Survival Analysis: Introduction

Survival Analysis typically focuses on time to event data.

In the most general sense, it consists of techniques for positivevalued random variables, such as
time to death
time to onset (or relapse) of a disease
length of stay in a hospital
duration of a strike
money paid by health insurance
viral load measurements
time to finishing a doctoral dissertation!

Kinds of survival studies include:

clinical trials
prospective cohort studies
retrospective cohort studies
retrospective correlative studies
Typically, survival data are not fully observed, but rather
are censored.
1

In this course, we will:

describe survival data
compare survival of several groups
explain survival with covariates
design studies with survival endpoints

Some knowledge of discrete data methods will be useful,

since analysis of the time to event uses information from
the discrete (i.e., binary) outcome of whether the event occurred or not.
Some useful references:
Collett: Modelling Survival Data in Medical Research
Cox and Oakes: Analysis of Survival Data
Kalbfleisch and Prentice: The Statistical Analysis of
Failure Time Data
Lee: Statistical Methods for Survival Data Analysis
Fleming & Harrington: Counting Processes and Survival Analysis
Hosmer & Lemeshow: Applied Survival Analysis
Kleinbaum: Survival Analysis: A self-learning text
2

 Klein & Moeschberger: Survival Analysis: Techniques

for censored and truncated data
Cantor: Extending SAS Survival Analysis Techniques
for Medical Research
Allison: Survival Analysis Using the SAS System
Jennison & Turnbull: Group Sequential Methods with
Applications to Clinical Trials
Ibrahim, Chen, & Sinha: Bayesian Survival Analysis

Some Definitions and notation

Failure time random variables are always non-negative.
That is, if we denote the failure time by T , then T 0.
T can either be discrete (taking a finite set of values, e.g.
a1, a2, . . . , an) or continuous (defined on (0, )).
A random variable X is called a censored failure time
random variable if X = min(T, U ), where U is a nonnegative censoring variable.

In order to define a failure time random variable,

we need:
(1) an unambiguous time origin
(e.g. randomization to clinical trial, purchase of car)
(2) a time scale
(e.g. real time (days, years), mileage of a car)
(3) definition of the event
(e.g. death, need a new car transmission)

Illustration of survival data

X
X
y
y

X
y

X
y

study
opens

study
closes

= censored observation
X = event

The illustration of survival data on the previous page shows

several features which are typically encountered in analysis
of survival data:
individuals do not all enter the study at the same time
when the study ends, some individuals still havent had
the event yet
other individuals drop out or get lost in the middle of
the study, and all we know about them is the last time
they were still free of the event

The first feature is referred to as staggered entry

The last two features relate to censoring of the failure
time events.

Types of censoring:
Right-censoring :
only the r.v. Xi = min(Ti, Ui) is observed due to
loss to follow-up
drop-out
study termination
We call this right-censoring because the true unobserved
event is to the right of our censoring time; i.e., all we
know is that the event has not happened at the end of
follow-up.
In addition to observing Xi, we also get to see the failure indicator:

1 if Ti Ui
i =
0 if T > U
i
i
Some software packages instead assume we have a
censoring indicator:

ci =

0 if Ti Ui
1 if Ti > Ui

Right-censoring is the most common type of censoring

assumption we will deal with in survival analysis.
7

 Left-censoring
Can only observe Yi = max(Ti, Ui) and the failure indicators:

1 if Ui Ti
i =
0 if U > T
i
i
e.g. (Miller) study of age at which African children learn
a task. Some already knew (left-censored), some learned
during study (exact), some had not yet learned by end
of study (right-censored).

Interval-censoring
Observe (Li, Ri) where Ti (Li, Ri)
Ex. 1: Time to prostate cancer, observe longitudinal
PSA measurements
Ex. 2: Time to undetectable viral load in AIDS studies,
based on measurements of viral load taken at each clinic
visit
Ex. 3: Detect recurrence of colon cancer after surgery.
Follow patients every 3 months after resection of primary
tumor.

Independent vs informative censoring

We say censoring is independent (non-informative) if
Ui is independent of Ti.
Ex. 1 If Ui is the planned end of the study (say, 2
years after the study opens), then it is usually independent of the event times.
Ex. 2 If Ui is the time that a patient drops out
of the study because he/she got much sicker and/or
had to discontinue taking the study treatment, then
Ui and Ti are probably not independent.
An individual censored at U should be representative of all subjects who survive to U .
This means that censoring at U could depend on prognostic characteristics measured at baseline, but that among
all those with the same baseline characteristics, the probability of censoring prior to or at time U should be the
same.
Censoring is considered informative if the distribution of Ui contains any information about the parameters
characterizing the distribution of Ti.

Suppose we have a sample of observations on n people:

(T1, U1), (T2, U2), ..., (Tn, Un)
There are three main types of (right) censoring times:
Type I: All the Uis are the same
e.g. animal studies, all animals sacrificed after 2 years

Type II: Ui = T(r) , the time of the rth failure.

e.g. animal studies, stop when 4/6 have tumors

Type III: the Uis are random variables, is are failure

indicators:

i =

1 if Ti Ui
0 if Ti > Ui

Type I and Type II are called singly censored data,

Type III is called randomly censored (or sometimes progressively censored).

10

Some example datasets:

Example A. Duration of nursing home stay
(Morris et al., Case Studies in Biometry, Ch 12)
The National Center for Health Services Research studied
36 for-profit nursing homes to assess the effects of different
financial incentives on length of stay. Treated nursing
homes received higher per diems for Medicaid patients, and
bonuses for improving a patients health and sending them
home.
Study included 1601 patients admitted between May 1, 1981
and April 30, 1982.
Variables include:
LOS - Length of stay of a resident (in days)
AGE - Age of a resident
RX - Nursing home assignment (1:bonuses, 0:no bonuses)
GENDER - Gender (1:male, 0:female)
MARRIED - (1: married, 0:not married)
HEALTH - health status (2:second best, 5:worst)
CENSOR - Censoring indicator (1:censored, 0:discharged)
First few lines of data:
37 86 1 0 0 2 0
61 77 1 0 0 4 0
11

Example B. Fecundability
Women who had recently given birth were asked to recall
how long it took them to become pregnant, and whether or
not they smoked during that time. The outcome of interest (summarized below) is time to pregnancy (measured in
menstrual cycles).
19 subjects were not able to get pregnant after 12 months.
Cycle Smokers Non-smokers
1
29
198
2
16
107
3
17
55
4
4
38
5
3
18
6
9
22
7
4
7
8
5
9
9
1
5
10
1
3
11
1
6
12
3
6
12+
7
12

12

Example C: MAC Prevention Clinical Trial

ACTG 196 was a randomized clinical trial to study the effects
of combination regimens on prevention of MAC (mycobacterium avium complex), one of the most common opportunistic infections in AIDS patients.
The treatment regimens were:
clarithromycin (new)
rifabutin (standard)
clarithromycin plus rifabutin
Other characteristics of trial:
Patients enrolled between April 1993 and February 1994
Follow-up ended August 1995
In February 1994, rifabutin dosage was reduced from 3
pills/day (450mg) to 2 pills/day (300mg) due to concern
over uveitis1
The main intent-to-treat analysis compared the 3 treatment
arms without adjusting for this change in dosage.

Uveitis is an adverse experience resulting in inflammation of the

uveal tract in the eyes (about 3-4% of patients reported uveitis).
1

13

Example D: HMO Study of HIV-related Survival

This is hypothetical data used by Hosmer & Lemeshow (described on pages 2-17) containing 100 observations on HIV+
subjects belonging to an Health Maintenance Organization
(HMO). The HMO wants to evaluate the survival time of
these subjects. In this hypothetical dataset, subjects were
enrolled from January 1, 1989 until December 31, 1991.
Study follow up then ended on December 31, 1995.
Variables:
ID
TIME
ENTDATE
ENDDATE
CENSOR
AGE
DRUG

Subject ID (1-100)
Survival time in months
Entry date
Date follow-up ended due to death or censoring
Death Indicator (1=death, 0=censor)
Age of subject in years
History of IV Drug Use (0=no,1=yes)

This dataset is used by Hosmer & Lemeshow to motivate

some concepts in survival analysis in Chap. 1 of their book.

14

Example E: UMARU Impact Study (UIS)

This dataset comes from the University of Massachusetts
AIDS Research Unit (UMARU) IMPACT Study, a 5-year
collaborative research project comprised of two concurrent
randomized trials of residential treatment for drug abuse.
(1) Program A: Randomized 444 subjects to a 3- or 6month program of health education and relapse prevention. Clients were taught to recognize high-risk situations that are triggers to relapse, and taught skills to
cope with these situations without using drugs.
(2) Program B: Randomized 184 participants to a 6- or
12-month program with highly structured life-style in a
communal living setting.
Variables:
ID
AGE
BECKTOTA
HERCOC
IVHX
NDRUGTX
RACE
TREAT
SITE
LOT
TIME
CENSOR

Subject ID (1-628)
Age in years
Beck Depression Score
Heroin or Cocaine Use prior to entry
IV Drug use at Admission
Number previous drug treatments
Subjects Race (0=White, 1=Other)
Treatment Assignment (0=short, 1=long)
Treatment Program (0=A,1=B)
Length of Treatment (days)
Time to Return to Drug Use (days)
Indicator of Drug Use Relapse (1=yes,0=censored)
15

Example F: Atlantic Halibut Survival Times

One conservation measure suggested for trawl fishing is a
minimum size limit for halibut (32 inches). However, this size
limit would only be effective if captured fish below the limit
survived until the time of their release. An experiment was
conducted to evaluate the survival rates of halibut caught by
trawls or longlines, and to assess other factors which might
contribute to survival (duration of trawling, maximum depth
fished, size of fish, and handling time).
An article by Smith, Waiwood and Neilson, Survival Analysis for Size Regulation of Atlantic Halibut in Case Studies
in Biometry compares parametric survival models to semiparametric survival models in evaluating this data.

Obs
#
100
109
113
116
....

Survival
Time
(min)
353.0
111.0
64.0
500.0

Tow
Censoring Duration
Indicator
(min.)
1
30
1
100
0
100
1
100

16

Diff
in
Depth
15
5
10
10

Length
of Fish
(cm)
39
44
53
44

Handling
Time
(min.)
5
29
4
4

Total
log(catch)
ln(weight)
5.685
8.690
5.323
5.323

More Definitions and Notation

There are several equivalent ways to characterize the probability distribution of a survival random variable. Some of
these are familiar; others are special to survival analysis. We
will focus on the following terms:
The density function f (t)
The survivor function S(t)
The hazard function (t)
The cumulative hazard function (t)

Density function (or Probability Mass Function) for discrete r.v.s

Suppose that T takes values in a1, a2, . . . , an.
f (t) = P r(T = t)
=

fj if t = aj , j = 1, 2, . . . , n
0

if t 6= aj , j = 1, 2, . . . , n

Density Function for continuous r.v.s

1
f (t) = lim
P r(t T t + t)
t0 t
17

 Survivorship Function: S(t) = P (T t).

In other settings, the cumulative distribution function,
F (t) = P (T t), is of interest. In survival analysis, our
interest tends to focus on the survival function, S(t).
For a continuous random variable:
S(t) =

f (u)du

For a discrete random variable:

S(t) =
=
=

ut

f (u)

f (aj )

fj

aj t
aj t

Notes:
From the definition of S(t) for a continuous variable,
S(t) = 1 F (t) as long as F (t) is absolutely continuous
w.r.t the Lebesgue measure. [That is, F (t) has a density
function.]
For a discrete variable, we have to decide what to do if
an event occurs exactly at time t; i.e., does that become
part of F (t) or S(t)?
To get around this problem, several books define
S(t) = P r(T > t), or else define F (t) = P r(T < t)
(eg. Collett)
18

 Hazard Function (t)

Sometimes called an instantaneous failure rate, the
force of mortality, or the age-specific failure rate.
Continuous random variables:
1
P r(t T < t + t|T t)
(t) = lim
t0 t
1 P r([t T < t + t]
= lim
t0 t
P r(T t)
= lim

t0

1 P r(t T < t + t)
t
P r(T t)

f (t)
S(t)

Discrete random variables:

(aj ) j = P r(T = aj |T aj )
=

P (T = aj )
P (T aj )

f (aj )
S(aj )

f (t)
P
k:ak aj f (ak )

19

[T t])

 Cumulative Hazard Function (t)

Continuous random variables:
(t) =

Z t

(u)du

Discrete random variables:

(t) =

k:ak <t

20

Relationship between S(t) and (t)

Weve already shown that, for a continuous r.v.

(t) =

f (t)
S(t)

For a left-continuous survivor function S(t), we can show:

f (t) = S 0(t)

or S 0(t) = f (t)

We can use this relationship to show that:

d
[log S(t)] =
dt
=
=

1 0
S (t)
S(t)

f (t)
S(t)

f (t)
S(t)

So another way to write (t) is as follows:

(t) =

d
[log S(t)]
dt

21

Relationship between S(t) and (t):

Continuous case:
(t) =

Z t

(u)du

Z t

f (u)
du
S(u)

Z t

d
log S(u)du
du

= log S(t) + log S(0)

S(t) = e(t)
Discrete case:
Suppose that aj < t aj+1. Then
S(t) = P (T a1 , T a2 , . . . , T aj+1 )
= P (T a1 )P (T a2 |T a1 ) P (T aj+1 |T aj )
= (1 1 ) (1 j )
=

k:ak <t

(1 k )
P

Cox defines (t) = k:ak <t log(1 k ) so that S(t) =

e(t) in the discrete case, as well.

22

Measuring Central Tendency in Survival

Mean survival - call this
=
=

n
X

j=1

uf (u)du for continuous T

aj fj for discrete T

Median survival - call this , is defined by

S( ) = 0.5
Similarly, any other percentile could be defined.
In practice, we dont usually hit the median survival
at exactly one of the failure times. In this case, the
estimated median survival is the smallest time such
that
) 0.5
S(

23

Some hazard shapes seen in applications:

increasing

e.g. aging after 65

decreasing

e.g. survival after surgery

bathtub

e.g. age-specific mortality

constant

e.g. survival of patients with advanced chronic disease

24

Estimating the survival or hazard function

We can estimate the survival (or hazard) function in two
ways:
by specifying a parametric model for (t) based on a
particular density function f (t)
by developing an empirical estimate of the survival function (i.e., non-parametric estimation)

If no censoring:

The empirical estimate of the survival function, S(t),

is the
proportion of individuals with event times greater than t.
With censoring:
is not a good
If there are censored observations, then S(t)
estimate of the true S(t), so other non-parametric methods
must be used to account for censoring (life-table methods,
Kaplan-Meier estimator)

25

Some Parametric Survival Distributions

The Exponential distribution (1 parameter)
f (t) = et for t 0

S(t) =

f (u)du

t
t

= e

f (t)
S(t)
=
constant hazard!

(t) =

(t) =
=

Z t

Z t

(u) du
du

= t

Check: Does S(t) = e(t) ?

median: solve 0.5 = S( ) = e :
=
mean:

log(0.5)

ueudu =

26

 The Weibull distribution (2 parameters)

Generalizes exponential:
S(t) = et
f (t) =

S(t) = t1 et
dt

(t) = t1
(t) =

Z t

(u)du = t

- the scale parameter

- the shape parameter

The Weibull distribution is convenient because of its simple form. It includes several hazard shapes:
= 1 constant hazard

0 < < 1 decreasing hazard

> 1 increasing hazard

27

 Rayleigh distribution
Another 2-parameter generalization of exponential:
(t) = 0 + 1t

compound exponential
T exp(), g

f (t) =

etg()d

log-normal, log-logistic:
Possible distributions for T obtained by specifying for
log T any convenient family of distributions, e.g.
log T normal (non-monotone hazard)
log T logistic

28

Why use one versus another?

technical convenience for estimation and inference
explicit simple forms for f (t), S(t), and (t).
qualitative shape of hazard function

One can usually distinguish between a one-parameter model

(like the exponential) and two-parameter (like Weibull or
log-normal) in terms of the adequacy of fit to a dataset.
Without a lot of data, it may be hard to distinguish between
the fits of various 2-parameter models (i.e., Weibull vs lognormal)

29

Plots of estimates of S(t)

Based on KM, exponential, Weibull, and log-normal
for study of protease inhibitors in AIDS patients
(ACTG 320)

0.96

0.98
0

0.92

0.94

Probability

0.96
0.94

KM
Exponential
Weibull
Lognormal

0.90

0.92

0.90

Probability

0.98

1.00

KM Curves for Time to PCP

- 3 Drug Arm

1.00

KM Curves for Time to PCP

- 2 Drug Arm

100

200

300

400

days

KM
Exponential
Weibull
Lognormal

100

200
days

30

300

400

Plots of estimates of S(t)

Based on KM, exponential, Weibull, and log-normal
for study of protease inhibitors in AIDS patients
(ACTG 320)

0.96

0.98
0

0.92

0.94

Probability

0.96
0.94

KM
Exponential
Weibull
Lognormal

0.90

0.92

0.90

Probability

0.98

1.00

KM Curves for Time to MAC

- 3 Drug Arm

1.00

KM Curves for Time to MAC

- 2 Drug Arm

100

200

300

400

days

KM
Exponential
Weibull
Lognormal

100

200
days

31

300

400

Plots of estimates of S(t)

Based on KM, exponential, Weibull, and log-normal
for study of protease inhibitors in AIDS patients
(ACTG 320)

0.96

0.98
0

0.92

0.94

Probability

0.96
0.94

KM
Exponential
Weibull
Lognormal

0.90

0.92

0.90

Probability

0.98

1.00

KM Curves for Time to CMV

- 3 Drug Arm

1.00

KM Curves for Time to CMV

- 2 Drug Arm

100

200

300

400

days

KM
Exponential
Weibull
Lognormal

100

200
days

32

300

400

Preview of Coming Attractions

Next we will discuss the most famous non-parametric approach for estimating the survival distribution, called the
Kaplan-Meier estimator.
To motivate the derivation of this estimator, we will first
consider a set of survival times where there is no censoring.

The following are times to relapse (weeks) for 21 leukemia

patients receiving control treatment (Table 1.1 of Cox &
Oakes):
1, 1, 2, 2, 3, 4, 4, 5, 5, 8, 8, 8, 8, 11, 11, 12, 12, 15, 17, 22, 23

How would we estimate S(10), the probability that an individual survives to time 10 or later?

What about S(8)?

Is it

12
21

33

or

8
21 ?


Lets construct a table of S(t):

Values of t
S(t)
t 1 21/21=1.000
1 < t 2 19/21=0.905
2 < t 3 17/21=0.809
3<t4
4<t5
5<t8
8 < t 11
11 < t 12
12 < t 15
15 < t 17
17 < t 22
22 < t 23
Empirical Survival Function:
When there is no censoring, the general formula is:
= # individuals with T t
S(t)
total sample size

34

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.

Example for leukemia data (control arm):

35

Stata Commands for Survival Estimation

.use leukem
.stset remiss status if trt==0
(21 observations deleted)
. sts list
failure _d:
analysis time _t:

(to keep only untreated patients)

status
remiss

Beg.
Net
Survivor
Std.
Time
Total
Fail
Lost
Function
Error
[95% Conf. Int.]
---------------------------------------------------------------------1
21
2
0
0.9048
0.0641
0.6700
0.9753
2
19
2
0
0.8095
0.0857
0.5689
0.9239
3
17
1
0
0.7619
0.0929
0.5194
0.8933
4
16
2
0
0.6667
0.1029
0.4254
0.8250
5
14
2
0
0.5714
0.1080
0.3380
0.7492
8
12
4
0
0.3810
0.1060
0.1831
0.5778
11
8
2
0
0.2857
0.0986
0.1166
0.4818
12
6
2
0
0.1905
0.0857
0.0595
0.3774
15
4
1
0
0.1429
0.0764
0.0357
0.3212
17
3
1
0
0.0952
0.0641
0.0163
0.2612
22
2
1
0
0.0476
0.0465
0.0033
0.1970
23
1
1
0
0.0000
.
.
.
---------------------------------------------------------------------.sts graph

36

SAS Commands for Survival Estimation

data leuk;
input t;
cards;
1
1
2
2
3
4
4
5
5
8
8
8
8
11
11
12
12
15
17
22
23
;
proc lifetest data=leuk;
time t;
run;

37

SAS Output for Survival Estimation

The LIFETEST Procedure
Product-Limit Survival Estimates

Survival

0.0000
1.0000
1.0000
2.0000
2.0000
3.0000
4.0000
4.0000
5.0000
5.0000
8.0000
8.0000
8.0000
8.0000
11.0000
11.0000
12.0000
12.0000
15.0000
17.0000
22.0000
23.0000

1.0000
.
0.9048
.
0.8095
0.7619
.
0.6667
.
0.5714
.
.
.
0.3810
.
0.2857
.
0.1905
0.1429
0.0952
0.0476
0

Failure
0
.
0.0952
.
0.1905
0.2381
.
0.3333
.
0.4286
.
.
.
0.6190
.
0.7143
.
0.8095
0.8571
0.9048
0.9524
1.0000

38

Survival
Standard
Error
0
.
0.0641
.
0.0857
0.0929
.
0.1029
.
0.1080
.
.
.
0.1060
.
0.0986
.
0.0857
0.0764
0.0641
0.0465
0

Number
Failed

Number
Left

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0

SAS Output for Survival Estimation (contd)

Summary Statistics for Time Variable t
Quartile Estimates

Percent

Point
Estimate

95% Confidence Interval

[Lower
Upper)

75
50
25

12.0000
8.0000
4.0000

8.0000
4.0000
2.0000

Mean

Standard Error

8.6667

1.4114

17.0000
11.0000
8.0000

Summary of the Number of Censored and Uncensored Values

Total

Failed

Censored

Percent
Censored

21

21

0.00

39

Does anyone have a guess regarding how to calculate the standard error of the estimated survival?

+) = P (T > 8) = 8 = 0.381
S(8
21
(at t = 8+, we count the 4 events at time=8 as already
having failed)

+)] = 0.106
se[S(8

40

S-Plus Commands for Survival Estimation

> t_c(1,1,2,2,3,4,4,5,5,8,8,8,8,11,11,12,12,15,17,22,23)

> surv.fit(t,status=rep(1,21))
95 percent confidence interval
time n.risk n.event
survival
1
21
2 0.90476190
2
19
2 0.80952381
3
17
1 0.76190476
4
16
2 0.66666667
5
14
2 0.57142857
8
12
4 0.38095238
11
8
2 0.28571429
12
6
2 0.19047619
15
4
1 0.14285714
17
3
1 0.09523810
22
2
1 0.04761905
23
1
1 0.00000000

is of type "log"
std.dev lower 95% CI upper 95% CI
0.06405645
0.78753505
1.0000000
0.08568909
0.65785306
0.9961629
0.09294286
0.59988048
0.9676909
0.10286890
0.49268063
0.9020944
0.10798985
0.39454812
0.8276066
0.10597117
0.22084536
0.6571327
0.09858079
0.14529127
0.5618552
0.08568909
0.07887014
0.4600116
0.07636035
0.05010898
0.4072755
0.06405645
0.02548583
0.3558956
0.04647143
0.00703223
0.3224544
NA
NA
NA

41

Estimating the Survival Function

One-sample nonparametric methods:
We will consider three methods for estimating a survivorship
function
S(t) = P r(T t)
without resorting to parametric methods:

(1) Kaplan-Meier
(2) Life-table (Actuarial Estimator)
(3) via the Cumulative hazard estimator

42

(1) The Kaplan-Meier Estimator

The Kaplan-Meier (or KM) estimator is probably

the most popular approach. It can be justified
from several perspectives:
product limit estimator
likelihood justification

redistribute to the right estimator

We will start with an intuitive motivation based

on conditional probabilities, then review some of
the other justifications.

43

Motivation:
First, consider an example where there is no censoring.
The following are times of remission (weeks) for 21 leukemia
patients receiving control treatment (Table 1.1 of Cox &
Oakes):
1, 1, 2, 2, 3, 4, 4, 5, 5, 8, 8, 8, 8, 11, 11, 12, 12, 15, 17, 22, 23
How would we estimate S(10), the probability that an individual survives to time 10 or later?

What about S(8)?

Is it

12
21

or

8
21 ?

Lets construct a table of S(t):

Values of t
S(t)
t 1 21/21=1.000
1 < t 2 19/21=0.905
2 < t 3 17/21=0.809
3<t4
4<t5
5<t8
8 < t 11
11 < t 12
12 < t 15
15 < t 17
17 < t 22
22 < t 23

44

Empirical Survival Function:

When there is no censoring, the general formula is:
= # individuals with T t
S(t)
total sample size

Example for leukemia data (control arm):

45

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(6)= 21/21, because everyone survived at least

until time 6 or greater. But, we cant say S(7) = 17/21,
because we dont 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 nonparametrically estimate S(t), even in the presence of censoring.
The method is based on the ideas of conditional probability.

46

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 probability of all k events happening together can be written as
a product of conditional probabilities:
P (A1 A2... Ak ) = P (Ak |Ak1 ... A1)
P (Ak1 |Ak2 ... A1)
...
P (A2|A1)
P (A1)

47

Now, lets apply these ideas to estimate S(t):

Suppose ak < t ak+1 . Then
S(t) = P (T ak+1 )
= P (T a1, T a2, . . . , T ak+1 )
= P (T a1)
=
=

so S(t)
=

k
Y

j=1

P (T aj+1|T aj )

k
Y

[1 P (T = aj |T aj )]

k
Y

[1 j ]

j=1

j=1

k
Y

j=1

dj

1
rj

dj

=
rj
j:aj <t
Y

dj is the number of deaths at aj

rj is the number at risk at aj

48

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

Using the law of conditional probability,

P r(T t) =

Y
j

P r(survive j-th interval Ij | survived to start of Ij )

where the product is taken over all the intervals including or

preceding time t.

49

4 possibilities for each interval:

(1) No events (death or censoring) - conditional probability of surviving the interval is 1
(2) Censoring - assume they survive to the end of the interval, so that the conditional probability of surviving
the interval is 1
(3) Death, but no censoring - conditional probability
of not surviving the interval is # deaths (d) divided by #
at risk (r) at the beginning of the interval. So the conditional probability of surviving the interval is 1 (d/r).
(4) Tied deaths and censoring - assume censorings last
to the end of the interval, so that conditional 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 conditional
probability of surviving the j-th interval that holds for all 4
cases:
dj
1
rj

50

We could use the same approach by grouping the event times

into intervals (say, one interval for each month), and then
counting up the number of deaths (events) in each to estimate the probability of surviving the interval (this is called
the lifetable estimate).
However, the assumption that those censored last until the
end of the interval wouldnt be quite accurate, so we would
end up with a cruder approximation.
As the intervals get finer and finer, the approximations made
in estimating the probabilities of getting through each interval become smaller and smaller, so that the estimator converges to the true S(t).
This intuition clarifies why an alternative name for the KM
is the product limit estimator.

51

The Kaplan-Meier estimator of the survivorship

function (or survival probability) S(t) = P r(T t)
is:

S(t)
=

j:j <t

rj dj
rj

j:j <t

dj
rj

where
1, ...K is the set of K distinct death times observed in
the sample
dj is the number of deaths at j
rj is the number of individuals at risk right before the
j-th death time (everyone dead or censored at or after
that time).
cj is the number of censored observations between the
j-th and (j + 1)-st death times. Censorings tied at j
are included in cj
Note: two useful formulas are:
(1) rj = rj1 dj1 cj1
(2) rj =

lj

(cl + dl )

52

Calculating the KM - Cox and Oakes example

Make a table with a row for every death or censoring time:
j
6
7
9
10
11
13
16
17
19
20
22
23

dj
3
1
0

cj
1
0
1

1 (dj /rj )
18
21 = 0.857

rj
21
17
16

j+)
S(

Note that:
+) only changes at death (failure) times
S(t
+) is 1 up to the first death time
S(t

+) only goes to 0 if the last event is a death

S(t

53

KM plot for treated leukemia patients

Note: most statistical software packages summarize the KM survival function at j+, i.e., just after the time of the j-th failure.

+).
In other words, they provide S(
j
When there is no censoring, the empirical survival estimate
would then be:
+) = # individuals with T > t
S(t
total sample size
54

Output from STATA KM Estimator:

failure time:
failure/censor:

weeks
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
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

55

Two Other Justifications for KM Estimator

I. Likelihood-based derivation (Cox and Oakes)
For a discrete failure time variable, define:
dj number of failures at aj
rj number of individuals at risk at aj
(including those censored at aj ).
j Pr(death) in j-th interval
(conditional on survival to start of interval)
The likelihood is that of g independent binomials:
L() =

g
Y

j=1

j j (1 j )rj dj

Therefore, the maximum likelihood estimator of j

is:
j = dj /rj

Now we plug in the MLEs of to estimate S(t):

S(t)
=

j:aj <t

j)
(1

dj

1
=
rj
j:aj <t
Y

56

II. Redistribute to the right justification

(Efron, 1967)
is just the proportion of
In the absence of censoring, S(t)
individuals with T t. The idea behind Efrons approach
is to spread the contributions of censored observations out
over all the possible times to their right.
Algorithm:
Step (1): arrange the n observed times (deaths or censorings) in increasing order. If there are ties, put censored
after deaths.
Step (2): Assign weight (1/n) to each time.
Step (3): Moving from left to right, each time you encounter a censored observation, distribute its mass to all
times to its right.
Step (4): Calculate Sj by subtracting the final weight
for time j from Sj1

57

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)
Times
2
2.5+
3
4
4.5+
5
6
7

Step 2
Step 3a
1/9=0.11
1/9=0.11
0
2/9=0.22
0.25
1/9=0.11
0.13
1/9=0.11
0.13
1/9=0.11
0.13
1/9=0.11
0.13
1/9=0.11
0.13

Step 3b

0
0.17
0.17
0.17

(Step 4)
j)
S(
0.889
0.889
0.635
0.508
0.508
0.339
0.169
0.000

This comes out the same as the product limit approach.

58

Properties of the KM estimator

In the case of no censoring:
= S(t)
= # deaths at t or greater
S(t)
n
where n is the number of individuals in the study.
This is just like an estimated probability from a binomial
distribution, so we have:
' N (S(t), S(t)[1 S(t)]/n)
S(t)
How does censoring affect this?
is still approximately normal
S(t)

converges to the true S(t)

The mean of S(t)
The variance is a bit more complicated (since the denominator n includes some censored observations).
Once we get the variance, then we can construct (pointwise)

(1 )% confidence intervals (NOT bands) about S(t):

z1/2 se[S(t)]

S(t)

59

Greenwoods formula (Collett 2.1.3)

We can think of the KM estimator as
=
S(t)

j:j <t

j)
(1

j = dj /rj .
where
j s are just binomial proportions, we can apply
Since the
j is approxistandard likelihood theory to show that each
mately normal, with mean the true j , and
j (1
j)

var(j )
rj
j s are independent in large enough samples.
Also, the
is a function of the j s, we can estimate its variSince S(t)
ance using the delta method:
Delta method: If Y is normal with mean and
variance 2, then g(Y ) is approximately normally
distributed with mean g() and variance [g 0()]2 2.

60

Two specific examples of the delta method:

(A) Z = log(Y )

2
1

then Z N log(), 2

(B) Z = exp(Y )

2 2

then Z N e , [e ]

The examples above use the following results from calculus:

1
d
log u =
dx
u
d u
e = eu
dx

61

du

dx
du

dx

Greenwoods formula (continued)

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

log[S(t)]
=

j:j <t

j)
log(1

j s,
Thus, by approximate independence of the

var(log[S(t)])
=
by (A)

=
=
=
=

j:j <t

j )]
var[log(1

j)

var(

j:j <t 1 j

2
1
X
j )/rj

j (1

j:j <t 1 j
j

X
j )rj
j:j <t (1
dj
X
j:j <t (rj dj )rj
X

= exp[log[S(t)]].

Now, S(t)
Thus by (B),

2var log[S(t)]

var(S(t))
= [S(t)]

Greenwoods Formula:

2 Pj: <t dj
var(S(t))
= [S(t)]
j
(rj dj )rj
62

Back to confidence intervals

For a 95% confidence interval, we could use
z1/2 se[S(t)]

S(t)

where se[S(t)]
is calculated using Greenwoods formula.
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 proper range when we transform back.
To see why this works, note the following:
is an estimated probability
Since S(t)
1
0 S(t)

has bounds:
Taking the log of S(t)

log[S(t)]
0
Taking the opposite:

0 log[S(t)]

Taking the log again:

log log[S(t)]

To transform back, reverse steps with S(t) = exp( exp(L(t))

63

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 L(t),

A, L(t)
+ A]
yielding [L(t)
(3) Since S(t) = exp( exp(L(t)), the confidence bounds
for the 95% CI on S(t) are:

[exp(e(L(t)+A) ), exp(e(L(t)A) )]
(note that the upper and lower bounds switch)
= log( log(S(t)))

(4) Substituting L(t)

back into the above
bounds, we get confidence bounds of
eA , [S(t)]
eA )
([S(t)]

64

What is A?

A is 1.96 se(L(t))
To calculate this, we need to calculate

var(L(t))
= var log( log(S(t)))

From our previous calculations, we know

var(log[S(t)])
=

dj
j:j <t (rj dj )rj
X

Applying the delta method as in example (A), we get:

var(L(t))
= var(log( log[S(t)]))
dj
1
X
=
2 j:j <t (rj dj )rj
[log S(t)]

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

and then form the confidence intervals as:

e1.96 se(L(t))

S(t)

This is the approach that Stata uses. Splus gives an option to calculate these bounds (use conf.type=log-log
in surv.fit).

65

Summary of Confidence Intervals on S(t)

1.96 se[S(t)]

Calculate S(t)
where se[S(t)]
is calculated using Greenwoods formula, and replace negative
lower bounds by 0 and upper bounds greater than 1 by
1.
Recommended by Collett
This is the default using SAS
not very satisfactory
Use a log transformation to stabilize the variance and
allow for non-symmetric confidence intervals. This is
what is normally done for the confidence interval of an
estimated odds ratio.
d
P

Use var[log(S(t))]
= j:j <t (rj dj j )rj already calculated as part of Greenwoods formula

This is the default in Splus

Use the log-log transformation just described
Somewhat complicated, but always yields proper bounds
This is the default in Stata.

66

Software for Kaplan-Meier Curves

Stata - stset and sts commands
SAS - proc lifetest
Splus - surv.fit(time,status)

Defaults for Confidence Interval Calculations

1.96 se[L(t)]

Stata - log-log L(t)

where L(t) = log[ log(S(t))]

1.96 se[S(t)]

SAS - plain S(t)

Splus - log log S(t) 1.96 se[log(S(t))]

but Splus will also give either of the other two options if
you request them.

67

Stata Commands
Create a file called leukemia.dat with the raw data, with
a column for treatment, weeks to relapse (i.e., duration of
remission), and relapse status:
.infile trt remiss status using leukemia.dat
.stset remiss status

(sets up a failure time dataset,

with failtime status in that order,
type help stset to get details)

.sts list

(estimated S(t), se[S(t)], and 95% CI)

.sts graph, saving(kmtrt)

(creates a Kaplan-Meier plot, and

saves the plot in file kmtrt.gph,
type help gphdot to get some
printing instructions)

.graph using kmtrt

(redisplays the graph at any later time)

If the dataset has already been created and loaded into Stata,
then you can substitute the following commands for initializing the data:
.use leukem

(finds Stata dataset leukem.dta)

.describe

(provides a description of the dataset)

.stset remiss status

(declares data to be failure type)

.stdes

(gives a description of the survival dataset)

68

STATA Output for Treated Leukemia Patients:

.use leukem
.stset remiss status if trt==1
.sts list
failure time:
failure/censor:

remiss
status

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
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

69

SAS Commands for Kaplan Meier Estimator PROC LIFETEST

The SAS command for the Kaplan-Meier estimate is:
or

time failtime*censor(1);
time failtime*failind(0);

The first variable is the failure time, and the second is the
failure or censoring indicator. In parentheses you need to put
the specific numeric value that corresponds to censoring.
are included
The upper and lower confidence limits on S(t)
in the data set OUTSURV when specified. The upper and
lower limits are called: sdf ucl, sdf lcl.
data leukemia;
input weeks remiss;
label weeks=Time to Remission (in weeks)
remiss=Remission indicator (1=yes,0=no);
cards;
6 1
6 1
........... ( lines edited out here)
34 0
35 0
;
proc lifetest data=leukemia outsurv=confint;
time weeks*remiss(0);
title Leukemia data from Table 1.1 of Cox and Oakes;
run;
proc print data=confint;
title 95% Confidence Intervals for Estimated Survival;

70

Output from SAS Proc Lifetest

Note: this information is not printed if you use NOPRINT.
Leukemia data from Table 1.1 of Cox and Oakes
The LIFETEST Procedure
Product-Limit Survival Estimates

WEEKS
0.0000
6.0000
6.0000
6.0000
6.0000*
7.0000
9.0000*
10.0000
10.0000*
11.0000*
13.0000
16.0000
17.0000*
19.0000*
20.0000*
22.0000
23.0000
25.0000*
32.0000*
32.0000*
34.0000*
35.0000*

Survival
1.0000
.
.
0.8571
.
0.8067
.
0.7529
.
.
0.6902
0.6275
.
.
.
0.5378
0.4482
.
.
.
.
.

Failure

Survival
Standard
Error

0
0
.
.
.
.
0.1429
0.0764
.
.
0.1933
0.0869
.
.
0.2471
0.0963
.
.
.
.
0.3098
0.1068
0.3725
0.1141
.
.
.
.
.
.
0.4622
0.1282
0.5518
0.1346
.
.
.
.
.
.
.
.
.
.
* Censored Observation

71

Number
Failed

Number
Left

0
1
2
3
3
4
4
5
5
5
6
7
7
7
7
8
9
9
9
9
9
9

21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0

Output from printing the CONFINT file

95% Confidence Intervals for Estimated Survival
OBS

WEEKS

_CENSOR_

SURVIVAL

SDF_LCL

SDF_UCL

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

0
6
6
7
9
10
10
11
13
16
17
19
20
22
23
25
32
32
34
35

0
0
1
0
1
0
1
1
0
0
1
1
1
0
0
1
1
1
1
1

1.00000
0.85714
0.85714
0.80672
0.80672
0.75294
0.75294
0.75294
0.69020
0.62745
0.62745
0.62745
0.62745
0.53782
0.44818
.
.
.
.
.

1.00000
0.70748
.
0.63633
.
0.56410
.
.
0.48084
0.40391
.
.
.
0.28648
0.18439
.
.
.
.
.

1.00000
1.00000
.
0.97711
.
0.94178
.
.
0.89955
0.85099
.
.
.
0.78915
0.71197
.
.
.
.
.

The output dataset will have one observation for each unique
combination of weeks and censor . It will also add an
observation for failure time equal to 0.

72

Splus Commands
Create a file called leukemia.dat with the variables names
in the first row, as follows:
t
c
6
1
6
1
etc ...
In Splus, type
y_read.table(leukemia.dat,header=T)
surv.fit(y$t,y$c)
plot(surv.fit(y$t,y$c))
(the plot command will also yield 95% confidence intervals)
To specify the type of confidence intervals, use the conf.type=
option in the surv.fit statements: e.g. conf.type=log-log
or conf.type=plain

73

>surv.fit(y$t,y$c)
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

> surv.fit(y$t,y$c,conf.type="log-log")
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

> surv.fit(y$t,y$c,conf.type="plain")
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

74

0.0

0.2

Survival
0.4
0.6

0.8

1.0

KM Survival Estimate and Confidence intervals

(SPlus)

10

15

20
Time

75

25

30

35

Means, Medians, Quantiles based on the KM

Mean:

Pk

j=1 j

P r(T = j )

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

S( ) = 0.5. However, in practice, it is defined as the
) 0.5. The median is more
smallest time such that S(
appropriate for censored survival data than the mean.
For the treated leukemia patients, we find:

S(22)
= 0.5378

S(23)
= 0.4482
The median is thus 23. This can also be seen visually on
the graph to the left.
Lower quartile (25th percentile):

the smallest time (LQ) such that S(LQ)

0.75
Upper quartile (75th percentile):
Q) 0.25
the smallest time (UQ) such that S(U

76

The (2) Lifetable Estimator of Survival:

We said that we would consider the following three methods
for estimating a survivorship function
S(t) = P r(T t)
without resorting to parametric methods:

(1)

Kaplan-Meier

(2) = Life-table (Actuarial Estimator)

(3) = Cumulative hazard estimator

77

(2) 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 complicated by the fact that we
dont know exactly when during each time interval an event
occurs.

