Survival Analysis

Stochastic processes in event history analysis

Marta Fiocco & Hein Putter, Mathematical Institute, University

of Leiden.

Discover the world at Leiden University

Survival Analysis
Introduction

Table of Contents 1 | 66

Stochastic processes in discrete time

Processes in continuous time

Processes with continuous sample paths

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Martingales in discrete time

Martingales in discrete time 2 | 66

• Let M = {M0 , M1 , M2 , . . .} be a stochastic process in

discrete time
• The process M is a martingale if

E(Mn | M0 , M1 , . . . , Mn−1 ) = Mn−1

for each n ≥ 1
• Fair game
• It will be tacitly assumed that M0 = 0 throughout
• What does this imply for EMn ?

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Martingales in discrete time

The past 3 | 66
• The “past” can be seen as a wider amount of information as
time progresses
• Often includes some external information (covariates) in
addition to the previous values of the process itself
• In stochastic process theory the past is formulated as a
σ-algebra of events
• Informal definition: family of events that can be decided to
have happened or not by observing the past
• Denoted by Fn , formal way of representing what is known at
time n
• We will refer to Fn as the history at time n
• Family has to increase, since our past knowledge increases as
time passes
Discover the world at Leiden University
Survival Analysis
Stochastic processes in discrete time: Martingales in discrete time

Back to martingales 4 | 66
• Assume that Fn , for each n, is generated by M1 , . . . , Mn plus
possibly external information
• Technical formulation would be that the process
M = {M0 , M1 , M2 , . . .} is adapted to the history {Fn }
• This means that, for each n, the random variables M1 , . . . , Mn
are measurable with respect to the σ-algebra Fn
• Practical implication:
E(Mm | Fn ) = Mm for all m ≤ n.

• The process M = {M0 , M1 , M2 , . . .} is a martingale with

respect to the history {Fn } if
E(Mn | Fn−1 ) = Mn−1 for all n ≤ 1.
Discover the world at Leiden University
Survival Analysis
Stochastic processes in discrete time: Martingales in discrete time

More martingales 5 | 66
• So we have:
E(Mm | Fn ) = Mm for all m ≤ n.

• The process M = {M0 , M1 , M2 , . . .} is a martingale with

respect to the history {Fn } if
E(Mn | Fn−1 ) = Mn−1 for all n ≤ 1.

• This is equivalent to the more general statement

E(Mn | Fm ) = Mm for all n > m.

• The first displayed formula states that we know the past and
present of the process M, the last states that the expected
value of the process in the future equals its present value
Discover the world at Leiden University
Survival Analysis
Stochastic processes in discrete time: Martingales in discrete time

Properties of martingales 6 | 66

• As a consequence we get, using double expectations and the

assumption M0 = 0,

E(Mn ) = E(Mn | F0 ) = E(M0 ) = 0

• Since the martingale has mean zero for all n, we say that it is
a mean zero martingale
• By similar argument one may show that

Cov(Mm , Mn − Mm ) = 0 for all n > m,

that is, the martingale has uncorrelated increments

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Martingales in discrete time

Martingale differences 7 | 66

• Denote the martingale differences ∆Mn = Mn − Mn−1 ,

n = 1, 2, . . .
• These martingale differences satisfy

E(∆Mn | Fn−1 ) = 0, for all n ≤ 1

• Last equality would also hold if the process had independent

zero-mean increments
• In this sense the concept of martingale differences is a
weakening of the independent increment concept

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Variation processes

Variation processes 8 | 66
• Two processes describe the variation of a martingale
M = {M0 , M1 , M2 , . . .}
• The predictable variation process, denoted ⟨M⟩, is defined as
the sum of conditional variances of the martingale differences
n
X n o n
X
⟨M⟩n = E (Mi − Mi −1 )2 | Fi −1 = Var(∆Mi | Fi −1 ),
i =1 i =1

with ⟨M⟩0 = 0
• The optional variation process, denoted [M], is defined by
n
X n
X
[M]n = (Mi − Mi −1 )2 = (∆Mi )2 ,
i =1 i =1

with [M]0 = 0
Discover the world at Leiden University
Survival Analysis
Stochastic processes in discrete time: Variation processes

Properties of variation processes 9 | 66

• The following statements can be proved:
1. M 2 − ⟨M⟩ is a mean zero martingale
2. M 2 − [M] is a mean zero martingale

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Stopping times and transformations

