Upload
matthew-bickley
View
70
Download
1
Tags:
Embed Size (px)
Citation preview
Mathematical Modelling of Infectious Disease: A Stochastic Approach
0903642
October 20
University of Wolverhampton
Faculty of Science & Engineering
School of Mathematics & Computer Science
Dissertation Title: Mathematical Modelling of Infectious
Disease: A Stochastic Approach
Student Name: Matthew Bickley
Student ID: 0903642
Supervisor: Nabeil Maflahi
Award Title: MSc Mathematics
Presented in partial fulfilment of the assessment requirements for the above award.
This work or any part thereof has not previously been presented in any form to the University
or to any other institutional body whether for assessment or for other purposes. Save for any
acknowledgements, references and/or bibliographies cited in the work, I confirm that the
intellectual content of the work is the result of my own efforts and of no other person.
It is acknowledged that the author of any dissertation work shall own the copyright. However,
by submitting such copyright work for assessment, the author grants to the University a
perpetual royalty-free licence to do all or any of those things referred to in section 16(i) of the
Copyright Designs and Patents Act 1988 (viz: to copy work; to issue copies to the public; to
perform or show or play the work in public; to broadcast the work or to make an adaptation of
the work).
Signature:
Date: 03/10/2014
1
i. Dissertation Declaration
This document must accompany all dissertation document submissions
PLEASE READ THIS VERY CAREFULLY.
The University considers seriously all acts of Academic Misconduct, which by definition are dishonest and in direct opposition to the values of a learning community. Misconduct may result in penalties ranging from the failure of the assessment to exclusion from the University.
Further help and guidance can be obtained from your academic tutor or from our guide on How to avoid Academic Misconduct – available at http://www.wlv.ac.uk/skills
By submitting this document for assessment you are confirming the following statements
I declare that this submission is my own work and has not been copied from someone else or commissioned to another to complete.
Any materials used in this work (whether from published sources, the internet or elsewhere) have been fully acknowledged and referenced and are without fabrication or falsification of data.
I have adhered to relevant ethical guidelines and procedures in the completion of this assignment.
I have not allowed another student to have access to or copy from this work.This work has not been submitted previously.
By this declaration I confirm my understanding and acceptance that –
1. The University may use this work for submission to the national plagiarism detection facility. This searches the internet and an extensive database of reference material, including other students’ work and available essay sites, to identify any duplication with the work you have submitted. Once your work has been submitted to the detection service it will be stored electronically in a database and compared against work submitted from this and other Universities. The material will be stored in this manner indefinitely.
2. In the case of project module submissions, not subject to third party confidentiality agreements, exemplars may be published by the University Learning Centre. I have read the above, and declare that this is my work only, and it adheres to the standards above.
Signature: Date: 03/10/2014
Print Name: MATTHEW BICKLEY Student ID: 0903642
2
I have submitted this digitally and will provide a signed copy prior to marking
3
ii. Acknowledgments
I would like to thank my family for putting up with me during my academic career so far, in
the last five years since I started my undergraduate degree at Keele University, but especially
the past two years during my PGCE and postgraduate studies which has been difficult for not
just me, but my family as a whole. I would also like to thank Nabeil Maflahi for supervising
this dissertation and making sure my ideas stay on track and that I get to the point of what I
am trying to show, alongside putting up with my random chats and lengthy meetings when I
am worrying that all is wrong. I would also like to thank the other lecturers at the University
of Wolverhampton who have taught me various high-level topics in the past year to add on to
my always improving mathematical knowledge. I would like to thank my lecturers at Keele
University, specifically Martin Parker, David Bedford, Neil Turner and Douglas Quinney,
amongst others who all taught me the main areas of mathematics that I love. Finally, I would
also like to thank my Bro, Ian. He has been there for me since I meet him at Keele throughout
all of my trials and tribulations. He has been and still is in a similar position to me since we
met, on a personal, academic and career basis. We have been through a lot together and he has
always been there when things were awesome and when they were not so great. Without him,
I may not have had the confidence and vision to make sure I get what I want from my career.