78

Lee (section 4.2) provides a good description of lifetable

methods, and distinguishes several types 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 information 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 Tables.

Clinical Life tables - applies to grouped survival data

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

79

Notation
the j-th time interval is [tj1, 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 (Lee, p.91)

Year after
Diagnosis
[0, 1)
[1, 2)
[2, 3)
[3, 4)
[4, 5)
[5, 6)
[6, 7)
etc..

j
1
2
3
4
5
6
7

d j cj
456 0
226 39
152 22
171 23
135 24
125 107
83 133

rj
2418
1962
1697
1523
1329
1170
938

80

rj0 = rj cj /2
2418.0
1942.5 (1962 - 39
2)
1686.0
1511.5
1317.0
1116.5
871.5

Estimating the survivorship function

We could apply the K-M formula directly to the numbers in
the table on the previous page, estimating S(t) as

dj

S(t)
=
rj
j:j <t
Y

However, this approach is unsatisfactory for grouped data....

it treats the problem as though it were in discrete time, with
events happening only at 1 yr, 2 yr, etc. In fact, what we
are trying to calculate here is the conditional probability of
dying within the interval, given survival to the beginning of
it.
What should we do with the censored people?
We can 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 interval.
81

Constructing the lifetable

First, some additional notation for the j-th interval, [t j1, tj ):
Midpoint (tmj ) - useful for plotting the density and
the hazard function
Width (bj = tj tj1) needed for calculating the hazard
in the j-th interval
Quantities estimated:
Conditional probability of dying

qj = dj /rj0

Conditional probability of surviving

pj = 1 qj
Cumulative probability of surviving at tj :
j) =
S(t

`j

p`

d`
1
=
r `0
`j
Y

82

Some important points to note:

Because the intervals are defined as [tj1, tj ), the first
interval typically starts with t0 = 0.
Stata estimates the survival function at the right-hand
endpoint of each interval, i.e., S(tj )
However, SAS estimates the survival function at the lefthand endpoint, S(tj1 ).
0 ) = 1 and S(t
1 ) = p1
The implication in SAS is that S(t

83

Other quantities estimated at the

midpoint of the j-th interval:
Hazard in the j-th interval:
mj ) =
(t
=

dj
bj (rj0 dj /2)
qj
bj (1 qj /2)

the number of deaths in the interval divided by the average number of survivors at the midpoint
density at the midpoint of the j-th interval:
j1 ) S(t
j)
S(t

f (tmj ) =
bj
j1 ) qj
S(t
=
bj
Note: Another way to get this is:
mj )S(t
mj )
f(tmj ) = (t
mj )[S(t
j ) + S(t
j1 )]/2
= (t

84

Constructing the Lifetable using Stata

Uses the ltable command.
If the raw data are already grouped, then the freq statement
must be used when reading the data.
. infile years status count using angina.dat
(32 observations read)
. ltable years status [freq=count]

Beg.
Std.
Interval
Total Deaths Lost
Survival
Error
[95% Conf. Int.]
------------------------------------------------------------------------0
1
2418
456
0
0.8114
0.0080
0.7952
0.8264
1
2
1962
226
39
0.7170
0.0092
0.6986
0.7346
2
3
1697
152
22
0.6524
0.0097
0.6329
0.6711
3
4
1523
171
23
0.5786
0.0101
0.5584
0.5981
4
5
1329
135
24
0.5193
0.0103
0.4989
0.5392
5
6
1170
125
107
0.4611
0.0104
0.4407
0.4813
6
7
938
83
133
0.4172
0.0105
0.3967
0.4376
7
8
722
74
102
0.3712
0.0106
0.3505
0.3919
8
9
546
51
68
0.3342
0.0107
0.3133
0.3553
9
10
427
42
64
0.2987
0.0109
0.2775
0.3201
10
11
321
43
45
0.2557
0.0111
0.2341
0.2777
11
12
233
34
53
0.2136
0.0114
0.1917
0.2363
12
13
146
18
33
0.1839
0.0118
0.1614
0.2075
13
14
95
9
27
0.1636
0.0123
0.1404
0.1884
14
15
59
6
23
0.1429
0.0133
0.1180
0.1701
15
16
30
0
30
0.1429
0.0133
0.1180
0.1701
-------------------------------------------------------------------------------

85

j,
It is also possible to get estimates of the hazard function,
and its standard error using the hazard option:
. ltable years status [freq=count], hazard

Beg.
Cum.
Std.
Std.
Interval
Total
Failure
Error
Hazard
Error
[95% Conf Int]
-------------------------------------------------------------------------0
1
2418
0.1886 0.0080
0.2082
0.0097
0.1892 0.2272
1
2
1962
0.2830 0.0092
0.1235
0.0082
0.1075 0.1396
2
3
1697
0.3476 0.0097
0.0944
0.0076
0.0794 0.1094
3
4
1523
0.4214 0.0101
0.1199
0.0092
0.1020 0.1379
4
5
1329
0.4807 0.0103
0.1080
0.0093
0.0898 0.1262
5
6
1170
0.5389 0.0104
0.1186
0.0106
0.0978 0.1393
6
7
938
0.5828 0.0105
0.1000
0.0110
0.0785 0.1215
7
8
722
0.6288 0.0106
0.1167
0.0135
0.0902 0.1433
8
9
546
0.6658 0.0107
0.1048
0.0147
0.0761 0.1336
9
10
427
0.7013 0.0109
0.1123
0.0173
0.0784 0.1462
10
11
321
0.7443 0.0111
0.1552
0.0236
0.1090 0.2015
11
12
233
0.7864 0.0114
0.1794
0.0306
0.1194 0.2395
12
13
146
0.8161 0.0118
0.1494
0.0351
0.0806 0.2182
13
14
95
0.8364 0.0123
0.1169
0.0389
0.0407 0.1931
14
15
59
0.8571 0.0133
0.1348
0.0549
0.0272 0.2425
15
16
30
0.8571 0.0133
0.0000
.
.
.
-------------------------------------------------------------------------

There is also a failure option which gives the number of

failures (like the default), and also provides a 95% confidence
interval on the cumulative failure probability.

86

Constructing the lifetable using SAS

If the raw data are already grouped, then the FREQ statement must be used when reading the data.
SAS requires that the interval endpoints be specified, using
one of the following (see SAS manual or online help for more
detail):
intervals - specify the the interval endpoints
width - specify the width of each interval
ninterval - specify the number of intervals
Title Actuarial Estimator for Angina Pectoris Example;
data angina;
input years status count;
cards;
0.5 1 456
1.5 1 226
2.5 1 152
/* angina cases */
3.5 1 171
4.5 1 135
5.5 1 125
.
.
0.5 0
0
1.5 0
39
2.5 0
22
/* censored */
3.5 0
23
4.5 0
24
5.5 0 107
.
.
proc lifetest data=angina outsurv=survres intervals=0 to 15 by 1 method=act;
time years*status(0);
freq count;

87

SAS output:
Actuarial Estimator for Angina Pectoris Example
The LIFETEST Procedure
Life Table Survival Estimates

Interval
[Lower, Upper)
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
.

Interval
[Lower, Upper)
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
.

Number
Failed

Number
Censored

Effective
Sample
Size

456
226
152
171
135
125
83
74
51
42
43
34
18
9
6
0

0
39
22
23
24
107
133
102
68
64
45
53
33
27
23
30

2418.0
1942.5
1686.0
1511.5
1317.0
1116.5
871.5
671.0
512.0
395.0
298.5
206.5
129.5
81.5
47.5
15.0

Survival

Failure

Survival
Standard
Error

1.0000
0.8114
0.7170
0.6524
0.5786
0.5193
0.4611
0.4172
0.3712
0.3342
0.2987
0.2557
0.2136
0.1839
0.1636
0.1429

0
0.1886
0.2830
0.3476
0.4214
0.4807
0.5389
0.5828
0.6288
0.6658
0.7013
0.7443
0.7864
0.8161
0.8364
0.8571

0
0.00796
0.00918
0.00973
0.0101
0.0103
0.0104
0.0105
0.0106
0.0107
0.0109
0.0111
0.0114
0.0118
0.0123
0.0133

88

Conditional
Probability
of Failure
0.1886
0.1163
0.0902
0.1131
0.1025
0.1120
0.0952
0.1103
0.0996
0.1063
0.1441
0.1646
0.1390
0.1104
0.1263
0

Conditional
Probability
Standard
Error
0.00796
0.00728
0.00698
0.00815
0.00836
0.00944
0.00994
0.0121
0.0132
0.0155
0.0203
0.0258
0.0304
0.0347
0.0482
0

Median
Residual
Lifetime

Median
Standard
Error

5.3313
6.2499
6.3432
6.2262
6.2185
5.9077
5.5962
5.1671
4.9421
4.8258
4.6888
.
.
.
.
.

0.1749
0.2001
0.2361
0.2361
0.1853
0.1806
0.1855
0.2713
0.2763
0.4141
0.4183
.
.
.
.
.

j)
more SAS output: (estimated density fj and hazard
Evaluated at the Midpoint of the Interval

Interval
[Lower, Upper)
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
.

PDF

PDF
Standard
Error

Hazard

Hazard
Standard
Error

0.1886
0.0944
0.0646
0.0738
0.0593
0.0581
0.0439
0.0460
0.0370
0.0355
0.0430
0.0421
0.0297
0.0203
0.0207
.

0.00796
0.00598
0.00507
0.00543
0.00495
0.00503
0.00469
0.00518
0.00502
0.00531
0.00627
0.00685
0.00668
0.00651
0.00804
.

0.208219
0.123531
0.09441
0.119916
0.108043
0.118596
0.1
0.116719
0.10483
0.112299
0.155235
0.17942
0.149378
0.116883
0.134831
.

0.009698
0.008201
0.007649
0.009154
0.009285
0.010589
0.010963
0.013545
0.014659
0.017301
0.023602
0.030646
0.03511
0.038894
0.054919
.

Summary of the Number of Censored and Uncensored Values

Total

Failed

Censored

%Censored

2418

1625

793

32.7957

89

Suppose we wish to use the actuarial method, but the data

do not come grouped.
Consider the treated nursing home patients, with length of
stay (los) grouped into 100 day intervals:

.use nurshome
.drop if rx==0
(881 observations deleted)

(keep only the treated patients)

.stset los fail

.ltable los fail, intervals(100)
Beg.
Std.
Interval
Total Deaths
Lost
Survival
Error
[95% Conf. Int.]
-----------------------------------------------------------------------0
100
710
328
0
0.5380
0.0187
0.5006
0.5739
100
200
382
86
0
0.4169
0.0185
0.3805
0.4529
200
300
296
65
0
0.3254
0.0176
0.2911
0.3600
300
400
231
38
0
0.2718
0.0167
0.2396
0.3050
400
500
193
32
1
0.2266
0.0157
0.1966
0.2581
500
600
160
13
0
0.2082
0.0152
0.1792
0.2388
600
700
147
13
0
0.1898
0.0147
0.1619
0.2195
700
800
134
10
30
0.1739
0.0143
0.1468
0.2029
800
900
94
4
29
0.1651
0.0143
0.1383
0.1941
900 1000
61
4
30
0.1508
0.0147
0.1233
0.1808
1000 1100
27
0
27
0.1508
0.0147
0.1233
0.1808
-------------------------------------------------------------------------

90

SAS Commands for lifetable analysis - grouping

data

Title Actuarial Estimator for nursing home data;

data morris ;
infile ch12.dat ;
input los age trt gender marstat hltstat cens ;
data morristr;
set morris;
if trt=1;
proc lifetest data=morristr outsurv=survres
intervals=0 to 1100 by 100 method=act;
time los*cens(1);
run ;
proc print data=survres;
run;

91

Actuarial estimator for treated nursing home patients

Actuarial Estimator for Nursing Home Patients
The LIFETEST Procedure
Life Table Survival Estimates

Interval
[Lower, Upper)
0
100
200
300
400
500
600
700
800
900
1000

100
200
300
400
500
600
700
800
900
1000
1100

Interval
[Lower, Upper)
0
100
200
300
400
500
600
700
800
900
1000

100
200
300
400
500
600
700
800
900
1000
1100

Number
Failed

Number
Censored

330
86
65
38
32
13
13
10
4
4
0

Conditional
Probability
Standard
Error
0.0187
0.0214
0.0241
0.0244
0.0268
0.0216
0.0234
0.0254
0.0245
0.0415
0

0
0
0
0
1
0
0
30
29
30
27

Effective
Sample
Size
712.0
382.0
296.0
231.0
192.5
160.0
147.0
119.0
79.5
46.0
13.5

Conditional
Probability
of Failure
0.4635
0.2251
0.2196
0.1645
0.1662
0.0813
0.0884
0.0840
0.0503
0.0870
0

Survival

Failure

Survival
Standard
Error

1.0000
0.5365
0.4157
0.3244
0.2711
0.2260
0.2076
0.1893
0.1734
0.1647
0.1503

0
0.4635
0.5843
0.6756
0.7289
0.7740
0.7924
0.8107
0.8266
0.8353
0.8497

0
0.0187
0.0185
0.0175
0.0167
0.0157
0.0152
0.0147
0.0143
0.0142
0.0147

92

Median
Residual
Lifetime
130.2
306.2
398.8
617.0
.
.
.
.
.
.
.

Actuarial estimator for treated nursing home patients, contd

Evaluated at the Midpoint
of the Interval

Interval
[Lower, Upper)
0
100
200
300
400
500
600
700
800
900
1000

100
200
300
400
500
600
700
800
900
1000
1100

Median
Standard
Error
15.5136
30.4597
65.7947
74.5466
.
.
.
.
.
.
.

PDF

PDF
Standard
Error

Hazard

Hazard
Standard
Error

0.00463
0.00121
0.000913
0.000534
0.000451
0.000184
0.000184
0.000159
0.000087
0.000143
0

0.000187
0.000122
0.000108
0.000084
0.000078
0.00005
0.00005
0.00005
0.000043
0.00007
.

0.006033
0.002537
0.002467
0.001792
0.001813
0.000847
0.000925
0.000877
0.000516
0.000909
0

0.000317
0.000271
0.000304
0.00029
0.000319
0.000235
0.000256
0.000277
0.000258
0.000454
.

Summary of the Number of Censored and Uncensored Values

Total

Failed

Censored

%Censored

712

595

117

16.4326

93

Actuarial estimator for treated nursing home patients, contd

Output from SURVRES dataset
Actuarial Estimator for Nursing Home Patients
OBS

LOS

SURVIVAL

SDF_LCL

SDF_UCL

MIDPOINT

PDF

1
2
3
4
5
6
7
8
9
10
11

0
100
200
300
400
500
600
700
800
900
1000

1.00000
0.53652
0.41573
0.32444
0.27107
0.22601
0.20764
0.18928
0.17337
0.16465
0.15033

1.00000
0.49989
0.37953
0.29005
0.23842
0.19528
0.17783
0.16048
0.14536
0.13677
0.12157

1.00000
0.57315
0.45193
0.35883
0.30372
0.25674
0.23745
0.21808
0.20139
0.19253
0.17910

50
150
250
350
450
550
650
750
850
950
1050

.0046348
.0012079
.0009129
.0005337
.0004506
.0001836
.0001836
.0001591
.0000872
.0001432
.0000000

OBS

PDF_LCL

PDF_UCL

HAZARD

HAZ_LCL

HAZ_UCL

1
2
3
4
5
6
7
8
9
10
11

.0042685
.0009685
.0007014
.0003686
.0002981
.0000847
.0000847
.0000617
.0000027
.0000069
.

.0050011
.0014472
.0011245
.0006988
.0006031
.0002825
.0002825
.0002565
.0001717
.0002794
.

.0060329
.0025369
.0024668
.0017925
.0018130
.0008469
.0009253
.0008772
.0005161
.0009091
.0000000

.0054123
.0020050
.0018717
.0012248
.0011874
.0003869
.0004228
.0003340
.0000105
.0000191
.

.0066535
.0030687
.0030619
.0023601
.0024386
.0013069
.0014277
.0014203
.0010218
.0017991
.

94

Examples for Nursing home data:

Estimated Survival:

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

95

Estimated hazard:

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

96

(3) Estimating the cumulative hazard

(Nelson-Aalen estimator)
Suppose we want to estimate (t) =
tive hazard at time t.

Rt

(u)du, the cumula-

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

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

(t) can then be approximated by a sum:

= X j
(t)
j

where the sum is over intervals, j is the value of the hazard

in the j-th interval and is the width of each interval. Since
is approximately the probability of dying in the interval,

we can further approximate by

= X dj /rj
(t)
j

It follows that (t) will change only at death times, and

hence we write the Nelson-Aalen estimator as:
N A(t) =

j:j <t

97

dj /rj

rj

D
n

dj
cj

0
0

0
0

1
0

j) 0
(t

1/n 0

j) 0
(t

1/n 1/n 1/n 1/n 1/n

n1
0
0

C
n1
0
1
0

n-2

C
n-2

D
n-3

D
n-4

0
0

0
1

1
0

1
0

1
n3

1
n4

N A(t), we can also find another estimator of

Once we have
S(t) (Fleming-Harrington):
N A(t))
SF H (t) = exp(
In general, this estimator of the survival function will be
close to the Kaplan-Meier estimator, SKM (t)
We can also go the other way ... we can take the KaplanMeier estimate of S(t), and use it to calculate an alternative
estimate of the cumulative hazard function:
KM (t) = log SKM (t)

98

Stata commands for FH Survival Estimate

Say we want to obtain the Fleming-Harrington estimate of
the survival function for married females, in the healthiest
initial subgroup, who are randomized to the untreated group
of the nursing home study.
First, we use the following commands to calculate the NelsonAalen cumulative hazard estimator:
. use nurshome
. keep if rx==0 & gender==0 & health==2 & married==1
(1579 observations deleted)
. sts list, na
failure _d:
analysis time _t:

fail
los

Beg.
Net
Nelson-Aalen
Std.
Time
Total
Fail
Lost
Cum. Haz.
Error
[95% Conf. Int.]
---------------------------------------------------------------------14
12
1
0
0.0833
0.0833
0.0117
0.5916
24
11
1
0
0.1742
0.1233
0.0435
0.6976
25
10
1
0
0.2742
0.1588
0.0882
0.8530
38
9
1
0
0.3854
0.1938
0.1438
1.0326
64
8
1
0
0.5104
0.2306
0.2105
1.2374
89
7
1
0
0.6532
0.2713
0.2894
1.4742
113
6
1
0
0.8199
0.3184
0.3830
1.7551
123
5
1
0
1.0199
0.3760
0.4952
2.1006
149
4
1
0
1.2699
0.4515
0.6326
2.5493
168
3
1
0
1.6032
0.5612
0.8073
3.1840
185
2
1
0
2.1032
0.7516
1.0439
4.2373
234
1
1
0
3.1032
1.2510
1.4082
6.8384
----------------------------------------------------------------------

99

After generating the Nelson-Aalen estimator, we manually

have to create a variable for the survival estimate:
. sts gen nelson=na
. gen sfh=exp(-nelson)
. list sfh

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

sfh
.9200444
.8400932
.7601478
.6802101
.6002833
.5203723
.4404857
.3606392
.2808661
.2012493
.1220639
.0449048

Additional built-in functions can be used to generate 95%

confidence intervals on the FH survival estimate.

100

We can compare the Fleming-Harrington survival estimate

to the KM estimate by rerunning the sts list command:
. sts list
. sts gen skm=s
. list skm sfh

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

skm
.91666667
.83333333
.75
.66666667
.58333333
.5
.41666667
.33333333
.25
.16666667
.08333333
0

sfh
.9200444
.8400932
.7601478
.6802101
.6002833
.5203723
.4404857
.3606392
.2808661
.2012493
.1220639
.0449048

In this example, it looks like the Fleming-Harrington estimator is slightly higher than the KM at every time point, but
with larger datasets the two will typically be much closer.

101

Splus Commands for Fleming-Harrington Estimator:

(Nursing home data: females, untreated, married, healthy)
Fleming-Harrington:
>fh<-surv.fit(los,cens,type="f",conf.type="log-log")
>fh
95 percent confidence interval is of type "log-log"
time n.risk n.event
survival
std.dev lower 95% CI upper 95% CI
14
12
1 0.9200444 0.08007959 0.5244209125
0.9892988
24
11
1 0.8400932 0.10845557 0.4750041174
0.9600371
25
10
1 0.7601478 0.12669130 0.4055610500
0.9200425
38
9
1 0.6802101 0.13884731 0.3367907188
0.8724502
64
8
1 0.6002833 0.14645413 0.2718422278
0.8187596
89
7
1 0.5203723 0.15021856 0.2115701242
0.7597900
113
6
1 0.4404857 0.15045450 0.1564397006
0.6960354
123
5
1 0.3606392 0.14723033 0.1069925657
0.6278888
149
4
1 0.2808661 0.14043303 0.0640979523
0.5560134
168
3
1 0.2012493 0.12990589 0.0293208029
0.4827590
185
2
1 0.1220639 0.11686728 0.0058990525
0.4224087
234
1
1 0.0449048 0.06216787 0.0005874321
0.2740658

Kaplan-Meier:
>km<-surv.fit(los,cens,conf.type="log-log")
>km
95 percent confidence interval
time n.risk n.event survival
14
12
1 0.91666667
24
11
1 0.83333333
25
10
1 0.75000000
38
9
1 0.66666667
64
8
1 0.58333333
89
7
1 0.50000000
113
6
1 0.41666667
123
5
1 0.33333333
149
4
1 0.25000000
168
3
1 0.16666667
185
2
1 0.08333333
234
1
1 0.00000000

is of type "log-log"
std.dev lower 95% CI upper 95% CI
0.07978559 0.538977181
0.9878256
0.10758287 0.481714942
0.9555094
0.12500000 0.408415913
0.9117204
0.13608276 0.337018933
0.8597118
0.14231876 0.270138924
0.8009402
0.14433757 0.208477143
0.7360731
0.14231876 0.152471264
0.6653015
0.13608276 0.102703980
0.5884189
0.12500000 0.060144556
0.5047588
0.10758287 0.026510427
0.4129803
0.07978559 0.005052835
0.3110704
NA
NA
NA
102

Comparison of Survival Curves

We spent the last class looking at some nonparametric ap
proaches for estimating the survival function, S(t),
over time
for a single sample of individuals.
Now we want to compare the survival estimates between two
groups.
Example: Time to remission of leukemia patients

103

How can we form a basis for comparison?

At a specific point in time, we could see whether the confidence intervals for the survival curves overlap.
However, the confidence intervals we have been calculating
are pointwise they correspond to a confidence inter ) at a single point in time, t.
val for S(t
In other words, we cant say that the true survival function
S(t) is contained between the pointwise confidence intervals
with 95% probability.

(Aside: if youre interested, the issue of confidence bands

for the estimated survival function are discussed in Section
4.4 of Klein and Moeschberger)

104

 ) overlap
Looking at whether the confidence intervals for S(t
between the 6MP and placebo groups would only focus on
comparing the two treatment groups at a single point in
time, t. We want an overall comparison.
on:
Should we base our overall comparison of S(t)
the furthest distance between the two curves?
the median survival for each group?
the average hazard? (for exponential distributions, this
would be like comparing the mean event times)
adding up the difference between the two survival estimates over time?
X

jA) S(t
jB )
S(t

a weighted sum of differences, where the weights reflect

the number at risk at each time?
a rank-based test? i.e., we could rank all of the event
times, and then see whether the sum of ranks for one
group was less than the other.

105

Nonparametric comparisons of groups

All of these are pretty reasonable options, and well see that
there have been several proposals for how to compare the
survival of two groups. For the moment, we are sticking to
nonparametric comparisons.

Why nonparametric?
fairly robust
efficient relative to parametric tests
often simple and intuitive

Before continuing the description of the two-sample comparison, Im going to try to put this in a general framework to
give a perspective of where were heading in this class.

106

General Framework for Survival Analysis

We observe (Xi, i, Zi) for individual i, where
Xi is a censored failure time random variable
i is the failure/censoring indicator
Zi represents a set of covariates
Note that Zi might be a scalar (a single covariate, say treatment or gender) or may be a (p 1) vector (representing
several different covariates).
These covariates might be:
continuous
discrete
time-varying (more later)
If Zi is a scalar and is binary, then we are comparing the
survival of two groups, like in the leukemia example.
More generally though, it is useful to build a model that
characterizes the relationship between survival and all of the
covariates of interest.

107

Well proceed as follows:

Two group comparisons
Multigroup and stratified comparisons - stratified logrank
Failure time regression models
Cox proportional hazards model
Accelerated failure time model

108

Two sample tests

Mantel-Haenszel logrank test
Peto & Petos version of the logrank test
Gehans Generalized Wilcoxon
Peto & Petos and Prentices generalized Wilcoxon
Tarone-Ware and Fleming-Harrington classes
Coxs F-test (non-parametric version)

References:
Hosmer & Lemeshow
Collett
Klein & Moeschberger
Kleinbaum
Lee

Section 2.4
Section 2.5
Section 7.3
Chapter 2
Chapter 5

109

Mantel-Haenszel Logrank test

The logrank test is the most well known and widely used.
It also has an intuitive appeal, building on standard methods for binary data. (Later we will see that it can also be
obtained as the score test from a partial likelihood from the
Cox Proportional Hazards model.)

First consider the following (2 2) table classifying those

with and without the event of interest in a two group setting:
Event
Group
0
1
Total

Yes
d0
d1
d

No
n0 d 0
n1 d 1
nd

110

Total
n0
n1
n

If the margins of this table are considered fixed, then d0

?
distribution. Under
follows a
the null hypothesis of no association between the event and
group, it follows that
E(d0 ) =
V ar(d0) =

n0 d
n
n0 n1 d(n d)
n2(n 1)

Therefore, under H0:

2M H =

[d0 n0 d/n]2
n0 n1 d(nd)
n2 (n1)

21

This is the Mantel-Haenszel statistic and is approximately

equivalent to the Pearson 2 test for equality of the two
groups given by:
2p

(o e)2
e

Note: recall that the Pearson 2 test was derived for the
case where only the row margins were fixed, and thus the
variance above was replaced by:
V ar(d0

n0(d0 + d1)
n0 n1 d(n d)
) =
n
n3

111

Example: Toxicity in a clinical trial with two treatments

Group
0
1
Total

Toxicity
Yes
No
8
42
2
48
10
90

Total
50
50
100

2p = 4.00

(p = 0.046)

2M H = 3.96

(p = 0.047)

112

Now suppose we have K (22) tables, all independent, and

we want to test for a common group effect. The CochranMantel-Haenszel test for a common odds ratio not equal to
1 can be written as:

2CM H

2
[ K
j=1 (d0j n0j dj /nj )]
= PK
2
j=1 n1j n0j dj (nj dj )/[nj (nj 1)]

where the subscript j refers to the j-th table:

Event
Group
0
1
Total

Yes
d0j
d1j
dj

No
n0j d0j
n1j d1j
nj d j

Total
n0j
n1j
nj

This statistic is distributed approximately as 21.

113

How does this apply in survival analysis?

Suppose we observe
Group 1: (X11 , 11) . . . (X1n1 , 1n1 )
Group 0: (X01 , 01) . . . (X0n0 , 0n0 )
We could just count the numbers of failures:
j=1 1j

PK

eg., d1 =

Example: Leukemia data, just counting up the number

of remissions in each treatment group.
Fail
Group
0
1
Total

Yes
21
9
30

No
0
12
12

2p = 16.8
2M H = 16.4

Total
21
21
42

(p = 0.001)
(p = 0.001)

But, this doesnt account for the time at risk.

Conceptually, we would like to compare the KM survival
curves. Lets put the components side-by-side and compare.
114

Cox & Oakes Table 1.1 Leukemia example

Ordered
Death Times
1
2
3
4
5
6
7
8
9
10
11
12
13
15
16
17
19
20
22
23
25

Group
dj
cj
2
0
2
0
1
0
2
0
2
0
0
0
0
0
4
0
0
0
0
0
2
0
2
0
0
0
1
0
0
0
1
0
0
0
0
0
1
0
1
0
0
0

0
rj
21
19
17
16
14
12
12
12
8
8
8
6
4
4
3
3
2
2
2
1
0

Group
dj
cj
0
0
0
0
0
0
0
0
0
0
3
1
1
0
0
0
0
1
1
1
0
1
0
0
1
0
0
0
1
0
0
1
0
1
0
1
1
0
1
0
0
1

1
rj
21
21
21
21
21
21
17
16
16
15
13
12
12
11
11
10
9
8
7
6
5

Note that I wrote down the number at risk for Group 1 for times
1-5 even though there were no events or censorings at those times.

115

Logrank Test: Formal Definition

The logrank test is obtained by constructing a (2 2) table at each distinct death time, and comparing the death
rates between the two groups, conditional on the number at
risk in the groups. The tables are then combined using the
Cochran-Mantel-Haenszel test.
Note: The logrank is sometimes called the Cox-Mantel test.

Let t1, ..., tK represent the K ordered, distinct death times.

At the j-th death time, we have the following table:

Group
0

Die/Fail
Yes
No
d0j
r0j d0j

Total
r0j

d1j

r1j d1j

r1j

Total

dj

rj d j

rj

where d0j and d1j are the number of deaths in group 0 and
1, respectively at the j-th death time, and r0j and r1j are
the number at risk at that time, in groups 0 and 1.

116

The logrank test is:

2logrank =

2
[ PK
j=1 (d0j r0j dj /rj )]
r1j r0j dj (rj dj )
j=1
[rj2 (rj 1)]

PK

Assuming the tables are all independent, then this statistic

will have an approximate 2 distribution with 1 df.

Based on the motivation for the logrank test,

which of the survival-related quantities are we
comparing at each time point?

j=1 wj S1(tj ) S2(tj )

PK

1(tj )
2(tj )
j=1 wj

1(tj )
2(tj )
j=1 wj
PK

PK

117

First several tables of leukemia data

CMH analysis of leukemia data
TABLE 1 OF TRTMT BY REMISS
CONTROLLING FOR FAILTIME=1

TABLE 3 OF TRTMT BY REMISS

CONTROLLING FOR FAILTIME=3

TRTMT

TRTMT

REMISS

Frequency|
Expected |
0|
1|
---------+--------+--------+
0 |
19 |
2 |
|
20 |
1 |
---------+--------+--------+
1 |
21 |
0 |
|
20 |
1 |
---------+--------+--------+
Total
40
2

REMISS

Frequency|
Expected |
0|
1|
---------+--------+--------+
0 |
16 |
1 |
| 16.553 | 0.4474 |
---------+--------+--------+
1 |
21 |
0 |
| 20.447 | 0.5526 |
---------+--------+--------+
Total
37
1

Total
21

21

42

TABLE 2 OF TRTMT BY REMISS

CONTROLLING FOR FAILTIME=2

TABLE 4 OF TRTMT BY REMISS

CONTROLLING FOR FAILTIME=4

TRTMT

TRTMT

REMISS

Frequency|
Expected |
0|
1|
---------+--------+--------+
0 |
17 |
2 |
| 18.05 |
0.95 |
---------+--------+--------+
1 |
21 |
0 |
| 19.95 |
1.05 |
---------+--------+--------+
Total
38
2

19

21

40

118

17

21

38

REMISS

Frequency|
Expected |
0|
1|
---------+--------+--------+
0 |
14 |
2 |
| 15.135 | 0.8649 |
---------+--------+--------+
1 |
21 |
0 |
| 19.865 | 1.1351 |
---------+--------+--------+
Total
35
2

Total

Total

Total
16

21

37

CMH statistic = logrank statistic

SUMMARY STATISTICS FOR TRTMT BY REMISS
CONTROLLING FOR FAILTIME

Cochran-Mantel-Haenszel Statistics (Based on Table Scores)

Statistic
Alternative Hypothesis
DF
Value
Prob
----------------------------------------------------------------1
Nonzero Correlation
1
16.793
0.001
2
Row Mean Scores Differ
1
16.793
0.001
3
General Association
1
16.793
0.001 <===LOGRANK
TEST

Note: Although CMH works to get the correct logrank test,

it would require inputting the dj and rj at each time of death
for each treatment group. Theres an easier way to get the
test statistic, which Ill show you shortly.

119

Calculating logrank statistic by hand

Leukemia Example:
Ordered
Death Times
1
2
3
4
5
6
7
8
10
11
12
13
15
16
17
22
23
Sum

Group 0
d0j
r0j
2
21
2
19
1
17
2
16
2
14
0
12
0
12
4
12
0
8
2
8
2
6
0
4
1
4
0
3
1
3
1
2
1
1

Combined
dj
rj
2
42
2
40
1
38
2
37
2
35
3
33
1
29
4
28
1
23
2
21
2
18
1
16
1
15
1
14
1
13
2
9
2
7

ej
1.00
0.95
0.45
0.86

oj e j
1.00
1.05
0.55
1.14

vj
0.488

10.251

6.257

oj = d0j
ej = dj r0j /rj
vj = r1j r0j dj (rj dj )/[rj2(rj 1)]
2logrank

(10.251)2
= 16.793
=
6.257
120

Notes about logrank test:

The logrank statistic depends on ranks of event times
only
If there are no tied deaths, then the logrank has the form:
[

r0j 2
j=1 (d0j rj )]
PK
2
j=1 r1j r0j /rj

PK

Numerator can be interpreted as (o e) where o is

the observed number of deaths in group 0, and e is
the expected number, given the risk set. The expected
number equals #deaths proportion in group 0 at risk.
The (o e) terms in the numerator can be written as
r0j r1j
0j )
(1j
rj
It does not matter which group you choose to sum over.
To see this, note that if we summed up (o-e) over the death
times for the 6MP group we would get -10.251, and the sum of
the variances is the same. So when we square the numerator,
the test statistic is the same.

121

Analogous to the CMH test for a series of tables at different

levels of a confounder, the logrank test is most powerful when
odds ratios are constant over time intervals. That is, it is
most powerful for proportional hazards.

Checking the assumption of proportional hazards:

check to see if the estimated survival curves cross - if
they do, then this is evidence that the hazards are not
proportional
more formal test: any ideas?

What should be done if the hazards are not

proportional?
If the difference between hazards has a consistent sign,
the logrank test usually does well.
Other tests are available that are more powerful against
different alternatives.

122

Getting the logrank statistic using Stata:

After declaring data as survival type data using
the stset command, issue the sts test command
. stset remiss status
data set name:
id:
entry time:
exit time:
failure/censor:

leukem
--remiss
status

(meaning each record a unique subject)

(meaning all entered at time 0)

. sts list, by(trt)

Beg.
Net
Survivor
Std.
Time
Total
Fail
Lost Function
Error
[95% Conf. Int.]
---------------------------------------------------------------------trt=0
1
21
2
0
0.9048
0.0641
0.6700
0.9753
2
19
2
0
0.8095
0.0857
0.5689
0.9239
3
17
1
0
0.7619
0.0929
0.5194
0.8933
4
16
2
0
0.6667
0.1029
0.4254
0.8250
.
.
(etc)
. sts test trt
Log-rank test for equality of survivor functions
-----------------------------------------------| Events
trt
| observed
expected
------+------------------------0
|
21
10.75
1
|
9
19.25
------+------------------------Total |
30
30.00
chi2(1) =
Pr>chi2 =

16.79
0.0000
123

Getting the logrank statistic using SAS

Still use PROC LIFETEST
Add STRATA command, with treatment variable
Gives the chi-square test (2-sided), but also gives you
the terms you need to calculate the 1-sided test; this is
useful if we want to know which of the two groups has
the higher estimated hazard over time.
The STRATA command also gives the Gehan-Wilcoxon
test (which we will talk about next)
Title Cox and Oakes example;
data leukemia;
input weeks remiss trtmt;
cards;
6
0
1
6
1
1
6
1
1
6
1
1
/* data for 6MP group */
7
1
1
9
0
1
etc
1
1
0
1
1
0
/* data for placebo group */
2
1
0
2
1
0
etc
;
proc lifetest data=leukemia;
time weeks*remiss(0);
strata trtmt;
title Logrank test for leukemia data;
run;

124

Output from leukemia example:

Logrank test for leukemia data
Summary of the Number of Censored and Uncensored Values
TRTMT

Total

Failed

Censored

%Censored

6-MP
Control

21
21

9
21

12
0

57.1429
0.0000

Total

42

30

12

28.5714

Testing Homogeneity of Survival Curves over Strata

Time Variable FAILTIME
Rank Statistics
TRTMT
6-MP
Control

Log-Rank

Wilcoxon

-10.251
10.251

-271.00
271.00

Covariance Matrix for the Log-Rank Statistics

TRTMT
6-MP
Control

6-MP

Control

6.25696
-6.25696

-6.25696
6.25696

Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

16.7929
13.4579
16.4852

1
1
1

0.0001
0.0002
0.0001

125

<== Heres the one we want!!

Getting the logrank statistic using Splus:

Instead of the surv.fit command, use the

surv.diff command with a group (treatment)
variable.

Mantel-Haenszel logrank:
> logrank<-surv.diff(weeks,remiss,trtmt)
> logrank
N Observed Expected (O-E)^2/E
0 21
21
10.75
9.775
1 21
9
19.25
5.458
Chisq= 16.8

on 1 degrees of freedom, p= 4.169e-05

126

Generalization of logrank test

= Linear rank tests

The logrank and other tests can be derived by assigning

scores to the ranks of the death times, and are members of
a general class of linear rank tests (for more detail, see
Lee, ch 5)
First, define
=
(t)

dj
j:tj <t rj
X

where dj and rj are the number of deaths and the number

at risk, respectively at the j-th ordered death time.
Then assign these scores (suggested by Peto and Peto):
Event
Death at tj

Score
j)
wj = 1 (t
j)
wj = (t

Censoring at tj

To calculate the logrank test, simply sum up the scores for

group 0.

127

Example

Group 0: 15, 18, 19, 19, 20

Group 1: 16+, 18+, 20+, 23, 24+

Calculation of logrank as a linear rank statistic

j ) score wj
Ordered Data Group dj
rj (t
15
0
1
10 0.100 0.900
16+
1
0
9 0.100 -0.100
18
0
1
8 0.225 0.775
18+
1
0
7 0.225 -0.225
19
0
2
6 0.558 0.442
20
0
1
4 0.808 0.192
20+
1
0
3 0.808 -0.808
23
1
1
2 1.308 -0.308
24+
1
0
1 1.308 -1.308
The logrank statistic S is sum of scores for group 0:
S = 0.900 + 0.775 + 0.442 + 0.442 + 0.192 = 2.75
The variance is:
P

n0n1 nj=1 wj2

V ar(S) =
n(n 1)
In this case, V ar(S) = 1.210, so
2.75
Z=
= 2.50 = 2logrank = (2.50)2 = 6.25
1.210
128

Why is this form of the logrank equivalent?

P

The logrank statistic S is equivalent to (o e) over the

distinct death times, where o is the observed number of
deaths in group 0, and e is the expected number, given
the risk sets.
At deaths:
At censorings:

weights are 1

weights are

So we are summing up 1s for deaths (to get d0j ), and sub at both deaths and censorings. This amounts to
tracting
subtracting dj /rj at each death or censoring time in group
0, at or after the j-th death. Since there are a total of r0j of
these, we get e = r0j dj /rj .
Why is it called the logrank test?
Since S(t) = exp((t)), an alternative estimator of S(t)
is:
X dj
= exp((t))

S(t)
= exp(
)
j:tj <t rj
= log(S(t))

So, we can think of (t)

as yielding the logsurvival scores used to calculate the statistic.

129

Comparing the CMH-type Logrank and

Linear Rank logrank

A. CMH-type Logrank:
We motivated the logrank test through the CMH statistic
for testing Ho : OR = 1 over K tables, where K is the
number of distinct death times. This turned out to be what
we get when we use the logrank (default) option in Stata or
the strata statement in SAS.

B. Linear Rank logrank:

The linear rank version of the logrank test is based on adding
up scores for one of the two treatment groups. The particular scores that gave us the same logrank statistic were
j ). This
= P (t
based on the Nelson-Aalen estimator, i.e.,
is what you get when you use the test statement in SAS.
Here are some comparisons, with a new example to show
when the two types of logrank statistics will be equal.