Stopping times 10 | 66

• One major advantage of the martingale assumption is that one

can make certain manipulations of the process without
destroying the martingale property
• First example of this is concept of optional stopping time
• Example: the first time M passes above a certain limit
• Stopping time denoted by T and the value of the process at
time T by MT
• Definition: a time T is called an optional stopping time if the
event {T = t} is only dependent on what has been observed
up to and including time t
• First passage time is stopping time, last passage time is not

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Stopping times and transformations

Stopped martingales 11 | 66

• Stopping a fair game at an optional stopping time T preserves

the fairness of the game
• For a martingale M, define M stopped at T as

MnT = Mn∧T

Then it can be proved that M T is a martingale

• The idea of fairness of a game is strongly connected to the
idea of unbiasedness in statistics
• Stopping will commonly occur in event history analysis in the
form of censoring (stopping observation of events)
• Preserving martingale property ensures that estimates and tests
remain essentially unbiased

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Stopping times and transformations

Transformations 12 | 66
• Let X = {X0 , X1 , X2 , . . .} be some general process with a
history {Fn }, and let H = {H0 , H1 , H2 , . . .} be a
predictable process
• That is a sequence of random variables where each Hn is
measurable with respect to Fn−1 and hence is known one step
ahead of time
• The process Z defined by

Zn = H0 X0 + H1 (X1 − X0 ) + . . . + Hn (Xn − Xn−1 )

is denoted the transformation of X by H and written

Z =H•X

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Stopping times and transformations

Transformations and martingales 13 | 66

• The following calculation shows that if M is a martingale, ten

so is X = H • M

E(Zn − Zn−1 | Fn−1 ) = E(Hn (Mn − Mn−1 ) | Fn−1 )

= Hn E(Mn − Mn−1 | Fn−1 ) = 0

• So a transformation preserves the martingale property

• Moreover, since Z0 = H0 M0 = 0, the transformation is a
mean zero martingale
• Later on we will consider continuous versions of these
transformations

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: Stopping times and transformations

Transformations and variation processes 14 | 66

• The variation processes obey the following rules under

transformation

⟨H • M⟩ = H 2 • ⟨M⟩ and [H • M] = H 2 • [M]

• Formulated as sums:
n
X
⟨H • M⟩n = Hs2 ∆⟨M⟩s
s=1

n
X
[H • M]n = Hs2 ∆[M]s
s=1

Discover the world at Leiden University

Survival Analysis
Stochastic processes in discrete time: The Doob decomposition

Signal plus noise 15 | 66

• Martingales arise naturally when we try to explain the
developments in a stochastic process as a function of its
previous development
• It is possible to decompose a stochastic process into a
sequence of martingale differences (noise) and a predictable
process (signal)
• Let X = {X0 , X1 , X2 , . . .} be some general process, with
X0 = 0, with respect to a history {Fn }
• Define a process M = {M0 , M1 , M2 , . . .} by
M0 = X0 , Mn − Mn−1 = Xn − E(Xn | Fn−1 )

• It is clear that the ∆Mn = Mn − Mn−1 are martingale

differences (why?)
Discover the world at Leiden University
Survival Analysis
Stochastic processes in discrete time: The Doob decomposition

The Doob decomposition 16 | 66

• We can therefore write (Doob decomposition)
Xn = E(Xn | Fn−1 ) + ∆Mn

• The quantity E(Xn | Fn−1 ) is a function of the past only,

hence constitutes a predictable process
• The martingale differences ∆Mn are often termed innovations,
since they represent what is new and unexpected compared to
past experience
• Hence, above formula decomposes a process into what can be
predicted from the past and what is new and “surprising”
• Time-continuous generalization of the Doob decomposition,
the Doob-Meyer decompositition, is key to the counting
process approach to survival analysis
Discover the world at Leiden University
Survival Analysis
Processes in continuous time
Continuous time processes 17 | 66

• We will consider processes in continuous time, see how many

of the concepts and results of previous section carry over to
continuous time
• We will assume throughout the course that the
time-continuous stochastic processes we consider are defined
on the finite interval [0, τ ]
• Formally, we say that a stochastic process
X = {X(t); t ∈ [0, τ ]} is adapted to a history {Ft } (an
increasing family of σ-algebras) if X(t) is Ft -measurable for
each t
• This means that at time t we know the value of X(s) for all
s≤t

Discover the world at Leiden University

Survival Analysis
Processes in continuous time
Continuous time processes 18 | 66

• A realization of X is a function of t and is called a sample path

• If the sample paths of a stochastic process are right-continuous
and have left-hand limits, we say that the process is cadlag
• Unless otherwise stated, all time-continuous processes we
encounter are assumed to be cadlag
• We say that T is a stopping time if the event {T ≤ t} is
Ft -measurable for each t
• This means that at time t we know whether T ≤ t or T > t

