——————————————————————–

Survival Analysis
Regression models

Marta Fiocco & Hein Putter, Mathematical Institute, University

of Leiden.

Discover the world at Leiden University

Survival Analysis
Regression models

Table of Contents 1 | 40

Regression models

Discover the world at Leiden University

Survival Analysis
Regression models

2 | 40
• Counting processes are registering the occurrences of a
specific event of interest
• When the population is grouped into a few subpopulations
according to the value of one or two categorical covariates
data may be summarized into one or a few counting processes
• Usually more covariates (some are numeric) are of interested
and then regression models need to be used

Discover the world at Leiden University

Survival Analysis
Regression models

3 | 40
For regression models need to consider one counting
process for each individual under study

• n: number of individuals
• N1 , . . . , Nn : counting processes
• Ni (t): counts the number of occurrences of the event of
interest for individual i in the interval [0, t]
• In general the event of interest may occur more than once for
each individual i
• Recurrent event: the process Ni (t) may take integer values
larger than 1
• For survival data Ni (t) = 1 if by time t the event has been
observed to occur for individual i , otherwise Ni (t) = 0

Discover the world at Leiden University

Survival Analysis
Regression models

4 | 40
How to model the effect of the covariates on the intensity
process of a counting process?

• xi (t) = (xi 1 (t), ..., xip (t))T : covariate vector at time t for
individual i
• The components are fixed or time-varying; for this lecture only
time-fixed covariates are discussed (next lecture
time-dependent)
• Intensity process of Ni

λi (t) = Yi (t)α(t|xi )

• Yi (t) is 1 if individual i is at risk for the event of interest just

before time t and 0 otherwise
Discover the world at Leiden University
Survival Analysis
Regression models

• α(t|xi ): intensity or hazard rate of individual i

5 | 40
• It is defined conditional on the values of the fixed covariates as
well as on the observed paths of the time-varying
• To obtain a regression model, we need to specify how α(t|xi )
depends on xi (t)
• Use of the relative risk regression models and the additive
regression model requires the collection of information on the
covariate values for all individuals in a study

Discover the world at Leiden University

Survival Analysis
Regression models

6 | 40
• Assumption in the model λi (t) = Yi (t)α(t|xi ): censoring and
truncation are independent as defined in the first lecture (i.e.
for right censoring)

P(t ≤ T
e i < t + dt, Di = 1|T̃i ≥ t, past) =

= P(t≤ Ti < t + dt|Ti ≥ t)

• Di = 1: if the actual survival time is observed (T
e i = Ti )
• Di = 0: if only right-censoring time is observed (T e i < Ti )
• The intensity α(t|xi ) is the same as one would have in the
situation without censoring and truncation

Discover the world at Leiden University

Survival Analysis
Regression models

7 | 40
• The independent censoring assumption implies that censoring
may depend on information in the past but not on future events
• Independent censoring implies that censoring may depend on
covariates that are included in the model but not on covariates
that are not included

Discover the world at Leiden University

Survival Analysis
Regression models

Type of covariates 8 | 40
• External (or exogenous)
• Time-fixed (ex stage of disease, age at the start of treatment)
• Time-dependent
• Internal (or endogenous) (see Chapters 8 and 9)
• Examples: values of biomarkers measured on the individuals
during follow-up

Discover the world at Leiden University

Survival Analysis
Regression models: Relative risk regression models

• xi (t): covariates for individual i

9 | 40
• α(t|xi ): hazard rate

α(t|xi ) = α0 (t)r (β, xi (t)

• r (β, xi (t)): relative risk function

• β = (β1 , . . . , βp )T : vector of regression coefficients which
describe the effect of the covariates
• α0 (t): baseline hazard rate (left unspecified)
• r (β, xi (t) is the parametric part, α0 (t) is the nonparametric
part

Discover the world at Leiden University

Survival Analysis
Regression models: Relative risk regression models

10 | 40
• Assume r (β, 0) = 1; α0 (t): hazard rate of an individual with
all covariates identically equal to zero
• Exponential relative risk function r (β, xi (t)) = exp{β T xi (t)}
(Cox regression model (1972))
• Ratio of hazard rates bewteen 2 individuals
α(t|x2 ) r (β, x2 (t))
=
α(t|x1 ) r (β, x1 (t))