130

First, lets go back to our example from Chapter 5 of Lee:

Example

Group 0: 15, 18, 19, 19, 20

Group 1: 16+, 18+, 20+, 23, 24+

A. The CMH-type logrank statistic:

(using the strata statement)
Rank Statistics
TRTMT
Control
Treated

Log-Rank

Wilcoxon

2.7500
-2.7500

18.000
-18.000

Covariance Matrix for the Log-Rank Statistics

TRTMT
Control
Treated

Control

Treated

1.08750
-1.08750

-1.08750
1.08750

Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

6.9540
5.5479
3.3444

1
1
1

0.0084
0.0185
0.0674

131

This is exactly the same chi-square test that you would get
P
if you calculated the numerator of the logrank as (oj ej )
and the variance as vj = r1j r0j dj (rj dj )/[rj2(rj 1)]

Ordered
Death Times
15
18
19
20
23
Sum

Group 0
d0j
r0j
1
5
1
4
2
3
1
1
0
0

2logrank

Combined
dj
rj
1
10
1
8
2
6
2
4
1
2

ej
0.50
0.50
1.00
0.25
0.00

(2.75)2
=
= 6.954
1.0875

132

oj e j
0.50
0.50
1.00
0.75
0.00
2.75

vj
0.2500
0.2500
0.4000
0.1870
0.0000
1.0875

B. The linear rank logrank statistic:

(using the test statement)
Univariate Chi-Squares for the LOG RANK Test

Variable
GROUP

Test
Statistic

Standard
Deviation

2.7500

1.0897

Chi-Square
6.3684

Pr >
Chi-Square
0.0116

Covariance Matrix for the LOG RANK Statistics

Variable
TRTMT

TRTMT
1.18750

This is actually very close to what we would get if we use

the Nelson-Aalen based scores:
Calculation of logrank as a linear rank statistic
j ) score wj
Ordered Data Group dj
rj (t
15
0
1
10 0.100 0.900
16+
1
0
9 0.100 -0.100
18
0
1
8 0.225 0.775
18+
1
0
7 0.225 -0.225
19
0
2
6 0.558 0.442
20
0
1
4 0.808 0.192
20+
1
0
3 0.808 -0.808
23
1
1
2 1.308 -0.308
24+
1
1
1 1.308 -1.308
Sum(grp 0)
2.750
133

Note that the numerator is the exact same number (2.75)

in both versions of the logrank test. The difference in the
denominator is due to the way that ties are handled.

CMH-type variance:
r1j r0j dj (rj dj )
rj2(rj 1)
r1j r0j dj (rj dj )
X
=
rj (rj 1)
rj

var =

Linear rank type variance:

P

n0n1 nj=1 wj2

var =
n(n 1)

134

Now consider an example where there are no tied

death times
Example I

Group 0: 15, 18, 19, 21, 22

Group 1: 16+, 17+, 20+, 23, 24+

A. The CMH-type logrank statistic:

(using the strata statement)
Rank Statistics
TRTMT
Control
Treated

Log-Rank

Wilcoxon

2.5952
-2.5952

15.000
-15.000

Covariance Matrix for the Log-Rank Statistics

TRTMT
Control
Treated

Control

Treated

1.21712
-1.21712

-1.21712
1.21712

Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

5.5338
4.3269
3.1202

1
1
1

0.0187
0.0375
0.0773

135

B. The linear rank logrank statistic:

(using the test statement)
Univariate Chi-Squares for the LOG RANK Test

Variable
TRTMT

Test
Statistic

Standard
Deviation

2.5952

1.1032

Chi-Square
5.5338

Pr >
Chi-Square
0.0187

Covariance Matrix for the LOG RANK Statistics

Variable
TRTMT

TRTMT
1.21712

Note that this time, the variances of the two logrank statistics are exactly the same, equal to 1.217.

If there are no tied event times, then the

two versions of the test will yield identical results. The more ties we have, the
more it matters which version we use.

136

Gehans Generalized Wilcoxon Test

First, lets review the Wilcoxon test for uncensored data:
Denote observations from two samples by:
(X1 , X2, . . . , Xn) and (Y1, Y2, . . . , Ym)
Order the combined sample and define:
Z(1) < Z(2) < < Z(m+n)
Ri1 = rank of Xi
R1 =

m+n
X
i=1

Ri1

Reject H0 if R1 is too big or too small, according to

R1 E(R1)
r

V ar(R1)

N (0, 1)

where
m(m + n + 1)
2
mn(m + n + 1)
V ar(R1) =
12
E(R1) =

137

The Mann-Whitney form of the Wilcoxon is defined as:

and

+1 if Xi > Yj
U (Xi, Yj ) = Uij =
0 if Xi = Yj

1
if Xi < Yj
U=

m
n X
X

i=1 j=1

Uij .

There is a simple correspondence between U and R1:

R1 = m(m + n + 1)/2 + U/2
so

U = 2R1 m(m + n + 1)

Therefore,
E(U ) = 0
V ar(U ) = mn(m + n + 1)/3

138

Extending Wilcoxon to censored data

The Mann-Whitney form leads to a generalization for censored data. Define

xi > yj or x+
i yj
xi = yi or lower value censored
xi < yj or xi yj+

+1 if
U (Xi, Yj ) = Uij =
0 if

1
if
Then define
W =

m
n X
X

i=1 j=1

Uij

Thus, there is a contribution to W for every comparison

where both observations are failures (except for ties), or
where a censored observation is greater than or equal to a
failure.

Looking at all possible pairs of individuals between the two

treatment groups makes this a nightmare to compute by
hand!

139

Gehan found an easier way to compute the above. First,

pool the sample of (n + m) observations into a single group,
then compare each individual with the remaining n + m 1:
For comparing the i-th individual with the j-th, define

ti > tj or t+
i tj
ti < tj or ti t+
j
otherwise

+1 if
Uij =
1 if

0
Then
Ui =

m+n
X
j=1

Uij

Thus, for the i-th individual, Ui is the number of observations which are definitely less than ti minus the number of
observations that are definitely greater than ti. We assume
censorings occur after deaths, so that if ti = 18+ and tj = 18,
then we add 1 to Ui.
The Gehan statistic is defined as
U =

m+n
X
i=1

Ui 1{i in group 0}

= W
U has mean 0 and variance
m+n
mn
X
var(U ) =
Ui2
(m + n)(m + n 1) i=1

140

Example from Lee:

Group 0:

15, 18, 19, 19, 20

Group 1:

16+, 18+, 20+, 23, 24+

Ui
-9
1
-6
2
-2
-2
1
5
4
6
-18

Time Group
15
0
16+
1
18
0
18+
1
19
0
19
0
20
0
20+
1
23
1
24+
1
SUM

Ui2
81
1
36
4
4
4
1
25
16
36
208

U = 18
V ar(U ) =

(5)(5)(208)
(10)(9)

= 57.78
and 2 = (18)2/57.78 = 5.61

141

SAS code:
data leedata;
infile lee.dat;
input time cens group;
proc lifetest data=leedata;
time time*cens(0);
strata group;
run ;
SAS OUTPUT: Gehans Wilcoxon test
Rank Statistics
TRTMT
Control
Treated

Log-Rank

Wilcoxon

2.7500
-2.7500

18.000
-18.000

Covariance Matrix for the Wilcoxon Statistics

TRTMT
Control
Treated

Control

Treated

58.4000
-58.4000

-58.4000
58.4000

Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

6.9540
5.5479
3.3444

1
1
1

Pr >
Chi-Square
0.0084
0.0185 **this is Gehans test
0.0674

142

Notes about SAS Wilcoxon Test:

SAS calculates the Wilcoxon as U instead of U , probably
so that the sign of the test statistic is consistent with the
logrank.
SAS gets something slightly different for the variance, and
this does not seem to depend on whether there are ties.
For example, the hypothetical dataset on p.6 without ties
P
yields U = 15 and Ui2 = 182, so

(5)(5)(182)
(15)2
2
V ar(U ) =
= 50.56 and =
= 4.45
(10)(9)
50.56
while SAS gives the following:
Rank Statistics
TRTMT
Control
Treated

Log-Rank

Wilcoxon

2.5952
-2.5952

15.000
-15.000

Covariance Matrix for the Wilcoxon Statistics

TRTMT
Control
Treated

Control

Treated

52.0000
-52.0000

-52.0000
52.0000

Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

5.5338
4.3269
3.1202

1
1
1

0.0187
0.0375
0.0773

143

Obtaining the Wilcoxon test using Stata

Use the sts test statement, with the appropriate option

sts test

varlist [if exp] [in range]

[, [logrank|wilcoxon|cox] strata(varlist) detail
mat(matname1 matname2) notitle noshow ]

logrank, wilcoxon, and cox specify which test of equality is desired.

logrank is the default, and cox yields a likelihood ratio test
under a cox model.

Example: (leukemia data)

. stset remiss status
. sts test trt, wilcoxon
Wilcoxon (Breslow) test for equality of survivor functions
---------------------------------------------------------| Events
Sum of
trt
| observed
expected
ranks
------+-------------------------------------0
|
21
10.75
271
1
|
9
19.25
-271
------+-------------------------------------Total |
30
30.00
0
chi2(1) =
Pr>chi2 =

13.46
0.0002

144

Generalized Wilcoxon
(Peto & Peto, Prentice)
Assign the following scores:
For a death at t:
For a censoring at t:

S(t+)
+ S(t)
1

S(t+)
1

The test statistic is (scores) for group 0.

Time Group
15
0
16+
1
18
0
18+
1
19
0
20
0
20+
1
23
1
24+
1

dj
1
0
1
0
2
1
0
1
0

rj
10
9
8
7
6
4
3
2
1

S(t+)
0.900
0.900
0.788
0.788
0.525
0.394
0.394
0.197
0.197

score wj
0.900
-0.100
0.688
-0.212
0.313
-0.081
-0.606
-0.409
-0.803

wj 1{j in group 0} = 0.900 + 0.688 + 2 (0.313) + (0.081)

= 2.13
P

n0n1 nj=1 wj2

V ar(S) =
= 0.765
n(n 1)
so

Z = 2.13/0.765 = 2.433
145

The Tarone-Ware class of tests:

This general class of tests is like the logrank test, but adds
weights wj . The logrank test, Wilcoxon test, and PetoPrentice Wilcoxon are included as special cases.

2tw =

2
w
(d

d
/r
)]
[PK
j
1j
1j
j
j
j=1
wj2 r1j r0j dj (rj dj )
l=1
rj2 (rj 1)

PK

Test
Logrank

Weight wj
wj = 1

Gehans Wilcoxon

w j = rj

Peto/Prentice

c
wj = nS(t
j)

Fleming-Harrington

j )]
wj = [S(t

Tarone-Ware

wj =

rj

Note: these weights wj are not the same as the scores wj weve been
talking about earlier, and they apply to the CMH-type form of the
P
test statistic rather than (scores) over a single treatment group.
146

Which test should we used?

CMH-type or Linear Rank?
If there are not a high proportion of ties, then it doesnt
really matter since:
The two Wilcoxons are similar to each other
The two logrank tests are similar to each other
Note: personally, I tend to use the CMH-type test, which you get with the strata
statement in SAS and the test statement in STATA.

Logrank or Wilcoxon?
Both tests have the right Type I power for testing the
null hypothesis of equal survival, Ho : S1(t) = S2(t)
The choice of which test may therefore depend on the
alternative hypothesis, which will drive the power of the
test.

147

 The Wilcoxon is sensitive to early differences between

survival, while the logrank is sensitive to later ones. This
can be seen by the relative weights they assign to the test
statistic:
numerator =

(oj ej )

WILCOXON numerator =

rj (oj ej )

LOGRANK

The logrank is most powerful under the assumption of

proportional hazards, which implies an alternative in
terms of the survival functions of Ha : S1(t) = [S2(t)]
The Wilcoxon has high power when the failure times
are lognormally distributed, with equal variance in both
groups but a different mean. It will turn out that this is
the assumption of an accelerated failure time model.
Both tests will lack power if the survival curves (or hazards) cross. However, that does not necessarily make
them invalid!

148

Comparison between TEST and STRATA in SAS

for 2 examples:

Data from Lee (n=10):

from STRATA:
Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

6.9540
5.5479
3.3444

1
1
1

Pr >
Chi-Square
0.0084
0.0185 **this is Gehans test
0.0674

from TEST:
Univariate Chi-Squares for the WILCOXON Test

Variable
GROUP

Test
Statistic

Standard
Deviation

1.8975

0.7508

Chi-Square
6.3882

Pr >
Chi-Square
0.0115

Univariate Chi-Squares for the LOG RANK Test

Variable
GROUP

Test
Statistic

Standard
Deviation

2.7500

1.0897

Chi-Square
6.3684

149

Pr >
Chi-Square
0.0116

Previous example with leukemia data:

from STRATA:
Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

16.7929
13.4579
16.4852

1
1
1

0.0001
0.0002
0.0001

from TEST:
Univariate Chi-Squares for the WILCOXON Test

Variable
GROUP

Test
Statistic

Standard
Deviation

6.6928

1.7874

Chi-Square

Pr >
Chi-Square

14.0216

0.0002

Univariate Chi-Squares for the LOG RANK Test

Variable
GROUP

Test
Statistic

Standard
Deviation

10.2505

2.5682

150

Chi-Square

Pr >
Chi-Square

15.9305

0.0001

P -sample and stratified logrank tests

We have been discussing two sample problems. In practice,
more complex settings often arise:
There are more than two treatments or groups, and the
question of interest is whether the groups differ from each
other.
We are interested in a comparison between two groups,
but we wish to adjust for another factor that may confound the analysis
We want to adjust for lots of covariates.

We will first talk about comparing the survival distributions

between more than 2 groups, and then about adjusting for
other covariates.

151

P -sample logrank
Suppose we observe data from P different groups, and the
data from group p (p = 1, ..., P ) are:
(Xp1, p1) . . . (Xpnp , pnp )
We now construct a (P 2) table at each of the K distinct
death times, and compare the death rates between the P
groups, conditional on the number at risk.
Let t1, ....tK represent the K ordered, distinct death times.
At the j-th death time, we have the following table:

Group
1

Die/Fail
Yes
No
d1j
r1l d1j

Total
r1j

P
Total

dP j
dj

rP j d P j
rj d j

rP j
rj

where dpj is the number of deaths in group p at the j-th

death time, and rpj is the number at risk at that time.
The tables are then combined using the CMH approach.

152

If we were just focusing on this one table, then a 2(P 1) test

statistic could be constructed through a comparison of os
and es, like before.
Example: Toxicity in a clinical trial with 3 treatments
TABLE OF GROUP BY TOXICITY
GROUP

TOXICITY

Frequency|
Row Pct |no
|yes
|
---------+--------+--------+
1 |
42 |
8 |
| 84.00 | 16.00 |
---------+--------+--------+
2 |
48 |
2 |
| 96.00 |
4.00 |
---------+--------+--------+
3 |
38 |
12 |
| 76.00 | 24.00 |
---------+--------+--------+
Total
128
22

Total
50

50

50

150

STATISTICS FOR TABLE OF GROUP BY TOXICITY

Statistic
DF
Value
Prob
-----------------------------------------------------Chi-Square
2
8.097
0.017
Likelihood Ratio Chi-Square
2
9.196
0.010
Mantel-Haenszel Chi-Square
1
1.270
0.260

Cochran-Mantel-Haenszel Statistics (Based on Table Scores)

Statistic
Alternative Hypothesis
DF
Value
Prob
---------------------------------------------------------1
Nonzero Correlation
1
1.270
0.260
2
Row Mean Scores Differ
2
8.043
0.018
3
General Association
2
8.043
0.018

153

Formal Calculations:

Let Oj = (d1j , ...d(P 1)j )T be a vector of the observed number of failures in groups 1 to (P 1), respectively, at the
j-th death time. Given the risk sets r1j , ... rP j , and the fact
that there are dj deaths, then Oj has a distribution like a
multivariate version of the Hypergeometric. Oj has mean:
Ej = (

dj r(P 1)j T
dj r1j
, ... ,
)
rj
rj

and variance covariance matrix:

v11j v12j ...

v22j ...

Vj =

...
...

v1(P 1)j
v2(P 1)j
...
v(P 1)(P 1)j

where the `-th diagonal element is:

v``j = r`j (rj r`j )dj (rj dj )/[rj2(rj 1)]
and the `m-th off-diagonal element is:
v`mj = r`j rmj dj (rj dj )/[rj2(rj 1)]

154

The resulting 2 test for a single (P 1) table would have

(P-1) degrees and is constructed as follows:
(Oj Ej )T Vj1 (Oj Ej )

Generalizing to K tables
Analogous to what we did for the two sample logrank, we
replace the Oj , Ej and Vj with the sums over the K distinct
P
P
death times. That is, let O = kj=1 Oj , E = kj=1 Ej , and
P
V = kj=1 Vj . Then, the test statistic is:
(O E)T V1 (O E)

155

Example:
Time taken to finish a test with 3 different noise distractions.
All tests were stopped after 12 minutes.

Noise Level
Group Group
1
2
9.0
10.0
9.5
12.0
9.0
12+
8.5
11.0
10.0
12.0
10.5
10.5

156

Group
3
12.0
12+
12+
12+
12+
12+

Lets start the calculations ...

Observed data table
Ordered
Times
8.5
9.0
9.5
10.0
10.5
11.0
12.0

Group 1
d1j
r1j
1
6
2
5
1
3
1
2
1
1
0
0
0
0

Group 2
d2j
r2j
0
6
0
6
0
6
1
6
1
5
1
4
2
3

Group 3
d3j
r3j
0
6
0
6
0
6
0
6
0
6
0
6
1
6

Combined
dj
rj

Group 2
o2j
e2j

Group 3
o3j
e3j

Combined
oj
ej

Expected table
Ordered
Times
8.5
9.0
9.5
10.0
10.5
11.0
12.0

Group 1
o1j
e1j

Doing the P -sample test by hand is cumbersome ...

Luckily, most statistical packages will do it for you!

157

P -sample logrank in Stata

.sts graph, by(group)
.sts test group, logrank
Log-rank test for equality of survivor functions
-----------------------------------------------| Events
group | observed
expected
------+------------------------1
|
6
1.57
2
|
5
4.53
3
|
1
5.90
------+------------------------Total |
12
12.00
chi2(2) =
Pr>chi2 =

20.38
0.0000

. sts test group, wilcoxon

Wilcoxon (Breslow) test for equality of survivor functions
---------------------------------------------------------| Events
Sum of
group | observed
expected
ranks
------+-------------------------------------1
|
6
1.57
68
2
|
5
4.53
-5
3
|
1
5.90
-63
------+-------------------------------------Total |
12
12.00
0
chi2(2) =
Pr>chi2 =

18.33
0.0001

158

SAS program for P -sample logrank

Title Testing with noise example;
data noise;
input testtime finish group;
cards;
9
1
1
9.5
1
1
9.0
1
1
8.5
1
1
10
1
1
10.5
1
1
10.0
1
2
12
1
2
12
0
2
11
1
2
12
1
2
10.5
1
2
12
1
3
12
0
3
12
0
3
12
0
3
12
0
3
12
0
3
;
proc lifetest data=noise;
time testtime*finish(0);
strata group;
run;

159

Testing Homogeneity of Survival Curves over Strata

Time Variable TESTTIME

Rank Statistics
GROUP

Log-Rank

Wilcoxon

4.4261
0.4703
-4.8964

68.000
-5.000
-63.000

1
2
3

Covariance Matrix for the Log-Rank Statistics

GROUP
1
2
3

1.13644
-0.56191
-0.57454

-0.56191
2.52446
-1.96255

-0.57454
-1.96255
2.53709

Covariance Matrix for the Wilcoxon Statistics

GROUP
1
2
3

284.808
-141.495
-143.313

-141.495
466.502
-325.007

-143.313
-325.007
468.320

Test of Equality over Strata

Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

20.3844
18.3265
5.5470

2
2
2

0.0001
0.0001
0.0624

160

Note: do not use Test in SAS PROC LIFETEST if you

want a P -sample logrank. Test will interpret the group
variable as a measured covariate (i.e., either ordinal or continuous).
In other words, you will get a trend test with only 1 degree
of freedom, rather than a P-sample test with (p-1) df.
For example, heres what we get if we use the TEST statement on the noise example:
proc lifetest data=noise;
time testtime*finish(0);
test group;
run;

SAS OUTPUT:
Univariate Chi-Squares for the LOG RANK Test
Test
Standard
Variable
Statistic
Deviation
Chi-Square
GROUP

9.3224

2.2846

Pr >
Chi-Square

16.6503

0.0001

Covariance Matrix for the LOG RANK Statistics

Variable
GROUP
GROUP

5.21957

Forward Stepwise Sequence of Chi-Squares for the LOG RANK Test

Pr >
Chi-Square
Pr >
Variable
DF
Chi-Square
Chi-Square
Increment
Increment
GROUP

16.6503

0.0001

161

16.6503

0.0001

The Stratified Logrank

Sometimes, even though we are interested in comparing two
groups (or maybe P ) groups, we know there are other factors
that also affect the outcome. It would be useful to adjust for
these other factors in some way.
Example: For the nursing home data, a logrank test comparing length of stay for those under and over 85 years of
age suggests a significant difference (p=0.03).
However, we know that gender has a strong association with
length of stay, and also age. Hence, it would be a good idea
to STRATIFY the analysis by gender when trying to assess
the age effect.
A stratified logrank allows one to compare groups, but
allows the shapes of the hazards of the different groups to
differ across strata. It makes the assumption that the group
1 vs group 2 hazard ratio is constant across strata.
In other words:
(s = 1, ..., S).

1s (t)
2s (t)

= where is constant over the strata

This method of adjusting for other variables is not as flexible

as that based on a modelling approach.

162

General setup for the stratified logrank:

Suppose we want to assess the association between survival
and a factor (call this X) that has two different levels. Suppose however, that we want to stratify by a second factor,
that has S different levels.
First, divide the data into S separate groups. Within group
s (s = 1, ..., S), proceed as though you were constructing
the logrank to assess the association between survival and
the variable X. That is, let t1s, ..., tKss represent the Ks
ordered, distinct death times in the s-th group.
At the j-th death time in group s, we have the following
table:

X
1
2
Total

Die/Fail
Yes
No
ds1j
rs1j ds1j
ds2j
dsj

rs2j ds2j
rsj dsj

163

Total
rs1j
rs2j
rsj

Let Os be the sum of the os obtained by applying the

logrank calculations in the usual way to the data from group
s. Similarly, let Es be the sum of the es, and Vs be the
sum of the vs.
The stratified logrank is
Z=

PS

(Os Es)
s=1
r
PS
s=1 (Vs )

164

Stratified logrank using Stata:

. use nurshome
. gen age1=0
. replace age1=1 if age>85
. sts test age1, strata(gender)
failure _d:
analysis time _t:

cens
los

Stratified log-rank test for equality of survivor functions

----------------------------------------------------------| Events
age1 | observed
expected(*)
------+------------------------0
|
795
764.36
1
|
474
504.64
------+------------------------Total |
1269
1269.00
(*) sum over calculations within gender
chi2(1) =
Pr>chi2 =

3.22
0.0728

165

Stratified logrank using SAS:

data pop1;
set pop;
age1=0;
if age >85 then age1=1;
proc lifetest data=pop1 outsurv=survres;
time stay*censor(1);
test age1;
strata gender;
RESULTS (just the logrank part .... you can also do a stratified
Wilcoxon)
The LIFETEST Procedure
Rank Tests for the Association of LSTAY with Covariates
Pooled over Strata
Univariate Chi-Squares for the LOG RANK Test

Variable

Test
Statistic

Standard
Deviation

29.1508

17.1941

AGE1

Pr >
Chi-Square

Chi-Square
2.8744

0.0900

Covariance Matrix for the LOG RANK Statistics

Variable
AGE1

AGE1
295.636

Forward Stepwise Sequence of Chi-Squares for the LOG RANK Test

Variable
AGE1

DF
1

Chi-Square
2.8744

Pr >
Chi-Square
0.0900

166

Chi-Square
Increment
2.8744

Pr >
Increment
0.0900

Modeling of Survival Data

Now we will explore the relationship between survival and
explanatory variables by modeling. In this class, we consider
two broad classes of regression models:
Proportional Hazards (PH) models
(t; Z) = 0(t)(Z)
Most commonly, we write the second term as:
(Z) = eZ
Suppose Z = 1 for treated subjects and Z = 0 for untreated subjects. Then this model says that the hazard
is increased by a factor of e for treated subjects versus
untreated subjects (c might be < 1).
This is an example of a semi-parametric model.

Accelerated Failure Time (AFT) models

log(T ) = + Z + w
where w is an error distribution. Typically, we place
a parametric assumption on w:
exponential, Weibull, Gamma
lognormal
167

Covariates:
In general, Z is a vector of covariates of interest.
Z may include:
continuous factors (eg, age, blood pressure),

discrete factors (gender, marital status),

possible interactions (age by sex interaction)

Discrete Covariates:
Just as in standard linear regression, if we have a discrete
covariate A with a levels, then we will need to include (a1)
dummy variables (U1, U2, . . . , Ua) such that Uj = 1 if A =
j. Then
i(t) = 0(t) exp(2 U2 + 3U3 + + aUa)
(In the above model, the subgroup with A = 1 or U1 = 1 is
the reference group.)
Interactions:
Two factors, A and B, interact if the hazard of death depends on the combination of levels of A and B.
We usually follow the principle of hierarchical models, and
only include interactions if all of the corresponding main
effects are also included.

168

The example I just gave was based on a proportional hazards

model, but the description of the types of covariates we might
want to include in our model applies to both the AFT and
PH model.
Well start out by focusing on the Cox PH model, and address some of the following questions:
What does the term 0(t) mean?

Whats proportional about the PH model?

How do we estimate the parameters in the model?

How do we interpret the estimated values?

How can we construct tests of whether the covariates

have a significant effect on the distribution of survival
times?

How do these tests compare to the logrank test or the

Wilcoxon test?

169

The Cox Proportional Hazards model

(t; Z) = 0(t) exp(Z)
This is the most common model used for survival data.
Why?
flexible choice of covariates

fairly easy to fit

standard software exists

References: Collett, Chapter 3*
Lee, Chapter 10*
Hosmer & Lemeshow, Chapters 3-7
Allison, Chapter 5
Cox and Oakes, Chapter 7
Kleinbaum, Chapter 3
Klein and Moeschberger, Chapters 8 & 9
Kalbfleisch and Prentice
Note: some books (like Collett and H & L) use h(t; X) as
their standard notation for the hazard instead of (t; Z), and
H(t) for the cumulative hazard instead of (t).

170

Why do we call it proportional hazards?

Think of the first example, where Z = 1 for treated and Z =
0 for control. Then if we think of 1(t) as the hazard rate
for the treated group, and 0(t) as the hazard for control,
then we can write:
1(t) = (t; Z = 1) = 0(t) exp(Z)
= 0(t) exp()
This implies that the ratio of the two hazards is a constant,
, which does NOT depend on time, t. In other words, the
hazards of the two groups remain proportional over time.
=

1(t)
= e
0(t)

is referred to as the hazard ratio.

What is the interpretation of here?

171

The Baseline Hazard Function

In the example of comparing two treatment groups, 0(t) is
the hazard rate for the control group.
In general, 0(t) is called the baseline hazard function,
and reflects the underlying hazard for subjects with all covariates Z1, ..., Zp equal to 0 (i.e., the reference group).
The general form is:
(t; Z) = 0(t) exp(1 Z1 + 2Z2 + + pZp)
So when we substitute all of the Zj s equal to 0, we get:
(t, Z = 0) = 0(t) exp(1 0 + 2 0 + + p 0)
= 0(t)

In the general case, we think of the i-th individual having a

set of covariates Zi = (Z1i, Z2i, ..., Zpi), and we model their
hazard rate as some multiple of the baseline hazard rate:
i(t, Zi) = 0(t) exp(1Z1i + + pZpi)

172

This means we can write the log of the hazard ratio for the
i-th individual to the reference group as:

log

i(t)
= 1 Z1i + 2 Z2i + + p Zpi
0(t)

The Cox Proportional Hazards model is a

linear model for the log of the hazard ratio
One of the biggest advantages of the framework of the Cox
PH model is that we can estimate the parameters which
reflect the effects of treatment and other covariates without
having to make any assumptions about the form of 0(t).
In other words, we dont have to assume that 0(t) follows
an exponential model, or a Weibull model, or any other particular parametric model.
Thats what makes the model semi-parametric.
Questions:
1. Why dont we just model the hazard ratio,
= i(t)/0 (t), directly as a linear function of the
covariates Z?
2. Why doesnt the model have an intercept?

173

How do we estimate the model parameters?

The basic idea is that under PH, information about can

be obtained from the relative orderings (i.e., ranks) of the
survival times, rather than the actual values. Why?
Suppose T follows a PH model:
(t; Z) = 0(t)eZ

Now consider T = g(T ), where g is a monotonic increasing

function. We can show that T also follows the PH model,
with the same multiplier, eZ.
Therefore, when we consider likelihood methods for estimating the model parameters, we only have to worry about the
ranks of the survival times.

174

Likelihood Estimation for the PH Model

Kalbfleisch and Prentice derive a likelihood involving only
and Z (not 0(t)) based on the marginal distribution of
the ranks of the observed failure times (in the absence of
censoring).
Cox (1972) derived the same likelihood, and generalized it
for censoring, using the idea of a partial likelihood
Suppose we observe (Xi, i, Zi) for individual i, where
Xi is a censored failure time random variable

i is the failure/censoring indicator (1=fail,0=censor)

Zi represents a set of covariates

The covariates may be continuous, discrete, or time-varying.

175

Suppose there are K distinct failure (or death) times, and

let 1, ....K represent the K ordered, distinct death times.
For now, assume there are no tied death times.
Let R(t) = {i : xi t} denote the set of individuals who
are at risk for failure at time t.
More about risk sets:
I will refer to R(j ) as the risk set at the jth failure time

I will refer to R(Xi) as the risk set at the failure time of

individual i
There will still be rj individuals in R(j ).

rj is a number, while R(j ) identifies the actual subjects

at risk

176

What is the partial likelihood?

Intuitively, it is a product over the set of observed death
times of the conditional probabilities of seeing the observed
deaths, given the set of individuals at risk at those times.
At each death time j , the contribution to the likelihood is:
Lj () = P r(individual j fails|1 failure from R(j ))
=

P r(individual j fails| at risk at j )

P
`R(j ) P r(individual ` fails| at risk at j )

(j ; Zj )
P
`R(j ) (j ; Z` )

Under the PH assumption, (t; Z) = 0(t)eZ , so we get:

partial

() =

0(j )eZj

K
Y

j=1

K
Y

`R(j ) 0 (j )e

eZj

j=1

`R(j ) e

177

Z`

Z`

Another derivation:
In general, the likelihood contributions for censored data fall
into two categories:
Individual is censored at Xi:
Li() = S(Xi) = exp[

Z X
i

i(u)du]

Individual fails at Xi:

Li() = S(Xi)i(Xi) = i(Xi) exp[

Z X
i

i(u)du]

Thus, everyone contributes S(Xi) to the likelihood, and only

those who fail contribute i(Xi).
This means we get a total likelihood of:
L() =

n
Y

i=1

i(Xi)

exp[

Z X
i

i(u)du]

The above likelihood holds for all censored survival data,

with general hazard function (t). In other words, we havent
used the Cox PH assumption at all yet.

178

Now, lets multiply and divide by the term

n
Y

i (Xi )

P
L() =

(X
)
i
jR(Xi ) i
i=1

jR(Xi )

i (Xi )

jR(Xi ) i (Xi )

exp[

Z X
i
0

i (u)du]

Cox (1972) argued that the first term in this product contained almost all of the information about , while the second two terms contained the information about 0(t), i.e.,
the baseline hazard.
If we just focus on the first term, then under the Cox PH
assumption:
L() =

n
Y

i=1
n
Y

i=1

n
Y

i=1

i(Xi)

(X
)
i
jR(Xi ) i

0(Xi) exp(Zi)

jR(Xi ) 0 (Xi ) exp(Zj )

exp(Zi)

exp(Z
)
j
jR(Xi )

This is the partial likelihood defined by Cox. Note that it

does not depend on the underlying hazard function 0().
Cox recommends treating this as an ordinary likelihood for
making inferences about in the presence of the nuisance
parameter 0().
179

A simple example:
individual
1
2
3
4

Xi
9
8
6
10

i
1
0
1
1

Zi
4
5
7
3

Now lets compile the pieces that go into the partial likelihood contributions at each failure time:

ordered
failure
j time Xi

Likelihood contribution
R(Xi)

ij

Zi

jR(Xi ) e

Zj i

{1,2,3,4}

e7 /[e4 + e5 + e7 + e3 ]

{1,2,4}

{1,4}

e4 /[e4 + e3 ]

10

{4}

e3 /e3 = 1

The partial likelihood would be the product of these four

terms.
180

Notes on the partial likelihood:

L() =

n
Y

j=1

K
Y

j=1

eZj
`R(Xj ) e

Z`

e Zj
P

`R(j ) e

Z`

where the product is over the K death (or failure) times.

contributions only at the death times

the partial likelihood is NOT a product of independent
terms, but of conditional probabilities
There are other choices besides (Z) = eZ , but this
is the most common and the one for which software is
generally available.

181

Partial Likelihood inference

Inference can be conducted by treating the partial likelihood
as though it satisfied all the regular likelihood properties.
The log-partial likelihood is:

n
Y

`() = log

j=1

K
Y

= log
=

j=1

K
X

j=1
K
X

j=1

eZj
P

`R(j ) e

Z`

eZj
P

`R(j ) e

Zj log[

Z`
X

`R(j )

eZ` ]

lj ()

where lj is the log-partial likelihood contribution at the j-th

ordered death time.

182

Suppose there is only one covariate ( is one-dimensional):

The partial likelihood score equations are:

Z`eZ`

`R(
)
j

j Zj P
`() =
U () =

Z
`

j=1
`R(j ) e
n
X

We can express U () intuitively as a sum of observed minus expected values:

n

X
j (Zj Zj )
U () =
`() =

j=1

where Zj is the weighted average of the covariate Z over

all the individuals in the risk set at time j . Note that is
involved through the term Zj .
The maximum partial likelihood estimators can be found by
solving U () = 0.

183

Analogous to standard likelihood theory, it can be shown

(though not easily) that
(c )
N (0, 1)

se()

The variance of can be obtained by inverting the second

derivative of the partial likelihood,

var()
`()
2

From the above expression for U (), we have:

n
(Zj Zj )2eZ`

`R(
)
j

j
`() =

P
2
Z

j=1
`R(j ) e `
2

Note:
The true variance of ends up being a function of , which
is unknown. We calculate the observed information by
substituting in our partial likelihood estimate of into the
above formula for the variance

184

Simple Example for 2-group comparison: (no ties)

Group 0:

4+, 7, 8+, 9, 10+ = Zi = 0

Group 1:

3, 5, 5+, 6, 8+

ordered failure
j
time Xi

= Zi = 1

Number at risk
Group 0
Group 1

Likelihood contribution
ii
h
P
eZi / jR(Xi ) eZj

e /[5 + 5e ]

e /[4 + 4e ]

e /[4 + 2e ]

e /[4 + 1e ]

e0 /[2 + 0] = 1/2

Again, we take the product over the likelihood contributions,

then maximize to get the partial MLE for .
What does represent in this case?

185

Notes
The observed information matrix is generally used because in practice, people find it has better properties.
Also, the expected is very hard to calculate.
There is a nice analogy with the score and information matrices from more standard regression problems,
except that here we are summing over observed death
times, rather than individuals.
Newton Raphson is used by many of the computer packages to solve the partial likelihood equations.

186

Fitting Cox PH model with Stata

Uses the stcox command.
First, try typing help stcox
---------------------------------------------------------------------help for stcox
---------------------------------------------------------------------Estimate Cox proportional hazards model
--------------------------------------stcox [varlist] [if exp] [in range]
[, nohr strata(varnames) robust cluster(varname) noadjust
mgale(newvar) esr(newvars)
schoenfeld(newvar) scaledsch(newvar)
basehazard(newvar) basechazard(newvar) basesurv(newvar)
{breslow | efron | exactm | exactp} cmd estimate noshow
offset level(#) maximize-options ]
stphtest [, km log rank time(varname) plot(varname) detail
graph-options ksm-options]
stcox is for use with survival-time data; see help st. You must
have stset your data before using this command; see help stset.
Description
----------stcox estimates maximum-likelihood proportional hazards models on st data.
Options (many more!)
------nohr reports the estimated coefficients rather than hazard ratios; i.e.,
b rather than exp(b). Standard errors and confidence intervals are
similarly transformed. This option affects how results are displayed,
not how they are estimated.

187

Ex. Leukemia Data

. stcox trt
Iteration 0:
log likelihood
Iteration 1:
log likelihood
Iteration 2:
log likelihood
Iteration 3:
log likelihood
Refining estimates:
Iteration 0:
log likelihood

= -93.98505
= -86.385606
= -86.379623
= -86.379622
= -86.379622

Cox regression -- Breslow method for ties

No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

42
30
541
-86.379622

Number of obs

42

LR chi2(1)
Prob > chi2

=
=

15.21
0.0001

-----------------------------------------------------------------------------_t |
_d | Haz. Ratio
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------trt |
.2210887
.0905501
-3.685
0.000
.0990706
.4933877
-----------------------------------------------------------------------------. stcox trt , nohr
(same iterations for log-likelihood)
Cox regression -- Breslow method for ties
No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

42
30
541
-86.379622

Number of obs

42

LR chi2(1)
Prob > chi2

=
=

15.21
0.0001

-----------------------------------------------------------------------------_t |
_d |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------trt | -1.509191
.4095644
-3.685
0.000
-2.311923
-.7064599
------------------------------------------------------------------------------

188

Fitting PH models in SAS - PROC PHREG

Ex. Leukemia data
Title Cox and Oakes example;
data leukemia;
input weeks remiss trtmt;
cards;
6
0
1
6
1
1
6
1
1
6
1
1
/* data for 6MP group */
7
1
1
9
0
1
etc
1
1
0
1
1
0
/* data for placebo group */
2
1
0
2
1
0
etc
;
proc phreg data=leukemia;
model weeks*remiss(0)=trtmt;
title Cox PH Model for leukemia data;
run;

189

PROC PHREG Output:

The PHREG Procedure
Data Set: WORK.LEUKEM
Dependent Variable: FAILTIME
Censoring Variable: FAIL
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to Relapse

Summary of the Number of

Event and Censored Values

Total

Event

Censored

Percent
Censored

42

30

12

28.57

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

187.970
.
.

172.759
.
.

Model Chi-Square
15.211 with 1 DF (p=0.0001)
15.931 with 1 DF (p=0.0001)
13.578 with 1 DF (p=0.0002)

Analysis of Maximum Likelihood Estimates

Variable DF

Parameter
Estimate

TRTMT

-1.509191

Standard
Wald
Pr >
Error Chi-Square Chi-Square
0.40956

190

13.57826

0.0002

Risk
Ratio
0.221

Fitting PH models in S-plus: coxph function

Here are some of the data in leuk.dat:
t
1
1
2
2
3

f x
1 0
1 0
1 0
1 0
1 0
...
19 0 1
20 0 1
22 1 1
23 1 1
25 0 1
32 0 1
32 0 1
34 0 1
35 0 1
leuk_read.table("leuk.dat",header=T)
#specify Breslow handling of ties
print(coxph(Surv(t,f) ~ x, leuk, method="breslow"))
#specify Efron handling of ties (default)
print(coxph(Surv(t,f) ~ x, leuk))
191

coxph Output:

Call:
coxph(formula = Surv(t, f) ~ x, data = leuk, method = "breslow")

coef exp(coef) se(coef)

z
p
x -1.51
0.221
0.41 -3.68 0.00023
Likelihood ratio test=15.2

on 1 df, p=0.0000961

n= 42

Call:
coxph(formula = Surv(t, f) ~ x, data = leuk)

coef exp(coef) se(coef)

z
p
x -1.57
0.208
0.412 -3.81 0.00014
Likelihood ratio test=16.4

on 1 df, p=0.0000526

192

n= 42

Compare this with the logrank test

from Proc Lifetest
(Using the Test statement)
The LIFETEST Procedure
Rank Tests for the Association of FAILTIME with Covariates
Pooled over Strata
Univariate Chi-Squares for the LOG RANK Test

Variable
TRTMT

Test
Statistic

Standard
Deviation

10.2505

2.5682

Chi-Square

Pr >
Chi-Square

15.9305

0.0001

Notes:
The logrank test=score test from Proc phreg!
In general, the score test would be for all of the variables
in the model, but in this case, we have only trtmt.
Stata does not provide a score test in its output from
the Cox model. However, the stcox command with
the breslow option for ties yields the same LR test as
the CMH-version logrank test from the sts test, cox
command.

