Você está na página 1de 46

University of Oxford

The SEIR Epidemic Model


by
Kim-Huat Lim
Keble College
A transfer of status thesis submitted in partial fullment of the degree of
Doctor of Philosophy in Statistics.
Department of Statistics. 1 South Parks Road,
Oxford OX1 3TG
June 2004
Contents
1 Introduction 2
1.1 Historical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Outline of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Epidemic Modelling 4
2.1 The role of analytical epidemiology . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Types of Epidemic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 The SIR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4 Recent developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4.1 Markovian models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4.2 Non-Markovian models . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 An Introduction to Steins method 9
3.1 Steins method for normal approximations . . . . . . . . . . . . . . . . . . . 9
3.2 The Local Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Steins method in general . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4 The weak law of large numbers . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.5 The weak law in measure space . . . . . . . . . . . . . . . . . . . . . . . . . 14
4 Model Extension: SEIR 16
4.1 Model description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3 Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.5 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5 Research Direction 41
5.1 Model review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Generalisation of the model . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.3 Parameters estimation and model testing . . . . . . . . . . . . . . . . . . . . 42
Bibliography 44
1
Chapter 1
Introduction
1.1 Historical background
The study of epidemics can be dated back to the mid-17th century, when Graunt and Petty
carried out basic statistical analysis on the London Bill Of Mortality. In one of the rst
examples of deducing the source of a disease, Snow (1855) found that contaminated water
was spreading illness, and a few years later Budd discovered typhoid was spread in the
same way. At the same time, Farr was hoping to nd laws directly governing the spread of
epidemics and in particular the smallpox that was aicting Britain at the time.
For many years, statistical predictions continued to be made concerning epidemics, but
were generally ignored due to their inaccuracy. The signicant advance within the science of
epidemiology was by Ross (1908) in the early 20th century when he introduced the idea of
probabilities as the basis of epidemic models. These probabilities were derived purely based
on available data, and this gave the rst accurate prediction of the spread of an epidemic.
This was the breakthrough that established epidemiology as a eld of study.
In the 1920-30s, Kermack and McKendrick developed a deterministic epidemic model
using rates of infection and recovery which changed during the epidemic, as well as the
threshold theory whereby a disease will not spread epidemically until the number of infected
people has reached a certain critical value.
Since then, modelling of epidemics has become ever more complex and accurate. Accord-
ing to the traits of the contagious disease, statisticians have created a number of epidemic
models to help them draw inferences on these diseases. Among these models, the one that
has attracted vast attention is a compartmental model known as the Susceptible-Infected-
Removed (SIR) model; the main reason for this is its simplicity and generality, as well as its
expansibility into more sophisticated variants. We will look into this model in greater detail
in the next chapter.
1.2 Outline of Contents
The report is organized in the following order. Chapter 2 gives a review on epidemic mod-
elling. A key focus here is to investigate the quality of the models as well as the underlying
assumptions as they form the bases for our work improving the accuracy of epidemic mod-
2
els and the removal of unrealistic assumptions. Chapter 3 gives a short review of Steins
Method, which is the fundamental tool used. We describe some technical background for
measure-valued random elements; the main results of Steins method for mean-eld limits are
taken from Reinert [18]. In Chapter 4, we introduce the SEIR model. This is an extension
of the SIR model to include an exposed state (to model the incubation period of diseases).
Here, we also derive the quality of the limiting deterministic approximation to an epidemic
process. In Chapter 5, we end the report by discussing directions for future research.
3
Chapter 2
Epidemic Modelling
In this chapter, we will review the modelling of epidemics. We begin by looking at the
motivation behind epidemic modelling and describe some of these models. We end the
chapter by looking at some recent development in this eld.
2.1 The role of analytical epidemiology
Epidemiological analysis
1
aims to give insights into epidemic quantities. These include key
parameters (e.g., R
0
- the basic reproduction number, incubation period, etc.) of diseases,
how they are transmitted, factors aecting the rate of transmission, and eectiveness of
disease control or risk reduction policies (e.g., 30 month ban of cattle in the food chain
to counter transmissible spongiform encephalopathy (TSE) and home quarantine of people
during the 2003 Severe Acute Respiratory Syndrome (SARS) epidemic in Asia).
The mathematical aspect of modelling epidemics attempts to represent the knowledge
and assumptions about disease transmission as simplied models which only capture the key
aspects of the disease transmission, rather than every detail of the disease. Associated with
the models, statistical methods are employed to estimate key parameters from epidemiologi-
cal data, test hypotheses and make robust predictions (with condence bounds) on potential
courses of epidemics.
Such knowledge is important during disease outbreaks as insights into the transmission
dynamics oer good means to control the disease. Moreover, if there are alternative control
measures, modelling oers a mode of comparison. This is particularly important in the
course of an outbreak when critical decisions have to be reached.
1
This section is based on a lecture series presented by Professor Christl Donnelly at the University of
Oxford, 2004.
4
2.2 Types of Epidemic Models
Epidemic modelling can be broadly classied into two main categories deterministic mod-
els and stochastic models.
Deterministic models typically assume that the spread of a disease is governed by a set
of dierential equations. This simplies the mathematical portion of the modelling work.
Thus, deterministic models can be fairly complex, yet still possible to analyse since solutions
can often be found via numerical methods. Examples of deterministic models can be found
in Anderson and May [1].
In contrast, stochastic models, as the name suggest, attempt to model epidemics by
allowing factors aecting spread of the diseases (e.g., whether an infected individual will
transmit the disease when in contact with a healthy individual) to be stochastic. This
immediately restricts the complexity of mathematically viable stochastic models. However,
stochastic models are, in general, more realistic since the spread of diseases is stochastic in
nature. A possible approach to the analysis of complex stochastic models is to show that they
converge to some deterministic models as the population increases. Other approaches include
simulations and moment closure techniques. Simulations oer insights into the dynamics of
the epidemic without getting overly involved with the technical details. The moment closure
technique (see Isham [14] and Keeling [15]) approximates higher order moments in terms
of known lower order ones, so that the dynamics of a complex stochastic system can be
captured by a few deterministic equations, allowing simple mathematical analysis of what
was previously a highly complex problem.
The texts by Bailey [3], Anderson and May [1] give good outline of the development of
mathematical theories for the spread of epidemics using deterministic models, though Bailey
[3] also covered some grounds on stochastic models. The monograph by Daley and Gani [11]
focuses on the techniques used in the mathematical modelling of epidemics and the recent
text Andersson and Britton [2] gives a good survey of stochastic models.
In this report, we will focus on stochastic models, rather than deterministic ones. An
example of a stochastic model is given in the next section.
2.3 The SIR Model
In 1949, Bartlett [10] introduced the General Stochastic Epidemic (GSE). This is a birth-
death-process in a closed population where the temporal evolution of one individual de-
pendsuniformly on those of the others. In a closed population with K individuals, at time
t = 0 a proportion a of the individuals is infected by a certain disease (and infective); the
remaining bK = (1 a)K individuals are susceptible to that disease. Infectious individuals
will get removed after some time, e.g., by lifelong immunity or death, and are then no longer
aected by that disease. Thus we have an SIR model.
In Bartlett [10], the general stochastic epidemic is dened as a Markov process
(Z
K
(t))
t0
, where Z
K
(t) = (X
K
(t), Y
K
(t)), taking values in the set
S = (r, s) : r (1 a)K, r +s K, r, s N,
5
with Z
K
(0) = ((1 a)K, aK) and transition probabilities (, > 0)
_
_
_
P[Z
K
(t + t) = (r 1, s + 1)[Z
K
(t) = (r, s)] = rst +o(t)
P[Z
K
(t + t) = (r, s 1)[Z
K
(t) = (r, s)] = st +o(t)
P[Z
K
(t + t) = (r, s)[Z
K
(t) = (r, s)] = 1 (r +)st +o(t)
(2.1)
for (r, s), (r, s1), (r1, s+1) S. Here, X
K
(t) is interpreted as the number of susceptibles
and Y
K
(t) is the number of infectives at time t respectively. An infective individual is assumed
to be infectious during its infected period, and no multiple infections are allowed.
It is often convenient to re-parameterize (2.1) using the so-called the relative removal
rate :=

. One may then choose = 1 or =


1
K
; this could be viewed as just a change in
the time scale.
From this SIR model, we can investigate the basic reproduction number, R
0
. This is the
mean number of susceptibles that a typical infective spread the disease to during the initial
stages of the epidemic. This parameter R
0
plays a crucial role in determining whether an
epidemic will take place. The Limit Threshold Theorem states that a major outbreak in a
large population is possible if and only if R
0
> 1.
2.4 Recent developments
Here, we focus on contributions made by Barbour ([6] and [7]), Wang([25] & [26]) and Reinert
([19] & [21]) to the SIR model as these lead us to the model that we will focus on in Chapter
4.
2.4.1 Markovian models
Let X(t), Y (t) denote the number of susceptibles and infectives present at time t during
an epidemic in a closed population. The evolution of the epidemic can be modelled by a a
continuous time Markov process on a two-dimensional non-negative lattice, with transition
probabilities given by
(X, Y ) (X 1, Y + 1) at rate
1
N
XY,
(X, Y ) (X, Y 1) at rate Y,
(2.2)
with initial condition is X(0), Y (0) = (N, I), and where is the relative removal rate and
is independent of N. Here, the population size K = N + I; but it is more convenient to
adopt the representation (N, I).
While distributions associated with this model are intractable in detail, approximations
based on large population sizes, N have been successfully carried out. For example, Daniels
[12] and Nagaev & Startsev [17] have derived approximations of the distribution of X(),
the number of susceptibles that escape the disease in an epidemic. Their approaches depend
on the structure of the embedded jump chain in the epidemic model, and are thus not suitable
for investigating time related quantities such as the duration of epidemics.
To counter this, Barbour [6] developed a dierent approach which incorporates a time-
dependent structure. Here, the asymptotic behaviour of the proportions of susceptibles and
6
infectives,
_
X(t)
N
,
Y (t)
N
_
(2.3)
is described in terms of its deterministic limit () and its uctuations around the limit.
Barbour showed that the rate of convergence of the quantity

