whether
hethe the survival
s i al functions
f nctions are
a e
statistically equivalent
• is a large-sample chi-square test that uses the
observed and expected cell counts across the
event times
• has maximum power when the ratio of hazards
is constant over time.

12
 More Statistics tutorial at www.dumblittledoctor.com

Wilcoxon Test

The Wilcoxon test

• weights
g the observed number of events
minus the expected number of events by
the number at risk across the event times
• can be biased if the pattern of censoring is
different between the groups.

13
 More Statistics tutorial at www.dumblittledoctor.com

Log-rank versus Wilcoxon

Log-
Test
Log-rank test
• is more sensitive than the Wilcoxon test to
d ff
differences between
b groups in later
l points in
time.
Wilcoxon test
• is more sensitive than the log-rank test to
differences between groups that occur in
early points in time.

14
 More Statistics tutorial at www.dumblittledoctor.com

4. Two Parametric
Distributions
y Here we p present two most notable
models for the distribution of T.
y Exponential distribution:λ (t ) = λ
y Weibull distribution:
λ (t ) = λp(λt ) = pλ × t
p −1 p p −1

◦ Its survival function:

⎛ t ⎞
(
S (t ) = exp⎜⎜ − ∫ pλ u du ⎟⎟ = exp − (λt ) p
p p −1
)
⎝ 0 ⎠
◦ Thus:
Th l (− ln(
ln l (S (t ) ) = p(ln( l (λ ) )
l (t ) + ln(
15
 More Statistics tutorial at www.dumblittledoctor.com

Weibull Hazard Function,

Plot

16
 More Statistics tutorial at www.dumblittledoctor.com

5 Regression Models
5.
y The Exponential
p and the Weibull
distribution inspired two
pparametric regression
g
approaches:
1. Parametric proportional hazard
model – this model can be
generalized to a semi-parametric
semi parametric
model: the Cox proportional
hazard model
2. Accelerated failure time model
17
 More Statistics tutorial at www.dumblittledoctor.com

Proportional Hazard Model

y In a regression model for survival
analysis one can try to model the
dependence on the explanatory
variables
i bl bby ttaking
ki ththe ((new)
)
hazard rate to be:
λ = λ0 × c( β 0 + β1 xi1 + β 2 xi 2 + ... + β k xik )

y Hazardd rates bbeing

H i positive
i i iit iis
natural to choose the function c
such that c(β x)
c(β,x) is positive
irrespective the values of x.
18
 Proportional Hazard Model
y Thus a good choice is: c (.) = exp(.)
y The resulting proportional hazard
model is:
λ = λ0 × exp(β 0 + β1 xi1 + β 2 xi 2 + ... + β k xik )

y For the Weibull distribution we

λ = pλ0 × t × exp( β 0 + β1 xi1 + β 2 xi 2 + ... + β k xik )
have:
p p −1

λ = λ0 × exp((β 0 + β1 xi1 + β 2 xi 2 + ... + β k xik )

y For the Exponential distribution we
have: 19
 More Statistics tutorial at www.dumblittledoctor.com

Accelerated Failure Time

Model
y For the Weibull distribution
(including the Exponential
distribution), the proportional p p
hazard model is equivalent to a
logg linear model in survival time
l (T ) = α 0 + α1 xi1 + α 2 xi 2 + ... + α k xik + σε
ln
T:
ε

y Here the error term can be

shown to follow the 2-parameter 20

Extreme Vvalue distribution

 More Statistics tutorial at www.dumblittledoctor.com

Apply Both Models

Simultaneously
y If the underlying
y g distribution for
T is Weibull or Exponential, one
pp y both regression
can apply g models
simultaneously to reflect
different aspects of the survival
process. That is
y Prediction of degree of decline
using the Weibull proportional
hazard model
y Prediction of time of decline
21
using the accelerated failure time
 More Statistics tutorial at www.dumblittledoctor.com

An Example
y In a recent p
paper
p (Reisberg
g et
al., 2010), we applied both
g
regression models to a dementia
study conducted at NYU:
Λ(T) = Λ0 (T)exp(α1 *Group +α2 * Age +α3 *Gender +α4 * Education +α5 * FollowUp)

log T = β 0 + β1 * Group + β 2 * Age + β 3 * Gender + β 4 * Education + β 5 * FollowUp + σε

y The results are shown next

22
More Statistics tutorial at www.dumblittledoctor.com

23
 More Statistics tutorial at www.dumblittledoctor.com

6 Cox Proportional Hazards

6.
Model

24
 More Statistics tutorial at www.dumblittledoctor.com

Parametric versus
Nonparametric Models

Parametric models require that

• the distribution of survival time is known
• the hazard function is completely specified
exceptt for
f the
th values
l off the
th unknown
k
parameters.
Examples include the Weibull model
model, the
exponential model, and the log-normal model.