193

More Notes:
The Cox Proportional hazards model has the advantage
over a simple logrank test of giving us an estimate of
the risk ratio (i.e., = 1(t)/0 (t)). This is more
informative than just a test statistic, and we can also
form confidence intervals for the risk ratio.
In this case, = 0.221, which can be interpreted to mean
that the hazard for relapse among patients treated with
6-MP is less than 25% of that for placebo patients.
From the sts list command in Stata or Proc lifetest
in SAS, we were able to get estimates of the entire sur for each treatment group; we cant
vival distribution S(t)
immediately get this from our Cox model without further assumptions. Why not?

194

Adjustments for ties

The proportional hazards model assumes a continuous hazard ties are not possible. There are four proposed modifications to the likelihood to adjust for ties.
(1) Coxs (1972) modification: discrete method
(2) Peto-Breslow method
(3) Efrons (1977) method
(4) Exact method (Kalbfleisch and Prentice)
(5) Exact marginal method (stata)

Some notation:
1, ....K

the K ordered, distinct death times

dj

the number of failures at j

Hj

the history of the entire data set, up to the

j-th death or failure time, including the time
of the failure, but not the identities of the dj
who fail there.

ij1, ...ijdj

the identities of the dj individuals who fail at j

195

(1) Coxs (1972) modification: discrete method

Coxs method assumes that if there are tied failure times,
they truly happened at the same time. It is based on a
discrete likelihood.
The partial likelihood is:
L() =

K
Y

j=1

P r(ij1, ...ijdj fail | dj fail at j , from R)

P r(ij1, ...ijdj fail | in R(j ))

P
j=1 `s(j,dj ) P r(`1 , ....`dj fail | in R(j ))
K
Y

K
Y

j=1
K
Y

j=1

exp(Zij1 ) exp(Zijdj )

`s(j,dj ) exp(Z`1 ) exp(Z`dj )

exp(Sj )
`s(j,dj ) exp(Sj` )

where
s(j, dj ) is the set of all possible sets of dj individuals that
can possibly be drawn from the risk set at time j
Sj is the sum of the Zs for all the dj individuals who
fail at j
Sj` is the sum of the Zs for all the dj individuals in the
`-th set drawn out of s(j, dj )
196

What does this all mean??!!

Lets modify our previous simple example to include ties.
Simple Example (with ties)
Group 0:
Group 1:
Ordered
failure
j time Xi

4+, 6, 8+, 9, 10+ = Zi = 0

3, 5, 5+, 6, 8+
= Zi = 1

Number at risk
Group 0
Group 1

Likelihood Contribution
P
eSj / `s(j,dj ) eSj`

e /[5 + 5e ]

e /[4 + 4e ]

e /[6 + 8e + e2 ]

e0 /2 = 1/2

The tie occurs at t = 6, when R(j ) = {Z = 0 : (6, 8+, 9, 10+),

Z = 1 : (6, 8+)}. Of the 62 = 15 possible pairs of subjects
at risk at t=6, there are 6 pairs formed where both are from
group 0 (Sj = 0), 8 pairs formed with one in each group
(Sj = 1), and 1 pairs formed with both in group 1 (Sj = 2).
Problem: With large numbers of ties, the denominator can
have many many terms and be difficult to calculate.

197

(2) Breslow method: (default)

Breslow and Peto suggested replacing the term

Z` dj

`s(j,dj ) e

Sj`

in the denominator by the term `R(j ) e

, so that the
following modified partial likelihood would be used:
L() =

K
Y

eSj

j=1

Sj`
`s(j,dj ) e

K
Y

eSj

j=1 P

`R(j )

eZ`

Justification:
Suppose individuals 1 and 2 fail from {1, 2, 3, 4} at time j .
Let (i) be the hazard ratio for individual i (compared to
baseline).
P

eSj
`s(j,dj ) e

Sj`

(1)
(2)

(1) + (2) + (3) + (4) (2) + (3) + (4)

+

(1)
(2)

(1) + (2) + (3) + (4) (1) + (3) + (4)

2(1)(2)
[(1) + (2) + (3) + (4)]2

The Peto (Breslow) approximation will break down when

the number of ties are large relative to the size of the risk
sets, and then tends to yield estimates of which are biased
toward 0.

198

(3) Efrons (1977) method:

Efron suggested an even closer approximation to the discrete
likelihood:
L() =

K
Y

j=1

eSj

`R(j )

eZ`

!d
j
j1 P
Z
`
e
`D(j )
dj

Like the Breslow approximation, Efrons method will yield

estimates of which are biased toward 0 when there are
many ties.
However, Allison (1995) recommends the Efron approximation since it is much faster than the exact methods and tends
to yield much closer estimates than the default Breslow approach.

199

(4) Exact method (Kalbfleisch and Prentice):

The discrete option that we discussed in (1) is an exact
method based on a discrete likelihood (assuming that tied
events truly ARE tied).
This second exact method is based on the continuous likelihood, under the assumption that if there are tied events,
that is due to the imprecise nature of our measurement, and
that there must be some true ordering.
All possible orderings of the tied events are calculated, and
the probabilities of each are summed.
Example with 2 tied events (1,2) from riskset (1,2,3,4):

eSj
`s(j,dj ) e

Sj`

eS1
eS2
= S

e 1 + eS2 + eS3 + eS4 eS2 + eS3 + eS4

eS2
eS1
+ S

e 1 + eS2 + eS3 + eS4 eS1 + eS3 + eS4

200

Bottom Line: Implications of Ties

(See Allison (1995), p.127-137)
(1) When there are no ties, all options give exactly the
same results.
(2) When there are only a few ties, it wont make
much difference which method is used. However, since
the exact methods wont take much extra computing
time, you might as well use one of them.
(3) When there are many ties (relative to the number
at risk), the Breslow option (default) performs poorly
(Farewell & Prentice, 1980; Hsieh, 1995). Both of the
approximate methods, Breslow and Efron, yield coefficients that are attenuated (biased toward 0).
(4) The choice of which exact method to use should
be based on substantive grounds - are the tied event
times truly tied? ...or are they the result of imprecise
measurement?
(5) Computing time of exact methods is much longer
than that of the approximate methods. However, in most
cases it will still be less than 30 seconds even for the exact
methods.
(6) Best approximate method - the Efron approximation nearly always works better than the Breslow
method, with no increase in computing time, so use this
option if exact methods are too computer-intensive.
201

Example: The fecundability study

Women who had recently given birth (or had tried to get
pregnant for at least a year) were asked to recall how long
it took them to become pregnant, and whether or not they
smoked during that time. The outcome of interest is time to
pregnancy (measured in menstrual cycles).
data fecund;
input
smoke
cycle
status
count;
cards;
0
1
1
198
0
2
1
107
0
3
1
55
0
4
1
38
0
5
1
18
0
6
1
22
..........................................
1
1
1
1
;

10
11
12
12

1
1
1
0

1
1
3
7

proc phreg;
model cycle*status(0) = smoke /ties=breslow;
freq count;
proc phreg;
model cycle*status(0) = smoke /ties=discrete;
freq count;
proc phreg;
model cycle*status(0) = smoke /ties=exact;
freq count;
proc phreg;
model cycle*status(0) = smoke /ties=efron;
freq count;

202

/* default */

SAS Output for Fecundability study:

Accounting for Ties
***************************************************************************
Ties Handling: BRESLOW

Variable
SMOKE

DF
1

Parameter
Estimate
-0.329054

Standard
Error
0.11412

Wald
Chi-Square
8.31390

Pr >
Chi-Square
0.0039

Risk
Ratio
0.720

***************************************************************************
Ties Handling: DISCRETE

Variable
SMOKE

DF
1

Parameter
Estimate
-0.461246

Standard
Error
0.13248

Wald
Chi-Square
12.12116

Pr >
Chi-Square
0.0005

Risk
Ratio
0.630

***************************************************************************
Ties Handling: EXACT

Variable
SMOKE

DF
1

Parameter
Estimate
-0.391548

Standard
Error
0.11450

Wald
Chi-Square
11.69359

Pr >
Chi-Square
0.0006

Risk
Ratio
0.676

***************************************************************************
Ties Handling: EFRON

Variable
SMOKE

DF
1

Parameter
Estimate
-0.387793

Standard
Error
0.11402

Wald
Chi-Square
11.56743

Pr >
Chi-Square
0.0007

Risk
Ratio
0.679

***************************************************************************

For this particular dataset, does it seem like it

would be important to consider the effect of tied
failure times? Which method would be best?
203

Stata Commands for PH Model with Ties:

Stata also offers four options for adjustments with tied data:
breslow (default)
efron
exactp (same as the discrete option in SAS)
exactm - an exact marginal likelihood calculation
(different than the exact option in SAS)
Fecundability Data Example:
. stcox smoker, efron nohr
failure _d:
analysis time _t:

status
cycle

Iteration 0:
log likelihood = -3113.5313
Iteration 1:
log likelihood = -3107.3102
Iteration 2:
log likelihood = -3107.2464
Iteration 3:
log likelihood = -3107.2464
Refining estimates:
Iteration 0:
log likelihood = -3107.2464
Cox regression -- Efron method for ties
No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

586
567
1844
-3107.2464

Number of obs

586

LR chi2(1)
Prob > chi2

=
=

12.57
0.0004

-----------------------------------------------------------------------------_t |
_d |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------smoker | -.3877931
.1140202
-3.401
0.001
-.6112685
-.1643177
------------------------------------------------------------------------------

204

A special case: the two-sample problem

Previously, we derived the logrank test from an intuitive perspective, assuming that we have (X01 , 01) . . . (X0n0 , 0n0 ) from
group 0 and (X11, 11), . . . , (X1n1 , 1n1 ) from group 1.
Just as a 2 test for binary data can be derived from a logistic
model, we will see here that the logrank test can be derived
as a special case of the Cox Proportional Hazards model.
First, lets re-define our notation in terms of (Xi, i, Zi):
(X01 , 01 ), . . . , (X0n0 , 0n0 ) = (X1 , 1 , 0), . . . , (Xn0 , n0 , 0)
(X11 , 11 ), . . . , (X1n1 , 1n1 ) = (Xn0+1 , n0+1 , 1), . . . , (Xn0+n1 , n0+n1 , 1)

In other words, we have n0 rows of data (Xi, i, 0) for the

group 0 subjects, then n1 rows of data (Xi, i, 1) for the
group 1 subjects.
Using the proportional hazards formulation, we have
(t; Z) = 0(t) eZ

Group 0 hazard:

0(t)

Group 1 hazard:

0(t) e

205

The log-partial likelihood is:

K
Y

logL() = log
=

j=1

K
X

j=1

Zj

`R(j ) e

Zj log[

Z`

`R(j )

eZ` ]

Taking the derivative with respect to , we get:

U () =
=
=

`()

n
X

j=1
n
X

j=1

j Zj

where Zj =

Z`
`R(j ) Z` e
P
Z
`R(j ) e `

U () is called the score.

206

`R(j ) Z` e

P
Z
`R(j ) e `

j (Zj Zj )
P

Z`

As we discussed earlier in the class, one useful form of a

likelihood-based test is the score test. This is obtained by
using the score U () evaluated at Ho as a test statistic.
Lets look more closely at the form of the score:
j Zj
j Zj

observed number of deaths in group 1 at j

expected number of deaths in group 1 at j

Why? Under H0 : = 0, Zj is simply the number of

individuals from group 1 in the risk set at time j (call this
r1j ), divided by the total number in the risk set at that time
(call this rj ). Thus, Zj approximates the probability that
given there is a death at j , it is from group 1.
Thus, the score statistic is of the form:
n
X

j=1

(Oj Ej )

When there are ties, the likelihood has to be replaced by one

that allows for ties.
In SAS or Stata:
discrete/exactp Mantel-Haenszel logrank test
breslow

linear rank version of the logrank test

207

I already showed you the equivalence of the linear rank logrank test and the Breslow (default) Cox PH model in SAS
(p.24-25)
Here is the output from SAS for the leukemia data using the
method=discrete option:
Logrank test with proc lifetest - strata statement
Test of Equality over Strata
Test
Log-Rank
Wilcoxon
-2Log(LR)

Chi-Square

DF

Pr >
Chi-Square

16.7929
13.4579
16.4852

1
1
1

0.0001
0.0002
0.0001

The PHREG Procedure

Data Set: WORK.LEUKEM
Dependent Variable: FAILTIME
Censoring Variable: FAIL
Censoring Value(s): 0
Ties Handling: DISCRETE

Time to Relapse

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

165.339
.
.

149.086
.
.

208

Model Chi-Square
16.252 with 1 DF (p=0.0001)
16.793 with 1 DF (p=0.0001)
14.132 with 1 DF (p=0.0002)

More on the Cox PH model

I. Confidence intervals and hypothesis tests

Two methods for confidence intervals
Wald tests and likelihood ratio tests
Interpretation of parameter estimates
An example with real data from an AIDS
clinical trial

II. Predicted survival under proportional hazards

III. Predicted medians and P-year survival

209

I. Constructing Confidence intervals and tests for

the Hazard Ratio (see H & L 4.2, Collett 3.4):

Many software packages provide estimates of , but the hazard ratio HR= exp() is usually the parameter of interest.
We can use the delta method to get standard errors for

exp():
d
= exp(2)V
ar()

V ar(HR)
= V ar(exp())

Constructing confidence intervals for exp()

Two options: (assuming that is a scalar)
obtained above via the delta method as
I. Using se(expr)
= [V ar(exp())],
calculate the endpoints as:
se(exp )
d
d
d
d
[L, U ] = [OR
1.96 se(OR),
OR
+ 1.96 se(OR)]

and then exponentiate

II. Form a confidence interval for ,
the endpoints.

[L, U ] = [e1.96se() , e+1.96se()]

Which approach do you think would be the most

preferable?
210

Hypothesis Tests:
For each covariate of interest, the null hypothesis is
Ho : HRj = 1 j = 0
A Wald test2 of the above hypothesis is constructed as:
j
Z=
se(j )

or

j 2

2 =

se(j )

This test for j = 0 assumes that all other terms in the

model are held fixed.
Note: if we have a factor A with a levels, then we would need
to construct a 2 test with (a 1) df, using a test statistic
based on a quadratic form:
0

c
c
1 c
2(a1) =
A V ar( A ) A

where A = (2, ..., a)0 are the (a 1) coefficients corresponding to Z2, ..., Za (or Z1, ..., Za1 , depending on the
reference group).

The first follows a normal distribution, and the second follows a 2 with 1 df.
STATA gives the Z statistic, while SAS gives the 21 test statistic (the p-values
are also given, and dont depend on which form, Z or 2 , is provided)
2

211

Likelihood Ratio Tests:

Suppose there are (p + q) explanatory variables measured:
Z1, . . . , Zp, Zp+1, . . . , Zp+q
and proportional hazards are assumed.
Consider the following models:
Model 1: (contains only the first p covariates)
i(t, Z)
= exp(1 Z1 + + pZp)
0(t)

Model 2: (contains all (p + q) covariates)

i(t, Z)
= exp(1 Z1 + + p+q Zp+q )
0(t)

These are nested models. For such nested models, we can

construct a likelihood ratio test of
H0 : p+1 = = p+q = 0
as:
2LR

= 2 log(L(1))
log(L(2))

Under Ho, this test statistic is approximately distributed as

2 with q df.

212

Some examples using the Stata stcox command:

Model 1:
. use mac
. stset mactime macstat
. stcox karnof rif clari, nohr
failure _d:
analysis time _t:

macstat
mactime

Cox regression -- Breslow method for ties

No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

1151
121
489509
-754.52813

Number of obs

1151

LR chi2(3)
Prob > chi2

=
=

32.01
0.0000

----------------------------------------------------------------------_t |
_d |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+------------------------------------------------------------karnof | -.0448295
.0106355
-4.215 0.000 -.0656747
-.0239843
rif |
.8723819
.2369497
3.682 0.000
.4079691
1.336795
clari |
.2760775
.2580215
1.070 0.285 -.2296354
.7817903
-----------------------------------------------------------------------

213

Model 2:
. stcox karnof rif clari cd4, nohr
failure _d:
analysis time _t:

macstat
mactime

Cox regression -- Breslow method for ties

No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

1151
121
489509
-738.66225

Number of obs

1151

LR chi2(4)
Prob > chi2

=
=

63.74
0.0000

------------------------------------------------------------------------_t |
_d |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+--------------------------------------------------------------karnof | -.0368538
.0106652
-3.456
0.001
-.0577572
-.0159503
rif |
.880338
.2371111
3.713
0.000
.4156089
1.345067
clari |
.2530205
.2583478
0.979
0.327
-.253332
.7593729
cd4 | -.0183553
.0036839
-4.983
0.000
-.0255757
-.0111349
-------------------------------------------------------------------------

214

Notes:
If we omit the nohr option, we will get the estimated
hazard ratio along with 95% confidence intervals using
Method II (i.e., forming a CI for the log HR (beta), and
then exponentiating the bounds)
-----------------------------------------------------------------------_t |
_d | Haz. Ratio
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------karnof |
.9638171
.0102793
-3.456
0.001
.9438791
.9841762
rif |
2.411715
.5718442
3.713
0.000
1.515293
3.838444
clari |
1.28791
.3327287
0.979
0.327
.7762102
2.136936
cd4 |
.9818121
.0036169
-4.983
0.000
.9747486
.9889269
------------------------------------------------------------------------

We can also compute the hazard ratio ourselves, by exponentiating the coefficients:
HRcd4 = exp(0.01835) = 0.98
Why is this HR so close to 1, and yet still
highly significant?
What is the interpretation of this HR?
The likelihood ratio test for the effect of CD4 is twice
the difference in minus log-likelihoods between the two
models:
2LR = 2 (754.533 (738.66)) = 31.74
How does this test statistic compare to the Wald 2 test?
215

 In the mac study, there were three treatment arms (rif,

clari, and the rif+clari combination). Because we have
only included the rif and clari effects in the model,
the combination therapy is the reference group.
We can conduct an overall test of treatment using the
test command in Stata:
. test rif clari
( 1)
( 2)

rif = 0.0
clari = 0.0
chi2( 2) =
Prob > chi2 =

17.01
0.0002

for a 2 df Wald chi-square test of whether both treatment

coefficients are equal to 0. This test command can be
used to conduct an overall test for any number of effects.
The test command can also be used to test whether
there is a difference between the rif and clari treatment arms:
.

test rif=clari
( 1)

rif - clari = 0.0

chi2( 1) =
Prob > chi2 =

8.76
0.0031

216

Some examples using SAS PROC PHREG

proc phreg data=alloi;
model dthtime*dthstat(0)=mlogrna cd4grp1 cd4grp2 combther
/ risklimits;
cd4level: test cd4grp1, cd4grp2;
title1 Proportional hazards regression model for time to Death;
title2 Baseline viral load and CD4 predictors;
proc phreg data=alloi;
model dthtime*dthstat(0)=mlogrna cd4grp1 cd4grp2 combther decrs8 incrs8
/ risklimits;
cd4level: test cd4grp1, cd4grp2;
wk8resp: test decrs8, incrs8;

Notes:
The risklimits option on the model statement provides 95%
confidence intervals using Method II from page 2. (i.e., forming
a CI for the log HR (beta), and then exponentiating the bounds)
The test statement has the following form:
Label: test varname1, varname2, ..., varnamek;
for a k df Wald chi-square test of whether the k coefficients are
all equal to 0.
We can use the same approach described by Freedman to assess
the effects of intermediate endpoints (incrs8, decrs8) on the
treatment effect (i.e., assess their use as surrogate markers).
The percentage of treatment effect explained, , is estimated
by:
trt,M 2
= 1
trt,M 1
where M1 is the model without the intermediate endpoint and
M2 is the model with the marker.
217

OUTPUT FROM PROC PHREG (Model 1)

Proportional hazards regression model for time to Death
Baseline viral load and CD4 predictors
Data Set: WORK.ALLOI
Dependent Variable: DTHTIME
Censoring Variable: DTHSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to death (days)

Death status (1=died,0=censored)

Summary of the Number of

Event and Censored Values

Total

Event

Censored

Percent
Censored

690

89

601

87.10

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1072.543
.
.

924.167
.
.

Model Chi-Square
148.376 with 4 DF (p=0.0001)
189.702 with 4 DF (p=0.0001)
127.844 with 4 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

MLOGRNA
CD4GRP1
CD4GRP2
COMBTHER

0.833237
2.364612
1.171137
-0.497161

0.17808
0.32436
0.34434
0.24389

21.89295
53.14442
11.56739
4.15520

0.0001
0.0001
0.0007
0.0415

1
1
1
1

218

OUTPUT FROM PROC PHREG, continued

Output from risklimits and test statements

Analysis of Maximum Likelihood Estimates

Conditional Risk Ratio and
95% Confidence Limits

Variable

Risk
Ratio

Lower

MLOGRNA
CD4GRP1
CD4GRP2
COMBTHER

2.301
10.640
3.226
0.608

1.623
5.634
1.643
0.377

Upper Label
3.262
20.093
6.335
0.981

log baseline rna (roche assay)

CD4<=100
100<CD4<=200
Combination therapy with AZT/ddI/ddC/Nvp

Linear Hypotheses Testing

Label
CD4LEVEL

Wald
Chi-Square

DF

Pr >
Chi-Square

55.0794

0.0001

219

OUTPUT FROM PROC PHREG, (Model 2)

Proportional hazards regression model for time to Death
Baseline viral load and CD4 predictors
Data Set: WORK.ALLOI
Dependent Variable: DTHTIME
Censoring Variable: DTHSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to death (days)

Death status (1=died,0=censored)

Summary of the Number of

Event and Censored Values

Total

Event

Censored

Percent
Censored

690

89

601

87.10

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1072.543
.
.

912.009
.
.

Model Chi-Square
160.535 with 6 DF (p=0.0001)
198.537 with 6 DF (p=0.0001)
132.091 with 6 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

MLOGRNA
CD4GRP1
CD4GRP2
COMBTHER
DECRS8
INCRS8

0.893838
2.023005
1.001046
-0.456506
-0.410919
-0.834101

0.18062
0.33594
0.34907
0.24687
0.26383
0.32884

24.48880
36.26461
8.22394
3.41950
2.42579
6.43367

0.0001
0.0001
0.0041
0.0644
0.1194
0.0112

1
1
1
1
1
1

220

OUTPUT FROM PROC PHREG, continued

Output from risklimits and test statements

Analysis of Maximum Likelihood Estimates

Conditional Risk Ratio and
95% Confidence Limits

Variable

Risk
Ratio

Lower

MLOGRNA
CD4GRP1
CD4GRP2
COMBTHER
DECRS8
INCRS8

2.444
7.561
2.721
0.633
0.663
0.434

1.716
3.914
1.373
0.390
0.395
0.228

Upper Label
3.483
14.606
5.394
1.028
1.112
0.827

log baseline rna (roche assay)

CD4<=100
100<CD4<=200
Combination therapy with AZT/ddI/ddC/Nvp
Decrease>=0.5 log rna at week 8?
Increase>=50 CD4 cells, week 8?

Linear Hypotheses Testing

Label
CD4LEVEL
WK8RESP

Wald
Chi-Square

DF

Pr >
Chi-Square

37.6833
10.4312

2
2

0.0001
0.0054

The percentage of treatment effect explained by including

the RNA and CD4 response to treatment by Week 8 is:
0.456
= 1
0.08
0.497
or 8%. The percentage of treatment effect on time to first
opportunistic infection or death is much higher (about 24%).
221

II. Predicted Survival using PH

The Cox PH model says that i(t, Z) = 0(t) exp(Z).
What does this imply about the survival function, Sz (t), for
the i-th individual with covariates Zi?
For the baseline (reference) group, we have:
S0(t) =

R
0t 0 (u)du
e

= e0(t)

This is by definition of a survival function (see intro notes).

For the i-th patient with covariates Zi, we have:

Si(t) = e
= e
= e
"

R
0t i (u)du

= ei(t)

R
0t 0 (u) exp(Zi )du

R
exp(Zi ) 0t 0 (u)du

= e

#
R
0t 0 (u)du exp(Zi )

= [S0(t)]exp(Zi)
(This uses the mathematical relationship [eb]a = eab)

222

Say we are interested in the survival pattern for single males

in the nursing home study. Based on the previous formula,
if we had an estimate for the survival function in the reference group, i.e., S0(t), we could get estimates of the survival
function for any set of covariates Zi.
How can we estimate the survival function, S0(t)?
We could use the KM estimator, but there are a few disadvantages of that approach:
It would only use the survival times for observations contained in the reference group, and not all the rest of the
survival times.
It would tend to be somewhat choppy, since it would
reflect the smaller sample size of the reference group.
Its possible that there are no subjects in the dataset
who are in the reference group (ex. say covariates are
age and sex; there is no one of age=0 in our dataset).

223

Instead, we will use a baseline hazard estimator which takes

advantage of the proportional hazards assumption to get a
smoother estimate.

Si(t) = [S0(t)]exp(Zi)
c
Using the above formula, we substitute
based on fitting the
Cox PH model, and calculate S0(t) by one of the following
approaches:

Breslow estimator (Stata)

Kalbfleisch/Prentice estimator (SAS)

224

(1) Breslow Estimator:

S0(t) = exp0(t)

0(t) is the estimated cumulative baseline hazard:

where
=
(t)

j:j <t

dj

kR(j ) exp(1 Z1k + . . . p Zpk )

(2) Kalbfleisch/Prentice Estimator

S0(t) =

j:j <t

where
j , j = 1, ...d are the MLEs obtained by assuming that S(t; Z) satisfies
S(t; Z) = [S0(t)]

eZ

225

j:j <t

eZ

j:j <t

je

Breslow Estimator: further motivation

The Breslow estimator is based on extending the concept
of the Nelson-Aalen estimator to the proportional hazards
model.
Recall that for a single sample with no covariates, the NelsonAalen Estimator of the cumulative hazard is:
=
(t)

dj
j:j <t rj
X

where dj and rj are the number of deaths and the number

at risk, respectively, at the j-th death time.
When there are covariates and assuming the PH model above,
one can generalize this to estimate the cumulative baseline
hazard by adjusting the denominator:
=
(t)

j:j <t

dj

kR(j ) exp(1 Z1k + . . . p Zpk )

Heuristic: The expected number of failures in (t, t + t) is

dj t

kR(t)

Hence,
t 0(tj )

0(t)exp(zk )

dj
P

kR(t) exp(zk )
226

Kalbfleisch/Prentice Estimator: further motivation

This method is analogous to the Kaplan-Meier Estimator.
Consider a discrete time model with hazard (1 j ) at the
j-th observed death time.
(Note: we use j = (1 j ) to simplify the algebra!)
Thus, for someone with z=0, the survivorship function is
S0(t) =

j:j <t

and for someone with Z 6= 0, it is:

S(t; Z) = S0(t)

eZ

j:j <t

eZ

j:j <t

je

The likelihood contributions under this model are:

for someone censored at t:

S(t; Z)

for someone who fails at tj :

S(t(j1) ; Z) S(tj ; Z) =

k<j

ez

The solution for j satisfies:

X

kDj

exp(Zk )
1

exp(Z )
j k

kRj

(Note what happens when Z = 0)

227

[1 je ]

exp(Zk )

Obtaining S0(t) from software packages

Stata provides the Breslow estimator of S0(t; Z), but not

predicted survivals at specified covariate values..... you
have to construct these yourself

SAS uses the Kalbfleisch/Prentice estimator of the baseline hazard, and can provide estimates of survival at arbitrary values of the covariates with a little bit of programming.

In practice, they are incredibly close! (see Fleming and

Harrington 1984, Communications in Statistics)

228

Using Stata to Predict Survival

The Stata command basesurv calculates the predicted survival values for the reference group, i.e., those subjects with
all covariates=0.
(1) Baseline Survival:
To obtain the estimated baseline survival S0(t), follow
the example below (for the nursing home data):
. use nurshome
. stset los fail
. stcox married health, basesurv(prsurv)
. sort los
. list los prsurv

229

Estimating the Baseline Survival with Stata

los

prsurv

1.
2.
3.
4.
5.

1
1
1
1
1

.99252899
.99252899
.99252899
.99252899
.99252899

22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.

1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
3
3

.99252899
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98671824
.98362595
.98362595

.
.
.

.
.
.

Stata creates a predicted baseline survival estimate for

every observed event time in the dataset, even if there
are duplicates.
230

(2) Predicted Survival for Subgroups

To obtain the estimated survival Si(t) for any other subgroup (i.e., not the reference or baseline group), follow
the Stata commands below:
. predict betaz, xb
. gen newterm=exp(betaz)
. gen predsurv=prsurv^newterm
. sort married health los
. list married health los predsurv

231

Predicting Survival for Subgroups with Stata

married

health

los

predsurv

1.
0
2
1
.9896138
8.
0
2
2
.981557
11.
0
2
3
.9772769
13.
0
2
4
.9691724
16.
0
2
5
.9586483
................................................................
300.
0
3
1
.9877566
302.
0
3
2
.9782748
304.
0
3
3
.9732435
305.
0
3
4
.9637272
312.
0
3
5
.9513916
................................................................
768.
0
4
1
.9855696
777.
0
4
2
.9744162
779.
0
4
3
.9685058
781.
0
4
4
.9573418
785.
0
4
5
.9428996
.
.
.
1468.
1
4
1
.9806339
1469.
1
4
2
.9657326
1472.
1
4
3
.9578599
1473.
1
4
5
.9239448
................................................................
1559.
1
5
1
.9771894
1560.
1
5
2
.9596928
1562.
1
5
3
.9504684
1564.
1
5
4
.9331349

232

Using SAS to Predict Survival

The SAS command BASELINE calculates the predicted survival values at the event times for a given set of covariate
values.
(1) To get the estimated baseline survival S0(t), create a
dataset with 0s for values of all covariates in the model
(2) To get the estimated survival Si(t) for any other subgroup (i.e., not the reference or baseline group), create a
data set which inputs the baseline values of the covariates for the subgroup of interest.

For either case, we then supply the corresponding dataset

name to the BASELINE command under PROC PHREG.
By giving the input dataset several lines, each corresponding
to a different combination of covariate values, we can compute predicted survival values for more than one group at
once.

233

(1) Baseline Survival Estimate

(note that the baseline survival function does not correspond
to any observations in our sample, since health status values
range from 2-5)

*** Estimating Baseline Survival Function under PH;

data inrisks;
input married health;
cards;
0 0
;
proc phreg data=pop out=survres;
model los*fail(0)=married health;
baseline covariates=inrisks out=outph survival=ps/nomean;
proc print data=outph;
title1 Nursinghome data: Baseline Survival Estimate;

234

Estimating the Baseline Survival with SAS

Nursinghome data: Baseline Survival Estimate

OBS

MARRIED

HEALTH

LOS

PS

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

1.00000
0.99253
0.98672
0.98363
0.97776
0.97012
0.96488
0.95856
0.95361
0.94793
0.94365
0.93792
0.93323
0.92706
0.92049
0.91461
0.91017
0.90534
0.90048
0.89635
0.89220
0.88727
0.88270

.
.
.
235

(2) Predicted Survival Estimate for Subgroup

The following SAS commands will generate the predicted
survival probability for each combination of covariates, at
every observed event time in the dataset.
*** Estimating Baseline Survival Function under PH;
data inrisks;
input married health;
cards;
0 2
0 5
1 2
1 5
;
proc phreg data=pop out=survres;
model los*fail(0)=married health;
baseline covariates=inrisks out=outph survival=ps/nomean;
proc print data=outph;
title1 Nursinghome data: predicted survival by subgroup;

236

Survival Estimates by Marital and Health Status

Nursinghome data: Predicted Survival by Subgroup
OBS

MARRIED

HEALTH

LOS

PS

1
0
2
0
1.00000
2
0
2
1
0.98961
3
0
2
2
0.98156
4
0
2
3
0.97728
................................................................
171
0
2
184
0.50104
172
0
2
185
0.49984
................................................................
396
0
5
0
1.00000
397
0
5
1
0.98300
398
0
5
2
0.96988
399
0
5
3
0.96295
................................................................
474
0
5
78
0.50268
475
0
5
80
0.49991
................................................................
791
1
2
0
1.00000
792
1
2
1
0.98605
793
1
2
2
0.97527
794
1
2
3
0.96955
................................................................
897
1
2
108
0.50114
898
1
2
109
0.49986
................................................................
1186
1
5
0
1.00000
1187
1
5
1
0.97719
1188
1
5
2
0.95969
1189
1
5
3
0.95047
................................................................
1233
1
5
47
0.50519
1234
1
5
48
0.49875

237

We can get a visual picture of what the proportional hazards assumption implies by looking at
these four subgroups

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

Subgroup

300

400

500 600 700 800 900 1000

LOS
Single, healthy
Single, unhealth
Married, healthy
Married, unhealt

238

III. Predicted medians and P-year survival

Predicted Medians
Suppose we want to find the predicted median survival for an
individual with a specified combination of covariates (e.g., a
single person with health status 5).
Three possible approaches:
(1) Calculate the median from the subset of individuals with
the specified covariate combination (using KM approach)
(2) Generate predicted survival curves for each combination
of covariates, and obtain the medians directly
OBS

MARRIED

HEALTH

LOS

PREDSURV

171
172

0
0

2
2

184
185

0.50104
0.49984

474
475

0
0

5
5

78
80

0.50268
0.49991

897
898

1
1

2
2

108
109

0.50114
0.49986

1233
1234

1
1

5
5

47
48

0.50519
0.49875

Recall that previously we defined the median as the

0.5, so the medians
smallest value of t for which S(t)
from above would be 185, 80, 109, and 48 days for single
healthy, single unhealthy, married healthy, and married
unhealthy, respectively.
239

(3) Generate the predicted survival curve from the estimated

baseline hazard, as follows:
We want the estimated median (M ) for an individual
with covariates Zi. We know
S(M ; Z) = [S0(M )]e

Zi

= 0.5

Hence, M satisfies (multiplying both sides by eZi ):

S0(M ) = [0.5]e

Ex. Suppose we want to estimate the median survival

for a single unhealthy subject from the nursing home
data. The reciprocal of the hazard ratio for unhealthy
(health=5) is: e0.1655 = 0.4373, (where = 0.165 for
health status)
So, we want M such that S0(M ) = (0.5)0.4373 = 0.7385
So the median for single unhealthy subject is the 73.8th
percentile of the baseline group.
OBS

MARRIED

HEALTH

LOS

PREDSURV

79
80
81

0
0
0

0
0
0

78
80
81

0.74028
0.73849
0.73670

So the estimated median would still be 80 days. Note: similar logic can be followed to estimate other quantiles besides
the median.
240

Estimating P-year survival

Suppose we want to find the P-year survival rate for an indi ; Zi )
vidual with a specified combination of covariates, S(P
For an individual with Zi = 0, the P-year survival can be
obtained from the baseline survivorship function, S0(P )
For individuals with Zi 6= 0, it can be obtained as:
; Zi) = [S0(P )]
S(P

eZi

Notes:
Although I say P-year survival, the units of time in a
particular dataset may be days, weeks, or months. The
answer here will be in the same units of time as the
original data.
c
If Z
i is positive, then the P-year survival rate for the ith individual will be lower than for a baseline individual.

Why is this true?

241

Model Selection in Survival Analysis

Suppose we have a censored survival time that we want to
model as a function of a (possibly large) set of covariates.
Two important questions are:
How to decide which covariates to use
How to decide if the final model fits well

To address these topics, well consider a new example:

Survival of Atlantic Halibut - Smith et al

Obs
#
100
109
113
116
..
.

Survival
Time
(min)
353.0
111.0
64.0
500.0

Tow
Censoring Duration
Indicator
(min.)
1
30
1
100
0
100
1
100

Diff
in
Depth
15
5
10
10

Length
of Fish
(cm)
39
44
53
44

Handling
Time
(min.)
5
29
4
4

Hosmer & Lemeshow

Chapter 5:
Model Development
Chapter 6:
Assessment of Model Adequacy
(sections 6.1-6.2)

242

Total
log(catch)
ln(weight)
5.685
8.690
5.323
5.323

Process of Model Selection

Collett (Section 3.6) has an excellent discussion of various
approaches for model selection. In practice, model selection
proceeds through a combination of
knowledge of the science
trial and error, common sense
automatic variable selection procedures
forward selection
backward selection
stepwise selection

Many advocate the approach of first doing a univariate analysis to screen out potentially significant variables for consideration in the multivariate model (see Collett).
Lets start with this approach.

243

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

Survival Distribution Function

Survival Distribution Function

Univariate KM plots of Atlantic Halibut survival

(continuous variables have been dichotomized)
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 1100 1200
SURVTIME
STRATA: LENGTHGP=0 LENGTHGP=1

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

Survival Distribution Function

Survival Distribution Function

0 100 200 300 400 500 600 700 800 900 1000 1100 1200
SURVTIME
STRATA: TOWDUR=0 TOWDUR=1

Survival Distribution Function

0 100 200 300 400 500 600 700 800 900 1000 1100 1200
SURVTIME
STRATA: DEPTHGP=0 DEPTHGP=1

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 1100 1200
SURVTIME
STRATA: HANDLGP=0 HANDLGP=1

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 1100 1200
SURVTIME
STRATA: LOGCATGP=0 LOGCATGP=1

Which covariates look like they might be important?

244

Automatic Variable selection procedures

in Stata and SAS
Statistical Software:
Stata: sw command before cox command
SAS: selection= option on model statement of
proc phreg

Options:
(1) forward
(2) backward
(3) stepwise
(4) best subset (SAS only, using score option)
One drawback of these options is that they can only handle
variables one at a time. When might that be a disadvantage?