• If all covariates are fixed, this ratio is constant over time, and
the model is called a proportional hazards regression model

Discover the world at Leiden University

Survival Analysis
Regression models: Relative risk regression models

11 | 40
• Let all components of x1 (t) and x2 (t) be equal except for the
j th
• Assume x2j (t) = x1j (t) + 1
• The hazard ratio becomes
α(t|x2 )
= exp {β T (x2 (t) − x1 (t))} = e βj
α(t|x1 )

• Interpretation: increase in the j th covariate is to multiply the

hazard rate by e βj when all covariates are kept the same but
the the j th -ones
• e βj : common terminology relative risk of the j th -covariate

Discover the world at Leiden University

Survival Analysis
Regression models: Partial likelihood and inference for regression
coefficients

• Semiparametric model 12 | 40
α(t|xi ) = α0 (t)r (β, (xi (t))

• Cannot use the classical likelihood methods due to the

semiparamtric nature of the model
• Use the partial likelihood approach
• Rewrite the intensity process of Ni (t) as
λi (t|xi ) = Yi (t)α0 (t)r (β, (xi (t))

• Intensity process of the aggregated counting process

N (t) = nl=1 Nl (t) is as follows
P

n
X n
X
λ (t) = λl (t) = Yl (t)α0 (t)r (β, (xl (t))
l =1 l =1
Discover the world at Leiden University
Survival Analysis
Regression models: Partial likelihood and inference for regression
coefficients

• Factorize the intensity process of Ni (t) by using the

13 | 40
conditional probability

λi (t) Yi (t)r (β, (xi (t))

π(i |t) = = Pn
λ (t) l =1 Yl (t)r (β, (xl (t))

• π(i |t): conditional probability of observing an event for

individual i at time t, given the past and that an event is
observed at that specific time t
• Factorization
λi (t) = λ (t)π(i |t)

Discover the world at Leiden University

Survival Analysis
Regression models: Partial likelihood and inference for regression
coefficients

• T1 < T2 < . . .: distinct observed event times (no ties)

14 | 40
• ij : index of the individual who experiences an event at Tj
• Partial likelihood for β is obtained by multiplying the
conditional probabilities π(i |t) over all observed event time
Y Y Yij (Tj )r (β, xij (Tj ))
L(β) = π(ij |Tj ) = Pn
Tj Tj l =1 Yl (Tj )r (β, xl (Tj ))

• Rj = {l |Yl (Tj ) = 1}: risk set at Tj

Y r (β, xij (Tj ))

L(β) = Pn
Tj l ∈Rj r (β, xl (Tj ))

Discover the world at Leiden University

Survival Analysis
Regression models: Partial likelihood and inference for regression
coefficients

• β̂: maximum partial likelihood estimator of β

15 | 40
• β̂ is in large samples approximately multivariate normally
distributed around the true value of β (see Section 4.1.5)
• Covariance matrix is estimated by I(β̂)−1 where I(β) is the
observed information matrix
∂2
I(β) = − log L(β)
∂βh ∂βj

Discover the world at Leiden University

Survival Analysis
Regression models: Partial likelihood and inference for regression
coefficients
Test of hypothesis 16 | 40
• H0 : β = β0 , where β0 is known
• apply the likelihood-based tests with the vector of the score
∂
function U(β) = ∂β log Lβ
• Likelihood ratio test statistic:
χ2LR = 2{log L(β̂) − log L(β0 )}

• Score test statistic:

χ2SC = U(β0 )T F (β0 )−1 U(β0 )

• Wald test statistic:

(β̂ − β0 )T F (β̂)(β̂ − β0 )

Discover the world at Leiden University

Survival Analysis
Regression models: Partial likelihood and inference for regression
coefficients
17 | 40
Under the null hypothesis the three test statistics are
asymptotically equivalent, and they are all approximately
chi-squared distributed with p degrees of freedom