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Martingales in continuous time

Martingales in continuous time 19 | 66

• A stochastic process M = {M(t); t ∈ [0, τ ]} is a martingale
relative to the history {Ft } if it is adapted to the history and
satisfies the martingale property:
E(M(t) | Fs ) = M(s) for all t > s

• Note similarity with martingale property in discrete time

• A heuristic way of formulating the martingale property is to
say that M is a martingale if
E(dM(t) | Ft− ) = 0

• Here dM(t) is the increment of M over the small time interval

[t, t + dt), and Ft− denotes the history until just before time
t
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Martingales in continuous time

Martingale increments 20 | 66

• As for the time-discrete case, we will tacitly assume

throughout that M(0) = 0
• Then (why?) we have EM(t) = 0 for all t, i.e., M is a mean
zero martingale
• Similarly we may show that a martingale has uncorrelated
increments:

Cov(M(t) − M(s), M(v) − M(u)) = 0

for all 0 ≤ s < t < u < v ≤ τ

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Martingales in continuous time

Variation processes in continuous time 21 | 66

• Also for time-continuous martingale M we introduce the

predictable variation process ⟨M⟩ and the optional variation
process [M]
• These are defined as the appropriate limits (in probability) of
their time-discrete counterparts
• Partition [0, t] into n subintervals each of length t/n, and
define ∆Mk = M(kt/n) − M((k − 1)t/n) as the increment
of the martingale over the kth of these subintervals
• Then
n
X
⟨M⟩(t) = lim Var(∆Mk | F(k−1)t/n )
n→∞
k=1
n
X
[M](t) = lim (∆Mk )2
n→∞
k=1
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Martingales in continuous time

Martingales and variation processes 22 | 66

• Informally, the predictable variation process ⟨M⟩ can be seen

as providing

d⟨M⟩(t) = Var(dM(t) | Ft− )

• The increment d⟨M⟩(t) of the predictable variation process

over the infinitesimal interval [t, t + dt) is the conditional
variance of the increment of the martingale
• Similarly to discrete-time martingales, we have that both
M 2 − ⟨M⟩ and M 2 − [M] are mean zero martingales

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Martingales in continuous time

Multiple martingales 23 | 66
• Extension of the concept of covariance to martingales
• We may define the predictable covariation process ⟨M1 , M2 ⟩
for a pair of martingales M1 and M2 as the limit (in
probability) of the sum of conditional covariances
Cov(∆M1k , ∆M2k | F(k−1)t/n ) and the optional covariation
process [M1 , M2 ] as the limit of the sum of the products
∆M1k ∆M2k
• Informally, we may write
d⟨M1 , M2 ⟩(t) = Cov(dM1 (t), dM2 (t) | Ft− )

• Note that ⟨M, M⟩ = ⟨M⟩ and [M, M] = [M]

• Similar to before we have that both M1 M2 − ⟨M1 , M2 ⟩ and
M1 M2 − [M1 , M2 ] are mean zero martingales
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Martingales in continuous time

Multiple martingales 24 | 66
• As a consequence of these results
Cov(M1 (t), M2 (t)) = E(M1 (t)M2 (t)) =
= E⟨M1 , M2 ⟩(t) = E[M1 , M2 ](t)

• Rules for evaluating the (co)variation processes of linear

combinations of martingales are similar to those for linear
combinations for ordinary random variables
• For example, the predictable variation process of a sum of
martingales may be written as
⟨M1 + M2 ⟩ = ⟨M1 ⟩ + ⟨M2 ⟩ + 2⟨M1 , M2 ⟩

• Similar relation holds for the optional variation processes

Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Stochastic integrals

Stochastic integrals 25 | 66
• Stochastic integrals are analogues to the transformations for
discrete-time martingales
• Let H = {H(t); t ∈ [0, τ ]} be a a predictable stochastic
process
• This means that for any time t, the value of H(t) is known
just before t
• (Formal definition of predictability is a bit intricate, we will
not go into this.)
• Sufficient conditions for H to be predictable are
• H is adapted to the history {Ft }
• The sample paths of H are left-continuous
• Predictability may sound uninteresting and technical, but
turns out to be very important!
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Stochastic integrals

Stochastic integrals 26 | 66

• We can now introduce the stochastic integral

Z t
I(t) = H(s)dM(s)
0

• Although valid in a much broader context than counting

processes, in this course most martingales will arise from
counting processes
• Stochastic integration is the exact analogue of transformation
of discrete-time martingales