245

Colletts Model Selection Approach

Section 3.6.1
This approach assumes that all variables are considered to
be on an equal footing, and there is no a priori reason to
include any specific variables (like treatment).
Approach:
(1) Fit a univariate model for each covariate, and identify
the predictors significant at some level p1, say 0.20.
(2) Fit a multivariate model with all significant univariate
predictors, and use backward selection to eliminate nonsignificant variables at some level p2, say 0.10.
(3) Starting with final step (2) model, consider each of the
non-significant variables from step (1) using forward selection, with significance level p3, say 0.10.
(4) Do final pruning of main-effects model (omit variables
that are non-significant, add any that are significant),
using stepwise regression with significance level p4. At
this stage, you may also consider adding interactions between any of the main effects currently in the model,
under the hierarchical principle.
Collett recommends using a likelihood ratio test for all variable inclusion/exclusion decisions.
246

Stata Command for Forward Selection:

Forward Selection = use pe() option, where is the
significance level for entering a variable into the model.
. use halibut
. stset survtime censor
. sw cox survtime towdur depth length handling logcatch,
> dead(censor) pe(.05)
begin with empty model
p
p
p
p

=
=
=
=

0.0000
0.0000
0.0010
0.0003

<
<
<
<

0.0500
0.0500
0.0500
0.0500

adding
adding
adding
adding

handling
logcatch
towdur
length

Cox Regression -- entry time 0

Number of obs
chi2(4)
Prob > chi2
Pseudo R2

Log Likelihood = -1257.6548

=
294
= 84.14
= 0.0000
= 0.0324

--------------------------------------------------------------------------survtime |
censor |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+----------------------------------------------------------------handling |
.0548994
.0098804
5.556
0.000
.0355341
.0742647
logcatch | -.1846548
.051015
-3.620
0.000
.2846423
-.0846674
towdur |
.5417745
.1414018
3.831
0.000
.2646321
.818917
length | -.0366503
.0100321
-3.653
0.000
-.0563129
-.0169877
---------------------------------------------------------------------------

247

Stata Command for Backward Selection:

Backward Selection = use pr() option, where is
the significance level for a variable to remain in the model.
. sw cox survtime towdur depth length handling logcatch,
> dead(censor) pr(.05)

begin with full model

p = 0.1991 >= 0.0500

removing depth

Cox Regression -- entry time 0

Number of obs
chi2(4)
Prob > chi2
Pseudo R2

Log Likelihood = -1257.6548

=
294
= 84.14
= 0.0000
= 0.0324

-------------------------------------------------------------------------survtime |
censor |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+---------------------------------------------------------------towdur |
.5417745
.1414018
3.831
0.000
.2646321
.818917
logcatch | -.1846548
.051015
-3.620
0.000 -.2846423
-.0846674
length | -.0366503
.0100321
-3.653
0.000 -.0563129
-.0169877
handling |
.0548994
.0098804
5.556
0.000
.0355341
.0742647
--------------------------------------------------------------------------

248

Stata Command for Stepwise Selection:

Stepwise Selection = use both pe(.) and pr(.) options,
with pr(.) > pe(.)
. sw cox survtime towdur depth length handling logcatch,
> dead(censor) pr(0.10) pe(0.05)

begin with full model

p = 0.1991 >= 0.1000

removing depth

Cox Regression -- entry time 0

Number of obs
chi2(4)
Prob > chi2
Pseudo R2

Log Likelihood = -1257.6548

=
294
= 84.14
= 0.0000
= 0.0324

------------------------------------------------------------------------survtime |
censor |
Coef.
Std. Err.
z
P>|z| [95% Conf. Interval]
---------+--------------------------------------------------------------towdur |
.5417745
.1414018
3.831
0.000 .2646321
.818917
handling |
.0548994
.0098804
5.556
0.000 .0355341
.0742647
length | -.0366503
.0100321
-3.653
0.000 -.0563129
-.0169877
logcatch | -.1846548
.051015
-3.620
0.000 -.2846423
-.0846674
-------------------------------------------------------------------------

It is also possible to do forward stepwise regression by including both pr(.) and pe(.) options with forward option

249

SAS programming statements for model selection

data fish;
infile fish.dat;
input ID SURVTIME CENSOR TOWDUR DEPTH LENGTH HANDLING LOGCATCH;
run;
title Survival of Atlantic Halibut;
*** automatic variable selection procedures;
proc phreg data=fish;
model survtime*censor(0)= towdur depth length handling logcatch
/selection=stepwise slentry=0.1 slstay=0.1 details;
title2 Stepwise selection;
run;
proc phreg data=fish;
model survtime*censor(0)= towdur depth length handling logcatch
/selection=forward slentry=0.1 details;
title2 Forward selection;
run;
proc phreg data=fish;
model survtime*censor(0)= towdur depth length handling logcatch
/selection=backward slstay=0.1 details;
title2 Backward selection;
run;
proc phreg data=fish;
model survtime*censor(0)= towdur depth length handling logcatch
/selection=score;
title2 Best subsets selection;
run;

250

Final model for stepwise selection approach

Survival of Atlantic Halibut
Stepwise selection
The PHREG Procedure
Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

TOWDUR
LENGTH
HANDLING
LOGCATCH

1
1
1
1

0.007740
-0.036650
0.054899
-0.184655

0.00202
0.01003
0.00988
0.05101

14.68004
13.34660
30.87336
13.10166

0.0001
0.0003
0.0001
0.0003

1.008
0.964
1.056
0.831

Analysis of Variables Not in the Model

Variable

Score
Chi-Square

Pr >
Chi-Square

1.6661

0.1968

DEPTH

Residual Chi-square = 1.6661

with 1 DF (p=0.1968)

NOTE: No (additional) variables met the 0.1 level for entry into the
model.
Summary of Stepwise Procedure

Step
1
2
3
4

Variable
Entered
Removed
HANDLING
LOGCATCH
TOWDUR
LENGTH

Number
In

Score
Chi-Square

Wald
Chi-Square

Pr >
Chi-Square

1
2
3
4

47.1417
18.4259
11.0191
13.4222

.
.
.
.

0.0001
0.0001
0.0009
0.0002

251

Output from PROC SAS score option

NUMBER OF
VARIABLES

SCORE
VALUE

VARIABLES INCLUDED
IN MODEL

1
47.1417
HANDLING
1
29.9604
TOWDUR
1
12.0058
LENGTH
1
4.2185
DEPTH
1
1.4795
LOGCATCH
--------------------------------2
65.6797
HANDLING LOGCATCH
2
59.9515
TOWDUR HANDLING
2
56.1825
LENGTH HANDLING
2
51.6736
TOWDUR LENGTH
2
47.2229
DEPTH HANDLING
2
32.2509
TOWDUR LOGCATCH
2
30.6815
TOWDUR DEPTH
2
16.9342
DEPTH LENGTH
2
14.4412
LENGTH LOGCATCH
2
9.1575
DEPTH LOGCATCH
------------------------------------3
76.8829
LENGTH HANDLING LOGCATCH
3
76.3454
TOWDUR HANDLING LOGCATCH
3
75.5291
TOWDUR LENGTH HANDLING
3
69.0334
DEPTH HANDLING LOGCATCH
3
60.0340
TOWDUR DEPTH HANDLING
3
56.4207
DEPTH LENGTH HANDLING
3
55.8374
TOWDUR LENGTH LOGCATCH
3
52.4130
TOWDUR DEPTH LENGTH
3
34.7563
TOWDUR DEPTH LOGCATCH
3
24.2039
DEPTH LENGTH LOGCATCH
-------------------------------------------4
94.0062
TOWDUR LENGTH HANDLING LOGCATCH
4
81.6045
DEPTH LENGTH HANDLING LOGCATCH
4
77.8234
TOWDUR DEPTH HANDLING LOGCATCH
4
75.5556
TOWDUR DEPTH LENGTH HANDLING
4
59.1932
TOWDUR DEPTH LENGTH LOGCATCH
------------------------------------------------5
96.1287
TOWDUR DEPTH LENGTH HANDLING LOGCATCH
------------------------------------------------------

252

Best multivariate model for all 3 options

Survival of Atlantic Halibut
Best Multivariate Model
The PHREG Procedure
Data Set: WORK.FISH
Dependent Variable: TIME
Censoring Variable: CENSOR
Censoring Value(s): 0
Ties Handling: BRESLOW

Summary of the Number of

Event and Censored Values

Total

Event

Censored

Percent
Censored

294

273

21

7.14

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

2599.449
.
.

2515.310
.
.

Model Chi-Square
84.140 with 4 DF (p=0.0001)
94.006 with 4 DF (p=0.0001)
90.247 with 4 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

Standard
Error

TOWDUR
LENGTH
HANDLING
LOGCATCH

1
1
1
1

0.007740
-0.036650
0.054899
-0.184655

0.00202
0.01003
0.00988
0.05101

253

Wald
Pr >
Chi-Square Chi-Square
14.68004
13.34660
30.87336
13.10166

0.0001
0.0003
0.0001
0.0003

Risk
Ratio
1.008
0.964
1.056
0.831

Notes:
When the halibut data was analyzed with the forward,
backward and stepwise options, the same final model was
reached. However, this will not always be the case.
Variables can be forced into the model using the lockterm
option in Stata and the include option in SAS. Any
variables that you want to force inclusion of must be
listed first in your model statement.
Stata uses the Wald test for both forward and backward
selection, although it has an option to use the likelihood
ratio test instead (lrtest). SAS uses the score test to
decide what variables to add and the Wald test for what
variables to remove.
If you fit a range of models manually, you can apply the
AIC criteria described by Collett:
+ ( q)
minimize AIC = 2 log(L)
where q is the number of unknown parameters in the
model and is typically between 2 and 6 (they suggest
= 3).
The model is then chosen which minimizes the AIC (similar to maximizing log-likelihood, but with a penalty for
number of variables in the model)

254

Questions:
When might we want to force certain variables into the
model?
(1) to examine interactions
(2) to keep main effects in the model
(3) to calculate a score test for a paricular effect

Would it be possible to get different final models from

SAS and Stata?
Based on what weve seen in the behavior of Wald tests,
would SAS or Stata be more likely to add a covariate to
a model in a forward selection model?
If we use the AIC criteria with = 3, how does that
compare to the likelihood ratio test?

255

Assessing overall model fit

How do we know if the model fits well?
Always look at univariate plots (Kaplan-Meiers)
Construct a Kaplan-Meier survival plot for each of the important predictors, like the ones shown at the beginning of these
notes.

Check proportionality assumption (this will be the topic

of the next lecture)
Check residuals!
(a) generalized (Cox-Snell)
(b) martingale
(c) deviance
(d) Schoenfeld
(e) weighted Schoenfeld

256

Residuals for survival data are slightly different than for

other types of models, due to the censoring. Before we start
talking about residuals, we need an important basic result:
Inverse CDF:
If Ti (the survival time for the i-th individual) has
survivorship function Si(t), then the transformed
random variable Si(Ti) (i.e., the survival function
evaluated at the actual survival time Ti) should
be from a uniform distribution on [0, 1], and hence
log[Si(Ti)] should be from a unit exponential distribution
More mathematically:
If Ti Si(t)
then Si(Ti) U nif orm[0, 1]
and log Si(Ti) Exponential(1)

257

(a) Generalized (Cox-Snell) Residuals:

The implication of the last result is that if the model is correct, the estimated cumulative hazard for each individual at
the time of their death or censoring should be like a censored
sample from a unit exponential. This quantity is called the
generalized or Cox-Snell residual.
Here is how the generalized residual might be used. Suppose
we fit a PH model:
S(t; Z) = [S0(t)]exp(Z)
or, in terms of hazards:
(t; Z) = 0(t) exp(Z)
= 0(t) exp(1 Z1 + 2Z2 + + k Zk )

After fitting, we have:

1, . . . , k

S0(t)

258

So, for each person with covariates Zi, we can get

Zi) = [S0(t)]exp(Zi)
S(t;
This gives a predicted survival probability at each time t in
the dataset (see notes from the previous lecture).

Then we can calculate

i = log[S(T
i; Zi)]

In other words, first we find the predicted survival probability at the actual survival time for
an individual, then log-transform it.

259

Example: Nursing home data

Say we have
a single male
with actual duration of stay of 941 days (Xi = 941)
We compute the entire distribution of survival probabilities

for single males, and obtain S(941)

= 0.260.

log[S(941,
single male)] = log(0.260) = 1.347
We repeat this for everyone in our dataset. These should be
like a censored sample from an exponential (1) distribution
if the model fits the data well.
Based on the properties of a unit exponential model

plotting log(S(t))
vs t should yield a straight line
plotting log[ log S(t)] vs log(t) should yield a straight
line through the origin with slope=1.
To convince yourself of this, start with S(t) = et and
calculate log[ log S(t)]. What do you get for the slope and
intercept?
(Note: this does not necessarily mean that the underlying
distribution of the original survival times is exponential!)
260

Obtaining the generalized residuals from Stata

Fit a Cox PH model with the stcox command, along
with the mgale(newvar) option
Use the predict command with the csnell option
Define a survival dataset using the Cox-Snell residuals
as the pseudo failure times
Calculate the estimated KM survival
Take the log[ log(S(t))] based on the above
Generate the log of the Cox-Snell residuals
Graph log[ log S(t)] vs log(t)
. stcox towdur handling length logcatch, mgale(mg)
. predict csres, csnell
. stset csres censor
. sts list
. sts gen survcs=s
. gen lls=log(-log(survcs))
. gen loggenr=log(csres)
. graph lls loggenr

261

LLS
2
1
0
-1
-2
-3
-4
-5
-6
-5

-4

-3

-2
-1
Log of SURVIVAL

Allison states Cox-Snell residuals... are not very informative for

Cox models estimated by partial likelihood. He instead prefers
deviance residuals (later).

262

Obtaining the generalized residuals from SAS

The generalized residuals can be obtained from SAS

after fitting a PH model using the output statement with
the logsurv option.
proc phreg data=fish;
model survtime*censor(0) = towdur handling logcatch length;
output out=phres logsurv=genres;
*** take negative log Pr(survival) at each persons survtime;
data phres;
set phres;
genres=-genres;
*** Now we treat the generalized residuals as the input dataset;
*** to evaluate whether the assumption of an exponential;
*** distribution is appropriate;
proc lifetest data=phres outsurv=survres;
time genres*censor(0);
data survres;
set survres;
lls=log(-log(survival));
loggenr=log(genres);
proc gplot data=survres;
plot lls*loggenr;
run;

263

(b) Martingale Residuals

(see Fleming and Harrington, p.164)
Martingale residuals are defined for the i-th individual as:

i)
ri = i (T
Properties:
ris have mean 0
range of ris is between and 1
approximately uncorrelated (in large samples)
Interpretation: - the residual ri can be viewed as the
difference between the observed number of deaths (0 or
1) for subject i between time 0 and Ti, and the expected
numbers based on the fitted model.

264

The martingale residuals can be obtained from Stata

using the mgale option shown previously.
Once the martingale residual is created, you can plot it versus
the predicted log HR (i.e., Zi), or any of the individual
covariates.

. stcox towdur handling length logcatch, mgale(mg)

. predict betaz=xb
. graph mg betaz
. graph mg logcatch
. graph mg towdur
. graph mg handling
. graph mg length

265

The martingale residuals can be obtained from SAS

after fitting a PH model using the output statement with
the resmart option.
Once you have them, you can
plot against predicted values
plot against covariates

proc phreg data=fish;

model survtime*censor(0) = towdur handling logcatch length;
output out=phres resmart=mres xbeta=xb;
proc gplot data=phres;
plot mres*xb;
plot mres*towdur;
plot mres*handling;
plot mres*logcatch;
plot mres*length;
run;

/* predicted values */

Allison still prefers the deviance residuals (next)

266

Martingale Residuals
M 1
a
r 0
t
i -1
n
g -2
a
l
e -3

M 1
a
r 0
t
i -1
n
g -2
a
l
e -3

R -4
e
s -5
i
d
u -6
a 0
l

R -4
e
s -5
i
d
u -6
a 20
l

10 20 30 40 50 60 70 80 90 100 110 120

Towing Duration

M 1
a
r 0
t
i -1
n
g -2
a
l
e -3

M 1
a
r 0
t
i -1
n
g -2
a
l
e -3

R -4
e
s -5
i
d
u -6
a 2
l

R -4
e
s -5
i
d
u -6
a 0
l

5
6
Log(weight) of catch

M 1
a
r 0
t
i -1
n
g -2
a
l
e -3
R -4
e
s -5
i
d
u -6
a -3
l

-2

-1
Linear Predictor

267

30

40
Length of fish (cm)

50

60

10

20
Handling time

30

40

(c) Deviance Residuals

One problem with the martingale residuals is that they tend
to be asymmetric.
A solution is to use deviance residuals. For person i,
these are defined as a function of the martingale residuals
(ri):
r

i = sign(
D
ri) 2[
ri + ilog(i ri)]
In Stata, the deviance residuals are generated using the same
approach as the Cox-Snell residuals.
. stcox towdur handling length logcatch, mgale(mg)
. predict devres, deviance

and then they can be plotted versus the predicted log(HR)

or the individual covariates, as shown for the Martingale
residuals.
In SAS, just use resdev option instead of resmart.
Deviance residuals behave much like residuals from OLS regression (i.e., mean=0, s.d.=1). They are negative for observations with survival times that are smaller than expected.

268

Deviance Residuals
3
D
e 2
v
i
a 1
n
c 0
e

3
D
e 2
v
i
a 1
n
c 0
e

R -1
e
s
i -2
d
u
a -3
l 0

R -1
e
s
i -2
d
u
a -3
l 20

10 20 30 40 50 60 70 80 90 100 110 120

Towing Duration

3
D
e 2
v
i
a 1
n
c 0
e

3
D
e 2
v
i
a 1
n
c 0
e

R -1
e
s
i -2
d
u
a -3
l 2

R -1
e
s
i -2
d
u
a -3
l 0

5
6
7
Log(weight) of total catch
4
D
e
v 3
i
a
n 2
c
e
1
R
e
s 0
i
d
u
a -1
l 0.0

0.1

0.2

0.3

0.4 0.5 0.6

Linear Predictor

269

0.7

0.8

30

40
Length of fish (cm)

50

60

10

20
Handling time

30

40

0.9

(d) Schoenfeld Residuals

These are defined at each observed failure time as:
rs = Zij (ti) Zj (ti)
ij

Notes:
represent the difference between the observed covariate
and the average over the risk set at that time
calculated for each covariate

not defined for censored failure times.

useful for assessing time trend or lack or proportionality,

based on plotting versus event time
sum to zero, have expected value zero, and are uncorrelated (in large samples)

In Stata, the Schoenfeld residuals are generated in the stcox

command itself, using the schoenf(newvar(s)) option:
. stcox towdur handling length logcatch, schoenf(towres handres lenres
logres)
. graph towres survtime

In SAS, add to the output line

RESSCH=name1 name2 ... namek
for up to k regressors in the model.
270

Schoenfeld Residuals
20

50
40
30
20
10
0
-10
-20
-30
-40
-50
-60

10
0
-10
-20
0 100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

0 100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

30

5
4

20

3
2

10

1
0

0
-1

-10

-2
-3

-20
0

100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

0 100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

271

(e) Weighted Schoenfeld Residuals

These are actually used more often than the previous unweighted version, because they are more like the typical OLS
residuals (i.e., symmetric around 0).
They are defined as:
rijw = nVc rijs
The weighted residwhere Vc is the estimated variance of .
uals can be used in the same way as the unweighted ones to
assess time trends and lack of proportionality.
In Stata, use the command:
. stcox towdur length logcatch handling depth, scaledsch(towres2
> lenres2 logres2 handres2 depres2)
. graph logres2 survtime

In SAS, add to the output line

WTRESSCH=name1 name2 ... namek
for up to k regressors in the model.

272

Weighted Schoenfeld Residuals

0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
-0.01
-0.02
-0.03
-0.04
-0.05
-0.06
-0.07
-0.08

0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
-0.4
0 100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

0 100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
-0.4

3
2
1
0
-1
-2
0

100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

0 100 200 300 400 500 600 700 800 900 1000 1100 1200
Survival Time

273

Using Residual plots to explore relationships

If you calculate martingale or deviance residuals without any
covariates in the model and then plot against covariates, you
obtain a graphical impression of the relationship between the
covariate and the hazard.
In Splus, it is easy to do this (also possible in stata using the
estimate option)
** read in the dataset and fit a cox PH model
fish_read.table(fish.data,header=T)
x_fish$towdur
fishres_coxreg(fish$time, fish$censor, x, resid="martingale",iter.max=0)
** the 2 commands below set up the postscript file, with 4 graphs
postscript("fishres.plt",horizontal=F,height=10,width=7)
par(mfrow=c(2,2),oma=c(0,0,2,0))
** plot the martingale residuals vs each of the other covariates
** and add a lowess smoothed fit to the plot
plot(fish$depth, fishres$resid, xlab="depth")
lines(lowess(fish$depth,fishres$resid,iter=0))
plot(fish$length, fishres$resid, xlab="length")
lines(lowess(fish$length,fishres$resid,iter=0))
plot(fish$handling, fishres$resid, xlab="handling")
lines(lowess(fish$handling,fishres$resid,iter=0))
plot(fish$logcatch, fishres$resid, xlab="logcatch")
lines(lowess(fish$logcatch,fishres$resid,iter=0))

274

1
0
-3

-1

-2

-4

0
-1
-2

fishres$resid

fishres$resid

1
-3

-4

Splus Plots of Martingale Residuals for Cox Model

containing only towing duration as a predictor,
vs other covariates

0 10

30

50

30

40

-2

-1

-3

-4

Length

fishres$resid

-1
-2

-3
-4

fishres$resid

depth

50

logcatch

10

20
handling

275

30

(f ) Deletion diagnostics
Deletion diagnostics are defined generally as:
i = (i)
In other words, they are the difference between the estimated
regression coefficient using all observations and that without
the i-th individual. This can be useful for assessing the influence of an individual.
In SAS PROC PHREG, we use the dfbeta option:
(Note that there is a separate dfbeta calculated for each of
the predictors.)
proc phreg data=fish;
model survtime*censor(0)=towdur handling logcatch length;
id id;
output out=phinfl dfbeta=dtow dhand dlogc dlength ld=lrchange;
proc univariate data=phinfl;
var dtow dhand dlogc dlength lrchange;
id id;
run;

The proc univariate procedure will supply the 5 smallest values and the 5 largest values. The id statement means that
these will be labeled with the value of id from the dataset.

276

(g) Other Influence diagnostics

Other influence diagnostics:
The LD option is another method for checking influence. It
calculates how much the log-likelihood (x2) would change if
the i-th person was removed from the sample.

LDi = 2 logL() logL( i)

= MLE for all parameters with everyone included

c
i = MLE with i-th subject omitted

Again, the proc univariate procedure in SAS will identify

the observations with the largest and smallest values of the
lrchange diagnostic measure.

277

Can we improve the model?

The plots appear to have some structure, which indicate that
we could be leaving something out. It is always a good idea
to check for interactions:
In this case, there are several important interactions. I used
a backward selection model forcing all main effects to be
included, and considering all pairwise interactions. Here are
the results:
Parameter
Estimate

Standard
Wald
Error
Chi-Square

Pr >
Chi-Square

Risk
Ratio

Variable

DF

TOWDUR
DEPTH
LENGTH
HANDLING
LOGCATCH

1
1
1
1
1

-0.075452
0.123293
-0.077300
0.004798
-0.225158

0.01740
0.06400
0.02551
0.03221
0.07156

18.79679
3.71107
9.18225
0.02219
9.89924

0.0001
0.0541
0.0024
0.8816
0.0017

0.927
1.131
0.926
1.005
0.798

TOWDEPTH
TOWLNGTH
TOWHAND
DEPLNGTH
DEPHAND

1
1
1
1
1

0.002931
0.001180
0.001107
-0.006034
-0.004104

0.0004996
0.0003541
0.0003558
0.00136
0.00118

34.40781
11.10036
9.67706
19.77360
12.00517

0.0001
0.0009
0.0019
0.0001
0.0005

1.003
1.001
1.001
0.994
0.996

Interpretation:
Handling alone doesnt seem to affect survival, unless it is
combined with a longer towing duration or shallower trawling depths.

278

An alternative modeling strategy when we have

fewer covariates
With a dataset with only 5 main effects, it would make sense
to consider interactions from the start. How many would
there be?
Fit model with all main effects and pairwise interactions

Then use backward selection to eliminate non-significant

pairwise interactions (remember to force the main effects
into the model at this stage)
Once non-significant pairwise interactions have been eliminated, you could consider backwards selection to eliminate any non-significant main effects that are not involved in remaining interaction terms
After obtaining final model, use residuals to check fit of
model.

279

Assessing the PH Assumption

So far, weve been considering the following Cox PH model:
(t, Z) = 0(t) exp(Z)
X
= 0(t) exp ( j Zj )
where j is the parameter for the the j-th covariate (Zj ).

Important features of this model:

(1) the baseline hazard depends on t, but not on the covariates Z1, ..., Zp
(2) the hazard ratio, i.e., exp(Z), depends on the covariates
Z = (Z1, ..., Zp), but not on time t.

Assumption (2) is what led us to call this a proportional

hazards model. Thats because we could take the ratio of
the hazards for two individuals with covariates Zi and Zi0 ,
and write it as a constant in terms of the covariates.

280

Proportional Hazards Assumption

Hazard Ratio:
0(t) exp(Zi)
(t, Zi)
=
(t, Zi0 )
0(t) exp(Zi0 )
=

exp(Zi)
exp(Zi0 )

= exp[(Zi Zi0 )]
= exp[

j (Zij Zi0j )] =

In the last formula, Zij is the value of the j-th covariate for
the i-th individual. For example, Z42 might be the value of
gender (0 or 1) for the the 4-th person.
We can also write the hazard for the i-th person as a constant
times the hazard for the i0-th person:
(t, Zi) = (t, Zi0 )
Thus, the HR between two types of individuals is constant
(i.e., =) over time. These are mathematical ways of stating
the proportional hazards assumption.

281

There are several options for checking the assumption of proportional hazards:
I. Graphical
(a) Plots of survival estimates for two subgroups
vs log(t) for two subgroups
(b) Plots of log[ log(S)]
(c) Plots of weighted Schoenfeld residuals vs time
(d) Plots of observed survival probabilities versus expected under PH model (see Kleinbaum, ch.4)
II. Use of goodness of fit tests - we can construct
a goodness-of-fit test based on comparing the observed
survival probability (from sts list) with the expected
(from stcox) under the assumption of proportional hazards - see Kleinbaum ch.4
III. Including interaction terms between a covariate and t (time-dependent covariates)

282

How do we interpret the above?

Kleinbaum (and other texts) suggest a strategy of assuming
that PH holds unless there is very strong evidence to counter
this assumption:
estimated survival curves are fairly separated, then cross
estimated log cumulative hazard curves cross, or look
very unparallel over time
weighted Schoenfeld residuals clearly increase or decrease
over time (you could fit a OLS regression line and see if
the slope is significant)
test for time covariate interaction term is significant
(this relates to time-dependent covariates)
If PH doesnt exactly hold for a particular covariate but we
fit the PH model anyway, then what we are getting is sort
of an average HR, averaged over the event times.
In most cases, this is not such a bad estimate. Allison claims
that too much emphasis is put on testing the PH assumption,
and not enough to other important aspects of the model.

283

Implications of proportional hazards

Consider a PH model with a single covariate, Z:
(t; Z) = 0(t)eZ
What does this imply for the relation between the survivorship functions at various values of Z?
Under PH,
log[ log[S(t; Z)]] = log[ log[S0(t)]] + Z
In general, we have the following relationship:
i(t) =
=

Z t

Z t

i(u)du
0(u) exp(Zi)du

= exp(Zi)

Z t

0(u)du

= exp(Zi) 0 (t)
This means that the ratio of the cumulative hazards is the
same as the ratio of hazard rates:
i(t)
= exp(Zi) = exp(1Z1i + + pZpi)
0(t)

284

Using the above relationship, we can show that:

i(t)

Zi = log
0(t)
= log i(t) log 0(t)
= log[ log Si(t)] log[ log S0(t)]
so log[ log Si(t)] = log[ log S0(t)] + Zi
Thus, to assess if the hazards are actually proportional to
each other over time (using graphical option I(b))
calculate Kaplan Meier Curves for various levels of Z

Z))] (i.e., log cumulative hazard)

compute log[ log(S(t;
plot vs log-time to see if they are parallel (lines or curves)

Note: If Z is continuous, break into categories.

285

Question: Why not just compare the underlying

hazard rates to see if they are proportional?
Heres two simulated examples with hazards which are truly
proportional between the two groups:

Weibull-type hazard:

U-shaped hazard:

Plots of hazard function vs time

Simulated data with HR=2 for men vs women

Plots of hazard function vs time

Simulated data with HR=2 for men vs women

HAZARD
0.010
0.008
0.006
0.004
0.002
0.000

HAZARD
0.010
0.008
0.006
0.004
0.002
0.000
0 100 200 300 400 500 600 700 800 900 1000 1100
Length of Stay (days)
Gender Women
Men

0 100 200 300 400 500 600 700 800 900 1000 1100
Length of Stay (days)
Gender Women
Men

Reason 1: Its hard to eyeball these figures and

see that the hazard rates are proportional - it
would be easier to look for a constant shift between lines.

286

Reason 2: Estimated hazard rates tend to be

more unstable than the cumulative hazard rate
Consider the nursing home example (where we think PH is
reasonable). If we group the data into intervals and calculate
the hazard rate using actuarial method, we get these plots:
200 day intervals:

100 day intervals:

Plots of hazard function vs time

Plots of hazard function vs time

0.006
0.005

0.009
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0.000

0.004
0.003
0.002
0.001
0.000
0

100 200 300 400 500 600 700 800 900 1000
Length of Stay (days)
Gender Women
Men

50 day intervals:

100 200 300 400 500 600 700 800 900 1000
Length of Stay (days)
Gender Women
Men

25 day intervals:

Plots of hazard function vs time

Plots of hazard function vs time

0.012
0.010

0.014
0.012
0.010
0.008
0.006
0.004
0.002
0.000

0.008
0.006
0.004
0.002
0.000
0 100 200 300 400 500 600 700 800 900 1000 1100
Length of Stay (days)
Gender Women
Men

0 100 200 300 400 500 600 700 800 900 1000 1100
Length of Stay (days)
Gender Women
Men

287

In contrast, the log cumulative hazard plots are

easier to interpret and tend to give more stable
estimates
Ex: Nursing Home - gender and marital status
proc lifetest data=pop outsurv=survres;
time los*fail(0);
strata gender;
format gender sexfmt.;
title Duration of Length of Stay in nursing homes;
data survres;
set survres;
label log_los=Log(Length of stay in days);
if los > 0 then log_los=log(los);
if survival<1 then lls=log(-log(survival));
proc gplot data=survres;
plot lls*log_los=gender;
format gender sexfmt.;
title2 Plots of log-log KM versus log-time;
run;

The statements for marital status are similar, substituting married

for gender.

Note: This is equivalent to comparing plots of the log cumu

lative hazard, log((t)),
between the covariate levels, since
(t) =

Z t

(u; Z)du = log[S(t)]

288

Assessment of proportional hazards for gender

and marital status in nursing home data (Morris)
Plots of log-log KM versus log-time
LLS
1
0
-1
-2
-3
-4
-5
-6
0

3
4
5
Log(Length of stay in days)
Gender Women
Men

Plots of log-log KM versus log-time

LLS
1
0
-1
-2
-3
-4
-5
-6
0

3
4
5
Log(Length of stay in days)
Marital Status
Single Married
289

Assessing proportionality with several covariates

If there is enough data and you only have a couple of covariates, create a new covariate that takes a different value for
every combination of covariate values.
Example: Health status and gender for nursing home
data pop;
infile ch12.dat;
input los age rx gender married health fail;
if gender=0 and health=2 then hlthsex=1;
if gender=1 and health=2 then hlthsex=2;
if gender=0 and health=5 then hlthsex=3;
if gender=1 and health=5 then hlthsex=4;
proc format;
value hsfmt
1=Healthier Women
2=Healthier Men
3=Sicker Women
4=Sicker Men;
proc lifetest data=pop outsurv=survres;
time los*fail(0);
strata hlthsex;
format hlthsex hsfmt.;
title Length of Stay in nursing homes;
data survres;
set survres;
label log_los=Log(Length of stay in days);
label hlthsex=Health/Gender Status;
if los > 0 then log_los=log(los);
if survival<1 lls=log(-log(survival));
proc gplot data=survres;
plot lls*log_los=hlthsex;
format hlthsex hsfmt.;
title2 Plots of log-log KM versus log-time;
run;

290

Log[-log(survival)] Plots for Health status*gender

Plots of log-log KM versus log-time

LLS
1
0
-1
-2
-3
-4
-5
0

3
4
5
6
Log(Length of stay in days)
Health/Gender Status
Healthier Women
Healthier Men
Sicker Women
Sicker Men

If there are too many covariates (or not enough data) for this,
then there is a way to test proportionality for each variable,
one at a time, using the stratification option.

291

What if proportional hazards fails?

do a stratified analysis
include a time-varying covariate to allow changing hazard ratios over time
include interactions with time
The second two options relate to time-dependent covariates,
which will be covered in future lectures.
We will focus on the first alternative, and then the second
two options will be briefly described.

292

Stratified Analyses
Suppose:
we are happy with the proportionality assumption on Z1
proportionality simply does not hold between various
levels of a second variable Z2.

If Z2 is discrete (with a levels) and there is enough data, fit

the following stratified model:
(t; Z1, Z2) = Z2 (t)eZ1
For example, a new treatment might lead to a 50% decrease
in hazard of death versus the standard treatment, but the
hazard for standard treatment might be different for each
hospital.
A stratified model can be useful both for primary
analysis and for checking the PH assumption.

293

Assessing PH Assumption for Several Covariates

Suppose we have several covariates (Z = Z1, Z2, ... Zp), and
we want to know if the following PH model holds:
(t; Z) = 0(t) e1 Z1+...+pZp
To start, we fit a model which stratifies by Zk :
(t; Z) = 0Zk (t) e1Z1+...+k1 Zk1 +k+1 Zk+1 +...+pZp
Since we can estimate the survival function for any subgroup,
we can use this to estimate the baseline survival function,
S0Zk (t), for each level of Zk .
Then we compute log S(t) for each level of Zk , controlling
for the other covariates in the model, and graphically check
whether the log cumulative hazards are parallel across strata
levels.

294

Ex: PH assumption for gender (nursing home data):

include married and health as covariates in a Cox PH
model, but stratify by gender.
calculate the baseline survival function for each level of
the variable gender (i.e., males and females)
plot the log-cumulative hazards for males and females
and evaluate whether the lines (curves) are parallel
In the above example, we make the PH assumption for married
and health, but not for gender.
This is like getting a KM survival estimate for each gender without assuming PH, but is more flexible since we can
control for other covariates.
We would repeat the stratification for each variable for which
we wanted to check the PH assumption.

295

SAS Code for Assessing PH within Stratified Model:

data pop;
infile ch12.dat;
input los age rx gender married health fail;
if los<=0 then delete;
data inrisks;
input married health;
cards;
0 2
;
proc format;
value sexfmt
1=Male
0=Female;
proc phreg data=pop;
model los*fail(0)=married health;
strata gender;
baseline covariates=inrisks out=outph
loglogs=lls / nomean;
proc print data=outph;
title Log Cumulative Hazard Estimates by Gender;
title2 Controlling for Marital and Health Status;
data outph;
set outph;
if los>0 then log_los=log(los);
label log_los=Log(LOS)
lls=Log Cumulative Hazard;
proc gplot data=outph;
plot lls*log_los=gender;
format gender sexfmt.;
title1 Log-log Survival versus log-time by Gender;
run;

296

Log[-log(survival)] Plots for Gender

Controlling for Marital and Health Status

2
1
0
-1
-2
-3
-4
-5
5.8 5.9 6.0 6.1 6.2 6.3 6.4 6.5 6.6 6.7 6.8 6.9 7.0
Log(LOS)
GENDER
Female
Male

297

Models with Time-dependent Interactions

Consider a PH model with two covariates Z1 and Z2. The
standard PH model assumes
(t; Z) = 0(t) e1Z1+2Z2
However, if the log-hazards are not really parallel between
the groups defined by Z2, then you can try adding an interaction with time:
(t; Z) = 0(t) e1Z1+2Z2+3Z2t
A test of the coefficient 3 would be a test of the proportional
hazards assumption for Z2.
If 3 is positive, then the hazard ratio would be increasing
over time; if negative, then decreasing over time.
Changes in covariate status sometimes occur naturally during a study (ex. patient gets a kidney transplant), and are
handled by introducing time-dependent covariates.

298

Using Stata to Assess Proportional Hazards

Stata has two commands which can be used to graphically
assess the proportional hazards assumption, using graphical
options (b) and (d) described previously:
stphplot: plots log[ log((S(t))] curves for each
category of a nominal or ordinal independent variable
versus log(time). Optionally, these estimates can be adjusted for other covariates.
stcoxkm: plots Kaplan-Meier observed survival curves
and compares them to the Cox predicted curves for the
same variable. (No need to run stcox prior to this command, it will be done automatically)

For either command, you must have stset your data first.
You must specify by() with stcoxkm and you must specify
either by() or strata() with stphplot.

299

Assessing PH Assumption for a Single Covariate

by Comparing log[ log(S(t))] Curves
. use nurshome
. stset los fail
. stphplot, by(gender)

Note that the lines will be going from top left to bottom right,
rather than bottom left to top right, since we are plotting
log[ log(S(t))] rather than log[ log(S(t))].
This will give a plot similar to that on p.10 (top).
Of course, youll want to make your plot prettier by adding
titles and labels, as follows:
. stphplot, by(gender) xlab ylab b2(log(Length of Stay))
>title(Evaluation of PH Assumption) saving(phplot)

300

Assessing PH Assumption for Several Covariates

by Comparing log[ log(S(t))] Curves
. use nurshome
. stset los fail
. gen hlthsex=1
. replace hlthsex=2 if health==2 & gender==1
. replace hlthsex=3 if health==5 & gender==0
. replace hlthsex=4 if health==5 & gender==1
. tab hlthsex
. stphplot, by(hlthsex)

This will give a plot similar to that on p.12.

301

Assessing PH Assumption for a Single Covariate

Controlling for the Levels of Other Covariates
. use nurshome
. stset los fail
. stphplot, strata(gender) adjust(married health)

This will produce a plot similar to that on p.18.

302

Assessing PH Assumption for a Covariate

By Comparing Cox PH Survival to KM Survival
To construct plots based on option I(d), use the stcoxkm
command, either for a single covariate or for a newly generated covariate (like hlthsex) which represents combined
levels of more than one covariate.

. use nurshome
. stset los fail
. stcoxkm, by(gender)
. stcoxkm, by(hlthsex)

As usual, youll want to add titles, labels, and save your

graph for later use.

303

Time varying (or time-dependent) covariates

References:
Allison (*)
Hosmer & Lemeshow
Kalbfleisch & Prentice
Collett
Kleinbaum
Cox & Oakes
Andersen & Gill

p.138-153
Chapter 7, Section 3
Section 5.3
Chapter 7
Chapter 6
Chapter 8
Page 168 (Advanced!)

So far, weve been considering the following Cox PH model:

(t, Z) = 0(t) exp(Z)
X
= 0(t) exp( j Zj )
where j is the parameter for the the j-th covariate (Zj ).
Important features of this model:
(1) the baseline hazard depends on t, but not on the covariates Z1, ..., Zp
(2) the hazard ratio exp(Z) depends on the covariates Z1, ..., Zp,
but not on time t.
Now we want to relax the second assumption, and allow the
hazard ratio to depend on time t.

304

Example to motivate time-dependent covariates

Stanford Heart transplant example:
Variables:
survival - time from program enrollment until death or censoring
dead - indicator of death (1) or censoring (0)

transpl - whether patient ever had transplant

(1 if yes, 2 if no)
surgery - previous heart surgery prior to program
age - age at time of acceptance into program

wait - time from acceptance into program until transplant

surgery (=. for those without transplant)

Initially, a Cox PH model was fit for predicting survival time:

(t, Z) = 0 (t) exp(1 transpl + 2 surgery + 3 age)

However, this model could give misleading results, since patients who died more quickly had less time available to get
transplants. A model with a time dependent indicator of
whether a patient had a transplant at each point in time
might be more appropriate:
(t, Z) = 0 (t) exp(1 trnstime + 2 surgery + 3 age)
where trnstime = 1 if transpl=1 and wait> t

305

SAS code for these two models

Time-independent covariate for transpl:

proc phreg data=stanford;
model survival*dead(0)=transpl surgery age;
run;

Time-dependent covariate for transpl:

proc phreg data=stanford;
model survival*dead(0)=trnstime surgery age;
if wait>survival or wait=. then trnstime=0;
else trnstime=1;
run;

306

If we add time-dependent covariates or interactions with time

to the Cox proportional hazards model, then it is not proportional hazards model any longer.
We refer to it as an extended Cox model.
Comparison with a single binary predictor (like heart transplant):
A standard Cox PH model would compare the survival
distributions between those without a transplant (ever)
to those with a transplant. A subjects transplant status
at the end of the study would determine which category
they were put into for the entire study follow-up.
An extended Cox model would compare the risk of an
event between transplant and non-transplant at each
event time, but would re-evaluate which risk group each
person belonged in based on whether theyd had a transplant by that time.

307

Recidivism Example: (see Allison, p.42)

Recidivism study:
432 male inmates were followed for one year after release
from prison, to evaluate risk of re-arrest as function of financial aid (fin), age at release (age), race (race), full-time
work experience prior to first arrest (wexp), marital status (mar), parole status (paro=1 if released with parole,
0 otherwise), and number of prior convictions (prio). Data
were also collected on employment status over time during
the year.
Time-independent model:
A time independent model might include the employment
status of the individual at the beginning of the study (1 if
employed, 0 if unemployed), or perhaps at any point during
the year.
Time-dependent model:
However, employment status changes over time, and it may
be the more recent employment status that would affect the
hazard for re-arrest. For example, we might want to define
a time-dependent covariate for each month of the study that
indicates whether the individual was employed during the
past month.

308

Extended Cox Model

Framework:
For individual i, suppose we have their failure time, failure
indicator, and a summary of their covariate values over time:
(Xi, i, {Zi(t), t [0, Xi]}),
{Zi(t), t [0, Xi]} represents the covariate path for the
i-th individual while they are in the study, and the covariates
can take different values at different times.
Assumptions:
conditional on an individuals covariate history, the hazard for failure at time t depends only on the value of the
covariates at that time:
(t; {Zi(u), u [0, t]}) = (t; Zi(t))
the Cox model for the hazard holds:

(t; Zi(t)) = 0(t) eZi(t)

Survivor function:
S(t; Z) = exp{

Z t

exp(Z(u)) 0(u)du}

and depends on the values of the time dependent variables

over the interval from 0 to t.
This is the classic formulation of the time varying Cox regression survival model.
309

Kinds of time-varying covariates:

internal covariates:

variables that relate to the individuals, and can only be

measured when an individual is alive, e.g. white blood
cell count, CD4 count

external covariates:

variable which changes in a known way, e.g. age, dose

of drug
variable that exists totally independently of all individuals, e.g. air temperature

310

Applications and Examples

The extended Cox model is used:
I. When important covariates change during a study
Framingham Heart study

5209 subjects followed since 1948 to examine relationship between risk factors and cardiovascular disease. A
particular example:
Outcome: time to congestive heart failure
Predictors: age, systolic blood pressure, # cigarettes
per day

Liver Cirrhosis (Andersen and Gill, p.528)

Clinical trial comparing treatment to placebo for cirrhosis. The outcome of interest is time to death. Patients
were seen at the clinic after 3, 6 and 12 months, then
yearly.
Fixed covariates: treatment, gender, age (at diagnosis)
Time-varying covariates: alcohol consumption, nutritional status, bleeding, albumin, bilirubin, alkaline
phosphatase and prothrombin.