N
_
1
N
(X
N
(t), Y
N
(t)) (t)
_
to a Gaussian diusion is of O(N

1
2
), using a martingale argument. However, this result is
only valid for nite time intervals [0, T] for (), whereas a stochastic approximation to the
duration of the epidemic requires innite time.
Barbour furthered his work in [7] which removes the constraints of nite time interval.
With this, Barbour could estimate quantities such as infection times and the duration of an
epidemic.
To present his work formally, let X

N
() =
__
X
N
(), Y
N
()
__
N1
denote a sequence of
epidemic processes as in (2.2). The deterministic approximation to the process x

N
() =
N
1
X

N
() is then the solution to the equations
_
x = xy
y = (x )y
with x(0), y(0) = 1, N
1
I. (2.4)
In particular, the deterministic approximation
N
to N
1
X
N
() is the solution of
log = 1 +
I
N
(0 < < 1) (2.5)
and the deterministic approximation to the time which elapses while a N
1
X
N
(t) b is,
for < a < b 1,
J(a, b) =
_
b
a
_
u
_
1 +
I
N
+ log u u
_ _
1
du. (2.6)
Let T
N
denotes the rst time when there are no more infectives in a population of size N
with an initial proportion of infectives, I/N = h, a constant independent of N. It follows
from (2.4) that
N
= , a constant for all N. We have the following theorem:
Theorem 2.4.1 If I(N) = Nh, where h is independent of N, then, as N ,
( ) log N k
converges in distribution to W, where
k = lim
m
_
log m + ( )J
_

_
1 +
1
m( )
_
, 1
_
+ log
_
1

__
,
and W is a type I extremal random variable with P[W w] = exp(e
w
).
It should be noted that Barbours works imposed a Markovian structure, an assumption
that is necessary for his arguments. However, this restricts the applicability of his work to
epidemics as there are many diseases which clearly violate the Markovian assumption.
7
2.4.2 Non-Markovian models
Having seen that a Markovian structure is both restrictive and unrealistic in general, Wang
([25] & [26]) and Reinert ([19] & [21]) extended the SIR model. Before describing their work,
we rst introduce an elegant construction of the standard SIR epidemic model due to Sellke
[22].
We keep track of the total infection pressure generated by the infectives. Each individual
i is associated with a critical level of exposure to infection and as soon as the infection
pressure reaches this level, the susceptible individual becomes infected. We call this level
the threshold of individual i and denote it by l
i
. Associated with each initially susceptible
individual i is the time it will spend in the infected state if it does succumb to the disease.
We denote this time by r
i
.
First, Wang ([25] & [26]) relaxed the Markovian conditions by allowing the distribution
of the time spent in the infected state, r
i
, to be non-exponential; but the threshold, l
i
is still
assumed to be exponentially distributed.
In [19], Reinert used Steins Method to model Bartletts GSE without imposing a Marko-
vian or a martingale structure, thereby extending the work of Wang ([25] & [26]) on a semi-
Markovian epidemic. More signicantly, her method gives higher denition into the whole
epidemic process; quantities such as the number of susceptibles and infectives at any time
t, as well as the number of infectives present at time t = b who were infected before time
t = a, can be easily obtained via the deterministic approximation. Reinert [21] made further
improvements to the model by giving the rst known bound on the quality of mean-eld ap-
proximations in epidemics. At present, her formulation of epidemic models can be considered
as the most general (Daley & Gani [11]).
In her work, Reinert [21] described the epidemic via an empirical measure and the deter-
ministic approximation translates to a law of large numbers for such a measure. As Steins
method was the main tool used, we will present a short introduction in the next chapter.
8
Chapter 3
An Introduction to Steins method
Steins method was rst introduced by Stein [23] for proving normal approximations. This
method avoids characteristic functions, but instead relies on a characterizing equation for the
normal distribution. The distribution of any random variable would then be approximately
normal if it satises the characterizing equation approximately. Indeed, the deviation from
satisfying the characterizing equation exactly turns out to provide a measure on the distance
to the normal distribution.
In this chapter, we give an overview of Steins method and how it can be used to derive
the weak law of large number. The weak law is then extended to measure spaces. Most of
these works are presented in Reinert [21].
3.1 Steins method for normal approximations
To start with, we cite the original example, described in Stein [23] (and described in greater
details in Stein [24]), on normal approximations. The bounds were improved in Baldi et al.
[5] and can be sketched as follows.
1. A random variable Z has standard normal distribution, i.e., L(Z) = ^(0, 1), if and
only if for all smooth f,
Ef

(Z) Zf(Z) = 0.
2. Let L(Z) = ^(0, 1). For any smooth h, there is a function f = f
h
solving
h(x) Eh(Z) = f

(x) xf(x) (Stein equation)


with
|f|
_

2
|h Eh(Z)|,
|f

| (sup h inf h) and


|f
(2)
| 2|h

|.
Here, [[f[[ = sup
xR
[f(x)[ denotes the supremum norm.
9
3. It follows that for any random variable W and any smooth h, substituting W for x in
the Stein equation, followed by taking expectations on both sides yields
Eh(W) Eh(Z) = Ef

(W) EWf(W). (3.1)


Now if His a convergence-determining class for weak convergence such that, for all h H,
we have that f
h
H, then taking absolute values in Equation (3.1) gives
sup
hH
[Eh(W) Eh(Z)[ sup
fH
[Ef

(W) EWf(W)[.
If the RHS, for some quantity W = W
n
, tends to zero as n , then we obtain a central
limit theorem. In particular, we can choose H as the space of all continuous, bounded
functions with piecewise continuous, bounded rst derivatives.
3.2 The Local Approach
To see why this approach might be useful at all, consider a very classical example. Let
X, X
1
, . . . , X
n
i.i.d. random variables with EX = 0 and VarX =
1
n
. Then
W =
n

i=1
X
i
has mean zero and variance 1. Put
W
i
= W X
i
=

j=i
X
j
, (3.2)
so that W
i
is independent of X
i
. To evaluate the right-hand side of Equation (3.1) we rst
calculate
EWf(W) =
n

i=1
EX
i
f(W)
=
n

i=1
EX
i
f(W
i
) +
n

i=1
EX
2
i
f

(W
i
) +R
=
1
n
n

i=1
Ef

(W
i
) +R,
where we used Taylor expansion, and for the Taylor remainder term R we have
[R[ |f

|
n

i=1
E[X
i
[
3
.
So
Ef

(W) EWf(W) =
1
n
n

i=1
Ef

(W) f

(W
i
) +R.
Applying Taylor expansion again we obtain the following result (see, e.g. Stein [24]).
10
Theorem 3.2.1 Let L(Z) = ^(0, 1). For any smooth h
[Eh(W) Eh(Z)[ |h

|
_
2

n
+
n

i=1
E[X
3
i
[
_
.
Note that the bound on the RHS does not involve any asymptotic statement but is valid for
any number n. Thus, even if the sum to innity does not exist, a bound on the distance is
still be obtainable.
This approach can be generalised for locally dependent variables. In fact, Steins method has
proven to be particularly useful for proving results involving sums of dependent observations.
Consider, the case that X, X
1
, . . . , X
n
are random variables with
EX = 0 and VarX =
1
n
,
and for each X
i
there is a set S
i
such that X
i
is independent of (X
j
, j , S
i
). (A special
case would be m-dependent random variables.) For simplicity assume that [S
i
[ = for some
, and that X
i

C

n
for some constant C that does not depend on i. As before, let
W =
n

i=1
X
i
.
If is small, then the summands are approximately independent and so a normal approxi-
mation should hold. Stating this formally, we have the follow theorem.
Theorem 3.2.2 Let X, X
1
, . . . , X
n
be random variables with EX = 0 and VarX =
1
n
, such
that for each X
i
there is a set S
i
such that X
i
is independent of (X
j
, j , S
i
). Assume
that [S
i
[ = for some and X
i

C

n
for some constant C that is independent of i. Let
L(Z) = ^(0, 1). For all continuous, bounded functions h with piecewise continuous, bounded
rst derivatives,
[Eh(W) Eh(Z)[
|h

|
10
2
C
3

n
+ (sup h inf h)
n

i=1

jS
i
,j=i
E[X
i
X
j
[.
Note that this theorem gives an explicit bound in terms of neighbourhood size and number
of observations, as well as correlations.
A more detailed survey on Steins method for normal approximations is in Reinert [20].
3.3 Steins method in general
Steins method in general can briey be described as follows. Find a good characterization
(that is, an operator /) of the target distribution that is of the type
L(X) = E/f(X) = 0, for all smooth functions f,
11
where X is a random element, L(X) denotes the law of X and / is an operator associated
with the distribution . Then assume X to have distribution , and consider the Stein
equation
h(x) Eh(X) = /f(x), x R. (3.3)
For every smooth h, nd a corresponding solution f of this equation. Then, for any random
element W, we obtain
Eh(W) Eh(X) = E/f(W).
Hence, to estimate the proximity of W and X, it sucies to estimate E/f(W) for all possible
solutions f of (3.3). Moreover, bounding E/f(W) for all smooth f automatically yields a
bound on the distance of L(W) to in a smooth metric, regardless of asymptotics.
For being the standard normal distribution, the corresponding operator is
/f(x) = f

(x) xf(x).
Of course the operator could also be dened as a second-order operator, namely /f(x) =
f

(x) xf

(x), the Ornstein-Uhlenbeck generator.