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Stochastic integrals

Stochastic integrals 27 | 66
• The stochastic integral can be defined as a limit of such a
transformation in the following sense:
n
X
I(t) = lim Hk ∆Mk
n→∞
k=1

with partition of [0, t] into n subintervals of length t/n,

Hk = H((k − 1)t/n) and
∆Mk = M(kt/n) − M((k − 1)t/n)
• As for a transformation, the major interesting fact about a
stochastic integral is that I(t) is a mean zero martingale wrt
{Ft }
• Hence, the martingale property is preserved under stochastic
integration
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Stochastic integrals
Stochastic integrals and variation
28 | 66
processes
• In analogy with time-discrete case, the following rules hold for
evaluating the variation processes of a stochastic integral
Z  Z Z  Z
HdM = H 2 d⟨M⟩, HdM = H 2 d[M]

• The following rules hold for the covariation processes

Z Z  Z
H1 dM1 , H2 dM2 = H1 H2 d⟨M1 , M2 ⟩
Z Z  Z
H1 dM1 , H2 dM2 = H1 H2 d[M1 , M2 ]

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: The Doob-Meyer decomposition

Doob-Meyer decomposition 29 | 66
• In the discrete-time context we saw how a stochastic process
can be decomposed into a predictable process and a sequence
of martingale differences
• A similar result holds in continuous time and is known as the
Doob-Meyer decomposition
• An adapted process X = {X(t); t ∈ [0, τ ]} is called a
submartingale if it satisfies
E(X(t) | Fs ) ≥ X(s), for all t > s

• Thus a submartingale is a process that tends to increase as

time passes
• Any nondecreasing process, like a counting process, is a
submartingale
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: The Doob-Meyer decomposition

Doob-Meyer decomposition 30 | 66
• The Doob-Meyer decomposition states that any submartingale
X can be decomposed uniquely as
X = X ∗ + M,
where X ∗ is a nondecreasing predictable process, often
denoted as the compensator of X, and M is a means zero
martingale
• Heuristically we have that
dX ∗ (t) = E(dX(t) | Ft− )
and
dM(t) = dX(t) − E(dX(t) | Ft− )

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: The Doob-Meyer decomposition

Doob-Meyer decomposition 31 | 66

• Like the discrete-time Doob decomposition, the Doob-Meyer

decomposition tells us what can be predicted from the past,
dX ∗ (t), and what is the innovation (surprising element)
dM(t()
• In discrete-time case we noted that if M is a martingale, so is
M 2 − ⟨M⟩
• Here M 2 is a submartingale (by Jensen’s inequality) and ⟨M⟩
a nondecreasing predictable process (by construction)
• This shows that the predictable variation process ⟨M⟩ is the
compensator if M 2 , and this offers an alternative definition of
the predictable variation process

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: The Poisson process

Poisson process 32 | 66
• A homogeneous Poisson process describes the distribution of
events that occur independently of each other
• Imagine a basic rate of occurrence, denoted λ, such that the
probability of an event occurring in [t, t + dt) is λdt
• Well known properties:
• The time between events is exponentially distributed with
probability density λe −λt
• The expected value and the variance of the number of events in a
time interval of length h are both equal to λh
• The number of events in a time interval of length h is Poisson
distributed; the probability of exactly k events occurring is
(λh)k e −λh /k!
• The process has independent increments, that, is, the number of
events in nonoverlapping intervals are independent

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: The Poisson process

Poisson and counting processes 33 | 66

• We let N(t) be the number of events in [0, t] and introduce
the process
M(t) = N(t) − λt
obtained by centering the Poisson process
• Again Ft : information about all events in [0, t]
• Due to independent increments of a Poisson process, we have
for t > s
E{M(t) − M(s) | Fs } = E{M(t) − M(s)} =
= E{N(t) − N(s)} − λ(t − s) = 0,
which yields
E{M(t) | Fs } = M(s)

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: The Poisson process

Poisson processes and martingales 34 | 66

• E{M(t) | Fs } = M(s) means that M is a martingale
• It follows that λt is the compensator of the Poisson process N
• By a similar argument
E{M 2 (t) − λt | Fs } = M 2 (s) − λs,
so the process M 2 (t) − λt is a martingale
• Thus λt is also the compensator of M 2 (t)
• It follows that M has a predictable variation process
⟨M⟩(t) = λt

• We will see that similar relations are in general valid for

counting processes
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Counting processes