25
 More Statistics tutorial at www.dumblittledoctor.com

Parametric versus
Nonparametric Models

Properties
p of nonparametric
p models are
• the distribution of survival time is unknown
• the hazard function is unspecified.
unspecified
An example is the Cox proportional hazards
model.

26
 More Statistics tutorial at www.dumblittledoctor.com
...
Cox Proportional Hazards
Model
{ β1 X i 1 +...+ β k X ik }
hi (t ) = h0 (t )e

Baseline Hazard Linear function of a

function - involves set of predictor
time but not predictor variables - does
variables not involve time

β = 0 → hazard ratio = 1
Two ggroups
p have the same
survival experience
27
 More Statistics tutorial at www.dumblittledoctor.com

Popularity of the Cox

Model
The Cox proportional hazards model
• provides the primary information
d
desiredd ffrom a survivall analysis,
l h
hazard
d
ratios and adjusted survival curves, with
a minimum number of assumptions p
• is a robust model where the regression
coefficients closely approximate the
results from the correct parametric
model.

28
 Partial Likelihood

Partial likelihood differs from maximum

likelihood because
• it does not use the likelihoods for all subjects
• it only considers likelihoods for subjects that
experience the event
• it considers subjects as part of the risk set
until they are censored.

29
 More Statistics tutorial at www.dumblittledoctor.com

Partial Likelihood

Subject Survival Time Status

C 20
2.0 1
B 3.0 1
A 40
4.0 0
D 5.0 1
E 60
6.0 0

30
 More Statistics tutorial at www.dumblittledoctor.com

Partial Likelihood
hc (2)
Lc =
hc (2) + hb (2) + ha (2) + hd (2) + he (2)

hb (3)
Lb =
hb (3) + ha (3) + hd (3) + he (3)

hd (5)
Ld =
hd (5) + he (5)

31
 More Statistics tutorial at www.dumblittledoctor.com

Partial Likelihood
hd (5)
Ld =
hd (5) + he (5)

β1 X d 1 +β 2 X d 2 + .... + β k X dk
ho (5)e
Ld = β1 X d 1 +β 2 X d 2 + .... + β k X dk β1 X e1 +β 2 X e 2 + .... + β k X ek
ho (5)e + ho (5)e

β1 X d 1 +β 2 X d 2 + .... + β k X dk
e
Ld =
e β1 X d 1 +β 2 X d 2 + .... + β k X dk + e β1 X e1 +β 2 X e 2 + .... + β k X ek
32
 More Statistics tutorial at www.dumblittledoctor.com

Partial Likelihood
y The overall likelihood is the
product of the individual
likelihood. That is:
L = Lc * Lb * Ld

33
 More Statistics tutorial at www.dumblittledoctor.com

7. SAS Programs for Survival

Analysis
y There are three SAS procedures for analyzing
survival
ur i l data:
d t : LIFETEST,
LIFETEST PHREG,
PHREG and
nd LIFEREG.
LIFEREG
y PROC LIFETEST is a nonparametric procedure
for estimating the survivor function,
comparing the underlying survival curves of
two or more samples, and testing the
association of survival time with other
variables.
y PROC PHREG is a semiparametric procedure that
fits the Cox pproportional
p hazards model and
its extensions.
y PROC LIFEREG is a parametric regression
procedure for modeling the distribution of
survival time with a set of concomitant
variables. 34
 More Statistics tutorial at www.dumblittledoctor.com

Proc LIFETEST
y The Kaplan
Kaplan-
p -Meier(K-M) survival
curves and related tests (Log-
(Log-
Rank, Wilcoxon) can be generated
g
using SAS PROC LIFETEST

PROC LIFETEST DATA=SAS-data-set

<options>;
TIME variable <*censor(list)>;
STRATA variable <(list)> <...variable
<(list)>>;
TEST variables; 35
 More Statistics tutorial at www.dumblittledoctor.com

Proc PHREG
y The Cox (proportional
p p hazards)
regression is performed using SAS
PROC PHREG

proc phreg data=rsmodel.colon;

data=rsmodel colon;
model surv_mm*status(0,2,4) = sex
yydx / risklimits;
run;

36
 More Statistics tutorial at www.dumblittledoctor.com

Proc LIFEREG
y The accelerated failure time
regression is performed using SAS
PROC LIFEREG

proc lifereg data=subset

outest=OUTEST(keep=_scale_);
model (lower
(lower, hours) = yrs
yrs_ed
ed
yrs_exp / d=normal; output out=OUT
xbeta=Xbeta; run;
37
 More Statistics tutorial at www.dumblittledoctor.com

Selected References
y PD Allison (1995). Survival
Analysis Using SAS: A Practical
g
Guide. SAS Publishing.
y JD Kalbfleisch and RL Prentice
(2002).The Statistical Analysis
of Failure Time Data. Wiley-
Interscience.

38