Recidivism Study (Allison, p.42)

311

II. For cross-over studies, to indicate change in treatment

Stanford heart study (Cox and Oakes p.129)
Between 1967 and 1980, 249 patients entered a program
at Stanford University where they were registered to receive a heart transplant. Of these, 184 received transplants, 57 died while waiting, and 8 dropped out of the
program for other reasons. Does getting a heart transplant improve survival? Here is a sample of the data:
Waiting
transplant?
survival post
total
final
time
transplant
survival
status
-----------------------------------------------------------49
2
.
.
1
5
2
.
.
1
0
1
15
15
1
35
1
3
38
1
17
2
.
.
1
11
1
46
57
1
etc
(survival is not indicated above for those without transplants, but was available in the dataset)

Naive approach: Compare the total survival of transplanted and non-transplanted.

Problem: Length Bias!

312

III. For Competing Risks Analysis

For example, in cancer clinical trials, tumor response (or
shrinking of the tumor) is used as an outcome. However,
clinicians want to know whether tumor response correlates
with survival.
For this purpose, we can fit an extended Cox model for time
to death, with tumor response as a time dependent covariate.
IV. For testing the PH assumption
For example, we can fit these two models:
(1) Time independent covariate Z1
(t, Z) = 0(t) exp(1 Z1)

The hazard ratio for Z1 is exp(1 ).

(2) Time dependent covariate Z1

(t, Z) = 0(t) exp(1 Z1 + 2 Z1 t)

The hazard ratio for Z1 is exp(1 + 2t).

(note: we may want to replace t by (t t0 ), so that exp(1 )

represents HR at some convenient time, like the median survival
time.)

A test of the parameter 2 is a test of the PH assumption.

(how do we get the test? ...using the Wald test from the
output of second model, or LR test formed by comparing
the log-likelihoods of the two models)
313

Partial likelihood with time-varying covariates

Starting out just as before...
Suppose there are K distinct failure (or death) times, and
let (1 , ....K ) represent the K ordered, distinct death times.
For now, assume there are no tied death times.
Risk Set: Let R(t) = {i : xi t} denote the set of
individuals who are at risk for failure at time t.
Failure: Let ij denote the label or identity of the individual
who fails at time j , including the value of their time-varying
covariate during their time in the study
{Zij (t), t [0, j ]}
History: Let Hj denote the history of the entire data
set, up to the j-th death or failure time, including the time
of the failure, but not the identity of the one who fails, also
including the values of all covariates for everyone up to and
including time j .
Partial Likelihood: We have seen previously that the
partial likelihood can be written as
L() =
=

d
Y

j=1
d
Y

j=1

P (ij |Hj )
P

(j ; Zj (j ))
`R(j ) (j ; Z` (j ))
314

Under the PH assumption, this is:

L() =

exp(Zjj )
P
j=1 `R(j ) exp(Z`j )
d
Y

where Z`j is a short-cut way to denote the value of the covariate vector for the `-th person at the j-th death time,
ie:
Z`j = Z`(j )

What if Z is not measured for person ` at time j ?

use the most recent value (assumes step function)

interpolate

impute based on some model

Inference (i.e. estimating the regression coefficients, constructing score tests, etc.) proceeds similarly to standard
case. The main difference is that the values of Z will change
at each risk set.
Allison notes that it is very easy to write down a Cox model
with time-dependent covariates, but much harder to fit (computationally) and interpret.

315

Old Example revisited:

Group 0: 4+, 7, 8+, 9, 10+
Group 1: 3, 5, 5+, 6, 8+
Let Z1 be group, and add another fixed covariate Z2
ID fail censor Z1 Z2 e(1 Z1 +2 Z2 )
1
3
1
1 1
e1 +2
2
4
0
0 1
e 2
3
5
1
1 1
e1 +2
4
5
0
1 0
e 1
5
6
1
1 1
e1 +2
6
7
1
0 0
1
7
8
0
0 1
e 2
8
8
0
1 0
e 1
9
9
1
0 1
e 2
10 10
0
0 0
1

ordered
failure Individuals
time (j )
at risk
failure ID
3
5
6
7
9
316

Partial
Likelihood
contribution

Example continued:
Now suppose Z2 (a completely different covariate) is a time
varying covariate:

ID fail censor Z1
1
3
1
1
2
4
0
0
3
5
1
1
4
5
0
1
5
6
1
1
6
7
1
0
7
8
0
0
8
8
0
1
9
9
1
0
10 10
0
0

Z2 (t)
3 4 5 6 7 8 9
0
1 1
1 1 1
0 0 0
0 0 0 0
0 0 0 1 1
0 0 0 0 0 0
0 0 0 0 1 1
0 0 0 1 1 1 1
0 1 1 1 1 1 1

ordered
failure Individuals
time (j )
at risk
failure ID
3
5
6
7
9
317

Partial
Likelihood
contribution

SAS solution to previous examples

Title Ph regression: small class example;
data ph;
input time status group z3 z4 z5 z6 z7 z8 z9;
cards;
3
1
1 0 . . . . . .
4
0
0 1 1 . . . . .
5
1
1 1 1 1 . . . .
5
0
1 0 0 0 . . . .
6
1
1 0 0 0 0 . . .
7
1
0 0 0 0 1 1 . .
8
0
0 0 0 0 0 0 0 .
8
0
1 0 0 0 0 1 1 .
9
1
0 0 0 0 1 1 1 1
10 0
0 0 1 1 1 1 1 1
run;

proc phreg ;
model time*status(0)=group z3 ;
run;
proc phreg ;
model time*status(0)=group z ;
z=z3;
if (time >= 4) then z=z4;
if (time >= 5) then z=z5;
if (time >= 6) then z=z6;
if (time >= 7) then z=z7;
if (time >= 8) then z=z8;
if (time >= 9) then z=z9;
run;
318

SAS output from fitting both models

Model with z3:
Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

16.953
.
.

13.699
.
.

Model Chi-Square
3.254 with 2 DF (p=0.1965)
3.669 with 2 DF (p=0.1597)
2.927 with 2 DF (p=0.2315)

Analysis of Maximum Likelihood Estimates

Variable
GROUP
Z3

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1

1.610529
1.360533

1.21521
1.42009

1.75644
0.91788

0.1851
0.3380

5.005
3.898

Model with time-dependent Z:

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

16.953
.
.

14.226
.
.

Model Chi-Square
2.727 with 2 DF (p=0.2558)
2.725 with 2 DF (p=0.2560)
2.271 with 2 DF (p=0.3212)

Analysis of Maximum Likelihood Estimates

Variable
GROUP
Z

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1

1.826757
0.705963

1.22863
1.20630

2.21066
0.34249

0.1371
0.5584

6.214
2.026

319

The Stanford Heart Transplant data

Title Stanford heart transplant data: C & O Table 8.1;
data heart;
infile heart.dat;
input wait trans post surv status ;
run;
data heart;
set heart;
if trans=2 then surv=wait;
run;
*** naive analysis;
proc phreg;
model surv*status(2)=tstat;
tstat=2-trans;
*** analysis with time-dependent covariate;
proc phreg;
model surv*status(2)=tstat;
tstat = 0;
if (trans=1 and surv >= wait) then tstat = 1;
run;

The second model took about twice as long to run as the

first model, which is usually the case for models with timedependent covariates.

320

RESULTS for Stanford Heart Transplant data:

Naive model with fixed transplant indicator:
Criterion
-2 LOG L
Score
Wald

Covariates

Covariates

718.896
.
.

674.699
.
.

Model Chi-Square
44.198 with 1 DF (p=0.0001)
68.194 with 1 DF (p=0.0001)
51.720 with 1 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
TSTAT

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

-1.999356

0.27801

51.72039

0.0001

0.135

Model with time-dependent transplant indicator:

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1330.220
.
.

1312.710
.
.

Model Chi-Square
17.510 with 1 DF (p=0.0001)
17.740 with 1 DF (p=0.0001)
17.151 with 1 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
TSTAT

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

-0.965605

0.23316

17.15084

0.0001

0.381

321

Recidivism Example:
Hazard for arrest within one year of release from prison:
Model without employment status
Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1350.751
.
.

1317.496
.
.

Model Chi-Square
33.266 with 7 DF (p=0.0001)
33.529 with 7 DF (p=0.0001)
32.113 with 7 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
FIN
AGE
RACE
WEXP
MAR
PARO
PRIO

DF

Parameter
Estimate

Standard
Error

1
1
1
1
1
1
1

-0.379422
-0.057438
0.313900
-0.149796
-0.433704
-0.084871
0.091497

0.1914
0.0220
0.3080
0.2122
0.3819
0.1958
0.0287

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

3.931
6.817
1.039
0.498
1.290
0.188
10.200

0.0474
0.0090
0.3081
0.4803
0.2561
0.6646
0.0014

0.684
0.944
1.369
0.861
0.648
0.919
1.096

What are the important predictors of recidivism?

322

Recidivism Example: (contd)

Now, we use the indicators of employment status for each of
the 52 weeks in the study, recorded as emp1-emp52.
We can fit the model in 2 different ways:
proc phreg data=recid;
model week*arrest(0)=fin age race wexp mar parro prio employed
/ ties=efron;
array emp(*) emp1-emp52;
do i=1 to 52;
if week=i then employed=emp(i);
end;
run;
*** a shortcut;
proc phreg data=recid;
model week*arrest(0)=fin age race wexp mar parro prio employed
/ ties=efron;
array emp(*) emp1-emp52;
employed=emp(week);
run;

The second way takes 23% less time than the first
way, but the results are the same.

323

Recidivism Example: Output

Model WITH employment as time-dependent covariate

Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

FIN
AGE
RACE
WEXP
MAR
PARO
PRIO
EMPLOYED

1
1
1
1
1
1
1
1

-0.356722
-0.046342
0.338658
-0.025553
-0.293747
-0.064206
0.085139
-1.328321

0.1911
0.0217
0.3096
0.2114
0.3830
0.1947
0.0290
0.2507

3.484
4.545
1.197
0.015
0.488
0.109
8.644
28.070

0.0620
0.0330
0.2740
0.9038
0.4431
0.7416
0.0033
0.0001

0.700
0.955
1.403
0.975
0.745
0.938
1.089
0.265

Is current employment important?

Do the other covariates change much?
Can you think of any problem with using current
employment as a predictor?

324

Another option for assessing impact of employment

Allison suggests using the employment status of the past
week rather than the current week, as follows:
proc phreg data=recid;
where week>1;
model week*arrest(0)=fin age race wexp mar parro prio employed
/ ties=efron;
array emp(*) emp1-emp52;
employed=emp(week-1);
run;

The coefficient for employed changes from -1.33

to -0.79, so the risk ratio is about 0.45 instead of
0.27. It is still highly significant with 2 = 13.1.
Does this model improve the causal interpretation?
Other options for time-dependent covariates:
multiple lags of employment status (week-1, week-2, etc.)
cumulative employment experience (proportion of weeks
worked)

325

Some cautionary notes

Time-varying covariates must be carefully constructed
to ensure interpretability
There is no point adding a time-varying covariate whose
value changes the same as study time ..... you will get
the same answer as using a fixed covariate measured at
study entry. For example, suppose we want to study the
effect of age on time to death.
We could
1. use age at start of the study as a fixed covariate
2. age as a time varying covariate
However, the results will be the same! Why?

326

Using time-varying covariates to assess model fit

Suppose we have just fit the following model:
(t; Z) = 0(t) exp(1 Z1 + 2Z2 + . . . pZp)
E.g., the nursing home data with gender, marital status and
health.
Suppose we want to test the proportionality assumption on
health (Zp)
Create a new variable:
Zp+1(t) = Zp (t)
where (t) is a known function of time, such as
(t) =
or
or
or

t
log(t)
et
I{t>t}

Then testing H0 : p+1 = 0 is a test for non-proportionality

327

Illustration: Colon Cancer data

*** model without time*covariate interaction;
proc phreg data=surv;
model survtime*censs(1) = trtm stagen ;

Model without time*stage interaction

Event and Censored Values

Total

Event

Censored

Percent
Censored

274

218

56

20.44

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1959.927
.
.

1939.654
.
.

Model Chi-Square
20.273 with 2 DF (p=0.0001)
18.762 with 2 DF (p=0.0001)
18.017 with 2 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
TRTM
STAGEN

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1

0.016675
-0.701408

0.13650
0.16539

0.01492
17.98448

0.9028
0.0001

1.017
0.496

328

*** model WITH time*covariate interaction;

proc phreg data=surv ;
model survtime*censs(1) = trtm stagen tstage ;
tstage=stagen*exp(-survtime/1000);

Model WITH time*stage interaction

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1959.927
.
.

1902.374
.
.

Model Chi-Square
57.553 with 3 DF (p=0.0001)
35.960 with 3 DF (p=0.0001)
19.319 with 3 DF (p=0.0002)

Analysis of Maximum Likelihood Estimates

Variable
TRTM
STAGEN
TSTAGE

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1
1

0.008309
1.402244
-8.322371

0.13654
0.45524
2.04554

0.00370
9.48774
16.55310

0.9515
0.0021
0.0001

1.008
4.064
0.000

Like Cox and Oakes, we can run a few different models

329

Time-varying covariates in Stata

Create a data set with an ID column, and one line per person
for each different value of the time varying covariate.
. infile id time status group z using cox4_stata.dat
or
. input id time status group
z
1
3
1
1
0
2
5
0
1
0
3
5
1
1
1
4
6
1
1
0
5
6
0
1
0
5
8
0
1
1
6
4
0
0
1
7
5
0
0
0
7
7
1
0
1
8
8
0
0
0
9
5
0
0
0
9
9
1
0
1
10
3
0
0
0
10
10
0
0
1
. end

. stset time status

. cox time group z, dead(status) tvid(id)
-----------------------------------------------------------------------------time |
status |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------group |
1.826757
1.228625
1.487
0.137
-.5813045
4.234819
z |
.7059632
1.206304
0.585
0.558
-1.65835
3.070276
-----------------------------------------------------------------------------330

Time-varying covariates in Splus

Create a data set with start and stop values of time:

id start stop status group

1
0
3
1
1
2
0
5
0
1
3
0
5
1
1
4
0
6
1
1
5
0
6
0
1
5
6
8
0
1
6
0
4
0
0
7
0
5
0
0
7
5
7
1
0
8
0
8
0
0
9
0
5
0
0
9
5
9
1
0
10
0
3
0
0
10
3
10
0
0

331

z
0
0
1
0
0
1
1
0
1
0
0
1
0
1

Then the Splus commands and results are:

Commands:
y_read.table("cox4_splus.dat",header=T)
agreg(y$start,y$stop,y$status,cbind(y$group,y$z))

Results:
Alive Dead Deleted
9
5
0
coef exp(coef) se(coef)
z
p
[1,] 1.827
6.21
1.23 1.487 0.137
[2,] 0.706
2.03
1.21 0.585 0.558
exp(coef) exp(-coef) lower .95 upper .95
[1,]
6.21
0.161
0.559
69.0
[2,]
2.03
0.494
0.190
21.5
Likelihood ratio test= 2.73
Efficient score test = 2.73

332

on 2 df,
on 2 df,

p=0.256
p=0.256

Piecewise Cox Model: (Collett, Chapter 10)

A time dependent covariate can be used to create a piecewise
PH cox model. Suppose we are interested in comparing two
treatments, and:
HR=1 during the interval (0, t1 )
HR=2 during the interval (t1, t2)
HR=3 during the interval (t2, )
Define the following covariates:
X - treatment indicator
(X = 0 standard, X = 1 new treatment)
Z2 - indicator of change in HR during 2nd interval

Z2(t) =

1 if t (t1, t2) and X = 1

0 otherwise

Z3 - indicator of change in HR during 3rd interval

Z3(t) =

1 if t (t2 , ) and X = 1
0 otherwise

The model for the hazard for individual i is:

i(t) = 0(t) exp{1 xi + 2z2i(t) + 3z3i(t)}
What are the log hazard ratios for an individual on the new
treatment relative to one on the standard treatment?
333

Time varying (or time-dependent) covariates

Case Study of MAC Disease Trial
ACTG 196 was a randomized clinical trial to study the effects
of combination regimens on prevention of MAC (mycobacterium
avium complex) disease, which is one of the most common opportunistic infections in AIDS patients and is associated with high mortality and morbidity.
The treatment regimens were:
clarithromycin (new)
rifabutin (standard)
clarithromycin plus rifabutin
This trial enrolled patients between April 1993 and February 1994,
and followed patients through August 1995. In February of 1994, the
dosage of rifabutin was reduced from 3 capsules per day (450mg)
to 2 capsules per day (300mg) due to concern over uveitis, an
adverse experience resulting in inflammation of the uveal tract in
the eyes (about 3-4% of patients reported uveitis). All patients were
to reduce their dosage by March 8, 1994. However, some patients
had already discontinued the treatment, died, or discontinued the
study.
The main intent-to-treat analysis compared the 3 treatment arms
without adjusting for this change in dosage.
Other supporting analyses attempted to untangle the effect of this
study wide dose reduction (SWDR).

334

Proportion on each treatment arm with SWDR

Treatment by study wide dose reduction

TABLE OF TRTMT BY SWDRSTAT

TRTMT
SWDRSTAT(Study Wide Dose Reduction Status)
Frequency|
Row Pct |No
|Yes
| Total
---------+--------+--------+
R
|
125 |
266 |
391
| 31.97 | 68.03 |
---------+--------+--------+
C+R
|
170 |
219 |
389
| 43.70 | 56.30 |
---------+--------+--------+
C
|
124 |
274 |
398
| 31.16 | 68.84 |
---------+--------+--------+
Total
419
759
1178

STATISTICS FOR TABLE OF TRTMT BY SWDRSTAT

Statistic
DF
Value
Prob
-----------------------------------------------------Chi-Square
2
16.820
0.001
Likelihood Ratio Chi-Square
2
16.610
0.001
Mantel-Haenszel Chi-Square
1
0.067
0.795
Phi Coefficient
0.119
Contingency Coefficient
0.119
Cramers V
0.119
Sample Size = 1178

335

Original Logrank test Comparing 3 Treatment Arms

(How would you get pairwise tests?)
Dependent Variable: MACTIME
Censoring Variable: MACSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to MAC disease (days)

MAC status (1=yes,0=censored)

Summary of the Number of

Event and Censored Values
Total

Event

Censored

Percent
Censored

1178

121

1057

89.73

Testing Global Null Hypothesis: BETA=0

Criterion

Without
Covariates

With
Covariates

1541.064
.
.

1525.932
.
.

-2 LOG L
Score
Wald

Model Chi-Square
15.133 with 2 DF (p=0.0005)
15.890 with 2 DF (p=0.0004)
15.209 with 2 DF (p=0.0005)

Analysis of Maximum Likelihood Estimates

Variable
CLARI
RIF

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1

0.231842
0.826883

0.25748
0.23601

0.81074
12.27480

0.3679
0.0005

1.261
2.286

Variable

Label

CLARI
RIF

1=Clarithromycin arm, 0 otherwise

1=Rifabutin arm, 0 otherwise
Linear Hypotheses Testing

Label
TEST_TRT

Wald
Chi-Square

DF

Pr >
Chi-Square

15.2094

0.0005

336

Survival Distribution Function

Kaplan-Meier Survival Plot

Estimated Probabilities of Remaining MAC-free

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0

STRATA:

100

200

300 400 500 600 700 800 900

Time to MAC disease (days)
TRTMT=Clar + Rif
TRTMT=Clarithro
TRTMT=Rifabutin

%ps(mactrt.ps,mode=replace);
proc lifetest data=weighted noprint outsurv=survres
graphics nocens plots=(s);
time mactime*macstat(0);
strata trtmt;
title Time to MAC by Treatment Regimen;
format trtmt trtfmt.;
run;

337

How well does this model fit?

Lets take a look at the residual plots...
First, the deviance residuals:
4
D
e
v 3
i
a
n 2
c
e
1
R
e
s 0
i
d
u
a -1
l -1

4
D
e
v 3
i
a
n 2
c
e
1
R
e
s 0
i
d
u
a -1
l -1

0
1
1=Rifabutin arm, 0 otherwise

4
D
e
v 3
i
a
n 2
c
e
1
R
e
s 0
i
d
u
a -1
l 0.0

0.1

0.2

0.3

0.4 0.5 0.6

Linear Predictor

0
1
1=Clarithromycin arm, 0 otherwise

0.7

0.8

0.9

Plotting deviance residuals vs binary covariates

is not very useful.
338

How about the generalized residuals?

(Are they like a sample from a censored unit exponential?)

(i.e., is slope=1, intercept=0)

LLS
-1
-2
-3
-4
-5
-6
-7
-8
-8

-7

-6

-5
-4
-3
Log(generalized residual)

-2

intercept=0.056
slope=1.028
(based on fitting a regression line to residuals)

339

-1

We can also look at the log cumulative hazard

versus log time to see whether
plots (i.e., log[ log(S)])
the lines are parallel for the three treatment groups.

Plot of log-log KM versus log-time

-1
L -2
n
[
- -3
l
n
( -4
S
) -5
]
-6
3

MAC Prophylaxis Therapy

Log(Time to MAC)
Rifabutin
Clarithro

Clar + Rif

(I have joined the individual points using i=join in the symbol statement, to make them easier to see.)

340

Shouldnt we adjust for Baseline CD4 count?

Testing Global Null Hypothesis: BETA=0

Criterion

Without
Covariates

With
Covariates

1541.064
.
.

1488.737
.
.

-2 LOG L
Score
Wald

Model Chi-Square
52.328 with 3 DF (p=0.0001)
43.477 with 3 DF (p=0.0001)
43.680 with 3 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
CLARI
RIF
CD4

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1
1

0.198798
0.837240
-0.019641

0.25747
0.23598
0.00367

0.59619
12.58738
28.59491

0.4400
0.0004
0.0001

1.220
2.310
0.981

Analysis of Maximum Likelihood Estimates

Variable

Label

CLARI
RIF
CD4

1=Clarithromycin arm, 0 otherwise

1=Rifabutin arm, 0 otherwise
CD4 Cell Count

Is CD4 count a confounder?

(An analysis stratified by CD4 category gave almost identical results. Other important covariates included CTG (clinical trials
group) and Karnofsky status).

341

What do the deviance residuals look like versus

a continuous covariate, like CD4?

4
D
e 3
v
i
a 2
n
c 1
e
R 0
e
s
i -1
d
u
a -2
l 0

100

200

300

CD4 Cell Count

We might want to consider some kind of transformation of CD4

count (like log or square root). If we dont feel comfortable with
the linearity of CD4 count, we can also dichotomize it (CD4CAT).

342

Another way of checking the proportionality assumption is by using the Weighted Schoenfeld
residual plots for each covariate
Raw CD4 count

log CD4 count

0.25

0.20
1

0.15
0.10

0
0.05
0.00

-1

-0.05
-0.10

-2
0

100

200

300 400 500 600

Time to MAC disease (days)

700

800

900

100

200

300 400 500 600

Time to MAC disease (days)

Square root CD4 count

1.2
1.0
0.8
0.6
0.4
0.2
0.0
-0.2
-0.4
-0.6
-0.8
0

100

200

300 400 500 600

Time to MAC disease (days)

343

700

800

900

700

800

900

So far, the graphical techniques have not indicated any major departure from proportional hazards. However, we can
test this formally by creating a time dependent covariate for
rifabutin and clarithromycin:
riftd=rif*((mactime-365)/30);
claritd=clari*((mactime-365)/30);

Even though the dose reduction was only for rifabutin, patients on all 3 arms had to have the dose reduction ... they
just took 2 capsules of their placebo, and didnt know whether
it was placebo or active drug.
I have centered the time-dependent covariates at 365 days
(one year), so that the HR for rif alone and clari alone will
apply at one year. Then I have divided by 30, so that the
resulting HR can be interpreted as the change for each month
away from 365 days.
Question: Can we do this within a data step using the above statements, or do these statements
need to be given in the PROC PHREG procedure?

344

Time-dependent covariates for clari and rif

Testing Global Null Hypothesis: BETA=0

Criterion

Without
Covariates

With
Covariates

1541.064
.
.

1525.837
.
.

-2 LOG L
Score
Wald

Model Chi-Square
15.227 with 4 DF (p=0.0043)
16.033 with 4 DF (p=0.0030)
15.327 with 4 DF (p=0.0041)

Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

CLARI
RIF
CLARITD
RIFTD

1
1
1
1

0.229811
0.823227
0.003065
0.010627

0.25809
0.23624
0.04073
0.03765

0.79287
12.14274
0.00566
0.07965

0.3732
0.0005
0.9400
0.7778

1.258
2.278
1.003
1.011

Analysis of Maximum Likelihood Estimates

Variable

Label

CLARI
RIF

1=Clarithromycin arm, 0 otherwise

1=Rifabutin arm, 0 otherwise

Neither time-dependent covariate was significant.

345

This analysis also indicated that there are no major departures from proportional hazards for the three treatment
arms.
However, it may still be the case that having the study-wide
dose reduction had some relationship with MAC disease.
We can assess this by creating a time dependent variable for
the SWDR.

Well look at the following models:

(1) SWDRSTAT as a simple indicator
(2) SWDRSTAT and SWDRTD, with
swdrtd = swdrstat*((mactime-365)/30)

(3) SWDR as time dependent covariate

346

Naive model with fixed SWDR indicator (SWDRSTAT):

Dependent Variable: MACTIME
Censoring Variable: MACSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to MAC disease (days)

MAC status (1=yes,0=censored)

Testing Global Null Hypothesis: BETA=0

Criterion

Without
Covariates

With
Covariates

1541.064
.
.

1495.857
.
.

-2 LOG L
Score
Wald

Model Chi-Square
45.208 with 3 DF (p=0.0001)
51.497 with 3 DF (p=0.0001)
48.749 with 3 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

CLARI
RIF
SWDRSTAT

1
1
1

0.449936
1.006639
-1.125032

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

0.26142
0.23852
0.19283

2.96236
17.81114
34.04055

0.0852
0.0001
0.0001

1.568
2.736
0.325

Analysis of Maximum Likelihood Estimates

Variable

Label

CLARI
RIF
SWDRSTAT

1=Clarithromycin arm, 0 otherwise

1=Rifabutin arm, 0 otherwise
Study Wide Dose Reduction Status

Reduction of dosage from 450mg to 300mg appears to be

protective, which seems counter-intuitive
347

Predicted Baseline Survival Curves:

Another way to see this is through the predicted baseline
survival curves. The two lines are for those not on rifabutin,
while the xs and +s are for those on rifabutin. In each case,
the higher line (better prognosis) of the pair is for those who
did have the SWDR.

1.0
P
r 0.8
(
s
u 0.6
r
v
i 0.4
v
a 0.2
l
)
0.0
0

100

200

300 400 500

Time to MAC (days)
SWDR/Rifabutin Status
No SWDR/no RIF
No SWDR/RIF

348

600

700

SWDR/no RIF
SWDR/RIF

800

Test for proportionality:

proc phreg data=weighted;
model mactime*macstat(0) = clari rif swdrstat swdrtd;
*** create time by covariate interaction for swdr status;
swdrtd=swdrstat*((mactime-365)/30);
test_trt: test clari, rif;
title Test of treatment Differences;
title2 and test of proportionality at t=365 days;

Testing Global Null Hypothesis: BETA=0

Criterion

Without
Covariates

With
Covariates

1541.064
.
.

1492.692
.
.

-2 LOG L
Score
Wald

Model Chi-Square
48.372 with 4 DF (p=0.0001)
55.174 with 4 DF (p=0.0001)
50.719 with 4 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

CLARI
RIF
SWDRSTAT
SWDRTD

1
1
1
1

0.430051
1.005416
-1.126498
0.055550

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

0.26126
0.23845
0.19752
0.03201

2.70947
17.77884
32.52551
3.01112

0.0998
0.0001
0.0001
0.0827

1.537
2.733
0.324
1.057

Variable

Label

CLARI
RIF
SWDRSTAT
SWDRTD

1=Clarithromycin arm, 0 otherwise

1=Rifabutin arm, 0 otherwise
Study Wide Dose Reduction Status
swdrstat*((mactime-365)/30)

349

Interpretation of Hazard Ratios

swdrstat = 1.1265
swdrtd = 0.0556

Time
Time
Hazard
(months) (days) calculation
Ratio
6
182.5 exp[1.1265 + (6.08)(0.0556)] 0.231
12

365

exp[1.1265 + (0)(0.0556)]

0.324

18

547.5

exp[1.1265 + (6.08)(0.0556)]

0.454

24

730

exp[1.1265 + (12.17)(0.0556)]

0.637

30

912.5

exp[1.1265 + (18.25)(0.0556)]

0.893

36

1095

exp[1.1265 + (24.33)(0.0556)]

1.253

365)
HR = exp[swdrstat + swdrtd mactime
]
30

In the early period after randomization to treatment, reduction of randomized dosage from 450mg to 300mg is associated with a decreased risk of MAC disease. After taking the
higher dosage for about 32 months, dropping to the lower
dosage has no impact, and as the treatment time increases
beyond 32 months, a lower dosage tends to be associated
with increased risk of MAC.
350

3 different ways to code SWDR as time-dependent

covariate
proc phreg data=weighted;
model mactime*macstat(0) = clari rif swdr;
if (swdrtime>=mactime) then swdr=0;
else do;
if swdrstat=1 then swdr=1;
else swdr=0;
end;
test_trt: test clari, rif;
title2 I. Time-dependent indicator of dose reduction;

proc phreg data=weighted;

model mactime*macstat(0) = clari rif swdr;
if swdrstat=0 or (swdrtime>=mactime) then swdr=0;
else swdr=1;
test_trt: test clari, rif;
title2 II. Time-dependent indicator of dose reduction;

proc phreg data=weighted;

model mactime*macstat(0) = clari rif swdr;
if swdrstat=1 and (swdrtime<mactime) then swdr=1;
else swdr=0;
test_trt: test clari, rif;
title2 III. Time-dependent indicator of dose reduction;

351

Output is the same for all 3 cases:

Summary of the Number of
Event and Censored Values

Total

Event

Censored

Percent
Censored

1178

121

1057

89.73

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1541.064
.
.

1517.426
.
.

Model Chi-Square
23.639 with 3 DF (p=0.0001)
24.844 with 3 DF (p=0.0001)
24.142 with 3 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
CLARI
RIF
SWDR

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1
1

0.328849
0.905299
-0.648887

0.26017
0.23775
0.21518

1.59762
14.49956
9.09389

0.2062
0.0001
0.0026

1.389
2.473
0.523

SWDR is still protective? Does this make sense intuitively?

What other methods can we use to account for change in
dosage?

352

Weighted adjusted dose (WAD) analyses

To try to get a better idea of the effect of changing doses
of rifabutin on the hazard for MAC disease, I created the
following weighted dose of randomized rifabutin:
Between randomization date and SWDR date
= # Days at 450mg
Between SWDR date and off-study date
= # Days at 300mg
Between randomization date and Off-study date
= # Total Days
Weighted randomized dose
rifwadr = (days450 + days300)/totdays
Transformed to number of capsules per day;
rifwadr=rifwadr/150;
Also calculated weighted dose while on treatment by
starting with on treatment date, stopping with off-treatment
date, and dividing by the total days on study.

353

Weighted adjusted dose (WAD) analyses

Randomized assignment to rifabutin
Dependent Variable: MACTIME
Censoring Variable: MACSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to MAC disease (days)

MAC status (1=yes,0=censored)

Summary of the Number of

Event and Censored Values

Total

Event

Censored

Percent
Censored

1178

121

1057

89.73

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1541.064
.
.

1493.476
.
.

Model Chi-Square
47.588 with 3 DF (p=0.0001)
52.770 with 3 DF (p=0.0001)
50.295 with 3 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable

DF

Parameter
Estimate

CLARI
RIF
RIFWADR

1
1
1

0.453283
1.004846
1.530462

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

0.26119
0.23826
0.25681

3.01179
17.78681
35.51502

0.0827
0.0001
0.0001

1.573
2.731
4.620

For each additional capsule of rifabutin specified as randomized treatment, the HR for MAC increased by 4.6
times
354

Weighted adjusted dose (WAD) analyses

Actual dosage of rifabutin during the study
Dependent Variable: MACTIME
Censoring Variable: MACSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to MAC disease (days)

MAC status (1=yes,0=censored)

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1541.064
.
.

1489.993
.
.

Model Chi-Square
51.071 with 3 DF (p=0.0001)
55.942 with 3 DF (p=0.0001)
53.477 with 3 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
CLARI
RIF
RIFWAD

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1
1

0.489583
1.019675
-0.664689

0.26256
0.23873
0.10686

3.47693
18.24291
38.69332

0.0622
0.0001
0.0001

1.632
2.772
0.514

Here, higher values of RIFWAD probably reflect that the

patient was able to stay on treatment longer, which was
protective. The SWDR variable is also capturing whether a
patient had been able to tolerate the treatment long enough
to have the chance to have the protocol-mandated dose
reduction.

355

What happens if we add treatment discontinuation as a time dependent covariate?

Dependent Variable: MACTIME
Censoring Variable: MACSTAT
Censoring Value(s): 0
Ties Handling: BRESLOW

Time to MAC disease (days)

MAC status (1=yes,0=censored)

Summary of the Number of

Event and Censored Values

Total

Event

Censored

Percent
Censored

1178

121

1057

89.73

Testing Global Null Hypothesis: BETA=0

Criterion
-2 LOG L
Score
Wald

Without
Covariates

With
Covariates

1541.064
.
.

1501.595
.
.

Model Chi-Square
39.469 with 4 DF (p=0.0001)
42.817 with 4 DF (p=0.0001)
41.027 with 4 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Variable
CLARI
RIF
SWDR
RXSTOP

DF

Parameter
Estimate

Standard
Error

Wald
Chi-Square

Pr >
Chi-Square

Risk
Ratio

1
1
1
1

0.420447
0.984114
-0.139245
0.902592

0.26111
0.23847
0.23909
0.21792

2.59284
17.02975
0.33919
17.15473

0.1073
0.0001
0.5603
0.0001

1.523
2.675
0.870
2.466

SWDR is no longer significant!

356

Last of all, a comparison of some of these models:

Model terms
Clari, Rif

AIC
q 2 log L Criterion
2 1525.93 1531.93

Clari, Rif, Cd4cat

3 1497.57

1506.57

Clari, Rif, Cd4

3 1488.74

1497.74

Clari, Rif, Cd4cat, Ctg, Karnof

5 1482.67

1497.67

Clari, Rif, Swdrstat

3 1495.86

1504.86

Clari, Rif, Rifwadr

3 1493.48

1502.48

Clari, Rif, Swdrstat, Rifwadr

4 1493.44

1505.44

Clari, Rif, Rifwad

3 1489.99

1498.99

Models with time-dependent covariates

Clari, Rif, Claritd, Riftd
4 1525.84

1537.84

Clari, Rif, Swdrstat, Swdrtd

4 1492.69

1504.69

Clari, Rif, Swdr

3 1517.43

1526.43

Clari, Rif, Swdr, Rxstop

4 1501.60

1513.60

Clari, Rif, Cd4cat, Karnof, Rxstop 5 1461.90

1476.90

Clari, Rif, Cd4cat, Karnof, Rifwad 5 1448.14

1463.14

357

Parametric Survival Analysis

So far, we have focused primarily on nonparametric and
semi-parametric approaches to survival analysis, with heavy
emphasis on the Cox proportional hazards model:
(t, Z) = 0(t) exp(Z)

We used the following estimating approach:

We estimated 0(t) nonparametrically, using the KaplanMeier estimator, or using the Kalbfleisch/Prentice estimator under the PH assumption
We estimated by assuming a linear model between the
log HR and covariates, under the PH model
Both estimates were based on maximum likelihood theory.

358

There are several reasons why we should consider some alternative approaches based on parametric models:

The assumption of proportional hazards might not be

appropriate (based on major departures)
If a parametric model actually holds, then we would
probably gain efficiency
We may want to handle non-standard situations like
interval censoring
incorporating population mortality
We may want to make some connections with other familiar approaches (e.g. use of the Poisson likelihood)
We may want to obtain some estimates for use in designing a future survival study.

359

A simple start: Exponential Regression

Observed data: (Xi, i, Zi) for individual i,
Zi = (Zi1, Zi2, ..., Zip) represents a set of p covariates.
Right censoring: Assume that Xi = min(Ti, Ui)
Survival distribution: Assume Ti follows an exponential distribution with a parameter that depends on
Zi, say i = (Zi). Then we can write:
Ti exponential((Zi))
First, lets review some facts about the exponential distribution (from our first survival lecture):
f (t) = et

for t 0

S(t) = P (T t) =

f (u)du = et

F (t) = P (T < t) = 1 et
(t) =

f (t)
=
S(t)

constant hazard!

(t) =

Z t

Z t

(u) du =
360

du = t

Now, we say that is a constant over time t, but we want

to let it depend on the covariate values, so we are setting
i = (Zi)
The hazard rate would therefore be the same for any two
individuals with the same covariate values.
Although there are many possible choices for , one simple
and natural choice is:
(Zi) = exp[0 + Zi11 + Zi22 + ... + Zipp]

WHY?
ensures a positive hazard

for an individual with Z = 0, the hazard is e0 .

The model is called exponential regression because of
the natural generalization from regular linear regression

361

Exponential regression for the 2-sample case:

Assume we have only a single covariate Z = Z,

i.e., p = 1.
Hazard Rate:
(Zi) = exp(0 + Zi1)

Define:

Zi = 0 if individual i is in group 0
Zi = 1 if individual i is in group 1

What is the hazard for group 0?

What is the hazard for group 1?
What is the hazard ratio of group 1 to group
0?
What is the interpretation of 1?

362

Likelihood for Exponential Model

Under the assumption of right censored data, each person
has one of two possible contributions to the likelihood:
(a) they have an event at Xi (i = 1) contribution is
Li =

S(X
i)
{z

survive to Xi

(X
i)
{z

= eXi

fail at Xi

(b) they are censored at Xi (i = 0) contribution is

Li =

S(X
i)
{z

= eXi

survive to Xi

The likelihood is the product over all of the individuals:

L =

Li

Xi i
{z

events
i

Xi

363

Xi (1i )
{z

censorings

Maximum Likelihood for Exponential

How do we use the likelihood?
first take the log
then take the partial derivative with respect to

c
then set to zero and solve for

this gives us the maximum likelihood estimators

The log-likelihood is:

log L = log

[i log() Xi]

[i log()]

eXi

Xi

For the case of exponential regression, we now substitute the

hazard = (Zi) in the above log-likelihood:
log L =

[i log((Zi))]

364

(Zi)Xi

(1)

General Form of Log-likelihood

for Right Censored Data
In general, whenever we have right censored data, the likelihood and corresponding log likelihood will have the following
forms:

L =

log L =

[i(Xi)]i Si(Xi)
[i log (i(Xi))]

i(Xi)

where
i(Xi) is the hazard for the individual i who fails at Xi
i(Xi) is the cumulative hazard for an individual at their
failure or censoring time

For example, see the derivation of the likelihood for a Cox

model on p.11-13 of Lecture 4 notes. We started with the
likelihood above, then substituted the specific forms for (X i)
under the PH assumption.

365

Consider our model for the hazard rate:

= (Zi) = exp[0 + Zi11 + Zi22 + ... + Zipp]
We can write this using vector notation, as follows:
Let Zi = (1, Zi1, ...Zip)T
and = (0 , 1, ...p)
(Since 0 is the intercept (i.e., the log hazard rate for the
baseline group), we put a 1 as the first term in the vector
Zi.)
Then, we can write the hazard as:
(Zi) = exp[Zi]
Now we can substitute (Zi) = exp[Zi] in the log-likelihood
shown in (1):
log L =

n
X

i=1

i(Zi)

366

n
X

i=1

Xi exp(Zi)

Score Equations
Taking the derivative with respect to 0, the score equation
is:
n
log L
X
[i Xi exp(Zi)]
=
0
i=1
For k , k = 1, ...p, the equations are:
log L
=
k
=

n
X

[iZik XiZik exp(Zi)]

i=1

Zik [i Xi exp(Zi)]

i=1
n
X

To find the MLEs, we set the above equations to 0 and

solve (simultaneously). The equations above imply that
the MLEs are obtained by setting the weighted number of
P
failures ( i Zik i) equal to the weighted cumulative hazard
P
( i Zik (Xi)).