Barbour [8] suggests employing the generator of a Markov process as operator / in
Equation (3.3), since this provides a way to look for solutions of (3.3). This is what in
the following will be called the generator method. Suppose we can nd a Markov process
(X(t))
t0
with generator / and unique stationary distribution , such that L(X(t))
w

as t ) (
w
denotes weak convergence). Then, if a random variable X has distribution ,
E/f(X) = 0
for all f T(/), where T(/) is the domain of /.
A method for solving equation (3.3) is provided by Proposition 1.5 of Ethier and Kurtz
[13], p. 9 (for the argument, see Barbour [8]). Let (T
t
)
t0
be the transition semigroup of the
Markov process (X(t))
t0
. Then
T
t
h h = /
__
t
0
T
u
hdu
_
.
As (T
t
)
t0
is a strongly continuous contraction semigroup, / is closed (Ethier and Kurtz
[13], Corollary 1.6). We can thus formally take limits:
h(x) Eh(X) = /
__

0
T
u
hdu
_
.
Thus f =
_

0
T
u
hdu would be a solution of (3.3), if this expression exists and if f T(/).
This will in general be the case only for a certain class of functions h. However, the latter
conditions can usually be checked.
Steins method has proved to be an extremely versatile tool for establishing distributional
approximations. Some examples of such approximations are the Poisson, Uniform, Binomial
and Gamma distribution (see [20] for details).
12
3.4 The weak law of large numbers
Here we will be interested in applying Steins method to point mass in measure space, as
developed in Reinert [18], resulting in weak laws of large numbers. Point mass can be seen
as an extreme case of the normal distribution with zero variance. Hence we put
/f(x) = xf

(x), x R.
Note that / is the generator of the deterministic Markov process (Y (t))
t0
given by
P[Y (t) = xe
t
[ Y (0) = x] = 1, x R.
The corresponding transition semigroup is
T
t
h(x) = h(xe
t
),
and the unique stationary distribution is
0
.
According to the general equation (3.3), the Stein equation in this context is
h(x) h(0) = xf

(x), x R. (3.4)
Let C
2
b
(R) be the space of all bounded, twice continuously dierentiable real-valued functions
on R with bounded rst and second derivatives, and let D
2
b
(R) be the space of all twice
continuously dierentiable functions f : R R with bounded rst and second derivatives.
Using the semigroup approach, we have the following proposition:
Proposition 3.4.1 For any h C
2
b
(R), there is a function f = (h) D
2
b
(R) that solves the
Stein equation (3.4) for h. Furthermore, for the derivatives, |f

| |h

|, and |f

| |h

|.
With these ingredients, we can derive weak laws of large numbers.
Theorem 3.4.2 Let (X
i
)
iN
be a family of random variables on R, dened on the same
probability space, with nite variances. Put
Y
n
=
1
n
n

i=1
(X
i
EX
i
).
Then, for all h C
2
b
(R)
[Eh(Y
n
) h(0)[ |h

|Var
_
1
n
n

i=1
X
i
_
.
As C
2
b
(R) is convergence-determining for weak convergence of the laws of real-valued random
variables, a weak law of large numbers follows from Theorem 3.4.2 provided
Var
_
n

i=1
X
i
_
0 (n ).
The method of proving Theorem 3.4.2 can be generalised to local dependence. Suppose
(X
i
)
iI
is a family of random variables with nite rst and second moments. In addition,
assume that for each i 1, there is a set (neighbourhood) B
i
1 such that X
i
is independent
of (X
j
, j / B
i
). Then, we have the following proposition.
13
Proposition 3.4.3 For any h C
2
b
(R),
[Eh(W) h()[ [[h

[[

iI

jB
i
E[(X
i
EX
i
)(X
j
EX
j
)[.
This proposition can be generalised to allow for weak dependence between X
i
and (X
j
, j /
B
i
).
3.5 The weak law in measure space
In the next chapter, we will describe the epidemic via an empirical measure. As this is a
random element taking values in the space of substochastic measures, we need some technical
details on convergence in such a space (see, e.g., Reinert [18]). Let E be a locally compact
Hausdor space with a countable basis (for instance, E = R
3
), let c = B(E) be the Borel-
-eld of E, and let M
b
(E) the space of all bounded Radon measures on E, equipped with
the vague topology. Let C
c
(E) be the space of real-valued continuous functions on E with
support contained in a compact set. Abbreviate the integral
, ) =
_
E
d,
Convergence in the vague topology is dened as
Denition 3.5.1 (Vague convergence)

n
v
for all f C
c
(E) :
n
, f) , f) (n ).
For probability measures, this diers from weak convergence in that the class of test function
is not C
b
(E), the class of bounded continuous functions on E. For example, if E = R and

n
=
n
, then
n
v
0, but (
n
)
n
does not converge weakly. Here, 0 denotes the measure that
assigns measure 0 to any set. We use the notation
w
to mean weak convergence.
For M
b
(E), we dene [[ = sup
AE
[(A)[. Let
M
1
(E) = M
b
(E) : positive, [[ 1
be the space of all positive Radon measures with total mass smaller or equal to 1 (the space
of substochastic measures). As E has a countable basis, M
1
(E) is Polish with respect to
the vague topology. Moreover M
1
(E) is vaguely compact with a countable basis. Thus the
considerations about vague convergence are valid for both E = R
3
and E = M
1
(R
3
).
The nal ingredient needed for the weak law in measure spaces is a convergence-determining
class of functions (see Reinert [18]):
T :=
_
F C
b
(M
1
(R
3
)) : F has the form F() = f(,
1
), . . . , ,
m
))
for an m N, f C

b
(R
m
) with |f

| 1, |f

| 1, |f

| 1,
and for
i
C

b
(R
3
) with |
i
| 1, |

i
| 1, i = 1, . . . , m
_
. (3.5)
14
Here the superscript

indicates the total derivative, and the norm | | indicates the sum of
the supremum norms of the components. For example, if f C

b
(R
m
), the space of innitely
often dierentiable continuous functions from R
m
to R with bounded derivatives, then
[[f

[[ =
m

i=1
[[f
(j)
[[,
where f
(j)
is the partial derivative of f in direction x
j
.
From Reinert [18], the corresponding generator for the weak law of large numbers for
random measures with target measure

, for some M
1
(R
3
), is, for F T of the form
(3.5)
/F() =
m

i=1
f
(i)
(,
j
), j = 1, . . . , m) ,
i
). (3.6)
Moreover, the Stein equation has smooth solutions; we have
Proposition 3.5.2 For any function H T, there is a function (H) T that solves the
Stein equation with the operator (3.6) for H. If
H() = h(,
1
), . . . , ,
m
)),
then
(H)() = f(,
1
), . . . , ,
m
))
for a function f C

b
(R
m
) with |f
(k)
| |h
(k)
| for all k N.
Steins method then immediately yields the following theorem.
Theorem 3.5.3 Let
n
be a random element with values in M
1
(R
3
), let M
1
(R
3
), and
let / be as in (3.6). Then
sup[EH(
n
) H()[ : H T sup[E/F(
n
)[ : F T.
Many examples on how to apply this theorem are given in Reinert [18], including the
GSE mentioned in Chapter 1.
15
Chapter 4
Model Extension: SEIR
While the SIR model is able to capture most of the features of epidemic processes, its validity
is in doubt when applied to diseases where the incubation period is relatively long.
In this chapter, we will extend the SIR model to include anExposed state. In this
extended model, an initially susceptible individual who succumbs to the disease is consid-
ered as being exposed to the disease. In this state, the exposed individual is incubating
the disease and is not capable of spreading the disease. After the incubation period, the
individual moves on to the infectious state, and will subsequently be removed after some
time and is then no longer aected by the disease. Thus, we have an SEIR model instead.
4.1 Model description
From the view point of an individual, the epidemic process can be described as follows. Let
(l
i
, e
i
, r
i
)
iN
, ( e
i
)
iN
and ( r
i
)
iN
be three families of positive i.i.d random variables. At the
point when the epidemic is rst observed (at time t = 0), an initially exposed individual i
stays in the exposed state for a period of length e
i
before moving on to the infected state.
An initially infective individual i stays infected (and infectious) for a period of length r
i
;
then it is removed. Note that e
i
(correspondingly, r
i
) need not have the same distribution as
e
i
(correspondingly, r
i
), reecting the possibility that an exposed (correspondingly, infected)
individual has already been exposed (correspondingly, infected) for a certain period at the
point when we begin to observe the epidemic. An initially susceptible individual i, once
exposed, incubates the disease for a period of e
i
before being infected. Once infected, the
individual stays infected for a period of length r
i
, until it is removed. Furthermore, an initially
susceptible individual i accumulates exposure to the infection with a rate that depends on the
evolution of the epidemic; if the total exposure reaches l
i
, the individual becomes exposed.
The possible dependence between l
i
, e
i
and r
i
for each xed i reects the fact that both
the resistance to infection and the duration of the infection may, for a xed individual,
depend on its physical constitution.
More precisely, an initially susceptible individual i gets exposed as soon as a certain func-
tional, depending on the course of the epidemic, exceeds the individuals level of resistance,
l
i
; denote its exposed time by A
K
i
. If I
K
(t) denotes the proportion of infective individuals
16
Initially Susceptible

Size = cK
Threshold = l
i

Distribution =
Exposed

Duration = e
i

Distribution =
Infected

Duration = r
i

Distribution =

Removed

Initially Exposed

Size = aK
Duration =
i

Distribution =



Initially Infected

Size = bK
Duration =
i

Distribution =






























Figure 4.1: A Diagrammatic Representation of the SEIR Epidemic Model
present in the population at time t R
+
, then A
K
i
is given by
A
K
i
= inf
_
t R
+
:
_
(0,t]
(s, I
K
)ds = l
i
_
, (4.1)
for a certain function , acting on time as one variable and on a function (which will be
specied later) as the second variable.
17
Since, for epidemics, the total length of the exposed period and infected period of an
individual is usually very small compared to its life length, we will neglect births and removals
that are not caused by the disease, as well as any age-dependence of the infectivity or the
susceptibility. Thus, the model studied is a closed epidemic. Furthermore, the population is
idealized to be homogeneously mixing.
The quantity we will approximate is not only the vector of the proportion of susceptibles,
exposed, infected and removed, but rather the whole average path behaviour of the epidemic
process, given by the empirical measure in R
3
,