Counting processes 35 | 66
• Counting process N = {N(t); t ∈ [0, τ ]} is a
right-continuous process with jumps of size 1 at event times
and constant in between
• We assume it is adapted to the history {Ft }
• (Technical way of saying that the history is generated by N and
possibly additional external information)
• Intensity process λ(t) of N (wrt Ft ) is heuristically defined by
λ(t)dt = P(dN(t) = 1 | Ft− ) = E(dN(t) | Ft− )
• More precisely, counting process is nondecreasing, hence a
submartingale. So by the Doob-Meyer decomposition there
exists a unique predictable process Λ(t), called the cumulative
intensity process, such that M(t) = N(t) − Λ(t) is a mean
zero martingale
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Counting processes

Counting processes and intensities 36 | 66

• Throughout the course we will consider the case where the

cumulative intensity process is absolutely continuous
• Then there exists a predictable process λ(t) such that
Λ(t) = 0t λ(s)ds, and this gives a formal definition of λ(t)
R

• Further for the absolute continuous case, we have that

Z t
M(t) = N(t) − λ(s)ds
0

is a mean zero martingale

• This is a key relation that will be used over and over again

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Counting processes
Predictable and optional variation
37 | 66
processes
• The predictable and optional variation processes of M are
defined as
n
X
⟨M⟩(t) = lim Var(∆Mk | F(k−1)t/n )
n→∞
k=1
n
X
[M](t) = lim (∆Mk )2 ,
n→∞
k=1

with a partition
defining ∆Mk =[0, t] into n−subintervals
M(kt/n) each of
M((k − 1)t/n) as length
the t/n,
increment of the martingale over the kth of these subintervals
• For the latter we obtain [M](t) = N(t)

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Counting processes
Predictable and optional variation
38 | 66
processes
• For the former, heuristically we have that
d⟨M⟩(t) = Var(dM(t) | Ft− )
= Var(dN(t) − λ(t)dt | Ft− )
= Var(dN(t) | Ft− ) ≈ λ(t)dt

• This motivates the relation

Z t
⟨M⟩(t) = λ(s)ds
0
• Another key result that will be used many times later on
• Note that Λ(t) is the compensator of both N(t) and of M 2 (t)
• Similar to the homogeneous Poisson process
• Counting process has the same “local behavior” as a Poisson
process
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Counting processes

Two counting processes 39 | 66

• Often we have several counting processes (for different

individuals, different groups)
• Consider a pair N1 and N2 of counting processes, with
corresponding martingales M1 and M2
• Since the counting processes do not jump simultaneously, the
same applies for M1 and M2
• From this it follows that for all t:

⟨M1 , M2 ⟩(t) = 0, and [M1 , M2 ](t) = 0

• We say that the martingales are orthogonal

• Equivalent to the fact that the product M1 M2 is a martingale

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Stochastic integrals for counting process
martingales
Stochastic integrals 40 | 66
• In the context of counting processes the stochastic integral
Z t
I(t) = H(s)dM(s)
0
can be understood as
Z t Z t
I(t) = H(s)dN(s) − H(s)λ(s)ds
0 0

• The last integral is an ordinary (Riemann) integral

• The first is a sum of the values of H at every jump time of N:
Z t X
H(s)dN(s) = H(Tj ),
0 Tj ≤t

where T1 < T2 < . . . are the ordered jump times of N

Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Stochastic integrals for counting process
martingales
Predictable and optional variation
41 | 66
processes of stochastic integrals
• We get the following expressions for the predictable and
optional variation processes of a stochastic integral of a
counting process martingale:
Z  Z t
HdM (t) = H 2 (s)λ(s)ds,
0
Z  Z t
HdM (t) = H 2 (s)dN(s)
0

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Stochastic integrals for counting process
martingales
Multiple counting processes 42 | 66

• Consider multiple counting processes N1 , N2 , . . . , Nk with no

simultaneous jumps and intensity processes λ1 , λ2 , . . . , λk
(wrt the same history)
• The corresponding martingales are orthogonal and we obtain
* k Z + k Z t
X X
Hj dMj (t) = Hj2 (s)λj (s)ds
j =1 j =1 0

and  
Xk Z k Z t
X
 Hj dMj  (t) = Hj2 (s)dNj (s)
j =1 j =1 0

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: The innovation theorem

The innovation theorem 43 | 66

• The intensity process of a counting process N relative to a
history {Ft } is given informally by
λF (t)dt = E(dN(t) | Ft− )