All three tests can be generalized to the situation where

H0 : β1 = · · · = βr = 0 (r < p) p: number of regression
coefficients

χ2LR = 2{log L(β̂) − log L(β ∗ )}

• Under the null hypothesis the test statistics is approximately

chi-squared distributed with r degrees of freedom
Discover the world at Leiden University
Survival Analysis
Regression models: Example Mortality in three Norwegian counties

18 | 40
• Fit a Cox regression model with the 2 covariates to study the
effect of smoking habits on survival

• hazard rate ratios: e βj ;

• 95% confidence interval: [β̂j ± 1.96se(β̂j )]
• e 1.086 = 2.962; 95% CI: (2.190; 4.006)

Discover the world at Leiden University

Survival Analysis
Regression models: Estimation of cumulative hazards and survival
probabilities

• Cumulative baseline hazard

19 | 40
Z t
A0 (t) = α0 (u)d(u)
0

• Intensity process of the aggregated counting process

N (t) = nl=1 Nl (t)
P

n
X n
X
λ (t) = λl (t) = Yl (t)α0 (t)r (β, xl (t))
l =1 l =1

n
!
X
= Yl (t)r (β, xl (t)) α0 (t)
l =1

Discover the world at Leiden University

Survival Analysis
Regression models: Estimation of cumulative hazards and survival
probabilities

• If β is known estimate A0 (t) by the Nelson-Aalen estimator

20 | 40
Z t
dN (u)
A
b 0 (t; β) = Pn
0 l =1 Yl (u)r (β, xl (u))

• Replace β with βb to get an estimator for the cumulative

baseline hazard (called Breslow estimator)
Z t
dN (u)
A
b 0 (t) = Pn
0 l =1 Yl (u)r (β,
b xl (u))

X 1
= P
Tj ≤t l ∈Rj r (β,
b xl (Tj ))

Discover the world at Leiden University

Survival Analysis
Regression models: Estimation of cumulative hazards and survival
probabilities

• Fix all covariates; the (adjusted) cumulative hazard

21 | 40
corresponding to an individual with covariate vector x0
Z t
A(t|x0 ) = α(u|x0 )du = r (β, x0 )A0 (t)
0

• A(t|x0 ) can be estimated as

A(t|x
b 0 ) = r (β, x0 )A0 (t)
b b

• How to estimate the cumulative hazard when some of the

covariates are time-dependent?
• It makes no sense to estimate the cumulative hazard
corresponding to a fixed value of the vector of covariates
Discover the world at Leiden University
Survival Analysis
Regression models: Estimation of cumulative hazards and survival
probabilities

• Estimate the cumulative hazard over the interval [0,22

t]
| 40
corresponding to a given covariate path x0 (s); 0 < s ≤ t
• Cumulative hazard corresponding to such a path:
Z t
A(t|x0 ) = r (β, x0 (u))α0 (u)du
0

• Estimate of A(t|x0 )
Z t
A(t|x
b 0) = r (β,
b x0 (u))dA
b 0 (u)
0

X r (β,
b x0 (Tj ))
= P
Tj ≤t l ∈R| r (β,
b xl (Tj ))

Discover the world at Leiden University

Survival Analysis
Regression models: Estimation of cumulative hazards and survival
probabilities
23 to| 40
• Estimated the adjusted survival function corresponding
given values of time-fixed covariates or given covariate paths
of external time-dependent covariates
• In this context the survival function
S(t) = {1 − dA(u)} as product integral becomes
Ru≤t
S(t|x0) = {1 − dA(u|x0 )}
R
b
u≤t

• Insert A(t|x
b 0 ) to get the Kaplan-Meier-type estimator
Y
S(t|x0) = {1 − dA(u|x0 )} = {1 − ∆A(T
b j |x0 )}
R
b b
u≤t Tj ≤t

• where ∆A(T
b j ) = 1/Y (Tj ) is the increment of the
Nelson-Aalen estimator
Discover the world at Leiden University
Survival Analysis
Regression models: Estimation of cumulative hazards and survival
probabilities