K
=
1
K
aK

i=1

(0, e
i
,r
i
)
+
1
K
bK

i=1

(0,0, r
i
)
+
cK

i=1

(A
K
i
,A
K
i
+e
i
,A
K
i
+e
i
+r
i
)
,
where a, b and c are the initial proportion of exposed, infectives and susceptibles respectively.
(In general,
x
shall denote the Dirac measure at the point x).
As not necessarily (and hopefully not) every individual in the population gets infected
by the disease, the total mass of
K
can be less than 1. Thus
K
is a substochastic measure
on [0, )
3
. We could think of a point process with values in the positive quadrant [0, )
[0, ) [0, ), each point marking exposed time, infection time and removal time for one
individual. For a Borel set B, the empirical measure
K
then returns for B the number of
points of this process that are in B.
From this empirical measure, the proportion of susceptible, exposed, infected and re-
moved at time t can be expressed as:
S
K
(t) :=
K
((t, ] [0, ) [0, )) ,
E
K
(t) :=
K
([0, t] (t, ) [0, )) ,
I
K
(t) :=
K
([0, t] [0, t] (t, )) and
R
K
(t) :=
K
([0, t] [0, t] [0, t]) respectively.
Moreover, general quantities such as the number of infectives at time t = b who were
exposed before time t = a and removed only after time t = c can be easily derived via the
empirical measure
K
([0, a] [0, b] [c, ). If a control measure was introduced at time c,
one could read this o the proportion of individuals that were infected before time c and
still contribute to the overall infectiousness in the population from this empirical meaasure.
4.2 Assumptions
For the family of positive i. i. d. random variables (l
i
, e
i
, r
i
)
iN
, let , and be the common
distribution function of the (l
i
)
iN
, (e
i
)
iN
and (r
i
)
iN
respectively. Let the families of positive
i. i. d. random variable ( e
i
)
iN
and ( r
i
)
iN
have common distribution functions

and

respec-
tively, and laws
e
and
r
respectively. Assume that (l
i
, e
i
, r
i
)
iN
, ( e
i
)
iN
and ( r
i
)
iN
are mu-
tually independent. Let D
+
= x : [0, ) [1, 1] right continuous with left-hand limits,
and let : R
+
D
+
R
+
be the accumulation function.
We use the notation 1
C
(t) to denote the indicator function on the set C; the notation
I[t C] refers to the indicator of a set and is thus not considered a function. Then, for an
18
initially susceptible individual i, its infection time A
K
i
is given by
A
K
i
= inf
_
t R
+
:
_
(0,t]
(s, I
K
)ds = l
i
_
, (4.2)
with
I
K
(t) =
1
K
aK

j=1
1
[ e
j
, e
j
+r
j
)
(t) +
1
K
bK

j=1
1
[0, r
j
)
(t) +
1
K
cK

j=1
1
[A
K
j
+e
j
,A
K
j
+e
j
+r
j
)
(t)
being the proportion of infected individuals present in the population at time t R
+
.
This gives a recursive denition of the A
K
i
s: If l
(j)
is the j
th
order statistics of l
1
, . . . , l
bK
,
corresponding to the individual i
j
, say, then
A
K
i
j
= inf
_
t > 0 :
_
(0,t]
(s, I
K
(s)) ds = l
(j)
_
,
where
I
K
(s) =
1
K
aK

m=1
1
[ e
m
, e
m
+r
m
)
(s) +
1
K
bK

m=1
1
[0, r
m
)
(s) +
1
K
j1

k=1
1
[A
K
i
k
+e
i
k
,A
K
i
k
+e
i
k
+r
i
k
)
(s).
This completes the description of the model. Furthermore, we make some technical assump-
tions.
The function : R
+
D
+
R
+
satises for all t R
+
, x, y D
+
1. it is non-anticipating: (t, x) = (t, x
t
), where, for u R
+
, x
t
(u) = x(t u) ;
2. it is Lipschitz: there is a constant > 0 such that, for all t,
[(t, x) (t, y)[ sup
0st
[x(s) y(s)[ ; (4.3)
3. it is bounded: there is a constant > 0 such that, for all t,
sup
0st
(s, x) ;
4. we have that, for all t,
(t, x) = 0 x(t) = 0.
There is a constant > 0 such that, for each x, y R
+
, the conditional distribution
function
x,y
(t) := P[l
1
t[e
1
= x, r
1
= y] has a density
x,y
(t) that is uniformly
bounded from above by ;

x,y
(t) for all x, y R
+
, t R
+
. (4.4)
We assume that has a density . (In Reinert [19], it was only assumed that the con-
ditional distribution function is Lipschitz continuous; however, imposing the existence
of a density is more convenient.)
19
There is a constant > 0 such that, for each x R
+
, the conditional distribution
function
x
(t) := P[r
1
t[e
1
= x] has a density
x
(t) that is uniformly bounded from
above by ;

x
(t) for all x R
+
, t R
+
. (4.5)
We assume that

(0) = 0 and (0) = 0, so that infected individuals do not immediately
get removed.
For Bartletts GSE dened in Equations (2.1), the above assumptions are satised.
Choosing (t, x) = x(t), (l
i
) being i. i. d. exp(1), e
i
= 0 for all i N and ( r
i
), (r
i
) be-
ing i. i. d. exp(), where for each i, l
i
and r
i
are independent, we obtain
[(t, x) (t, y)[ sup
0st
[x(s) y(s)[.
Moreover is bounded by = 1, and P[l
1
t[e
1
= x, r
1
= y] = 1 e
t
has density that is
uniformly bounded by 1.
4.3 Heuristics
In this section, we present a heuristic argument on how we can obtain the limiting distribu-
tion. Let
F
K
(t) =
_
t
0
(s, I
K
(s))ds. (4.6)
Then,
A
K
i
= F
1
K
(l
i
).
Moreover, the proportion of infectives present at time t, I
K
(t), depends on F
K
; we have
I
K
(t) =
1
K
aK

i=1
1( e
i
t < e
i
+r
i
) +
1
K
bK

i=1
1( r
i
> t)
+
1
K
cK

i=1
1
_
F
1
K
(l
i
) +e
i
t < F
1
K
(l
i
) +e
i
+r
i
_
.
For f D(R
+
), the space of right-continuous functions from R
+
to R with left-hand limits,
and for t R+, we dene operators
Z
K
f(t) =
1
K
aK

i=1
1( e
i
t < e
i
+r
i
) +
1
K
bK

i=1
1( r
i
> t)
+
1
K
cK

i=1
1(f(t e
i
r
i
) < l
i
f(t e
i
)) (4.7)
and
L
K
f(t) =
_
(0,t]
(s, Z
k
f(s))ds. (4.8)
20
Then, L
K
F
K
= F
K
; thus F
K
can be described as a xed point of the operator L
K
.
Moreover, for the operator Z
K
, the weak law of large number suggests, as a limiting operator
for f C(R
+
), t R
+
,
Zf(t) := aP[ e
1
t < e
1
+r
1
] +b(1

(t)) +cP[f(t e
1
r
1
) < l
1
f(t e
1
)]. (4.9)
Accordingly, we dene
Lf(t) :=
_
(0,t]
(s, Zf(s))ds. (4.10)
Using the Contraction Theorem, we shall show that on each nite time interval, the
equation Lf = f has a unique solution, G. It thus heuristically follows that
F
K
G.
As G is deterministic, the desired limiting distribution of the empirical measure can then be
derived.
4.4 Results
In this section, we present two results on the SEIR model.
Theorem 4.4.1 For T R
+
, the equation
F(t) =
_
(0,t]
(s, ZF(s)) ds , 0 t T, (4.11)
has a unique solution G
T
. This solution can be obtained by an iteration procedure: Choose
an arbitrary f
0
C([0, T]) and put f
1
= Lf
0
, f
n
= Lf
n1
for n N. Then,
[[f
n
G
T
[[
T

(
c
2
)
n
1
c
2
[1 + 4T(n + 1)]

1
4T(n + 1)
[[f
0
Lf
0
[[
T
,
where
= sup
_
t T :
_
(0,T]
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds
1
2
_
.
Theorem 4.4.2 Fix T R
+
. For any given u, v, w (0, T], let u
T
M
1
(R
3
+
) be given by

T
([0, u] [0, v] [0, w])
= P(G
1
(l
1
) u, G
1
(l
1
) +e
1
v, G
1
(l
1
) +e + 1 +v + 1 w)
=
_
(0,[(vu)(wu)]0]
_
(0,(wux)0]

x,y
(G
T
(u)) d
x
(y) d(x)
+
_
(0(vu),v]
_
(0,(wv)0]

x,y
(G
T
(v x)) d
x
(y) d(x)
+
_
(0,u]
_
(0[(wv)(wux)],wx]

x,y
(G
T
(w x y)) d
x
(y) d(x).
21
Put

T
= a(
0

e

r
)
T
+b(
0

0

r
)
T
+c
T
.
Then, for all T R
+
and for all H T,
[E
T
H(
K
) H(
T
)[

a +

b +

K
+cT(T + 2) exp(b2T|)
_
(2a +b +c)
_
1
K
+
1
K
_
,
where x| is the smallest integer larger than x.
Remarks.
1. Theorem 4.4.1 gives an iterative procedure to obtain the deterministic approximation
to an epidemic, while Theorem 4.4.2 gives the quality of this approximation in terms
of its mean-eld limit.
2. Theorem 4.4.2 gives the result in terms of expectations. Using Markovs inequality, we
can obtain information for the distribution function:
P
_

E
T
H(
K
) H()

>
log K

K
_

1
log K
_
cT(T + 2) exp(b2T|)
_
(2a +b +c) +
_
1
K
_
+

a +

b +

c
_
.
3. The bound obtained in Theorem 4.4.2 tends to 0 as K tends to 0 for each nite value
of T; however, for any nite value of K, the bound tends to innity as T tends to
innity .
4. Compared to Reinert [21], the bound obtained in Theorem 4.4.2 is marginally sharper
the
1
K
term here has coecient 1, while in Reinert [21], the coecient is 2. This is
due to a slightly dierent argument in the proof.
In the next section, we will prove these two theorems.
4.5 Proofs
Proof of Theorem 4.4.1
To prove this theorem, we adopt a similar technique as in the proof of the Picard-Lindelof
Theorem. However, we need to make a slight modication here as do not have the Lipschitz
property for (s, Zf), as a function of f; we merely assumed that (s, f) is Lipschitz.
22
We rst prove that L is a contraction on a (small) compact interval in a complete metric
space and then employ the contraction mapping theorem in this interval to yield a unique
xed point (in this interval). Then, we extend the validity of our solution to [0, T] by
extending this small interval inductively.
Formally, let f be a solution of (4.11). Then, for s, t [0, T],
[f(t) f(s)[ =
_
(s,t]
(u, ZF(u)) du
[t s[.
Therefore, f C([0, T]). Similarly, Lf C([0, T]). Note also that (C([0, T]), [[ [[
T
) is a
complete metric space.
Next, we show that Lf is a contraction on small intervals and then employ the contraction
theorem.
Let f, g C([0, T]). Then, for all t R
+
,
[Lf(t) Lg(t)[
=