• Dependence on F in notation made explicit to point out that

the intensity process depends on the history (so if the history
is changed the intensity process may change as well)
• Let {Nt } be the history (or filtration) generated by the
counting process (the self-exiting filtration)
• Previously we assumed that a counting process is adapted to
the history
• {Nt } is the smallest history we may consider for the previous
results to be true
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: The innovation theorem

The innovation theorem 44 | 66

• Usually we consider histories not only generated by N, but
also by other counting processes, censoring processes,
covariates, etc.
• Consider {Ft } and {Gt } nested, i.e., {Ft } ⊆ {Gt }
• This means at any time t, all information contained in Ft is
also contained in Gt
• Using double expectations we then have
E(dN(t) | Ft− ) = E{E(dN(t) | Gt− ) | Ft− }

• It follows that the intensity processes of N wrt the two

histories are related by (innovation theorem)
λF (t) = E(λG (t) | Ft− )
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: The innovation theorem
Intensity processes and histories 45 | 66
• It is important to note that the innovation theorem applies only
to histories that are nested
• Also, in order for N to be adapted to the histories, both need
to contain {Nt }
• Thus, the innovation theorem holds provided that
Nt ⊆ Ft ⊆ Gt
• When considering more than one history, it is important to
realize that intensity processes of counting processes are
defined relative to a history, so to use notation that points out
their dependence on the history
• Usually it is clear from the context how the history is defined;
then dependence on the history will not be expressed in the
notation
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Independent censoring

Independent censoring 46 | 66
• In previous lecture we gave an informal discussion of the
concept of independent censoring
• The main point was that independent censoring preserves the
form of the intensity processes of the counting processes
• We will now discuss this more formally
• Three different models are of interest
(i) A model for the hypothetical situation without censoring, where
all occurrences of the event of interest are observed
(ii) A joint model for the situation where all occurrences of the event
of interest as well as the censoring processes are observed
(iii) A model for the situation with censoring, i.e., for the occurrences
of the event actually observed
• Parameters of interest defined according to (i), concept of
independent censoring by (ii), actual inference by (iii)
Discover the world at Leiden University
Survival Analysis
Processes in continuous time: Independent censoring

Complete observation counting processes47 | 66

• Consider the (hypothetical) situation where all occurrences of

the event of interest are observed
• Let N1c (t), . . . , Nnc (t) be the complete observation counting
processes
• Denote by {Ftc } the complete observation history; it is
generated by all the counting processes Nic (t) and covariate
processes that run in parallel
• We assume that the {Ftc }-intensity processes of the counting
processes take the form
c
λF c
i (t) = Yi (t)αi (t), i = 1, . . . , n

• Here Yic (t) is left-continuous, {Ftc }-adapted, taking the

value 1 if individual i may experience an event at time t, 0
Discover the world at Leiden University
otherwise
Survival Analysis
Processes in continuous time: Independent censoring

Observed counting processes 48 | 66

• The αi (t) are our key model parameters, object of inference

later on
• Important to note that they may depend on covariates as well
as previous occurrence of the event (through the past Ft− c )

• Study of αi (t) is complicated by incomplete observation of the

counting processes
• To handle this, for each i = 1, . . . , n we introduce a
left-continuous binary censoring process Yio (t) taking value 1
if individual i is under observation “just before” time t
• For the common special case of right censoring,
Yio (t) = I{t ≤ Ci } for a right-censoring time Ci

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Observed counting processes 49 | 66

• The observed counting processes are then given by

Z t
Ni (t) = Yio (u)dNic (u), i = 1, . . . , n
0

• The censoring creates extra randomness, causing Ni (t) not to

be adapted to the history {Ftc }
• Then we can not define the intensity processes for Ni (t)
relative to the complete history
• To handle this problem, we have to consider the joint model for
the Nic (t) and the censoring processes Yio (t)

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Joint model 50 | 66

• We consider the larger history {Gt } generated by the

complete history {Ftc } as well as the censoring processes
• This history corresponds to the hypothetical situation were all
occurrences of the event of interest are observed, and in
addition we observe the censoring processes
• Consider the {Gt }-intensity processes λGi (t) of the
completely observed counting processes N c (t)
• If the censoring processes carry information on the likelihood
of occurrence of event, the λGi (t) will differ from the
{Ftc }-intensity processes

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Independent censoring 51 | 66

• We assume that this is not the case, so that the intensity

processes relative to the two histories are the same:
c
λGi (t) = λF
i (t), i = 1, . . . , n

• This is the independent censoring assumption

• Consider right-censoring, where n individuals are followed
until observation stops at censoring or death
• For this situation we have Yio (t) = I{t ≤ Ci } for (potential)
right-censoring times Ci