• This estimator is asymptotically normally distributed around

24 | 40
its true value (Section 4.1.6)

Discover the world at Leiden University

Survival Analysis
Regression models: Example 4.4. Mortality in three Norwegian counties

• Effect of sex and smoking habit on mortality due to 25 | 40

any cause
in three Norwegian counties

• Use the Kaplan-Meier-type estimator to estimate the survival

function for given values of the covariates
Y
S(t|x0) = {1 − dA(u|x0 )} = {1 − ∆A(T
b j |x0 )}
R
b b
u≤t Tj ≤t

Discover the world at Leiden University

Survival Analysis
Regression models: Example 4.4. Mortality in three Norwegian counties

26 | 40

Discover the world at Leiden University

Survival Analysis
Regression models: Martingale residual processes and model check

27 | 40
• Use the martingale residuals to check the fit of the relative
risk regression model

α(t|xi ) = α0 (t)r (β, (xi (t))

• The martingale residuals are the difference between the

observed and expected numbers of events for each individual
over the full study time interval and they are therefore not
well suited to detect time-dependent deviations from the model

Discover the world at Leiden University

Survival Analysis
Regression models: Martingale residual processes and model check

• Focus on the martingale residual processes, which compare

28 | 40
the observed and expected numbers of events for each
individual as a function of time t
• To define the martingale residual processes, need to introduce
the cumulative intensity processes
Z t Z t
Λi (t) = λi (u) = Yi (u)r (β, (xi (u))α0 (u)du
0 0

• Martingales for i = 1, 2, . . . , n
Z t
Mi (t) = Ni (t) − λ(s)ds = Ni (t) − Λi (t)
0

Discover the world at Leiden University

Survival Analysis
Regression models: Martingale residual processes and model check

• Insert the maximum partial likelihood estimator βb for β and

29 | 40
the increment dAb 0 (u) of the Breslow estimator for α0 (u)du to
get the estimated cumulative intensity processes
Z t
Λ
b i (t) = Yi (u)r (β,
b (xi (u))dA
b 0 (u) =
0

X Yi (Tj )r (β,
b xi (Tj ))
= P
Tj ≤t
b
l ∈R| r (β, xl (Tj ))

Discover the world at Leiden University

Survival Analysis
Regression models: Martingale residual processes and model check

• Martingale residual processes

30 | 40

ci (t) = Ni (t) − Λ
M b i (t)

• τ : upper time limit for the study; evaluating the martingale

residual processes at t = τ we arrive at the martingale
residuals
ci (τ ) = Ni (τ ) − Λ
M b i (τ )

Discover the world at Leiden University

Survival Analysis
Regression models: Martingale residual processes and model check

• The individual martingale residual processes Mi (t) may be

31 | 40
useful for recurrent event data, where each individual may
experience a number of events
• For survival data each of the martingale residual processes
contains too little information to be of much use on its own
• Aggregate them over groups of individuals to obtain plots and
goodness-of-fit tests
• Grouping of the individuals, typically based on the values of
one or two covariates, and denote the groups by J = 1, . . . , G
• The grouping of the individuals may depend on time (an
individual may move from one group to another as time passes)
• The information used for grouping at time t is available just
before that time

Discover the world at Leiden University

Survival Analysis
Regression models: Martingale residual processes and model check

• J(u): set of all individuals who belong to group J at time u

32 | 40
• Group J martingale residual process:
P
Z t X X i ∈Rj ∩J(Tj ) r (β,
b xi (Tj ))
M
cJ (t) = dM
ci (u) = NJ (t)− P
0
i ∈J(u) Tj ≤t l ∈Rj r (β,
b xl (Tj ))

• NJ (t) = 0t i ∈J(u) dNi (u): observed number of events in

R P

group J in [0, t]
P
i ∈Rj ∩J(Tj )
r (β
b,xi (Tj ))
•
P
Tj ≤t P : estimate of the expected number
l ∈Rj
r (β
b,xl (Tj ))
of events in the group when the relative risk regression model
α(t|xi ) = α0 (t)r (β, (xi (t)) holds true