_
(0,t]
[(s, Zf(s)) (s, Zg(s))] ds

_
(0,t]
sup
0us
[Zf(u) Zg(u)[ ds
=
_
(0,t]
sup
0us

aP[ e
1
t < e
1
+r
1
] +b(1 +

(u))
+cP[f(u e
1
r
1
) < l
1
< f(u e
1
)]
aP[ e
1
t < e
1
+r
1
] b(1 +

(u))
cP[g(u e
1
r
1
) < l
1
< g(u e
1
)]

ds
= c
_
(0,t]
sup
0us

P[f(u e
1
r
1
) < l
1
< f(u e
1
)]
P[g(u e
1
r
1
) < l
1
< g(u e
1
)]

ds. (4.12)
Now, we calculate the two probabilities inside the integral.
23
First,
P[f(u e
1
r
1
) < l
1
< f(u e
1
)]
= P[f(u e
1
r
1
) < l
1
< f(u e
1
) ; e
1
+r
1
u]
+ P[f(u e
1
r
1
) < l
1
< f(u e
1
) ; e
1
< u < e
1
+r
1
]
=
_
(0,u]
_
(0,ux]
P[f(u x y) < l
1
< f(u x)[e
1
= x, r
1
= y] P[e
1
dx, r
1
dy]
+
_
(0,u]
_
(ux,)
P[0 < l
1
< f(u x)[e
1
= x, r
1
= y] P[e
1
dx, r
1
dy]
=
_
(0,u]
_
(0,ux]
[
x,y
(f(u x))
x,y
(f(u x y))] d
x
(y) d(x)
+
_
(0,u]
_
(ux,)

x,y
(f(u x)) d
x
(y) d(x)
=
_
(0,u]

x
(f(u x)) d(x)

_
(0,u]
_
(0,ux]

x,y
(f(u x y)) d
x
(y) d(x). (4.13)
Similarly, we obtain
P[g(u e
1
r
1
) < l
1
< g(u e
1
)]
=
_
(0,u]

x
(g(u x)) d(x)

_
(0,u]
_
(0,ux]

x,y
(g(u x y)) d
x
(y) d(x). (4.14)
Applying (4.13), (4.14) and the Lipschitz property of
x,y
to (4.12), we get
[Lf(t) Lg(t)[
c
_
(0,t]
sup
0us

_
(0,u]
[
x
(f(u x))
x
(g(u x))] d(x)

_
(0,u]
_
(0,ux]
[
x,y
(f(u x y))
x,y
(g(u x y))] d
x
(y) d(x)

ds
c
_
(0,t]
_
(0,s]
[f(s x) g(s x)[d(x)ds
+ c
_
(0,t]
_
(0,s]
_
(0,sx]
[f(s x y) g(s x y)[ d
x
(y) d(x) ds (4.15)
c[[f g[[
t
_
(0,t]
(s) ds +c
_
(0,t]
_
(0,s]
[[f g[[
sx
_
(0,sx]
d
x
(y) d(x) ds
c[[f g[[
t
_
(0,t]
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds.
24
Now, put
= sup
_
t T :
_
(0,T]
_
(s) +
_
(0,s]

x
(s x) d(x)
_
ds
1
2
_
.
Then, for t ,
[[Lf Lg[[
t

c
2
[[f g[[
t
.
So, L is a contraction on C([0, ]). It follows from the Contraction Theorem that (4.11) has
a unique solution, G

, in C([0, ]).
We proceed by induction. For k N, assume that L is a contraction on C

([0, k]).
Dene the space of continuous extension on C([0, (k +1)]) of a unique solution of (4.11) on
C([0, k]) as
C

[0, (k + 1)] = f C([0, (k + 1)]) : Lf(x) = f(x) for x kn.


Take any f, g C

[0, (k + 1)]. Then, for 0 t k,


[[Lf Lg[[
t
= 0
and for k < t (k + 1),
[Lf(t) Lg(t)[
c
_
(k,t]
sup
kus

_
(0,uk]
[
x
(f(u x))
x
(g(u x))]d(x)
+
_
(0,uk]
_
(0,ukx]
[
x,y
(f(u x y))
x,y
(g(u x y))] d
x
(y) d(x)

ds
since
f(u x) = g(u x) if u x k
and f(u x y) = g(u x y) if u x y k.
Using the Lipschitz property of
x,y
,
[Lf(t) Lg(t)[
c[[f g[[
t
_
(k,t]

_
(0,sk]
d(x) +
_
(0,sk]
_
(0,skx]
d
x
(y)d(x)

ds
= c[[f g[[
t
_
(k,t]
_
(s k) +
_
(0,sk]

x
(s k x)d(x)
_
ds
= c[[f g[[
t
_
(0,tk]
_
(w) +
_
(0,w]

x
(w x)d(x)
_
ds, where w = s k

c
2
[[f g[[
t
, (4.16)
since t k < . Hence, L is a contraction on C

([0, (k +1)]). By induction, for all k N,


the contraction property holds on every C

([0, k]).
25
Since k < T (k + 1) for some k N, we conclude that (4.11) has a unique solution
G
T
on C

([0, T]). As every solution of (4.11) must be in C

([0, T]), we also have uniqueness


on C([0, T]).
The Contraction Theorem also gives us the following iteration procedure. Choose an
arbitrary f
0
C([0, T]). Let f
1
= Lf
0
, f
n
= Lf
n1
for n N. Due to the Contraction
Theorem, we have
[[f
n
G
T
[[


(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[

.
Furthermore, for k N, to get a bound on [[f
n
G
T
[[
(k+1)
, put
g
0
(s) = G
T
(s)I(s k) +f
0
(s)I(s > k).
Let g
1
= Lg
0
, g
n
= Lg
n1
, n N. Then,
[[f
n
G
T
[[
(k+1)
[[f
n
g
n
[[
(k+1)
+[[g
n
G
T
[[
(k+1)
= [[Lf
n1
Lg
n1
[[
(k+1)
+[[g
n
G
T
[[
(k+1)
. (4.17)
Moreover, as g
0
C

[(0, (k + 1))],
[[g
n
G
T
[[
(k+1)

(
c
2
)
n
1
c
2
[[g
0
Lg
0
[[
(k+1)
. (4.18)
For s k, we have Lg
0
(s) = LG
T
(s) = G
T
(s) = g
0
(s). Hence,
[[g
0
Lg
0
(s)[[
(k+1)
= sup
s(k,(k+1)]
[f
0
(s) Lg
0
(s)[
[[f
0
Lf
0
[[
(k+1)
+ sup
s(k,(k+1)]
[Lf
0
(s) Lg
0
(s)[. (4.19)
Applying (4.18) and (4.19) to (4.17), we have
[[f
n
G
T
[[
(k+1)
[[f
n
G
T
[[
k
+ sup
s(k,(k+1)]
[Lf
n1
(s) Lg
n1
(s)[
+
(
c
2
)
n
1
c
2
_
[[f
0
Lf
0
[[
(k+1)
+ sup
s(k,(k+1)]
[Lf
0
(s) Lg
0
(s)[
_
. (4.20)
For k < t (k + 1), t T, similar to (4.15), we have
[Lf
n
(t) Lg
n
(t)[
(k+1)
c
__
(0,t]
_
(0,s]
[f
n
(s x) g
n
(s x)[ d(x) ds
+
_
(0,t]
_
(0,s]
_
(0,sx]
[f
n
(s x y) g
n
(s x y)[ d
x
(y)d(x)ds
_
(4.21)
26
We proceed to simplify the two integrals. The idea here is to split the integrals into sub-
integrals where [f
n
(s) g
n
(s)[ can be bounded by [[f
n
g
n
[[
k
or sup
k<s<t
[f
n
(s) g
n
(s)[.
We now bound the rst integral in (4.21).
_
(0,t]
_
(0,s]
[f
n
(s x) g
n
(s x)[d(x) ds
=
_
(0,k]
_
(0,s]
[f
n
(s x) g
n
(s x)[d(x) ds
+
_
(k,t]
_
(0,s]
[f
n
(s x) g
n
(s x)[d(x) ds
[[f
n
g
n
[[
k
_
(0,k]
(s) ds
+
_
(k,t]
_
(0,sk]
[f
n
(s x) g
n
(s x)[d(x) ds
+
_
(k,t]
_
(sk,s]
[f
n
(s x) g
n
(s x)[d(x) ds
[[f
n
g
n
[[
k
_
(0,k]
(s) ds
+ sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
(s k) ds
+[[f
n
g
n
[[
k
_
(k,t]
[(s) (s k)] ds
= [[f
n
g
n
[[
k
__
(0,k]
(s) ds +
_
(k,t]
(s) ds
_
(k,t]
(s k) ds
_
+ sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
(s k) ds
= [[f
n
g
n
[[
k
_
(tk,t]
(s) ds
+ sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
(s k) ds. (4.22)
27
The second integral in (4.21) can be bounded by
_
(0,t]
_
(0,s]
_
(0,sx]
[f
n
(s x y) g
n
(s x y)[d
x
(y) d(x) ds
=
_
(0,k]
_
(0,s]
_
(0,sx]
[f
n
(s x y) g
n
(s x y)[d
x
(y) d(x) ds
+
_
(k,t]
_
(0,s]
_
(0,sx]
[f
n
(s x y) g
n
(s x y)[d
x
(y) d(x) ds
[[f
n
g
n
[[
k
_
(0,k]
_
(0,s]

x
(s x)d(x) ds
+
_
(k,t]
__
(0,sk]
_
(0,skx]
[f
n
(s x y) g
n
(s x y)[d
x
(y) d(x)
+
_
(0,sk]
_
(skx,sx]
[f
n
(s x y) g
n
(s x y)[d
x
(y) d(x)
+
_
(sk,s]
_
(0,sx]
[f
n
(s x y) g
n
(s x y)[d
x
(y) d(x)
_
ds
[[f
n
g
n
[[
k
_
(0,k]
_
(0,s]

x
(s x)d(x) ds
+ sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
_
(0,sk]

x
(s k x)d(x) ds
+[[f
n
g
n
[[
k
_
(k,t]
_
(0,sk]
[
x
(s x)
x
(s k x)] d(x) ds
+[[f
n
g
n
[[
k
_
(k,t]
_
(sk,s]

x
(s x)d(x) ds
= [[f
n
g
n
[[
k
_
(0,t]
_
(0,s]