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Different types of censoring 52 | 66

• Censoring at fixed times, possibly different between

individuals, Ci = ci for given constants ci (type I censoring)
• Censoring of all individuals at the time T when a specified
number of occurrences have taken place, so Ci = T for all i
(type II censoring)
• Censoring of all individuals when the event has not occurred in
a certain time interval
• Censoring at random times Ci that may differ between
individuals and that are independent of the completely
observed counting processes Nic (t) (random censoring)

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Different types of censoring 53 | 66

• For the first three, Ci are stopping times relative to {Ftc } ⇒

no additional randomness and no need to enlarge the history
• As a result the histories {Gt } and {Ftc } are the same, and
the independent censoring assumption is automatically fulfilled
• For the last censoring scheme additional randomness is
introduced by the censoring
• But as the potential censoring times Ci are assumed
independent of the completely observed counting processes
Nic (t), the independent censoring assumption holds for this
censoring scheme as well

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Observed counting process 54 | 66

• Counting processes Ni (t) are observed, and these are adapted

to {Ft }
• However, we do not necessarily observe the censoring
processes, i.e., we do not observe censoring after death
• What is observed for each i is the left-continuous process

Yi (t) = Yic (t)Yio (t)

• This takes the value 1 if individual i is at risk “just before”

time t, 0 otherwise

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Observed counting process 55 | 66

• We can adopt similar argument as in the innovation theorem to
find the {Ft }-intensity processes λF
i (t) of the observed
counting processes
λF
i (t)dt = E(dNi (t) | Ft− )
= E(Yio (t)dNic (t) | Ft− )
= E{E(Yio (t)dNic (t) | Gt− ) | Ft− }
= E{Yio (t)E(dNic (t) | Gt− ) | Ft− }
= E{Yio (t)λGi (t)dt | Ft− }
= E{Yio (t)Yic (t)αi (t)dt | Ft− }
= Yi (t) E{αi (t) | Ft− }dt

Discover the world at Leiden University

Survival Analysis
Processes in continuous time: Independent censoring

Covariates 56 | 66
• In order to progress further, an assumption on the observation
of covariates is needed
• Typically, {Ftc } will be generated by the fully observed
counting processes as well as by covariate processes running
in parallel, and the αi (t) may depend on these covariates
• For statistical inference on αi (t), we must assume that these
covariates are available at time t, so they should be
Ft -predictable
• The intensity processes of the observed counting processes
take the form
λFi (t) = Yi (t)αi (t)

• We see that the form of the intensity processes is preserved

under independent right censoring
Discover the world at Leiden University
Survival Analysis
Processes with continuous sample paths
Wiener processes and Gaussian
57 | 66
martingales
• So far we have considered processes in continuous time, but
with discrete state space
• Now processes where both time and state space are continuous
• Examples are Wiener processes and Gaussian martingales
• These have applications as models for underlying, unobserved
processes and as limiting processes of stochastic integrals of
counting process martingales

Discover the world at Leiden University

Survival Analysis
Processes with continuous sample paths: The Wiener process and Gaussian
martingales
Wiener process 58 | 66
• Also called Brownian motion, model of completely random
noise
• In fact, the so-called white noise is a kind of derivative of the
Wiener process
• Let W (t) denote the value of the Wiener process at time t,
and consider (s, t]
• Then the increment W (t) − W (s) over (s, t] is normally
distributed with
E{W (t) − W (s)} = 0 and Var{W (t) − W (s)} = t − s

• Further, the Wiener process has continuous sample paths, and

the increment of the Wiener process over a time interval is
independent of its increments over all nonoverlapping intervals
Discover the world at Leiden University
Survival Analysis
Processes with continuous sample paths: The Wiener process and Gaussian
martingales
Extensions and use of Wiener process 59 | 66

Modifications
• Possible to add drift
• Possible to use increments of Wiener process as building
blocks (Ornstein-Uhlenbeck)

Use
• Wiener processes can be used as models for underlying
processes
• Disease is often the result of a long deterioration of a biological
process
• Time-transformations of the Wiener process arise as limits in a
number of applications (in particular for stochastic integrals of
counting process martingales, basis for asymptotic theory of
Discover the and
estimators worldtest
at statistics)
Leiden University
Survival Analysis
Processes with continuous sample paths: The Wiener process and Gaussian
martingales
Gaussian martingales 60 | 66