He is my Bro and that will never change. Thanks, Bro. Maybe Ellen too…
4
i. Abstract
Mathematical models are very useful for representing a real world problem. In particular,
these models can be used to simulate the spread of an infectious disease through a population.
However, standard approaches using deterministic rates for the flow of individuals from state
to state can pose problems when the disease in question becomes more complex. Numerical
methods can be used to help analyse a system, but this still does not take away from the fact
that some factors are ignored to allow for solutions to be found.
Stochastic processes have inherent properties that allow for random events to take place.
Markov chains, specifically, allow modelling of an individual within a system where
transitions are described by probabilities of changing state. The probabilities can be calculated
from real world data and due to this, will incorporate a multitude of extra influences that the
deterministic model had to assume insignificant.
After analysing three diseases using standard ordinary differential equations, these three
diseases were then also modelled using a Markov chain. Comparing the results of these
analyses, we find that the stochastic approach did not fundamentally give better or worse
results than that of the traditional method. Although the stochastic method does include
random effects and other factors not explicitly accounted for before, it appears as if the
original assumptions made were justified and made no real different to the results. One of the
adapted models was hugely inaccurate with respect to real life cases, but this was due to the
fact that the original model was also poor. Further work would be to stochastically model
diseases with Markov chains which cannot be solved analytically using standard techniques
and incorporating more and more complexity to each model to allow for the most accurate
results possible.
5
ii. Contents
i. Dissertation Declaration......................................................................................................3
ii. Acknowledgments...............................................................................................................4
iii. Abstract............................................................................................................................5
iv. Contents...........................................................................................................................6
1. Introduction.........................................................................................................................9
2. Deterministic Modelling – an overview............................................................................11
3. Deterministic Modelling – analysis...................................................................................13
3.1. Model 1 – Chickenpox...............................................................................................13
3.1.1. Compartmental model.........................................................................................13
3.1.2. Assumptions........................................................................................................14
3.1.3. Representing the model.......................................................................................15
3.1.4. Finding the steady states.....................................................................................16
3.1.5. Finding R 0..........................................................................................................18
3.1.6. Analysing the data...............................................................................................18
3.2. Model 2 – Measles.....................................................................................................21
3.2.1. Compartmental model.........................................................................................21
3.2.2. Assumptions........................................................................................................23
3.2.3. Representing the model.......................................................................................24
3.2.4. Finding the steady states.....................................................................................25
3.2.5. Finding R 0..........................................................................................................27
3.2.6. Analysing the data...............................................................................................28
3.3. Model 3 – H1N1.........................................................................................................31
3.3.1. Compartmental model.........................................................................................31
3.3.2. Assumptions........................................................................................................33
3.3.3. Representing the model.......................................................................................34
3.3.4. Finding the steady states.....................................................................................36
3.3.5. Finding R 0..........................................................................................................38
3.3.6. Analysing the data...............................................................................................39
4. Overview of deterministic models.....................................................................................42
6
5. Stochastic Processes – an overview...................................................................................45
6. Stochastic Modelling – adaptations and analysis..............................................................49
6.1. Model 1 adaptation – Chickenpox.............................................................................49
6.1.1. Stochastic model.................................................................................................49
6.1.2. Transition matrix.................................................................................................50
6.1.3. n-step transition...................................................................................................54
6.1.4. Limiting distribution...........................................................................................55
6.2. Model 2 adaptation – Measles....................................................................................57
6.2.1. Stochastic model.................................................................................................57
6.2.2. Transition matrix.................................................................................................58
6.2.3. n-step transition...................................................................................................63
6.2.4. Limiting distribution...........................................................................................65
6.3. Model 3 adaptation – H1N1.......................................................................................68