367

To find the variance of the MLEs, we need to take the second

derivatives:
n
2 log L X

Zik Zij Xi exp(Zi)

=
k j i=1

Some algebra (see Cox and Oakes section 6.2) reveals that
c

V ar() = I()

= Z(I )Z

T 1

where
Z = (Z1, . . . , Zn) is a (p + 1) n matrix
(p covariates plus the 1 for the intercept 0)
= diag(1, . . . , n) (this means that is a diagonal
matrix, with the terms 1, . . . , n on the diagonal)
i is the probability that the i-th person is censored, so
(1 i) is the probability that they failed.
Note: The information I() (inverse of the variance)
is proportional to the number of failures, not the sample
size. This will be important when we talk about study
design.

368

The Single Sample Problem (Zi = 1 for everyone):

First, what is the MLE of 0?

We set

log L
0

Pn

i=1 [i Xi exp(0 Zi )]

n
X

i=1

i =

n
X

i=1

[Xi exp(0)]

d = exp(0)
d
exp(
0) =

equal to 0 and solve:

n
X

i=1

Xi

d
Pn

i=1 Xi

= d

t
where d is the total number of deaths (or events), and t =
Xi is the total person-time contributed by all individuals.

If d/t is the MLE for , what does this imply

about the MLE of 0?

369

= Z(I )ZT
Using the previous formula V ar()
what is the variance of c0?:
With some matrix algebra, you can show that it is:

V ar(c0) =

1
1
=
Pn
d
i=1 (1 i )

= e0 ?
What about
By the delta method,
=
2 V ar(c0)
V ar()
=

370

The Two-Sample Problem:

Group 0:
Group 1:

Zi
Subjects Events Follow-up
P 0
Xi
Zi = 0
n0
d0
t0 = ni=1
Zi = 1

n1

d1

t1 =

P n1

i=1 Xi

The log-likelihood:

so

n
X

log L =

n
X

i=1

i(0 + 1Zi)

log L
=
0

n
X

[i Xi exp(0 + 1Zi)]

i=1

i=1

Xi exp(0 + 1Zi)

= (d0 + d1) (t0e0 + t1e0+1 )

log L
=
1

n
X

i=1

Zi[i Xi exp(0 + 1Zi)]

= d1 t1e0+1
This implies:

1 = e0+1 =?

0 = e0 =?

0 = ?
1 = ?

371

Important Result:

The maximum likelihood estimates

(MLEs) of the hazard rates under
the exponential model are the number of events divided by the personyears of follow-up!

(this result will be relied on heavily when we discuss study design)

372

Exponential Regression:
Means and Medians
Mean Survival Time
For the exponential distribution, E(T ) = 1/.
Control Group:

0 = 1/ exp(0)
T 0 = 1/

Treatment Group:

1 = 1/ exp(0 + 1)
T 1 = 1/

Median Survival Time

This is the value M at which S(t) = et = 0.5, so M =
median = log(0.5)

Control Group:

0 = log(0.5) = log(0.5)
M
0

exp(0)

Treatment Group:

1 = log(0.5) =
M
1

373

log(0.5)
exp(0 + 1)

Exponential Regression:
Variance Estimates and Test Statistics
We can also calculate the variances of the MLEs as simple
functions of the number of failures:
1
var(0) =
d0
1
1
+
d0 d1

var(1) =

So our test statistics are formed as:

For testing Ho : 0 = 0:
2w =

var(0 )

[log(d0 /t0 )]2

=
1/d0
For testing Ho : 1 = 0:
2w =

var(1 )

1
log( dd10 /t
/t0 )
1
1
d0 + d1

How would we form confidence intervals for the hazard

ratio?
374

The Likelihood Ratio Test Statistic:

(An alternative to the Wald test)
A likelihood ratio test is based on 2 times the log of the ratio
of the likelihoods under the null and alternative. We reject
H0 if 2 log(LR) > 21,0.05 , where
L(H1) L(b 0, b 1)
LR =
=
b
L(H0)
L()
For a sample of n independent exponential random variables
with parameter , the Likelihood is:
L =

n
Y

i=1

[i exp(xi)]

= d exp(

xi )

= d exp(n
x)
where d is the number of deaths or failures.
The log-likelihood is
` = d log() n
x
and the MLE is
b = d/(n
x)
375

2-Sample Case: LR test calculations

Data:
Group 0:
Group 1:

d0 failures among the n0 females

P
mean failure time is x 0 = ( ni 0 Xi)/n0

d1 failures among the n1 males

P
mean failure time is x 1 = ( ni 1 Xi)/n1

Under the alternative hypothesis:

L = d11 exp(1n1x1) d00 exp(0n0x0)
log(L) = d1 log(1) 1n1x1 + d0 log(0) 0n0x0
The MLEs are:
b 1 = d1/(n1x1)

b 0 = d0/(n0x0)

for males
for females

Under the null hypothesis:

L = d1+d0 exp[(n1x1 + n0x0)]
log(L) = (d1 + d0) log() [n1x1 + n0x0]
The corresponding MLE is
b = (d1 + d0)/[n1x1 + n0x0]
376

A likelihood ratio test can be constructed by taking twice the

difference of the log-likelihoods under the alternative and the
null hypotheses:

d0 + d 1
2 (d0 + d1) log
d1 log[d1/t1] d0 log[d0/t0]
t0 + t 1
Nursing home example:
For the females:
n0 = 1173
d0 = 902
t0 = 310754
x0 = 265
For the males:
n1 = 418
d1 = 367
t1 = 75457
x1 = 181
Plugging these values in, we get a LR test statistic of 64.20.

377

Hand Calculations using events and follow-up:

By adding up los for males to get t1 and for females to
get t0, I obtained:
d0 = 902 (females)
d1 = 367 (males)
t0 = 310754 (female follow-up)
t1 = 75457 (male follow-up)
This yields an estimated log HR:

d1/t1
367/75457
1 = log
= log
= log(1.6756) = 0.5162
d0/t0
902/310754
The estimated standard error is:
r

var(1 ) =

v
u
u
t

1
1
+
=
d1 d0

v
u
u
t

1
1
+
= 0.06192
902 367

So the Wald test becomes:

2W

12
(0.51619)2
= 69.51
=
=
0.061915
var(1 )

We can also calculate 0 = log(d0 /t0 ) =q5.842,

along with its standard error se(0 ) = (1/d0) = 0.0333
378

Exponential Regression in STATA

. use nurshome
. stset los fail
. streg gender, dist(exp) nohr
failure _d:
analysis time _t:

Iteration
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:
4:

log
log
log
log
log

fail
los

likelihood
likelihood
likelihood
likelihood
likelihood

=
=
=
=
=

-3352.5765
-3321.966
-3320.4792
-3320.4766
-3320.4766

Exponential regression -- log relative-hazard form

No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

1591
1269
386211
-3320.4766

Number of obs

1591

LR chi2(1)
Prob > chi2

=
=

64.20
0.0000

------------------------------------------------------------------------_t |
Coef. Std. Err.
z
P>|z|
[95% Conf. Interval]
---------|-------------------------------------------------------------gender | .516186 .0619148
8.337
0.000
.3948352
.6375368
_cons |-5.842142 .0332964 -175.459
0.000
-5.907402
-5.776883
-------------------------------------------------------------------------

Since Z = 8.337, the chi-square test is Z 2 = 69.51.

379

Exponential Regression in SAS - proc lifereg

proc format;
value censfmt

value grpfmt

1=Censored
0=Dead;
0=Group 0 (F)
1=Group 1 (M);

Title Exponential Hazard Model for Nursing Home Patients;

data morris;
infile ch12.dat;
input los age trt gender marstat hltstat cens;
data morris2;
set morris;
if los=0 then delete;
proc freq data=morris2;
table cens*gender/ norow nocol nopercent;
format cens censfmt. gender grpfmt.;
proc lifereg data=pop covout outest=survres;
model los*censor(1)=gender /dist=exponential;
run;
RESULTS:
TABLE OF CENS BY GENDER
CENS

GENDER

Frequency|Group 0 |Group 1 |
|(F)
|(M)
|
---------+--------+--------+
Event
|
902 |
367 |
---------+--------+--------+
Censored |
271 |
51 |
---------+--------+--------+
Total
1173
418

380

Total

1269
322
1591

PROC LIFEREG RESULTS:

Exponential Hazard Model for Nursing Home Patients
Lifereg

Procedure

Data Set
=WORK.MORRIS2
Dependent Variable=Log(LOS)
Censoring Variable=CENS
Censoring Value(s)=
1
Noncensored Values= 1269 Right Censored Values=
322
Left Censored Values=
0 Interval Censored Values=
0

Log Likelihood for EXPONENT -3320.476626

Lifereg
Variable
INTERCPT
GENDER
SCALE

DF

Estimate

Procedure

Std Err ChiSquare

1 5.84213388 0.033296
1 -0.5161878 0.061915
0
1
0

30786
69.50734

Pr>Chi Label/Value
0.0001 Intercept
0.0001
Extreme value scale

Note that the estimates for 0 and 1 above are

the opposites of what we calculated. Ill explain
why the output has this form when we get to
AFT models.

381

The Weibull Regression Model

At the beginning of the course, we saw that the survivorship
function for a Weibull random variable is:
S(t) = exp[(t)]
and the hazard function is:
(t) = t(1)

The Weibull regression model assumes that for someone with

covariates Zi, the survivorship function is
S(t; Zi) = exp[(Zi)(t)]
where (Zi) is defined as in exponential regression to be:
(Zi) = exp[0 + Zi11 + Zi22 + ...Zipp]

For the 2-sample problem, we have:

(Zi) = exp[0 + Zi11]

382

Weibull MLEs for the 2-sample problem:

Log-likelihood:
log L =

n
X

i=1

i log exp(0 + 1 Zi )Xi1

n
X

i=1

Xi exp(0 + 1 Zi )

exp(0) = d0/t0

exp(0 + 1) = d1/t1
where tj =

nj
X

i=1

Xi among nj subjects

0(t) =

exp(0) t1
1(t) =

exp(0 + 1) t1
d
1(t)/
0(t) = exp(1)
HR
=

= exp

383

d1/t1

d0/t0

Weibull Regression:
Means and Medians
Mean Survival Time
For the Weibull distribution, E(T ) = (1/) [(1/) + 1].
Control Group:

0(1/) [(1/
) + 1]
T0 =

Treatment Group:

1(1/) [(1/
) + 1]
T1 =

Median Survival Time

For the Weibull distribution, M = median =
Control Group:
0 =
M
Treatment Group:

1/

log(0.5)

1/

1 = log(0.5)
M
1

0 = exp(0 ) and
1 = exp(0 + 1).
where
384

log(0.5) 1/

Note: the symbol is the gamma function. If x is an

integer, then
(x) = (x 1)!
In cases where x is not an integer, this function has to be
evaluated numerically.

The Weibull regression model is very easy to fit:

In sas: use model option dist=weibull within the
proc lifereg procedure
In stata: Just specify dist(weibull) instead
of dist(exp) within the streg command

Note: to get more information on these modeling procedures,

use the online help facilities. For example, in Stata, you
can type:
.help streg

385

Weibull in Stata:
. streg gender, dist(weibull) nohr
failure _d:
analysis time _t:

fail
los

Fitting constant-only model:

Iteration
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:
4:

log
log
log
log
log

likelihood
likelihood
likelihood
likelihood
likelihood

= -3352.5765
= -3074.978
= -3066.1526
= -3066.143
= -3066.143

likelihood
likelihood
likelihood
likelihood
likelihood

=
=
=
=
=

Fitting full model:

Iteration
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:
4:

log
log
log
log
log

-3066.143
-3045.8152
-3045.2772
-3045.2768
-3045.2768

Weibull regression -- log relative-hazard form

No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

1591
1269
386211
-3045.2768

Number of obs

1591

LR chi2(1)
Prob > chi2

=
=

41.73
0.0000

-----------------------------------------------------------------------_t |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------gender |
.4138082
.0621021
6.663
0.000 .2920903
.5355261
_cons | -3.536982
.0891809 -39.661
0.000 -3.711773
-3.362191
---------+-------------------------------------------------------------/ln_p | -.4870456
.0232089 -20.985
0.00 -.5325343
-.4415569
-----------------------------------------------------------------------p |
.614439
.0142605
.5871152
.6430345
1/p |
1.627501
.0377726
1.555127
1.703243
-----------------------------------------------------------------------386

Weibull in SAS
proc lifereg data=morris2 covout outest=survres;
model los*censor(1)=gender / dist=weibull;
run;

Data Set
=WORK.MORRIS2
Dependent Variable=Log(LOS)
Censoring Variable=CENS
Censoring Value(s)=
1
Noncensored Values= 1269
Left Censored Values=
0

Right Censored Values=

322
Interval Censored Values=
0

Log Likelihood for WEIBULL -3045.276811

Lifereg
Variable
INTERCPT
GENDER
SCALE

DF

Estimate

Procedure

Std Err ChiSquare

1 5.75644118
0.0542
1 -0.6734732 0.101067
1 1.62750085 0.037773

11280.04
44.40415

387

Pr>Chi Label/Value
0.0001 Intercept
0.0001
Extreme value scale

In SAS, both the exponential and Weibull are special cases

of the general class of accelerated life models and the
parameter interpretations follow from this approach.
To translate the output of SAS (or Stata using the ereg
command) for Weibull regression, we have to take the negative of the numbers in the output, divided by the scale
parameter (, or 1/).
0 = intercpt/ scale

1 = covariate/ scale
Then we calculate the estimated HR as exp(1 ).
The MLEs are:
0 = intercpt/ scale = 5.756/1.627 = 3.537

1 = covariate/ scale = 0.6735/1.625 = 0.414

d
and the estimated HR is HR
= exp(1) = exp(0.414) =
1.513.

388

Weibull Regression:
Variance Estimates and Test Statistics
It is not so easy to get variance estimates from the output
of proc lifereg in SAS or weibull in stata, at least for
the parameters were interested in.
The variances depend on inverting a (3 3) matrix corresponding to the parameters 0, 1, and . The MLE for

has to be obtained numerically (i.e., no closed form), so the

standard errors also have to be obtained by computer.
Main objective: to obtain s.e.(1 ), so that we can form
tests and confidence intervals for the hazard ratio.
The output gives us 1 and s.e.(1), where 1 = 1/
. If
was a constant, then we could just compute
1
var(1)
2

but is also a random variable! Instead, you need to use

an approximation for the variance of a ratio of two random
variables:

1 2

var(1 ) = 4 var(1 ) + (1 ) var(

) 21 cov(1 , )

) by squaring the standard

where you get var() and var(
var(1 ) =

errors of the covariate term and scale term, respectively, from the proc lifereg or weibull output.
389

Comparison of Exponential with Kaplan-Meier

We can see how well the Exponential model fits by comparing the survival estimates for males and females under the

exponential model, i.e., P (T t) = e(z t), to the KaplanMeier survival estimates:

1.0
0.9
0.8
S 0.7
u
r 0.6
v 0.5
i
v 0.4
a
l 0.3
0.2
0.1
0.0
0

100 200 300 400 500 600 700 800 900 1000 1100
Length of Stay (days)

390

Comparison of Weibull with Kaplan-Meier

We can see how well the Weibull model fits by comparing

the survival estimates, P (T t) = e(z t ), to the KaplanMeier survival estimates.

1.0
0.9
0.8
S 0.7
u
r 0.6
v 0.5
i
v 0.4
a
l 0.3
0.2
0.1
0.0
0

100 200 300 400 500 600 700 800 900 1000 1100
Length of Stay (days)

Which do you think fits best?

391

Other useful plots for evaluating fit to exponential and Weibull models

log(S(t))
vs t

log[ log(S(t))]
vs log(t)
Why are these useful?
If T is exponential, then S(t) = exp(t))
so
and

log(S(t)) = t
(t) = t

a straight line in t with slope and intercept=0

If T is Weibull, then S(t) = exp((t))

and

so
log(S(t)) = t
then
(t) = t
log( log(S(t))) = log() + log(t)

a straight line in log(t) with slope and intercept log().

392

So we can calculate our estimated (t) and plot it versus t,

and if it seems to form a straight line, then the exponential
distribution is probably appropriate for our dataset.

Negative Log SDF

Plots for nursing home data: (t)

vs t

2.0
1.8
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200
LOS

393


Or we can plot log (t)
versus log(t), and if it seems to
form a straight line, then the Weibull distribution is probably
appropriate for our dataset.

Log Negative Log SDF

Plots for nursing home data: log[log(S(t))]

vs log(t)

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
-0.4
-0.5
4.50 4.75 5.00 5.25 5.50 5.75 6.00 6.25 6.50 6.75 7.00 7.25
Log of LOS

394

Comparison of Methods
for the Two-sample problem:
Data:
Group 0:
Group 1:

Zi
Subjects Events Follow-up
P 0
Xi
Zi = 0
n0
d0
t0 = ni=1
Zi = 1

n1

d1

t1 =

P n1

i=1 Xi

In General:
z (t) = (t, Z = z)

for z = 0 or 1.

The hazard rate depends on the value of the covariate Z.

In this case, we are assuming that we only have a single
covariate, and it is binary (Z = 1 or Z = 0)

395

MODELS
Exponential Regression:
z (t) = exp(0 + 1Z)
0 = exp(0)
1 = exp(0 + 1)
HR = exp(1)
Weibull Regression:
z (t) = exp(0 + 1Z) t1
0 = exp(0) t1

1 = exp(0 + 1) t1

HR = exp(1)
Proportional Hazards Model:
z (t) = 0(t) exp(1 )
0 = 0(t)
1 = 0(t) exp(1 )
HR = exp(1 )
396

Remarks

Exponential model is a special case of the Weibull model

with = 1 (note: Collett uses instead of )

Exponential and Weibull models are both special cases

of the Cox PH model.
How can you show this?

If either the exponential model or the Weibull model is

valid, then these models will tend to be more efficient
than PH (smaller s.e.s of estimates). This is because
they assume a particular form for 0(t), rather than estimating it at every death time.

397

For the Exponential model, the hazards are constant over

time, given the value of the covariate Zi:
Zi = 0
Zi = 1

0 = exp(0)

0 = exp(0 + 1 )

For the Weibull model, we have to estimate the hazard as a

function of time, given the estimates of 0, 1 and :
Zi = 0
Zi = 1

0(t) =

exp(0 ) t1
1(t) =

exp(0 + 1) t1

However, the ratio of the hazards is still just exp(1 ), since

the other terms cancel out.

398

Heres what the estimated hazards look like for

the nursing home data:

0.030

Exponential Hazard: Female

Exponential Hazard: Male
Weibull Hazard: Female
Weibull Hazard: Male

H 0.025
a
z 0.020
a
r
d 0.015
R
a 0.010
t
e 0.005
0.000
0

100 200 300 400 500 600 700 800 900 1000
Length of stay (days)

399

Comparison with Proportional Hazards Model

. stcox gender, nohr
failure _d:
analysis time _t:

fail
los

Iteration 0:
log likelihood
Iteration 1:
log likelihood
Iteration 2:
log likelihood
Iteration 3:
log likelihood
Refining estimates:
Iteration 0:
log likelihood

=
=
=
=

-8556.5713
-8537.8013
-8537.5605
-8537.5604

= -8537.5604

Cox regression -- Breslow method for ties

No. of subjects =
No. of failures =
Time at risk
=
Log likelihood

1591
1269
386211
-8537.5604

Number of obs

1591

LR chi2(1)
Prob > chi2

=
=

38.02
0.0000

----------------------------------------------------------------------_t |
_d |
Coef.
Std. Err.
z
P>|z| [95% Conf. Interval]
---------+------------------------------------------------------------gender |
.3943588
.0621004
6.350
0.000 .2726441
.5160734
-----------------------------------------------------------------------

d
For the PH model, 1 = 0.394 and HR
= e0.394 = 1.483.

400

Comparison with the Logrank and Wilcoxon Tests

. sts test gender
failure _d:
analysis time _t:

fail
los

Log-rank test for equality of survivor functions

-----------------------------------------------| Events
gender | observed
expected
-------+------------------------0
|
902
995.40
1
|
367
273.60
-------+------------------------Total |
1269
1269.00
chi2(1) =
Pr>chi2 =

41.08
0.0000

. sts test gender, wilcoxon

failure _d:
analysis time _t:

fail
los

Wilcoxon (Breslow) test for equality of survivor functions

---------------------------------------------------------| Events
Sum of
gender | observed
expected
ranks
-------+-------------------------------------0
|
902
995.40
-99257
1
|
367
273.60
99257
-------+-------------------------------------Total |
1269
1269.00
0
chi2(1) =
Pr>chi2 =

41.47
0.0000

401

Comparison of Hazard Ratios and Test Statistics

for effect of Gender

Wald
se(log HR) Statistic

Model/Method

HR

log(HR)

Exponential

0.0029

0.0049

1.676

0.5162

0.0619

69.507

Weibull
t = 50
t = 100
t = 500

0.0040
0.0030
0.0016

0.0060
0.0046
0.0025

1.513
1.513
1.513

0.4138

0.0636

42.381

Logrank

41.085

Wilcoxon

41.468

Cox PH
Ties=Breslow

1.483

0.3944

0.0621

40.327

Ties=Discrete

1.487

0.3969

0.0623

40.565

Ties=Efron

1.486

0.3958

0.0621

40.616

Ties=Exact

1.486

0.3958

0.0621

40.617

Score (Discrete)

41.085

402

Comparison of Mean and Median Survival

Times by Gender

Model/Method

Mean Survival
Female Male

Median Survival
Female Male

Exponential

344.5

205.6

238.8

142.5

Weibull

461.6

235.4

174.2

88.8

Kaplan-Meier

318.6

200.7

144

70

131

72

Cox PH
(Kalbfleisch/Prentice)

403

The Accelerated Failure Time Model

The general form of an accelerated failure time (AFT) model
is:
log(Ti) = AF T Zi +
where
log(Ti) is the log of a survival time
AF T is the vector of AFT model parameters corresponding to the covariate vector Zi
is a random error term
is a scale factor
In other words, we can model the log-survival
times as a linear function of the covariates.
proc lifereg in SAS and the streg command in stata
(without the exponential or weibull option) all use this loglinear model formulation for fitting parametric models.

404

By choosing different distributions for , we can obtain different parametric distributions:

Exponential
Weibull
Gamma
Log-logistic
Normal
Lognormal
We can compare the predicted survival under any of these
parametric distributions to the KM estimated survival to see
which one seems to fit best.
Once we decide on a certain class of model (say, Gamma),
we can evaluate the contributions of covariates by finding
the MLEs, and constructing Wald, Score, or LR tests of the
covariate effects.

405

We can motivate the AFT model by first demonstrating the

following two relationships:
1. For the Exponential Model:

If the failure times Ti = T (Zi) follow an exponential

distribution, i.e., Si(t) = eit with i = exp(Zi),
then
log(Ti) = Zi +

where follows an extreme value distribution (which just

means that e follows a unit exponential distribution).

2. For the Weibull Model:

If the failure times Ti = T (Zi) follow a Weibull distri

bution, i.e., Si(t) = eit with i = exp(Zi), then
log(Ti) = Zi +

where again follows an extreme value distribution, and

= 1/.
In other words, both the Exponential and Weibull model can
be written in the form of a log-linear model for the survival
times, if we choose the right distribution for .

406

The log-linear form for the exponential can be derived by:

(1) Creating a new variable T0 = TZ exp(Zi)
(2) Taking the log of TZ , yielding log(TZ ) = log

T0
exp(Zi )

Step (1): For an exponential model, recall that:

Si(t) = P r(TZ t) = et ,

with = exp(Zi)

It follows that T0 exp(1):

S0(t) = P r(T0 t) = P r(TZ exp(Z) t)
= P r(TZ t exp(Z))
= exp [ t exp(Z)]
= exp [ exp(Z) t exp(Z)]
= exp(t)
Step (2): Now take the log of the survival time:

T0

log(TZ ) = log
exp(Zi)

= log(T0) log (exp(Zi))

= Zi + log(T0)
= Zi +
where = log(T0) follows the extreme value distribution.
407

Relationship between Exponential and Weibull

If TZ has a Weibull distribution, i.e., S(t) = et

with = exp(Zi), then you can show that the new variable
TZ = TZ
follows an exponential distribution with parameter exp(Z i).
Based on the previous page, we can therefore write:
log(T ) = Z +
(where has an extreme value distribution.)
But since log(T ) = log(T ) = log(T ), we can write:
log(T ) = log(T )/

= (1/) (Zi + )
= Zi +
where = 1/.

408

This motivates the following general definition of the

Accelerated Failure Time Model by:
log(Ti) = AF T Zi +
where is a random error term, is a scale factor, Y is
the log of a survival random variable, and
AF T = e
where e came from the hazard = exp(Z).

The defining feature of an AFT model is:

S(t; Z) = Si(t) = S0( t)
That is, the effect of covariates is to accelerate
(stretch) or decelerate (shrink) the time-scale.
Effect of AFT on hazard:
i(t) = 0(t)

409

One way to interpret the AFT model is via its effect on

median survival times. If Si(t) = 0.5, then S0(t) = 0.5.
This means:
Mi = M0
Interpretation:
For < 1, there is an acceleration of the endpoint
(if M0 = 2yrs in control and = 0.5, then Mi = 1yr.
For > 1, there is a stretching or delay in endpoint
In general, the lifetime of individual i is times what
they would have experienced in the reference group
Since must be positive and a function of the covariates, we
model = exp(Zi).

410

When does Proportional hazards = AFT?

According to the proportional hazards model:
S(t) = S (t)exp(Zi)
0

and according to the accelerated failure time model:

S(t) = S0(t exp(Zi))
Say Ti W eibull(, ). Then (t) = t(1)
Under the AFT model:
i(t) = 0(t)
= eZi 0(eZi t)

= eZi 0 eZi t
=
=

eZi
eZi

!(1)

0t(1)
0(t)

But this looks just like the PH model:

i(t) = exp( Zi) 0(t)
It turns out that the Weibull distribution (and exponential,
since this is just a special case of a Weibull with = 1)
is the only one for which the accelerated failure time and
proportional hazards models coincide.
411

Special cases of AFT models

Exponential regression: = 1, following the extreme
value distribution.
Weibull regression: arbitrary, following the extreme
value distribution.
Lognormal regression: arbitrary, following the normal distribution.

Examples in stata: Using the streg command, one

has the following options of distributions for the log-survival
times:
. streg trt, dist(lognormal)
exponential
weibull
gompertz
lognormal
loglogistic
gamma

412

. streg gender, dist(exponential) nohr

-----------------------------------------------------------------------------_t |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------gender |
.516186
.0619148
8.337
0.000
.3948352
.6375368
-----------------------------------------------------------------------------. streg gender, dist(weibull) nohr
-----------------------------------------------------------------------------_t |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------gender |
.4138082
.0621021
6.663
0.000
.2920903
.5355261
1/p |
1.627501
.0377726
1.555127
1.703243
-----------------------------------------------------------------------------. streg gender, dist(lognormal)
-----------------------------------------------------------------------------_t |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------gender | -.6743434
.1127352
-5.982
0.000
-.8953002
-.4533866
_cons |
4.957636
.0588939
84.179
0.000
4.842206
5.073066
sigma |
1.94718
.040584
1.86924
2.028371
-----------------------------------------------------------------------------. streg gender, dist(gamma)
-----------------------------------------------------------------------------_t |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
---------+-------------------------------------------------------------------gender | -.6508469
.1147116
-5.674
0.000
-.8756774
-.4260163
_cons |
4.788114
.1020906
46.901
0.000
4.58802
4.988208
sigma |
1.97998
.0429379
1.897586
2.065951
------------------------------------------------------------------------------

413

This gives a good idea of the sensitivity of the test of gender

to the choice of model. It is also easy to get predicted survival curves under any of the parametric models using the
following:

. streg gender, dist(gamma)

. stcurv, survival

The options hazard and cumhaz can also be substituted

for survival above to obtain plots.

414

AFT models in SAS

proc lifereg

data=pop covout

outest=survres;

model los*censor(1)=gender /dist=exponential;

model los*censor(1)=gender /dist=weibull;
model los*censor(1)=gender /dist=gamma;
model los*censor(1)=gender /dist=normal;

Other options are lognormal, logistic, and log-logistic. The

default is to model log of response. Can specify NOLOG
for no log-transformation. In this case, normal is the same
as lognormal.

415

Lifereg

Procedure

Log Likelihood for EXPONENT -3320.476626

Variable

DF

Estimate

Std Err ChiSquare

Pr>Chi Label/Value

INTERCPT
1 5.84213388 0.033296
30786 0.0001 Intercept
GENDER
1 -0.5161878 0.061915 69.50734 0.0001
SCALE
0
1
0
Extreme value scale
Lagrange Multiplier ChiSquare for Scale 337.5998 Pr>Chi is 0.0001.

Log Likelihood for WEIBULL -3045.276811

Variable
INTERCPT
GENDER
SCALE

DF

Estimate

Std Err ChiSquare

1 5.75644118
0.0542
1 -0.6734732 0.101067
1 1.62750085 0.037773

11280.04
44.40415

Pr>Chi Label/Value
0.0001 Intercept
0.0001
Extreme value scale

Log Likelihood for GAMMA -2970.388508

Variable

DF

Estimate

INTERCPT
GENDER
SCALE
SHAPE

1
1
1
1

4.78811071
-0.6508468
1.97998063
-0.1906006

Std Err ChiSquare

0.104333
0.114748
0.043107
0.094752

2106.114
32.17096

Pr>Chi Label/Value
0.0001 Intercept
0.0001
Gamma scale parameter
Gamma shape parameter

Log Likelihood for NORMAL -9593.512838

Variable
INTERCPT
GENDER
SCALE

DF

Estimate

Std Err ChiSquare

1 303.824624 9.919629
1 -107.09585 18.97784
1 330.093584 6.918237

938.1129
31.84577

416

Pr>Chi Label/Value
0.0001 Intercept
0.0001
Normal scale parameter

Designing a Survival Study

We will focus on the power of tests based on the exponential
distribution and the logrank test.
As in standard designs, the power depends on
The Type I error (significance level)
The difference of interest, , under Ha.
A notable difference from the usual scenario is that power
depends on the number of failures, not the total
sample size.
In practice, designing a survival study involves deciding
how many patients or individuals to enter, as well as how
long they should be followed.
Designs may be fixed sample size or sequential
(More on this later!)

References:
Collett
Pocock
Williams

Chapter 12
Chapter 9 of Clinical Trials
Chapter 10 of AIDS Clinical Trials
(eds. Finkelstein and Schoenfeld)

417

Review of power calculations for 2-sample normal

Suppose we have the following data:
Group 1:

(Y11 , . . . Y1n1 )

Group 0:

(Y01 , . . . Y0n0 )

and make the following assumptions:

Y1j N (1, 2)

Y0j N (0 , 2)

Our objective is to test:

H0 : 1 = 0 H0 : 4 = 0

where 4 = 1 0

The standard test is based on the Z statistic:

Z=

Y1 Y 0
s2( n11 + n10 )

where s2 is the pooled sample variance (we are assuming

equal variances here). This test statistic follows a N (0, 1)
distribution under H0.
If the sample sizes are equal in the two arms, n0 = n1 = n/2,
(which will maximize the power), then we have the simpler
form:
Y1 Y 0
Y1 Y 0

Z = s
=
1
1
2
2s/
n
s( + )
n/2

n/2

418

The steps to follow in calculating the sample size are:

(1) Determine the critical value, c, for rejecting the null
when it is true.
(2) Calculate the probability of rejecting the null when the
alternative is true, substituting c from above.
(3) Rewrite the expression in terms of the sample size for a
given power.
Step (1):
Set the significance level, , equal to the probability of rejecting the null hypothesis when it is true:
= P r (|Y1 Y0| > c | H0)

= Pr

|Y1 Y0|
c

>
| H0
2s/ n
2s/ n

c
c
|Z| >

= 2

= Pr
2s/ n
2s/ n
so z1/2 =

2s/ n

z1/2 2 s

n
Note that z is the value such that (z ) = P r(Z < z ) =
.
or c =

419

Step (2):
Calculate the probability of rejecting the null when Ha is
true. Start out by writing down the probability of a Type II
error:
= P r (accept H0 | Ha)
so 1 = P r (reject H0 | Ha)
= P r (|Y1 Y0| > c | Ha)

c
|Y1 Y0|

| Ha
= Pr
>
2s/ n
2s/ n

c
Z >

= Pr
2s/ n
so we get

z = z1 =

2s/ n

Now we substitute c from Step (1):

z1

z1/2 2 s/ n

=
2s/ n
= z1/2
420

2s/ n

Step (3):
Now rewrite the equation in terms of sample size for a given
power, 1 , and significance level, :

2s/ n

n
=
2s

z1/2 + z1 =

(z1/2 + z1 )24s2
n =
2

Notes:
The power is an increasing function of the standardized difference:
4

T (4) =
2s/ n
This is just the number of standard errors between the two
means, under the assumption of equal variances.
1. As n increases, the power increases.
2. For fixed n, the power increases with 4.
3. For fixed n and 4, the power decreases with s.
4. Assigning equal numbers of patients to the two groups
(n1 = n0 = n/2) is best in terms of maximizing power.
421

An Example:

z1 2 + z1
n=
42

4s2

Say we want to derive the total sample size required to yield

90% power for detecting a difference of 0.5 standard deviations between means, based on a two-sided 0.05 level test.
= 0.05
z1 2 = 1.96
= 0.10
z1 = z0.90 = 1.28

(1.96 + 1.28)2 4s2

42 s2
n =

42
42
For a 0.5 standard deviation difference, /s = 0.5, so
n

42
= 168
(0.5)2

If you end up with n < 30, then you should be using the
t-distribution rather than the normal to calculate critical
values, and then the process is iterative.
422

Survival Studies: Comparing Proportions of Events

In some cases, the sample size for a survival trial is based
on a crude comparison of the proportion of events at some
fixed point in time.
In this case, we can apply the results just shown to get sample
sizes, based on the normal approximation to the binomial:
Define:
Pc

probability of event in control arm by time t

Pe

probability of event in experimental arm by time t

The number of patients required per treatment arm based

on a chi-square test comparing binomial proportions is:
q

{z1 2 2P (1 P ) + z1 Pe (1 Pe ) + Pc (1 Pc )}2
N =
(Pc Pe )2

where P = (Pe + Pc)/2

(This looks slightly different because the variance is not the
same under Ho and Ha, as was the case in the normal previous example.)
423

Notes on comparing proportions of failures:

Use of chi-square test is best when 0.2 < Pe, Pc < 0.8
Should have 15 patients in each cell of the (2x2) table
For smaller sample sizes, use Fishers exact test to motivate power calculations
Efficiency vs logrank test is near 100% for studies with
short durations relative to the median event time
What does this mean in terms of the event
rates? High or low?
Calculation of sample size for comparing proportions often provides an upper bound to those based on comparison of survival distributions

424

Sample size based on the logrank test

Recap: Consider a two group survival problem, with equal
numbers of individuals in the two groups (say n0 in group 0
and n1 in group 1). Let 1, ..., K represent the K ordered,
distinct failure times, and at the j-th event time:

Group
0

Die/Fail
Yes
No
Total
d0j
r0j d0j r0j

d1j

r1j d1j

r1j

Total

dj

rj d j

rj

where d0j and d1j are the number of deaths (events) in group
0 and 1, respectively, at the j-th event time, and r0j and r1j
are the corresponding numbers at risk.

The logrank test is: (z-statistic version)

ZLR =

PK

(d1j ej )
j=1
s
PK
j=1 vj

with ej = dj r1j /rj

vj = r1j r0j dj (rj dj )/[rj2(rj 1)]
425

Distribution of the logrank statistic

Suppose that the hazard rates in the two groups are 0(t)
and 1(t), with hazard ratio
1(t)
= e =
0(t)
and suppose we are interested in testing Ho : = ln() = 0
(which is equivalent to testing Ho : = 1.)
[Note: we will use ln() rather than in the following, so that there
is no confusion with the Type II error rate]

It is possible to show that

if there are no ties, and
we are near H0 :

then:
E(d1j ej |d1j , d0j , r1j , r0j ) ln()/4

vj 1/4

So, at a point ln() in the alternative, we get:

ZLR

PK

j=1 ln()/4

PK

j=1 1/4

d ln()/4
r

and ZLR N ( ln() d/2 , 1 )

426

d/4

d ln()
=
2

Heuristic Proof:

E(d1j |d1j , d0j , r1j , r0j ) = P r(d1j = 1|dj = 1, r1j , r0j )

=

r1j 0
r1j 0 + r0j 0

r1j
r1j + r0j

r1j
r1j r0j

+ ln()
r1j + r0j
(r1j + r0j )2

But ej = r1j /(r1j + r0j ), so:

r1j r0j

E(d1j |d1j , d0j , r1j , r0j ) ej = ln()

(r1j + r0j )2
If n0 = n1, then near H0 :, r1j r0j , hence,
E(d1j |d1j , d0j , r1j , r0j ) ej = ln()/4
Similarly, with no ties, we have
vj = r1j r0j /rj2 1/4

427

This can also be derived via the partial likelihood:

We can write the partial likelihood as:

j
Z
n
j
e
Y

P

log

Z
`
j=1
`R(j ) e

l() =

n
X

j=1

j Zj log

`R(j )

eZ`

and then the score (partial derivative of log-likelihood) becomes:

U () =

`()

Z`
`R(j ) Z` e

=
j Zj P
Z
`
e
j=1
`R( )
n
X

We can write the information (minus second partial derivative

of the log-likelihood) as:

n
X
2
2 `() =
j

j=1

Z`
`R(j ) e

Z` (
`R(j ) Z` e
P
Z`
`R( ) e
j

The logrank statistic (with no ties) is equivalent to the score statistic

for testing = 0:
U (0)
ZLR

I(0)

By a Taylor series expansion:

U
U (0)
(0)
= U ()

and I(0)
E[U (0)]
= d/4
= d/4
428

Z` )2
`R(j ) Z` e

Power of the Logrank Test

Using a similar argument to before, the power of the logrank
test (based on a two-sided level test) is approximately:

Power() 1 z

1 2

ln() d/2

Note: Power depends only on d and !

We can easily solve for the required number of events to
achieve a certain power at a specified value of :
To yield power() = 1 , we want d so that

1 = 1 z1 2 ln() d/2

z = z1 2 ln() d/2
d =
or d =

1 2

1 2

z
4 z
[ln()]2
4 z

+ z1
[ln()]2

429

Example:
Say we were planning a 2-arm study, and wanted to be able
to detect a hazard ratio of 1.5 with 90% power at a 2-sided
significance level of = 0.05.
Required number of events:

4 z1 2 + z1
d =
[ln()]2

4(1.96 + 1.282)2
=
[ln(1.5)]2

42
= 256
0.1644

# Events required for various Hazard Ratios

Hazard
Ratio
1.5
2.0
2.5
3.0

Power
80%
90%
191
256
66
88
38
50
26
35

Most studies are designed to detect a hazard ratio of 1.5-2.0.

430

Practical Considerations
How do we decide on ?
How do we translate numbers of failures to numbers of
patients?

Hazard ratios for the exponential distribution

The hazard ratio from two exponential distributions can be
easily translated into more intuitively interpretable quantities:
Median:
If Ti exp(i), then
Median(Ti) = ln(0.5)/i
It follows that
0
Median(T1)
1
=
= e =
Median(T0)
1

Hence, doubling the median survival of a treated compared

to a control group will correspond to halving the hazard.