• Let V (t) be a strictly increasing continuous function with

V (0) = 0, and consider U(t) = W (V (t))
• The process U inherits the following properties from the
Wiener process:
• The sample paths are continuous
• The increments over nonoverlapping intervals are independent
• The increment of (s, t] is normally distributed with mean zero
and variance V (t) − V (s)
• From these properties one may show that U is a mean zero
martingale with predictable variation process ⟨U⟩(t) = V (t)
• It is common to refer to U as a Gaussian martingale

Discover the world at Leiden University

Survival Analysis
Processes with continuous sample paths: Asymptotic theory for martingales

Intuitive discussion 61 | 66

• Martingales can be considered as processes of noise

• Often in statistics errors are approximately normally
distributed
• This is also true for martingales, in fact there are central limit
theorems for martingales, closely analogous to those known for
sums of independent random variables
• If we have a sequence of counting processes, where the
number of jumps increases, the properly normalized associated
martingales (or stochastic integrals wrt these martingales) will
converge to a limiting martingale with continuous sample path
• The limiting martingale is closely connected to a Wiener
process, or Brownian motion

Discover the world at Leiden University

Survival Analysis
Processes with continuous sample paths: Asymptotic theory for martingales

Intuitive discussion 62 | 66
• Two things to be taken care of to ensure that a sequence of
martingales converges to a Gaussian martingale
i The predictable variation processes of the martingales should
converge to a deterministic function
ii The sizes of the jumps of the martingales should go to zero
• The first assumption implies a stabilization of the sample
space of processes, the second one is a requirement on the
sample paths of the processes

Discover the world at Leiden University

Survival Analysis
Processes with continuous sample paths: Asymptotic theory for martingales

Central limit theorem for martingales 63 | 66

• Several versions of the CLT for martingales exist

• Very general and elegant theorem due to Rebolledo (1980)
• Present version taken from the book by Andersen, Borgan, Gill
& Keiding (1993)
• Since the relevant statistics may also be functionals of the
rocesses, many different probability distribution may arise from
the theory

Discover the world at Leiden University

Survival Analysis
Processes with continuous sample paths: Asymptotic theory for martingales

Central limit theorem for martingales 64 | 66

Theorem
Let M̃ (n) be a sequence of mean zero martingales defined on
[0, τ ], and let M̃ε(n) be the martingale containing all the jumps of
P
M̃ (n) larger than a given ε > 0. Let → denote convergence in
probability. Under the conditions:
P
(i) ⟨M̃ (n) ⟩(t) → V (t) for all t ∈ [0, τ ] as n → ∞, where V is
a strictly increasing continuous function with V (0) = 0
P
(ii) ⟨M̃ε(n) ⟩(t) → 0 for all t ∈ [0, τ ] and all ε > 0 as n → ∞
Then, as n → ∞, the sequence of martingales M̃ (n) converges in
distribution to the mean zero Gaussian martingale U given by
U(t) = W (V (t))

Discover the world at Leiden University

Survival Analysis
Processes with continuous sample paths: Asymptotic theory for martingales

Corollaries of the martingale CLT 65 | 66

• M̃ (n) (t) converges in distribution to a normally distributed
random variable with mean zero and variance V (t)
• Interested in the limit of stochastic integrals of the form
R t (n)
0 H (s)dM (n) (s), with H (n) (t) a predictable process and
t
M (n) (t) = N (n) (t) − 0 λ(n) (s)ds a counting process
R

martingale
• More generally, we will consider sums of stochastic integrals
k Z t
X (n) (n)
Hj (s)dMj (s),
j =1 0
(n)
where for each j , n, Hj (t) is a predictable process and
(n) (n) (n)
Mj t) = Nj (t) − 0t λj (s)ds is a counting process
R

martingale
Discover the world at Leiden University
Survival Analysis
Processes with continuous sample paths: Asymptotic theory for martingales

Martingale CLT corollaries 66 | 66

• For this situation, the two conditions of the martingale CLT
take the form
Pk Rt (n) (n) P
(i) j =1 0
(Hj (s))2 λj (s)ds → V (t) for all t ∈ [0, τ ],
Pk R t (n) (n) (n) P
(ii) j =1 0
(Hj (s))2 I{|Hj (s)| > ε}λj (s)ds → 0 for all
t ∈ [0, τ ]
• If we write V (t) = 0t v(s)ds, then apart from regularity
R

conditions, sufficient conditions for the above are that, taking

k fixed:
Pk (n) 2 (n) P
(i) j =1 (Hj (s)) λj (s) → v(s) for all s ∈ [0, τ ], as n → ∞
(n) P
(ii) Hj (s) → 0 for all j = 1, . . . , k and s ∈ [0, τ ], as n → ∞

Discover the world at Leiden University