x
(s x)d(x) ds
[[f
n
g
n
[[
k
_
(k,t]
_
(0,sk]

x
(s k x)d(x) ds
+ sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
_
(0,sk]

x
(s k x)d(x) ds. (4.23)
28
Applying (4.22) and (4.23) to (4.21), we have
[Lf
n
(t) Lg
n
(t)[
(k+1)
c
_
[[f
n
g
n
[[
k
_
(tk,t]
(s) ds + sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
(s k) ds
_
+c
_
[[f
n
g
n
[[
k
_
(0,t]
_
(0,s]

x
(s x)d(x) ds
[[f
n
g
n
[[
k
_
(k,t]
_
(0,sk]

x
(s k x)d(x) ds
+ sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
_
(0,sk]

x
(s k x)d(x) ds
_
= c[[f
n
g
n
[[
k
__
(tk,t]
(s) ds +
_
(0,t]
_
(0,s]

x
(s x)d(x) ds

_
(k,t]
_
(0,sk]

x
(s k x)d(x) ds
_
+c sup
k<st
[f
n
(s) g
n
(s)[
_
(k,t]
_
(s k) +
_
(0,sk]

x
(s k x)d(x)
_
ds
2T[[f
n
g
n
[[
k
+
c
2
sup
k<st
[f
n
(s) g
n
(s)[,
where we employed the the denition of , since t k . For n = 0, we get, by
construction, that for k < t (k + 1), t T,
[Lf
0
(t) Lg
0
(t)[ 2cT[[f
0
G
T
[[
k
.
Solving this recursion formula yields
sup
k<t(k+1)
[Lf
n
(t) Lg
n
(t)[ 2cT
n

i=0
[[f
ni1
G
T
[[
k
_
c
2
_
i
.
Combining these estimates, we get the recursion
[[f
n
G
T
[[
(k+1)
[[f
n
G
T
[[k
+2cT
n1

i=0
[[f
ni
G
T
[[
k
_
c
2
_
i
+
(
b
2
)
n
1
c
2
_
[[f
0
Lf
0
[[
(k+1)
+ 2cT[[f
0
G
T
[[
k
_
,
and [[f
n
G
T
[[


(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[

.
To get a bound on [[f
n
G
T
[[
(k+1)
, suppose that, for a c
k
R
+
,
[[f
n
G
T
[[
k

(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[
k
c
k
.
29
Then,
[[f
n
G
T
[[
(k+1)

(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[
k
c
k
+ 2cT
(
c
2
)
n1
1
c
2
[[f
0
Lf
0
[[
k
n1

i=0
c
k
+
(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[
(k+1)
+
(
c
2
)
n
(1
c
2
)
2
2cT[[f
0
Lf
0
[[
k
c
k

(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[
(k+1)
_
c
k
+ 4Tnc
k
+ 1 + 4Tc
k
_
=
(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[
(k+1)
_
c
k
[1 + 4T(n + 1)] + 1
_
.
Thus,
[[f
n
G
T
[[
(k+1)

(
c
2
)
n
1
c
2
[[f
0
Lf
0
[[
(k+1)
c
k+1
,
where
c
k+1
= c
k
[1 + 4T(n + 1)] + 1.
Finally, solving this recursion on c
k
with c
1
= 1, we get
c
k
=
k1

i=0
[1 + 4T(n + 1)]
i
=
[1 + 4T(n + 1)]
k
1
4T(n + 1)
.
Hence,
[[f
n
G
T
[[
T

(
c
2
)
n
1
c
2
[1 + 4T(n + 1)]

1
4T(n 1)
[[f
0
Lf
0
[[
T
.
This proves the iteration procedure and completes the proof of Theorem 4.4.1.
As the proof of Theorem 4.4.2 employs uniform convergence for the empirical distribution
function, we state the following lemma from Reinert [21].
Lemma 4.5.1 Let X
1
, . . . X
n
be i. i. d. real-valued random variables from a distribution with
distribution function F, and let F
n
denote the empirical distribution function
F
n
(t) =
n

i=1
1(X
i
t).
Then,
E
_
sup
xR
[F
n
(x) F(x)[
_

1

n
.
30
Proof of Theorem 4.4.2
We are now ready to prove Theorem 4.4.2. First, from Theorem 3.5.3, it suces to
bound, for all m N, f C

b
(R
m
), and for all
1
, . . . ,
m
C

b
([0, T]
3
) satisfying (3.5)
m

j=1
Ef
(j)
(
T
K
,
k
), k = 1, . . . , m)
T

T
K
,
j
). (4.24)
We use the notation
T
(x,y,z)
to mean
(x, y, z)
1(0 x T)1(0 y T)1(0 z T). Then,

T
K

T
= a
_
1
aK
aK

i=1

T
(0, e
i
,r
i
)
(
0

e

r
)
T
_
+ b
_
1
bK
bK

i=1

T
(0,0, r
i
)
(
0

0

r
)
T
_
+ c
_
1
cK
cK

i=1

T
(A
K
i
,A
K
i
+e
i
,A
K
i
+e
i
+r
i
)

T
_
.
Hence, we have

j=1
Ef
(j)
(
T
K
,
k
), k = 1, . . . , m)
T

T
K
,
j
)

a
_
m

j=1
[[f
(j)
[[ E

1
aK
aK

i=1

T
j
(0, e
i
, r
i
) E
T

T
j
(0, e
i
, r
i
)

_
+ b
_
m

j=1
[[f
(j)
[[ E

1
bK
bK

i=1

T
j
(0, 0, r
i
) E
T

T
j
(0, 0, r
i
)

_
+ c
_
m

j=1
[[f
(j)
[[ E

1
cK
cK

i=1

T
j
(A
T
i
, A
T
i
+e
i
, A
T
i
+e
i
+r
i
)
E
T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

_
.
Using the Cauchy-Schwartz inequality, we can bound the rst two summands by
a
_
m

j=1
[[f
(j)
[[ E

1
aK
aK

i=1

T
j
(0, e
i
, r
i
) E
T

T
j
(0, e
i
, r
i
)

_
a
m

j=1
[[f
(j)
[[ E
_
Var
_
1
aK
aK

i=1

T
j
(0, e
i
, r
i
)
_ _
1
2

K
(4.25)
31
and
b
_
m

j=1
[[f
(j)
[[ E

1
bK
bK

i=1

T
j
((0, 0, r
i
) E
T

T
j
(0, 0, r
i
)

_
b
m

j=1
[[f
(j)
[[ E
_
Var
_
1
bK
bK

i=1

T
j
(0, 0, r
i
)
_ _
1
2

K
. (4.26)
Finally, we bound
c
_
m

j=1
[[f
(j)
[[ E

1
cK
cK

i=1

T
j
(A
K
i
, A
K
i
+e
i
, A
K
i
+e
i
+r
i
)
E
T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

_
.
From (3.5) it suces to bound, for any as in (3.5),
cE

1
cK
cK

i=1

T
j
(A
K
i
, A
K
i
+e
i
, A
K
i
+e
i
+r
i
)
E(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

= c E

1
cK
cK

i=1

T
j
(A
K
i
, A
K
i
+e
i
, A
K
i
+e
i
+r
i
)

1
cK
cK

i=1

T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)
+
1
cK
cK

i=1

T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)
E(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

R
1
+R
2
,
where
R
1
= c E

1
cK
cK

i=1

T
j
(A
K
i
, A
K
i
+e
i
, A
K
i
+e
i
+r
i
)

1
cK
cK

i=1

T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

(4.27)
32
and
R
2
= c E

1
cK
cK

i=1

T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)
E(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

K
,
where the inequality follows by a similar argument as in (4.25).
Bounding R
1
Bounding R
1
is more complicated, partly because the derivative of G
1
T
is not necessar-
ily bounded. Instead, similar to Reinert [19], we mimic dierentiation making use of the
additional stochasticity introduced by the random point l
1
.
Note that F
K
and l
1
are dependent. In the spirit of (3.2), we introduce F
K,1
, which is similar
to F
K
, but with susceptible individual 1 omitted. As is non-anticipating, the course of the
epidemic up to time F
1
K
(l
1
) is not aected if individual 1 is omitted from the population.
To make this precise, let
H
T
= D([0, T]),
the space of right-continuous functions with left-hand limits from [0, T] to R. Similar to
(4.7) and (4.8), dene for h H
T
the operators
Z
K,1
h(t) =
1
K
aK

i=1
1( e
i
t < e
i
+r
i
) +
1
K
bK

i=1
1( r
i
> t)
+
1
K
cK

j=2
1(h(t e
j
r
j
) < l
j
h(t e
j
))
L
K,1
h(t) =
_
(0,t]
(s, Z
K,1
h(s)) ds.
and let F
K,1
be the unique xed point of L
K,1
h = h. Then we have F
1
K
(l
1
) = F
1
K,1
(l
1
) by
construction.
Thus,
R
1
= cE

1
cK
cK

i=1

T
j
(A
K
i
, A
K
i
+e
i
, A
K
i
+e
i
+r
i
)

1
cK
cK

i=1

T
j
(G
1
T
(l
i
), G
1
T
(l
i
) +e
i
, G
1
T
(l
i
) +e
i
+r
i
)

cE

(A
K
1
, A
K
1
+e
1
, A
K
1
+e
1
+r
1
)1(A
K
1
+e
1
+r
1
T)
(G
1
T
(l
1
), G
1
T
(l
1
) +e
1
, G
1
T
(l
1
) +e
1
+r
1
)1(G
1
T
(l
1
) +e
1
+r
1
T)