431

R-year survival rates

Suppose the R-year survival rate in group 1 is S1(R) and in
group 0 is S0(R). Under the exponential model:
Si(R) = exp(iR)
Hence,
ln(S1(R)) 1R
1
=
=
= e =
ln(S0(R)) 0R
0
Hence, doubling the hazard rate from group 1 to group 0 will
correspond to doubling the log of the R-year survival rate.
Note that this result does not depend on R!.
Example: Suppose the 5-year survival rate on treatment A
is 20% and we want 90% power to detect an improvement of
that rate to 30%. The corresponding hazard ratio of treated
to control is:
ln(0.3) 1.204
=
= 0.748
ln(0.2) 1.609
From our previous formula, the number of events (deaths)
needed to detect this improvement with 90% power, based
on a 2-sided 5% level test is:
4(1.96 + 1.282)2
d=
= 499
[ln(0.748)]2
432

Translating to Number of Enrolled Patients

First, suppose that we will enter N patients into our study
at time 0, and will then continue the study for F units of
time.
Under H0, the probability that an individual will fail during
the study is:
P r(f ail) =

Z F

0e0tdt

= 1 e0F
Hence, if our calculations say we need d failures, then to
decide how many patients to enter, we simply solve
d = (N/2)(1 e0F ) + (N/2)(1 e1F )
To solve the above equation for N , we need to supply values
of F and d. In other words, here we are already deciding
what HR we want to detect (with what power, etc), and for
how long we are going to follow patients. What we get is
the total number of patients we need to enroll in order to
observe the desired number of events in F units of follow-up
time.
433

Example: Suppose we want to detect a 50% improvement

in the median survival from 12 months to 18 months with
80% power at = 0.05, and we plan on following patients
for 3 years (36 months).
We can use the two medians to calculate both the parameters
0 and 1 and the hazard ratio, :
Median(Ti) = ln(0.5)/i
ln(0.5) 0.6931
=
= 0.0385
M1
18
ln(0.5) 0.6931
0 =
=
= 0.0578
M0
12
0.0385
1
12
=
=
=
= 0.667
0
0.0578
18

so 1 =

and from our previous table, # events required is d = 191

(same for = 1.5 as it is for 1/1.5=0.667).
So we need to solve:
191 = (N/2)(1 e0.057836 ) + (N/2)(1 e0.038536 )
= (N/2)(0.875) + (N/2)(0.7500) = (N/2)(1.625)
N = 235
(for practical reasons, we would probably round up to 236
and randomize 118 patients to each treatment arm)
434

A more realistic accrual pattern

In reality, not everyone will enter the study on the same day.
Instead, the accrual will occur in a staggered manner over
a period of time.
The standard assumption:
Suppose individuals enter the study uniformly over an accrual period lasting A units of time, and that after the accrual period, follow-up will continue for another F units of
time.
To translate d to N , we need to calculate the probability that
a patient fails under this accrual and follow-up scenario.
P r(fail) =

Z A

P r(fail|enter at a) f (a) da

= 1
Then solve:

RA

S(a + F ) da
A

(2)

d = (N/2)P r(fail; 0) + (N/2)P r(fail; 1)

= (N/2)Pc + (N/2)Pe
= (N/2)(Pc + Pe)

If we now solve for N (substituting in formula for d), we get:

2d
N =
(Pc + Pe)
N =

8 z

+ z1
[ln()]2

1 2

435

1
(Pc + Pe)

How can we get Pc and Pe from (2)?

If we assume that the exponential distribution holds, then
we can solve (2) to obtain:

Pi = 1

exp(iF )(1 exp(iA))

i A

(3)

(for i = c, e)

Freedman suggested an approximation for Pc and Pe, by

computing the probability of an event at the median duration
of follow-up, (A/2 + F ):

Pi = P r(fail; i) = 1 exp[i(A/2 + F )]

(4)

He showed that this approximation works pretty well for the

exponential distribution (i.e., it gives values close to (3)).

436

An alternative formulation
Rubenstein, Gail, and Santner (1981) suggest the following
approach for calculating the total sample size that must be
enrolled:

2 z1 2 + z1
N =
[ln()]2

1
1

+
Pc Pe

where Pc and Pe are the expected proportion of patients or

individuals who will fail (have an event) on the control and
treatment arms.

How do we calculate (estimate) Pc and Pe ?

using the general formula for a distribution S given in
(2)
using the exact formula for an exponential distribution
given in (3)
using the approximation given by (4)
Note: all of these formulas can be modified for unequal
assignment to treatment (or exposure) groups by changing
(N/2) in the formulas on p.17-19 to (qc N ) and (qe N ),
where qc and qe are the proportions assigned to the control
and exposed groups, respectively.
437

Freedmans Approach (1982)

Freedmans approach is based on the logrank statistic under
the assumption of proportional hazards, but does not require
the assumption of exponential survival distributions.
Total number of events:
d =

z1 2 + z1

+
1

Total sample size:

N =

2 z

+ z1
Pe + P c

1 2

+
1

where Pe and Pc are estimated using (4).

This approximation depends on the assumption of a constant ratio between the number of patients at risk in the two
treatment groups prior to each event time = r0j r1j (as
shown in the heuristic proof). When this assumption is
not satisfied, the required sample sizes tend to be overestimated.

Q. When would this assumption not be satisfied?

A. When the smallest detectable difference is large.

438

Some examples of study design

Example I:
A clinical trial in esophageal cancer will randomize patients
to radiotherapy alone (Rx A) versus radiotherapy plus chemotherapy (Rx B). The goal of the study is to compare the two
treatments with respect to survival, and we plan to use the
logrank test. From historical data, we know that the median
survival on RX A for this disease is around 9 months. We
want 90% power to detect an improvement in this median
to 18 months. Past studies have been able to accrue approximately 50 patients per year. Choose a suitable study
design.

439

Example II:
A clinical trial in early stage breast cancer will randomize
patients after their surgery to Tamoxifen (ARM A) versus
observation only (ARM B). The goal of the study is to compare the two treatments with respect to time to relapse, and
the logrank test will be used in the analysis. From historical
data, we know that after five years, 65% of the patients will
still be disease free. We would like to have 90% power to
detect an improvement in this disease free rate to 75%. Past
studies have been able to accrue approximately 200 patients
per year. Choose a suitable study design.

440

Example III:
Some investigators in the environmental health department
want to conduct a study to assess the effects of exposure to
toluene on time to pregnancy. They will conduct a cohort
study involving women who work in a chemical factory in
China. It is estimated that 20% of the women will have
workplace exposure to toluene. Furthermore, it is known
that among unexposed women, 80% will become pregnant
within a year. The investigators will be able to enroll 200
women per year into the study, and plan an additional year
of follow-up at the end of accrual. Assuming they have 2
years accrual, what reduction in the 1-year pregnancy rate
for exposed women will they be able to detect with 85%
power? What if they have 3 years of accrual?

441

Other important issues:

The approaches just described address the basic question of
calculating a sample size for study with a survival endpoint.
These approaches often need to be modified slightly to address the following complications:
Loss to follow-up
Non-compliance (or cross-overs)
Stratification
Sequential monitoring
Equivalence hypotheses
Next we summarize some of the main points.

442

Loss to follow-up
If some patients are lost to follow up (as opposed to censored
at the end of the trial without the event), the power will be
decreased.
There are two main approaches for dealing with this:
Simple inflation method - If `*100% of patients
are anticipated to be lost to follow up, calculate target
sample size to be
N

1
=
N
1`

Example: Say you calculate N = 200, and anticipate

losses of 20%. The simple inflation method would give
you a target sample size of N = (1/0.8) 200 = 250.

Warning: people often make the mistake of just inflating the original sample size by `*100%, which would
have given N = 240 for the example above.

Exponential loss assumption - the above approach

assumes that losses contribute NO information. But we
actually have information on them up until the time that
they are lost. Incorporate this by assuming that time to
loss also follows an exponential distribution, and modify
Pe and Pc.
443

Noncompliance
If some patients dont take their assigned treatments, the
power will be decreased. This issue has two sides:
Drop-outs (de) - patients who cannot tolerate the
medication stop taking it; their hazard rate would become the same as the placebo group (if included in study)
at that point.
Drop-ins (dc) - patients assigned to less effective therapy may not get relief from symptoms and seek other
therapy, or request to cross-over.
A conservative remedy adjust Pe and Pc as follows:

Pe = Pe(1 de) + Pc de
Pc = Pc(1 dc) + Pe dc

444

Design Strategy:
1. Decide on
Type I error (significance level)

clinically important difference (in terms of HR)

desired power

2. Determine the number of failures needed

3. Based on past experience
decide on a reasonable distribution for the controls
(usually exponential)
estimate anticipated accrual per unit time

estimate expected rate of loss to follow up

Vary the values of A and F until you get something practically feasible that gives the right number of failures.
4. Consider noncompliance, sequential monitoring, and other
issues impacting sample size

445

Included on the next several pages is a SAS program to calculate sample sizes for survival studies. It uses several of the
approaches weve discussed, including:
Rubenstein, Gail and Santner (RGS, 1981)

Freedman (1982)

Lachin and Foulkes (1986)

A copy of this program is shown on the next several pages.

The program requires entry of:
Significance level (alpha)

Power

Sides (1 for one-sided test, 2 for two-sided test)

Accrual period

Follow up period

Yearly rate of loss to follow-up

Proportion randomized to experimental treatment arm

One of the following:

Yearly event rate on control and experimental treatment arms

Yearly event rate on control arm, and the hazard
ratio
Median time to event on control and experimental
treatment arms
446

The SAS program rgsnew.sas

data rgs;
***********************************************************************;
*** enter the following information in this block;
alpha = 0.05;
/* significance level */
sides = 2;
/* one-sided or two-sided test */
power = 0.90;
/* Desired power */
accrual = 2;
/* Accrual period in years */
fu = 1.5;
/* Follow up after last patient is accrued */
loss = 0.0;
/* yearly rate of loss */
qe = 0.5;
/* proportion randomized to experimental arm */
*** either enter the median time to event in years on control;
*** or the yearly event rate - leave the other value missing;
medianc = 0.75;
/* median time to event on control arm */
probc = .;
/* yearly event rate in control arm */
*** either enter the yearly event rate in the experimental arm ;
*** or the hazard ratio for control vs experimental
;
*** or the median time to event on experimental arm in years
;
*** leave the other values missing (.);
mediane = 1.5;
/* median time to event on experimental */
probe = .;
/* yearly event rate in experimental arm */
rr=.;
/* hazard ratio */
***********************************************************************;
beta = 1 - power;
qc = 1 - qe;
zalpha = probit(1-alpha/sides);
zbeta = probit(1-beta);
*** calculate yearly event rate in both arms using medians, if supplied;
if medianc^=. then do;
hazc=-log(0.5)/medianc;
probc=1-exp(-hazc);
end;
if mediane^=. then do;
haze=-log(0.5)/mediane;
probe=1-exp(-haze);
end;
hazc = -log(1-probc);

447

*** calculate hazard in experimental group, using yearly event rate;

*** or hazard ratio;
if probe^=. then haze = -log(1-probe);
if rr^=. then haze = hazc/rr;
if probe^=. and haze^=. then do;
put "**************************************************************";
put "WARNING: both yearly event rate and hazard ratio (HR) have";
put "
been specified. Calculations will use the HR";
put "**************************************************************" /;
end;
*** calculate median survival times if not supplied;
medianc=-log(0.5)/hazc;
mediane=-log(0.5)/haze;
hazl = -log(1-loss);
avghaz = qc*hazc + qe*haze;
rr = hazc/haze;
log_rr = log(hazc/haze);
totloss=(accrual*0.5 + fu)*loss;
*** compute expected probability of death (event) during trial;
*** given staggered accrual but NO loss;
pc0loss = 1-((exp(-hazc*fu)-exp(-hazc*(accrual+fu)))/(hazc*accrual));
pe0loss = 1-((exp(-haze*fu)-exp(-haze*(accrual+fu)))/(haze*accrual));
*** compute expected probability of event during trial;
*** given staggered accrual AND loss;
pc = (1 - (exp(-(hazc+hazl)*fu)-exp(-(hazc+hazl)*(fu+accrual)))
/((hazc+hazl)*accrual))*(hazc/(hazc+hazl));
pe = (1 - (exp(-(haze+hazl)*fu)-exp(-(haze+hazl)*(fu+accrual)))
/((haze+hazl)*accrual))*(haze/(haze+hazl));
pbar = (1 - (exp(-(avghaz+hazl)*fu)-exp(-(avghaz+hazl)*(fu+accrual)))
/((avghaz+hazl)*accrual))*(avghaz/(avghaz+hazl));
*** compute total sample size assuming loss;
N = int(((zalpha+zbeta)**2)/(log_rr**2)*(1/(qc*pc)+1/(qe*pe))) + 1;
*** Compute sample size using method of Freedman (1982);
N_FRD = int((2*(((rr+1)/(rr-1))**2)*(zalpha+zbeta)**2)/(2*(qe*pe+qc*pc)))+1;

448

*** Compute sample size using method of Lachin and Foulkes (1986);
*** with rates under H0 given by pooled hazard;
N_LF = int((zalpha*sqrt((avghaz**2)*(1/pbar)*(1/qc + 1/qe)) +
zbeta*sqrt((hazc**2)*(1/(qc*pc)) + (haze**2)*(1/(qe*pe))))**2/
((hazc-haze)**2)) + 1;
*** compute total sample size assuming no loss;
n_0loss = int(((zalpha+zbeta)**2)/(log_rr**2)*
(1/(qc*pc0loss)+1/(qe*pe0loss))) + 1;
*** compute sample size using simple inflation method for loss;
naive = int(n_0loss/(1-totloss)) + 1;
*** verify that actual power is same as desired power;
newpower = probnorm(sqrt(((N)*(log_rr**2))/(1/(qc*pc) + 1/(qe*pe)))
- zalpha);
if abs(newpower-power)>0.001 then do;
put *** WARNING: actual power is not equal to desired power;
put Desired power: power Actual power: newpower;
end;
***********************************************************************;
*** compute number of events expected during trial;
***********************************************************************;
*** Compute expected number under null;
n_evth0 = int(n*pbar) + 1;
r=qc/qe;
*** Rubinstein, Gail and Santner (1981) method - simple approximation;
n_evtrgs = int((((r+1)**2)/r)*((zalpha + zbeta)**2)/(log_rr**2)) + 1;
*** Freedman (1982);
n_evtfrd = int((((rr+1)/(rr-1))**2) * (zalpha+zbeta)**2)+1;
*** Using
n_evt_c =
n_evt_e =
n_evtlf =

backtracking method of Lachin and Foulkes (1986);

int(N*qc*pc) + 1;
int(N*qe*pe) + 1;
n_evt_c + n_evt_e;

449

label sides=Sides
alpha=Alpha
power=Power
beta=Beta
zalpha=Z(alpha)
zbeta=Z(beta)
accrual=Accrual (yrs)
fu=Follow-up (yrs)
loss=Yearly Loss
totloss=Total Loss
probc=Yearly event rate: control
probe=Yearly event rate: active
rr=Hazard ratio
log_rr=Log(HR)
N=Total Sample size (RGS)
N_FRD=Total Sample size (Freedman)
N_LF=Total Sample size (L&F)
n_0loss=Sample size (no loss)
pc=Pr(event), control
pe=Pr(event), active
pc0loss=Pr(event| no loss), control
pe0loss=Pr(event| no loss), active
n_evth0=# events (Ho-pooled)
n_evtrgs=# events (RGS)
n_evtfrd=# events (Freedman)
n_evtlf=# events (L&F)
medianc=Median survival, control
mediane=Median survival, active
naive=Sample size (naive loss);

proc print data=rgs label noobs;

title Sample size & expected events for comparing two survival distributions;
title2 Using method of Rubinstein, Gail and Santer (RGS, 1981);
title3 Freedman (1982), or Lachin and Foulkes (L&F, 1986);
var sides alpha power accrual fu loss totloss
probc probe medianc mediane pc pe pc0loss pe0loss rr log_rr
n_evth0 n_evtrgs n_evtfrd n_evtlf N N_FRD N_LF n_0loss naive;
format power loss totloss f4.2 medianc mediane f5.3
probc probe rr log_rr pc pe pc0loss pe0loss f6.4;

450

Back to Example I:
A clinical trial in esophageal cancer will randomize patients
to radiotherapy alone (Rx A) versus radiotherapy plus chemotherapy (Rx B). The goal of the study is to compare the two
treatments with respect to survival, and we plan to use the
logrank test. From historical data, we know that the median
survival on Rx A for this disease is around 9 months. We
want 90% power to detect an improvement in this median
to 18 months. Past studies have been able to accrue approximately 50 patients per year. Choose a suitable study
design.

First, lets write down what we know:

desired significance level not stated, so use = 0.05
(assume a two-sided test)
assume equal randomization to treatment arms
(unless otherwise stated)
desired power is 90%
median survival on control is 9 months M 0 = 9
want to detect improvement to 18 months on Rx B
M 1 = 18
Maximum accrual per year is 50 patients
451

We have all of the information we need to run the program,

except the accrual and follow up times. We need to use trial
and error to get these.
Number of Total
Total
Accrual Follow-up Events Sample Study
Period
Period
Required
Size Duration
1

2.5

88

106

3.5

1.5

88

115

3.5

2.5

88

122

3.5

0.5

88

133

3.5

88

117

Shown on the next page is the output from rgsnew.sas using Accrual=2, Follow-up=1.5. Ive given the RGS numbers
above.
Which of the above are feasible designs?

452

Sample size & expected events for comparing two survival distributions
Using method of Rubinstein, Gail and Santer (RGS, 1981)
Freedman (1982), or Lachin and Foulkes (L&F, 1986)

Sides

Alpha

Power

Accrual
(yrs)

0.05

0.90

Follow-up
(yrs)

Yearly
Loss

Total
Loss

Yearly
event
rate:
control

1.5

0.00

0.00

0.6031

Yearly
event
rate:
active

Median
survival,
control

Median
survival,
active

Pr(event),
control

Pr(event),
active

Pr(event|
no loss),
control

0.3700

0.750

1.500

0.8860

0.6737

0.8860

Pr(event|
no loss),
active

Hazard
ratio

Log(HR)

# events
(Ho-pooled)

# events
(RGS)

# events
(Freedman)

0.6737

2.0000

0.6931

94

88

95

# events
(L&F)

Total
Sample
size (RGS)

Total
Sample
size
(Freedman)

Total
Sample
size (L&F)

Sample
size (no
loss)

Sample
size
(naive
loss)

90

115

122

121

115

115

453

How do we pick from the feasible designs?

The first 4 designs all have 3 1/2 years total duration, since
the follow-up period starts after the last patient has been
accrued. The shorter the follow-up period given this fixed
study duration, the more patients we have to enroll.
In some cases, it will be much more cost-effective to enroll
fewer patients and follow them for longer. This corresponds
to cases where the initial cost per patient is very high.
In other cases (where the initial cost per patient is lower), it
will be better to enroll more patients. The median follow-up
for the first 4 designs are 3, 2.5, 2.25, and 2 years, respectively. The total cost of treatment could be estimated by
multiplying the number of patients by the median follow-up
time.
Some prefer to keep the accrual period as short as possible, given how many patients can feasibly be enrolled. This
will tend to give the smallest number of patients among the
feasible designs. Which design would this correspond to?
Another issue to think about is whether the background conditions of the disease are changing rapidly (like AIDS) or are
fairly stable (like many types of cancer). For the former situation, it would be best to have a study with a short duration
so the results will have more interpretation.

454

Using the information given, there are a lot of other quantities we can calculate:
The hazard ratio of control to treated is:
median(Rx B)
18
=
= 2
median(Rx A)
9
The hazard rates for the two treatment arms are:
log(0.5)
log(0.5)
=
= 0.0770
for Rx A: 0 =
median(Rx A)
9
for Rx B: 1 =

log(0.5)
log(0.5)
=
= 0.0385
median(Rx B)
18

The yearly probability of an event is:

for Rx A: P r(T < 1|0) = 1 e(0t)

= 1 e(0.077012) = 0.603
for Rx B: P r(T < 1|1) = 1 e(1t)
= 1 e(0.038512) = 0.370

What would happen above if we used time t

in years (i.e., t=1) instead of months?
What would happen if we calculated both the
hazard rate and yearly event probability using
time in years?
455

Based on a design with 2.5 years accrual and

1 year follow-up:
The median follow-up time
median FU = A/2 + F
= 30/2 + 12 = 27 months
The probability of an event during the entire study is:
(using the approximation in notes)
for Rx A: Pc = 1 exp(0 [A/2 + F ])
= 1 exp(0.0770 27) = 0.875
for Rx B: Pe = 1 exp(0 [A/2 + F ])
= 1 exp(0.0385 27) = 0.646
(the above numbers differ from what youd get in the
printout from the program, since it calculates the exact
probability under the exponential distribution, instead
of using the approximation)
In the calculations above, all of the time periods were in
terms of months. You have to remember to keep the scale
the same throughout.
To use the program, you need to translate the time scale in
terms of years. So a median of 18 months survival would be
entered as median=1.5.

456

What happens if we add loss to follow-up?

Required sample size for A=2.5, FU=1 year
Yearly Number of Total
Total
Loss to
Events Sample Study
Follow-up Required
Size Duration
0

88

122

3.5

5%

88

128

3.5

10%

88

133

3.5

20%

88

147

3.5

457

Sequential Design and Analysis of survival studies

In clinical trials and other studies, it is often desirable to
conduct interim analyses of a study while it is still ongoing.
Rationale:
ethical: if one treatment is substantially worse than
another, then it is wrong to continue to give the inferior
treatment to patients.
timely reporting: if the hypothesis of interest has
been clearly established halfway through the study, then
science and the public may benefit from early reporting.

WARNING!!
Unplanned interim analyses can seriously inflate the true
type I error of a trial. If interim analyses are to be performed,
it is ESSENTIAL to carefully plan these in advance, and to
adjust all tests appropriately so the the type I error is of the
desired size.

458

How does the type I error become inflated?

Consider a two group study comparing treatments A and B.
Suppose the data are normally distributed (say Xi N (A, 2)
in group A, and similarly for group B), so that the null hypothesis of interest is
H0 : A = B
It is not too hard to figure out how the type I error can get
inflated if a naive approach is used.
Suppose we plan to do K interim analyses, and that exactly
m individuals will enter each treatment between each analysis. The test statistic at the kth analysis will be

Zk =

Pk

i=1

Pm

Aij
j=1 (X
r

XBij )/km

2/km

Pk

i=1 di /k

2/km

where di is the difference between the two group means at

the ith analysis,
di = X Ai X Bi
and X Ai and X Bi are the means in groups A and B of the
m individuals who entered in the i-th time period.
459

(naive) Interim monitoring procedure:

Allow m patients to enter on each treatment arm
(total of 2m additional patients)
Calculate Zk based on the current data
Reject the null hypothesis if |Zk | > z1/2 , where is
the desired type I error.
The overall type I error rate for the study is:
P r(|Z1 | > z1/2 or |Z2| > z1/2 ... or |ZK | > z1/2 )
If the test at each interim analysis is performed at level ,
then clearly this probability will exceed . The table below
shows the Type I error rate if each test is done at = 0.05
for various values of K:
Number of interim analyses (K)
1
2
3
4
5
10
25
5% 8.3% 10.7% 12.6% 14.2% 19.3% 26.6%
(from Lee, Statistical Methods for Survival Data, Table
12.9)

460

For survival data, the calculations become MUCH more complicated since the data collected within each time interval
continues to change as time goes on!

What can we do to protect against this type I error inflation?

Pocock Approach:
Pick a smaller significance level (say 0) to use at each interim
analysis so that the overall type I error stays at level .
A problem with the Pocock method is that even the very
last analysis has to be performed at level 0. This tends to
be very conservative at the final analysis.
OBrien and Fleming Approach:
A preferable approach would be to vary the alpha levels used
for each of the K interim analyses, and try to keep the very
last one close to the desired overall significance level. The
OBrien-Fleming approach does that.

461

Comments and notes:

There are several other approaches available for sequential design and analysis. The OBrien and Fleming
approach is probably the most popular in practice.
There are many variations on the theme of sequential
design. The type we have discussed here is called Group
sequential analysis.
There are other approaches that require continuous
analysis after each new individual enters the study!
There are also approaches where the randomization
itself is modified as the trial proceeds. E.g. Zelens Play the winner rule (New England Journal of
Medicine 300, 1979, page 1242) and Wares ECMO
study (Statistical Science, 4, 1989, page 298)
Some designs allow for early stopping in the absence of a
sufficient treatment effect as the trial progresses. These
procedures are referred to as stochastic curtailment or
conditional power calculations.

462

 Designing a group sequential trial for survival data requires sophisticated and highly specialized software. EaSt,
a package from CYTEL SOFTWARE that does standard (fixed) survival designs, as well as sequential designs.
Many non-statistical issues enter decisions about whether
or not to stop a trial early
P-values based on analyses of studies with sequential
designs are difficult to interpret.
Once you do 5 interim analyses, then adding more makes
little difference. Some clinical trials groups (HSPH AIDS
group) have large randomized Phase III studies monitored at least once per year (for safety reasons), and
most studies have 1-3 interim looks.
Going from a fixed to a group sequential design adds
only about 3-4% to the required maximum sample size.
This is a good rule of thumb to use in calculating the
sample size when you plan on doing interim monitoring.

463

Competing Risks and Multiple Failure Times

So far, weve been acting as if there was only one endpoint
of interest, and that censoring due to death (or some other
event) was independent of the event of interest.
However, in many contexts it is likely that the time to censoring is somehow correlated with the time to the event of
interest. In general, we often have several different types
of failure (death, relapse, opportunistic infection, etc) which
are related (i.e., dependent or competing risks).

Examples:
After a bone marrow transplantation, patients are followed to evaluate leukemia-free survival, so the endpoint is time to leukemia relapse or death. This endpoint
conists of two types of failures (competing risks):
leukemia relapse
non-relapse deaths
In cardiovascular studies, deaths from other causes (such
as cancer) are considered competing risks.
In actuarial analyses, we look at time to death, but want
to provide separate estimates of hazards for each cause
of death (multiple decrement lifetables).

464

Another example: For the MAC study, the analyses you

have been doing of time to MAC assume that the censoring
time is independent.

Recall:
T = time to event of interest (MAC)
U = time to censoring (death, loss to FU)
X = min(T, U )
= I(T U )

Observable Data: (X, )

What are the possiblities here?

(1) Failure T and censoring U are independent
(2) Failure T and censoring U are dependent

465

Case (1): Independent failure times

(this includes the case of independent censoring)
BOTTOM LINE NO PROBLEM
Nonparametric estimation:
In this case, we can use the Kaplan-Meier estimator to estimate ST (t) = P (T > t).

Parametric estimation:
If we know the joint distribution of (T, U ) has a certain
parametric form (exponential, Weibull, log-logistic), then we
can use the likelihood for (X, ) to get parameter estimates
of the marginal distribution of ST (t).

Semi-parametric estimation:
We can apply the Cox regression model to assess the effects
of covariates on the marginal hazard.

466

Case (2): Dependent failure times

BOTTOM LINE BIG PROBLEM
Tsiatis (1975) showed that ST (t) = P (T t) (i.e., the survival function for the event T of interest) cannot be identified from data of the form (X, ) for each subject.
In fact, observing (X, ) does not provide enough information to estimate the joint distribution of (T, U ) so that we
can even check whether the assumption of independence is
valid.
When is it reasonable to assume independent risks?
when censoring occurs because the study ends, or because the subject moves to a different state
and there is no trend over time in health status of enrolling patients
In the case of our MAC study, the fact that someone dies
may reflect that they would have been at greater risk of MAC
if they had not died than someone else who remained alive
at that point.
The assumption of independence means that the hazard for
someone who is censored at time t is exactly the same as that
for someone with the same covariates who is also at risk at
time t.
467

What is the impact of dependent competing risks?

Slud and Byar (1988) show that dependent causes of death
can potentially make risk factors appear protective:
If we have
T = death from cause of interest
and U = censoring, from death due to other cause
and a single binary covariate Z
Z =

1 if risk factor is present

0 otherwise

and we calculate the Kaplan-Meier survival estimates S1(t)

for Z = 1 and S0(t) for Z = 0 assuming independent censoring, then we could (in their hypothetical example) end up
reversing the sign of the survival functions:
True ordering between survival distributions:
S1(t) < S0(t)

for all t

Kaplan Meier estimates of survival distributions:

S1(t) > S0(t)

468

for all t

What can we do if we suspect dependent risks?

A lot of people have tried to tackle this problem!
References
Aly EAA, Kochar SC, and McKeague IW (1994). Some tests for
comparing cumulative incidence functions and cause-specific hazard
rates. JASA 89, 994-999.
Benichou J and Gail MH (1990). Estimates of absolute cause specific
risk in cohort studies. Biometrics 46, 813-826.
(*) Booth AL and Satchell SE (1995). The hazards of doing a PhD: an
analysis of completion and withdrawal rates of British PhD students
in the 1980s. JRSS-A, 297-318.
Farewell VT (1979). An application of Coxs proportional hazard
model to multiple infection data. Applied Statistics 28, 136-143.
Gail, M (1982). Competing risks. Encyclopedia of Statistical Sciences
2, 75-81.
Lin DY, Robins JM, and Wei LJ (1996). Comparing two failure time
distributions in the presence of dependent censoring. Biometrika
83, 381-393.
Lunn M and McNeil D (1995). Applying Cox regression to competing
risks. Biometrics 51, 524-532
Moeschberger ML and Klein JP (1988). Bounds on net survival probabilities for dependent competing risks. Biometrics 44, 529-538.
(*) Moeschberger ML and Klein JP (1995). Statistical methods for dependent competing risks. Lifetime Data Analysis 1, 193-204.

469

Pepe MS (1991). Inference for events with dependent risks in multiple

endpoint studies. JASA 86, 770-778.
(*) Pepe MS and Mori M (1993). Kaplan-Meier, marginal or conditional probability curves in summarizing competing risks failure
time data? Statistics in Medicine 12, 737-751.
Prentice RL, Kalbfleisch JD, Peterson AV, Flournoy N, Farewell VT,
and Breslow NE (1978). The analysis of failure times in the presence
of competing risks. Biometrics 34, 541-554.
Slud EV, Byar DP, and Schatzkin A (1988). Dependent competing
risks and the latent failure model. Biometrics 44, 1203-1205.
Slud E and Byar D (1988). How dependent causes of death can make
risk factors appear protective. Biometrics 44, 265-269.
Tsiatis, A. (1975). A nonidentifiability aspect of the problem of competing risks. Proceedings of the National Academy of Sciences 72,
20-22.

470

There has been a lively debate in the literature

about the best way to attack this problem. The
two sides are basically divided about which type
of model to use:
based on cause-specific hazard functions (observables)
based on latent variable models (unobservables)
The first approach focuses on what the observed survival is
due to a certain cause of failure, acknowleding that there are
other types of failures operating at the same time.
The second approach attempts to estimate what the survival
associated with a certain failure type would have been, if the
other types of failures had been removed.

471

General Case of Multiple Failure Types

In general, say we have m different types of failure (say,
causes of death), and the respective times to failure are:
T1 , T 2 , T 3 , , T m
and we observe T = min(T1, T2, . . . , Tm)
We can write the cause-specific hazard function for the j-th
failure type as:
1
j (t) = lim
P r(t T < t + t, J = j|T t)
t0 t
The overall hazard of death is the sum over the failure types:
(t) =

m
X

j=1

j (t)

where (t) = lim

t0

1
P r(t T < t + t|T t)
t

Q. Can we estimate these quantities? ...even if

the risks are dependent?
A. Yes, Prentice (1978) shows that probabilities
that can be expressed as a function of the causespecific hazards can be estimated.
472

For example, estimable quantities include:

(a) The overall survival probability3:
"

ST (t) = P (T t) = exp

= exp

Z t

(u)du

Z tX

0 j

j (u)du

(b) Conditional probability of failing from cause

j in the interval (i1 , i]
Q(i, j) = [ST (i1 )]

j (u) ST (u) du

(c) Conditional probability of surviving ith interval

i = 1

m
X

j=1

Q(i, j)

Note: previously I said you couldnt estimate ST (t), but that was when T was
the time to event of interest (possibly unobservable) and U was the possibly correlated censoring time. Here, ST (t) is the survival distribution for the minimum
of all failures, which can always be observed
3

473

Estimators:
(a) The MLE of Q(i, j) is simply
j) = dij
Q(i,
ri
i.e., the number of failures (deaths) due to cause j during
the i-th interval among the ri subjects at risk of failure
at the beginning of the interval.

(b) The MLE of i is::

i =

ri

Pm

j=1 dij

ri

= 1

(c) The MLE of ST (t) is based on i:

ST (i) =

474

i
Y

k=1

Pm

j=1 dij

ri

So what cant we estimate?

Compare the cause-specific hazard function:

j (t) = lim

t0

1
P r(t T < t + t, J = j|T t)
t

with the marginal hazard function:

j (t) = lim

t0

1
P r(t Tj < t + t|Tj t)
t

We can get estimates of the cause-specific hazard function,

since we can estimate ST (t) = P (T t) even if the failure times are dependent. (In other words, we can observe
whether each patient is still alive or not)
But unfortunately, we cant estimate the marginal hazard
function when the risks are dependent, since we cant estimate Sj (t) = P (Tj t). (we cant tell when they would
have had event Tj if they have a different event first)
This is the main tricky issue of competing risks analyses.

475

Back to original question...

What can we do if we suspect dependent risks?
Ex. Say we have two types of failures, T1 and T2, and we
think they are dependent. However, we are interested in the
first type of failure T1, and view the competing risk T2 as
censoring (like in bone marrow transplant example).
Methods of summarizing data with competing risks:
(1) Summarize the cause-specific hazard rate over time
(2) Use the Kaplan-Meier estimate anyway, ST1 (t)
(3) Report the complement to the KM, 1 ST1 (t)
(4) Use cumulative incidence curves (crude incidence
curve)
(5) Use the conditional probability function
(6) Give upper and lower bounds for the true marginal
survival function, in the absence of the competing risk

Pepe and Mori review the first 5 of these options, and recommend against option (2), but note that this is often what
people end up doing.

476

To make the example more concrete:

Say we are interested in time to MAC or death, whichever
occurs first. We define:
T = time to MAC or death
and U = censoring (assumed independent)
and the type of failure is denoted by j
j =

M if event is MAC

D if event is death from other causes

In the alternative latent variable framework, we would define

TM = time to MAC
and TD = time to Death
although we might not be able to observe TM if TD occurred
first.

477

Methods for competing risks:

(1) Summarizing the cause-specific hazard over
time
As mentioned above, this is one of the quantities that we
can estimate. Our basic estimator during time interval i is
j) = dij .
Q(i,
ri
j (t) over time, and get some insight as to
So we can plot
biological phenomona involved.
However, if you remember some of the plots I showed you of
hazards over time, they tended to be highly variable. Several
contributions have been made towards smoothing out the
inherent variability in the estimates of cause-specific hazards.
Efron (1988)
Ramlau-Hansen (1983)
Tanner and Wong (1983)
Drawback: the hazard functions alone do not give overall
effect of a covariate on survival.
Example: If the hazard functions for j (t) for two treatments cross, we cant say which treatment leads to lower
overall event rate.

478

Figure from Pepe and Mori for Leukemia Data

Kernel Estimates of Cause-Specific Hazards

479

Methods for competing risks:

(2) Applying Kaplan-Meier to cause-specific j s
Say we evaluate the MAC survival distribution by treating
all MAC cases as events
any deaths without MAC as censorings
and construct the Kaplan-Meier survival curve.
What are we estimating?

SM
(t)

"

= exp

Z t

M (u) du

where M (t) is the cause-specific hazard for MAC:

1
M (t) = lim
P r(t T < t + t, j = M |T t)
t0 t
i.e., the conditional probability that MAC occurs in a short
period of time, given that the subject is alive and MAC-free.
The interpretation of the Kaplan-Meier curve is as the
exponential of the negative cumulative causespecific hazard.
Clinicians (and others!) have difficulty understanding this
function, since it has no direct clinical interpretation.
480

Figure from Pepe and Mori for Leukemia Data

Kaplan Meier with Cause Specific Hazards

They thought this was such a bad idea, that they

did not include any plot of this!

481

Methods for competing risks:

(3) Using the complement of the Kaplan-Meier
Another function used fairly often in the competing risks
area is sometimes referred to as the pure probability
function:
1

Sj(t)

"

= 1 exp

Z t

j (u) du

(t) could be interprested as the

In our MAC example, 1SM
predictive probability of MAC by time t if the risk of death
could be removed.

If we were designing a new study for a miracle drug that

seemed so powerful that it would not only reduce MAC but
prevent all death from other causes in HIV-infected patients,

then we could use estimates 1 SM

(t) to help design our
new study.
This would be pretty optimistic, and there has been a lot of
work on the strict (and untestable) assumptions required to
interpret the KM curve in this manner.
Pepe and Mori contend that this function is irrelevant for
summarizing data from a competing risks study.

482

Figure from Pepe and Mori for Leukemia Data

Complement Kaplan-Meier Functions

483

Methods for competing risks:

(4) Using Cumulative Incidence Curves
This has also been termed the crude incidence curve, and
estimates the marginal probability of an event in the
setting where other competing risks are acknowledged to exist.
Description of Method: the method is described in
more detail in Kalbfleisch and Prentice (p.169).
Tests for covariates: Tests for comparing cumulative
incidence curves among treatment groups (or some other covariate) have been developed by Bob Gray (1988). They
are similar to logrank tests in that they are linear rank
statistics.

If we were able to follow up all subjects to time t, then the

cumulative incidence curves would relfect what proportion
of the total study population have had the particular event
(i.e., MAC) by time t.

484

Figure from Pepe and Mori for Leukemia Data

Cumulative Incidence Curves

485

Methods for competing risks:

(5) Conditional Probability Curves
This has the same flavor as the complement KM, but a more
natural interpretation. Pepe and Mori define the conditional
probability function as:
d
CP
M (t) = P (TM t|TD t)

PM (t)
=
1 PD (t)

where PM (t) =

Z t

d NM (u)
ST (u)
Y (u)

and PD (t) =

Z t

d ND (u)
ST (u)
Y (u)

In the above, ST (u) is the KM estimate of the overall macfree survival distribution, and the terms NM (u), ND (u), and
Y (u) reflect the counting process for the number of subjects with MAC, death, and at risk at time u, respectively.
In the absence of censoring, the interpretation is the proportion of patients who develop MAC among those who do not
die of other causes.
Tests: Pepe and Mori also present tests, which are sums
over time of weighted differences between the conditional
(or marginal) probabilities for two groups.
486

Figure from Pepe and Mori for Leukemia Data

Conditional Probability and Marginal Curves

487

Methods for competing risks:

(6) Bounds on Net Survival Curves
As noted previously, we cannot estimate Sj (t) = P (Tj t)
if the failure times are dependent (eg, we cant estimate the
survival function for MAC if time to MAC is correlated with
time to death without MAC).
However, we may be able to say something about the range
of Sj (t) by finding upper and lower bounds that contain
Sj (t).
Peterson (1976) obtained general bounds based on the
minimal and maximal dependence structure for (TM , TD ).
The bounds allow any possible dependence structure,
but can be very wide.
Slud and Rubenstein (1983) obtained tighter bounds
on Sj (t) by using additional information, but require the
user to specify reasonable bounds on a function . Once
is supplied, the marginal distribution Sj,(t) can be
obtained.
Klein and Moeshberger (1988) use the framework
of Clayton and Oakes for bivariate survival to obtain
tighter bounds than those of Peterson. Again, the user
has to supply bounds on a function , and once this is
given Sj, (t) can be obtained.

488

Figure from Klein and Moeschberger (1988)

Bounds on Net Survival Curves

489

One last example: Promotion of Faculty at HSPH

I was asked to analyze the schools data from 1980-1995 on
promotion of Faculty from Assistant Professor to Associate
Professor to assess whether there were differences between
males and females and among academic areas (social, laboratory, quantitative).
Problem:
Would you think that censoring (someone leaving their
tenure track position prior to getting promoted) is independent of the probability of promotion?
I considered 3 approaches for accounting for censoring:
Method I:

assumes those who departed would NOT have

been promoted

Method II:

assumes those who departed would have been

promoted at the same rate as those who stayed

Method III: assumes 50% of those who departed would not

have been promoted, and the other 50% would
have been promoted at the same rate as those
who stayed

Which Method corresponds to non-informative

(independent) censoring?
490

Results: Cumulative probabilities of promotion

Effect of Gender on Promotion

Method I:

Overall Males Females

0.631 0.719 0.451

Method II:

0.933

1.000

0.674

Method III:

0.736

0.825

0.531

Effect of Academic Area on Promotion

Method I:

Overall Quantitative Social Laboratory

0.631
0.703
0.238
0.701

Method II:

0.933

0.950

0.389

1.000

Method III:

0.736

0.803

0.287

0.801

491