Discover the world at Leiden University

Survival Analysis
Regression models: Use plots of the grouped martingale residual processes
to check the fit of model

• Data: Uranium miners (Colorado Plateau uranium miners’

33 | 40
cohort Example 1.6)
• Estimate the effect of radon exposure and smoking on the risk
of lung cancer death
• Covariates: x(t) = (xi 1 (t), xi 2 (t))T
• xi 1 (t): cumulative radon exposure measured in working level
months (WLM) up to two years prior to age t
• xi 2 (t): cumulative smoking in number of packs smoked up to
two years prior to t
• The lung cancer mortality rate for miner i is assumed to follow
the excess relative risk model

α(t|xi ) = α0 (t){1 + β1 xi 1 (t)}{1 + β2 xi 2 (t)}

Discover the world at Leiden University

Survival Analysis
Regression models: Use plots of the grouped martingale residual processes
to check the fit of model

• Both radon and smoking are seen to have a significant effect

34 | 40
on the risk of lung cancer death; βb1 = 0.40(0.12);
βb2 = 0.19(0.05);
• The effect of each 100 WLMs increase in the radon exposure is
to increase the relative lung cancer risk by 40%, compared to
an unexposed individual, while the relative risk increases by
19% for each additional 1000 packs of cigarettes smoked
• Grouped martingale residual processes
cJ (t) = NJ (t) − estimated expected ) can be used to
(M
check the fit of the excess relative risk model
(α(t|xi ) = α0 (t){1 + β1 xi 1 (t)}{1 + β2 xi 2 (t)})

Discover the world at Leiden University

Survival Analysis
Regression models: Use plots of the grouped martingale residual processes
to check the fit of model
35 | 40

Discover the world at Leiden University

Survival Analysis
Regression models: Use plots of the grouped martingale residual processes
to check the fit of model

• Grouped martingale residual processes

36 | 40
cJ (t) = NJ (t) − estimated expected
M
• More lung cancer deaths than expected occurred in the high
exposure group (group III) below the age of 60 years, fewer
cases than expected occur above this age
• The grouped martingale residual plot indicates that there is an
interaction between cumulative radon exposure and age

Discover the world at Leiden University

Survival Analysis
Regression models: Use plots of the grouped martingale residual processes
to check the fit of model
37 | 40
• Use chi-squared test to investigate whether the interaction
between age and radon exposure suggested by the visual
inspection is statistically significant

Discover the world at Leiden University

Survival Analysis
Regression models: Use plots of the grouped martingale residual processes
to check the fit of model

• The chi-squared goodness-of-fit statistic corresponding to the

38 | 40
observed and expected numbers of Table takes the value 10.2
(p-value = 0.038) (2(3-1) = 4 degrees of freedom)
• Results suggests there is an interaction with time
• The excess relative risk model
α(t|xi ) = α0 (t){1 + β1 xi 1 (t)}{1 + β2 xi 2 (t)} is maybe not
a good model to describe the phenomena

Discover the world at Leiden University

Survival Analysis
Regression models: Stratified models

• Assumption in the relative risk regression model

39 | 40

α(t|xi ) = α0 (t)r (β, (xi (t))

• Common baseline hazard for all individuals

• Sometimes it is not realistic
• Stratified version of the model: the study population is
grouped into k strata and for an individual i in stratum s, the
hazard is
α(t|xi ) = αs0 (t)r (β, (xi (t))

• The effects of the covariates are assumed to be the same across

strata, while the baseline hazard may vary between strata
Discover the world at Leiden University
Survival Analysis
Regression models: Stratified models

• The grouping into strata may depend on time

40 | 40
• The information used to group the individuals into strata only
depend on past information and not on future events
• Partial likelihood
k Y
Y r (β, xij (Tsj ))
L(β) = Pn
s=1 Tsj l ∈Rsj r (β, xl (Tsj ))

• Ts1 < Ts2 < . . .: times when events are observed in stratum
s
• Rsj : risk set in this stratum at timeTsj
• The statistical properties of the βb as before, same for the
martingale residual processes

Discover the world at Leiden University