.
33
Expanding the indicators, we have
(A
K
1
, A
K
1
+e
1
, A
K
1
+e
1
+r
1
)1(A
K
1
+e
1
+r
1
T)
(G
1
T
(l
1
), G
1
T
(l
1
) +e
1
, G
1
T
(l
1
) +e
1
+r
1
)1(G
1
T
(l
1
) +e
1
+r
1
T)
= (F
1
K,1
(l
1
), F
1
K,1
(l
1
) +e
1
, F
1
K,1
(l
1
) +e
1
+r
1
)1(F
1
K,1
(l
1
) +e
1
+r
1
T)
(G
1
T
(l
1
), G
1
T
(l
1
) +e
1
, G
1
T
(l
1
) +e
1
+r
1
)1(G
1
T
(l
1
) +e
1
+r
1
T)
= (F
1
K,1
(l
1
), F
1
K,1
(l
1
) +e
1
, F
1
K,1
(l
1
) +e
1
+r
1
)
1(F
1
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
T)
+(F
1
K,1
(l
1
), F
1
K,1
(l
1
) +e
1
, F
1
K,1
(l
1
) +e
1
+r
1
)
1(F
1
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
> T)
(G
1
T
(l
1
), G
1
T
(l
1
) +e
1
, G
1
T
(l
1
) +e
1
+r
1
)
1(F
1
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
T)
(G
1
T
(l
1
), G
1
T
(l
1
) +e
1
, G
1
T
(l
1
) +e
1
+r
1
)
1(F
1
K,1
(l
1
) +e
1
+r
1
> T, G
1
T
(l
1
) +e
1
+r
1
T).
(For simplicity of notation we omit the superscript T for F
K,1
and for the expectation).
Applying the mean value theorem and (3.5),
(x, x +e
1
, x +e
1
+r
1
) (y, y +e
1
, y +e
1
+r
1
) [[

[[ [x y[ [x y[.
Hence, we can bound R
1
by
1
c
R
1
E

_
F
1
K,1
(l
1
) G
1
T
(l
1
)
_
1(F
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
T)

+P
_
F
1
K,1
(l
1
) +e
1
+r
1
> T, G
1
T
(l
1
) +e
1
+r
1
T

+P
_
F
1
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
> T

. (4.28)
The middle term can be bounded by
P
_
F
1
K,1
(l
1
) +e
1
+r
1
> T, G
1
T
(l
1
) +e
1
+r
1
T

= P[F
K,1
(T e
1
r
1
) < l
1
< G
T
(T e
1
r
1
)]

G
T
(T e
1
r
1
) F
K,1
(T e
1
r
1
)

E[[G
T
F
K,1
[[
T
. (4.29)
Similarly, the last term can be bounded by
P
_
F
1
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
> T

E[[G
T
F
K,1
[[
T
. (4.30)
34
It remains to bound
E

_
F
1
K,1
(l
1
) G
1
T
(l
1
)
_
1(F
K,1
(l
1
) +e
1
+r
1
T, G
1
T
(l
1
) +e
1
+r
1
T)

_
F
1
K,1
(l
1
) G
1
T
(l
1
)
_
1(F
K,1
(l
1
) T, G
1
T
(l
1
) T)

= E
_
M
T
0

(F
K,1
)
1
(x) G
1
T
(x)

1(F
1
K,1
(x) T, G
1
T
(x) T)(x) dx
= E
_
M
T
0

G
1
T
(G
T
((F
K,1
)
1
(x))) G
1
T
(x)

1(F
1
K,1
(x) T, G
1
T
(x) T)(x) dx,
where M
T
= minF
K,1
(T), G
T
(T).
From Condition 4 on , we have G
T
to be strictly increasing until it stays constant thereafter.
Thus the derivative (G
1
T
)

exists for all t for which G


T
is not constant; in particular it exists
for all t < G
1
T
(x) when G
1
T
(x) T. Thus we may integrate as follows:
E
_
M
T
0

G
1
T
(G
T
((F
K,1
)
1
(x))) G
1
T
(x)

1(F
1
K,1
(x) T, G
1
T
(x) T)(x) dx
= E
_
M
T
0

_
G
T
(F
K,1
)
1
(x))
x
(G
1
T
)

(y)dy

(x) dx
= E
_
M
T
0
_
G
T
((F
K,1
)
1
(x))
x
(G
1
T
)

(y) dy 1(x < G


T
((F
K,1
)
1
(x)))(x) dx
+E
_
M
T
0
_
x
G
T
((F
K,1
)
1
(x))
(G
1
T
)

(y) dy 1(x > G


T
((F
K,1
)
1
(x)))(x) dx.
All integrals are nite and so we may interchange the order of integration to obtain
E
_
M
T
0

G
1
T
(G
T
((F
K,1
)
1
(x))) G
1
T
(x)

1(F
1
K,1
(x) T, G
1
T
(x) T)(x) dx
= E
_
G
T
(F
1
K,1
(M
T
))
0
_
G
T
(G
1
T
(y))
F
K,1
(G
1
T
(y))
1(x < G
T
((F
K,1
)
1
(x)))(x) dx(G
1
T
)

(y) dy
+E
_
G
T
(F
1
K,1
(M
T
))
0
_
F
K,1
(G
1
T
(y))
G
T
(G
1
T
(y))
1(x > G
T
((F
K,1
)
1
(x)))(x) dx(G
1
T
)

(y) dy
TE[[G
T
F
K,1
[[
T
. (4.31)
Applying (4.29), (4.30), (4.31) to (4.28), we can bound R
1
by
R
1
c(T + 2)E[[G
T
F
K,1
[[
T
. (4.32)
35
Bound on [[G
T
F
T
K,1
[[
T
To bound this quantity, we proceed by an inductive argument, similar to the proof of Theorem
4.4.1. For any t R
+
, we have
F
K,1
(t) G
T
(t) = L
K,1
F
K,1
(t) LG
T
(t)
= [L
K,1
F
K,1
(t) LF
K,1
(t)] + [LF
K,1
(t) LG
T
(t)].
So, for all t T,
E[[F
K,1
G
T
[[
t
E sup
hH
t
[[L
K,1
h Lh[[
T
+E[[LF
K,1
LG
T
[[
t
. (4.33)
First, we bound Esup
hH
t
[[L
K,1
hLh[[
T
. For h H
T
we have, due to the Lipschitz property
of , that
[[L
K,1
h Lh[[
T

_
T
0
sup
sx
[Z
K,1
h(s) Zh(s)[ dx
=
_
T
0
sup
sx

1
K
aK

i=1
1( e
i
s < e
i
+r
i
) +
1
K
bK

i=1
1( r
i
> s)
+
1
K
cK

i=2
1(h(s e
i
r
i
< l
i
< h(s e
i
))
aP[ e
1
s < e
1
+r
1
] b(1

(s))
cP[h(s e
i
r
i
) < l
i
< h(s e
i
)]

dx
T
_
a sup
s

1
K
aK

i=1
1( e
i
s < e
i
+r
i
) aP[ e
1
s < e
1
+r
1
]

+b sup
s

1
bK
bK

i=1
_
1 1( r
i
s)
_
1 +

(s)

+c sup
s

1
cK
cK

i=1
1(h(s e
i
r
i
) < l
i
< h(s e
i
))
P[h(s e
i
r
i
) < l
i
< h(s e
i
)]

1
K
1(h(s e
1
r
1
< l
1
< h(s e
1
))

_
T
_
R
3
+bR
4
+cR
5
+
1
K
_
,
where
R
3
= sup
s

1
K
aK

i=1
1( e
i
s < e
i
+r
i
) aP[ e
1
s < e
1
+r
1
]

, (4.34)
36
R
4
= sup
s

1
bK
bK

i=1
1( r s)

(s)

(4.35)
and
R
5
= sup
s

1
cK
cK

i=1
1(l
i
s) (s)

. (4.36)
From Lemma 4.5.1, E(R
4
) and E(R
5
) can be bounded by
_
1
K
. R
3
needs some simplication
before we can apply Lemma 4.5.1.
R
3
= sup
s

1
K
aK

i=1
1( e
i
s < e
i
+r
i
) aP[ e
1
s < e
1
+r
1
]

= a sup
s

1
aK
aK

i=1
1( e
1
s)
1
aK
aK

i=1
1( e
1
+r
i
s) P[ e
1
s] +P[ e
1
+r
1
s]

a sup
s

1
aK
aK

i=1
1( e
1
s) P[ e
1
s]

+a sup
s

aK

i=1
1( e
1
+r
i
s) P[ e
1
+r
1
s]

= a sup
s

1
aK
aK

i=1
1( e
1
s)

(s)

+a sup
s

aK

i=1
1(y
i
s) (s)

,
where y
i
= e
i
+r
i
and () is its distribution function.
Note that y
1
, y
2
, . . . , y
aK
is an i.i.d. sequence of random variables with common distribution
function . Applying Lemma 4.5.1, we have E(R
3
) to be bounded by 2a
_
1
K
.
Hence
E sup
hH
t
[[L
K,1
h Lh[[
T
T
_
(2a +b +c)
_
1
K
+
1
K
_
=: S(K). (4.37)
Next, we now bound E[[LF
K,1
LG
T
[[
t
. Similar to (4.15), we have
[LF
K,1
(t) LG
T
(t)[ c
_
(0,t]
[[F
K,1
G
T
[[
x
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds.
Thus we obtain
E[[F
K,1
G
T
[[
t
S(K) +c
_
(0,t]
E[[F
K,1
G
T
[[
x
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds. (4.38)
37
Fixed an arbitrary l c, l N and dene
=
1
2l
. (4.39)
Then we have that
2c
c
l
.
Also, as and are distribution functions,
(x) +
_
(0,x]
(x s)d(s) 2.
Hence
E[[F
K,1
G
T
[[

S(K) +
c
l
E[[F
K,1
G
T
[[

,
yielding
E[[F
K,1
G
T
[[


1
1
c
l
S(K)
=
l
l c
S(K).
We now prove by induction that, for any k N,
E[[F
K,1
G
T
[[
k

_
l
l c
_
k
S(K). (4.40)
The case k = 1 has already been proven above. Suppose (4.40) is true for some k N. Then,
from (4.38),
E[[F
K,1
G
T
[[
(k+1)
S(K) +c
_
(k+1)
0
E[[F
K,1
G
T
[[
x
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds
S(K) +c
k

j=1
_
j
(j1)
E[[F
K,1
G
T
[[
x
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds
+c
_
(k+1)
k
E[[F
K,1
G
T
[[
x
_
(s) +
_
(0,s]

x
(s x)d(s)
_
ds
S(K) +c
k

j=1
_
l
l c
_
l
S(K)
+c
_
(k+1)
k
E[[F
K,1
G
T
[[
x
_
(s) +
_
(0,s]

x
(s x)d(x)
_
ds,
38
where we used the induction hypothesis. Using the denition of in (4.39), we obtain
E[[F
K,1
G
T
[[
(k+1)
S(K) +
c
l
k

j=1
_
l
l c
_
j
S(K) +
c
l
E[[F
K,1
G
T
[[
(k+1)
.
Thus,
E[[F
K,1
G
T
[[
(k+1)

1
1
c
l
S(K)
_
1 +
c
l
k

j=1
_
l
l c
_
j
_
=
_
l
l c
_
k+1
S(K).
This proves (4.40).
Let
T

|, the smallest integer at least as large as


T

. Together with (4.39), we have shown


that,
E[[F
K,1
G
T
[[
T

_
l
l c
_

S(K)
= S(K)
_
l
l c
_
2lT
S(K)
_
1 +
c
l c
_
2T l
= S(K)
_
_
1 +
c
l c
_
lc
_
1 +
c
l c
_
c
_
2T
S(K) exp (c2T|) (l ).
As l > c was arbitrarily chosen, the limiting bound holds. Replacing S(K) by its denition
in (4.37), we obtain
E[[F
K,1
G
T
[[
T
S(K) exp (c2T|)
= T
_
(2a +b +c)
_
1
K
+
1
K
_
exp (c2T|) ,
Together with (4.32),
R
1
c(T + 2)E[[G
T
F
K,1
[[
T
cT(T + 2)
_
(2a +b +c)
_
1
K
+
1
K
_
exp (c2T|) .
39
Applying this bound together with those in (4.25) and (4.26) to the expression in (4.24), we
have
m

j=1
Ef
(j)
(
T
K
,
k
), k = 1, . . . , m)
T

T
K
,
j
)

K
+

K
+

K
+cT(T + 2)
_
(2a +b +c)
_
1
K
+
1
K
_
exp (c2T|) .
The completes the proof of Theorem 4.4.2.
40
Chapter 5
Research Direction
In this chapter, we review the quality of the model presented in Chapter 4 as this will shed
light into future research plans. We also present an outline of our future work.
5.1 Model review
While we have successfully extended the SIR model in Chapter 4 to the SEIR model, the
bound obtained for our approximation is a uniform bound and is rather crude. To give an
idea of the quality of the bound derived in Theorem (4.4.2), for c 1 and very large K, the
bound is less than 1 provided K is at least
()
2
T
4
e
4T
.
Thus, we expect the bound to be a rather weak one in general. Consequently, the deter-
ministic approximation is valid only during the initial stages of an epidemic or if is very
small, i.e., the epidemic process is approximately a simple death process.
This fact is manifested in the work of Marumo [16]. Based on the theories developed by
Reinert ([19] and [21]), Marumo [16] simulated the SIR model to compare the outbreak paths
by varying the distribution of the infectious period and the population size. The simulations
were carried out in the following settings:
Case 1 A serious outbreak in a small population.
Case 2 A serious outbreak in a large population.
Case 3 A small outbreak in a large population.
For Cases 1 and 3, the deterministic approximation and the simulated epidemic trajectories
deviate visibly; in contrast, the approximation is relatively good for Case 2. This is expected
because the SIR model under investigation is appropriate under the setting of Case 2 (i.e.,
large population size and that the epidemic has already taken o). This is consistent with
the results of Reinert [21]. She showed that if a deterministic approximation is used only
after the epidemic has taken o (i.e., when there is a sizable population of infected) and
before the epidemic trails o, the deterministic approximation in the SIR model is much
improved the bound on the distance to the mean-eld limit is linear in T, rather than
41
exponential in T (as in Theorem 4.4.2). We expect a similar result to hold for the SEIR
model.
Much of the uncertainty in approximating an epidemic lies in the initial and nal stages.
A natural strategy is therefore to partition the epidemic path into three the cut-o point
being a disease-dependent threshold value, d
0
. This threshold value d
0
is the minimum
number of infectives that will ensure that the disease persists.
For the initial and nal stages (when the number of infectives is below the threshold value
d
0
), we apply Theorem 4.4.2 or use a branching process approximation (Ball and Donnelly
[4]). Thus we do not expect to get a signicantly better bound than the exponential bound
that we have obtained in Theorem 4.4.2. However, we will improve on the error bounds for
the middle section of the epidemic process (similar to Reinert [21]). At present, the bound
is linear in T and it is hoped that this bound can be sharpened by removing the linearity in
T.
In addition, we will extend existing theories on Gaussian approximation in measure spaces
to obtain tighter bounds on the epidemic parameters. This will be very helpful in the
estimation of such parameters as we will then be able to construct condence bounds around
our deterministic approximation.
5.2 Generalisation of the model
One of the key assumptions in the model is that the population is idealised to be homoge-
neously mixing . This could be true in the initial part of an epidemic; however, as awareness
about the epidemic increases during the course of the epidemic, human intervention will
nullify the homogenous mixing assumption. From here, the spread of the disease could be
more closely modelled by taking into account spatial structure the spread of the disease
should occur in pockets where there are some infectives present, rather than uniformly
across the whole population. Thus the model has to be generalised further to accommodate
this phenomenon.
In addition, the virulence of dierent strains of diseases could also aect the course of
an epidemic. This has to be investigated further to determine if a more complex model is
required.
Another area of work would be to improve the sensitivity of the functional tracking the
infection pressure, which, at present is dependent only on time and the number of infectives.
Perhaps, it may be that individuals who are incubating the disease (exposed state) are
already capable of spreading the disease, but to a lesser degree than infective individuals.
Mathematically, we will then have to replace (4.1) by
A
K
i
= inf
_
t R
+
:
_
(0,t]
(s, E
K
(s), I
K
(s))ds = l
i
_
.
5.3 Parameters estimation and model testing
Before model testing can take place, we need to be able to estimate any unknown or im-
measurable variables in the analysis. This is particularly important in the case where the
42
observed data is incomplete.
For our model, we need to estimate the following:
the distribution of the disease incubation time (for the exposed), duration of the infec-
tive period. This is required for both the initially exposed or infective, as well as the
initially susceptibles. Medical experience could be used for the initial estimates.
the distribution of individual thresholds to the disease.
the parameters associated with these distributions.
the functional .
These estimations have to be based on empirical data and it is intended that data from
the 2003 Severe Acute Respiratory Syndrome (SARS) epidemic in Singapore be used for this
purpose.
Once we have all these estimates, we will be able test the accuracy of our model against
empirical data. Comparison can also be carried out against established models such as the
improved Reed-Frost model by Barbour and Utev [9].
43
Bibliography
[1] Anderson, R.M. & May, R.M. (1991). Infectious disease of humans; dynamics and
control. Oxford University Press, Oxford.
[2] Andersson, H. & Britton, T. (2000). Stochastic epidemic models and their sta-
tistical analysis. Springer Verlag.
[3] Bailey, N.T.J. (1975). The mathematical theory of infectious diseases and its impli-
cation. Grin, London.
[4] Ball, F. G. and P. Donnelly, P. (1995) . Strong approximations for epidemic
models. Stoch. Process. Appl. 55, 121.
[5] Baldi, P., Rinott, Y., and Stein, C. (1989). A normal approximation for the
number of local maxima of a random function on a graph. In Probability, Statistics and
Mathematics, Papers in Honor of Samuel Karlin. T.W. Anderson, K.B. Athreya and
D.L. Iglehart, eds., Academic Press, 5981.
[6] Barbour, A.D. (1974). On a functional central limit theorem for Markov population
processes. Adv. Appl. Probab. 6, 2139.
[7] Barbour, A.D. (1975). The duration of the closed stochastic epidemic. Biometrika
62, 477482.
[8] Barbour, A.D. (1990). Steins method for diusion approximations. Probability The-
ory and Related Fields 84, 297322.
[9] Barbour, A.D. & Utev, S. (2003). Approximating the Reed-Frost epidemic pro-
cess.Stoch. Procs. Applics. (to appear)
[10] Bartlett, M.S. (1949). Some evolutionary stochastic processes. J. R. Statist. Soc.,
Ser. B 11, 211229.
[11] Daley, D.J. & Gani, J. (1999). Epidemic modelling: an introduction. Cambridge
University Press.
[12] Daniels, H.E. (1965). The distribution of the total size of an epidemic. Proc. 5th
Berkeley Symp. 4, 281293.
[13] Ethier, S.N. and Kurtz, T.G. (1986). Markov processes: characterization and
convergence. Wiley, New York.
44
[14] Isham, V. (1995). Stochastic models of host-macroparasite interation. Ann. Appl.
Probab. 5, 720740.
[15] Keeling, M. (2000). Multiplicative moments and measures of persistence in ecology.
J. Theo. Biol. 205, 269281.
[16] Marumo, K. (2003). The quality of the deterministic approximation for the SIR
epidemic model - a simulation study. M. Sc. Dissertation, Oxford.
[17] Nagaev, A.V. & Startsev, A.N. (1970). The asymptotic analysis of a stochastic
model of an epidemic. Theory Prob. Applic. 15, 98107.
[18] Reinert, G. (1994). A weak law of large numbers for empirical measures via Steins
method. Ann. Probab. 23, 334354.
[19] Reinert, G. (1995). The asymptotic evolution of the General Stochastic Epidemic.
Ann. Appl. Probab. 5, 10611086.
[20] Reinert, G. (1998). Couplings for normal approximations with Steins method. In
DIMACS Series in Discrete Mathematics and Theoretical Computer Science 41, 193
207.
[21] Reinert, G. (2001). Steins method for epidemic processes. In Complex Stochastic
Systems. O.E. Barndor-Nielsen, D.R. Cox and C. Klueppelberg, eds., Chapman and
Hall, Boca Raton etc., 235 275.
[22] Sellke, T. (1983). On the asymptotic distribution of the size of a stochastic epidemic.
J. Appl. Prob. 20, 390394.
[23] Stein, C. (1972). A bound for the error in the normal approximation to the distribu-
tion of a sum of dependent random variables. Proc. Sixth Berkeley Symp. Math. Statist.
Probab. 2, 583602. Univ. California Press, Berkeley.
[24] Stein, C. (1986). Approximate computation of expectations. Institute of Mathematical
Statistics Lecture Notes - Monograph Series, Vol. 7, Hayward, California.
[25] Wang, F.S.J. (1975) Limit theorems for age and density dependent stochastic popu-
lation models. J. Math. Bio. 2, 373400
[26] Wang, F.S.J. (1977) A central limit theorem for age-and-density-dependent popula-
tion processes. Stoch. Proc. Appl. 5, 173193.
45

Você também pode gostar