6.3.1. Stochastic model.................................................................................................68
6.3.2. Transition matrix.................................................................................................70
6.3.3. n-step transition...................................................................................................74
6.3.4. Limiting distribution...........................................................................................75
7. Results and model comparisons........................................................................................78
7.1. Model 1 comparisons.................................................................................................80
7.2. Model 2 comparisons.................................................................................................81
7.3. Model 3 comparisons.................................................................................................81
8. Conclusions.......................................................................................................................83
8.1. Conclusion and critical evaluation.............................................................................83
8.2. Further work...............................................................................................................85
9. References.........................................................................................................................87
10. Bibliography..................................................................................................................93
11. Appendices.....................................................................................................................94
11.1. Appendix A.............................................................................................................94
11.1.1. Appendix A1.......................................................................................................94
11.1.2. Appendix A2.......................................................................................................95
11.1.3. Appendix A3.......................................................................................................96
11.1.4. Appendix A4.......................................................................................................98
11.1.5. Appendix A5.......................................................................................................99
7
11.1.6. Appendix A6.....................................................................................................101
11.1.7. Appendix A7.....................................................................................................102
11.1.8. Appendix A8.....................................................................................................103
11.1.9. Appendix A9.....................................................................................................104
11.1.10. Appendix A10...............................................................................................105
11.1.11. Appendix A11...............................................................................................106
11.2. Appendix B...........................................................................................................107
11.2.1. Appendix B1.....................................................................................................107
11.2.2. Appendix B2.....................................................................................................108
11.2.3. Appendix B3.....................................................................................................109
11.3. Appendix C...........................................................................................................110
8
1. Introduction
Mathematical models are a widely used system to represent problems in the real world. They
can be implemented represent almost anything from elements of the natural sciences; such as
physics, engineering and biology (PHAST 2011, Adeleye 2014), to computer science (Neves
and Teodoro 2010) and social sciences including business and economics (Reiter 2014) and
sociology (Inaba 2014). They can also represent how parts of language can be used (Pande
2012) and how education can adapt (Stohlmann et al. 2012). Mathematical models are
primarily used to study the effect of specific components and parameters involved in the
models, and use this and further analysis to predict behaviour and outcomes of some initial
input.
Models that represent biological situations can again be specified into particular areas; one of
which has been and will continue to be vitally important to the real world – mathematically
modelling infectious disease. Diseases have existed as long as life itself and will only
continue to be a problem to humans and other organisms for a long time to come. It is
therefore imperative that diseases can be outsmarted – that is to say we need to be able to
know, or at least predict with some certainty, exactly how a disease behaves. This is where
mathematical models come in. Using past data and specific discovered or known traits of a
particular disease, we can apply modelling knowledge to calculate how the disease will
spread, how infectious or deadly it may be, and how best to ensure the disease has the
minimal effect on the population.
However, these infections do not always behave the same from one outbreak to another, even
when looking at the same specified strain of the disease (Schmidt-Chanasit 2014, in Hille
2014). Other factors which may be very difficult or impossible to model can affect the
solutions in an unpredictable way. These may be but are not limited to environmental,
temporal, climate or population effects. This is where stochastic processes can be applied.
Stochastic processes are methods which inherently involve some amount of probability or
9
randomness, which diseases can show signs of when a small number of individuals infected
cause an outbreak (Spencer 2007). In this way, the solution to a stochastically-involved
system gives probabilities that certain events will have occurred, and exactly what this means
in the long-term. Although this can initially give the impression of misleading outcomes, it in
fact shows the variety that the system can produce, giving us a better overview of the situation
and then hopefully allowing us to pick the best real-world solution.
This study will look at a variety of diseases, from those which are not deadly but can cause
uncomfortable living, to those which have high mortality rates and or can reoccur in the same
individual over a period of time. At first, a standard albeit complex deterministic approach
will be used for each disease – a standard mathematical model. Each model will then be
adapted to include a stochastic basis for the disease in question, hopefully showing a similar
or if possible, better picture of the disease’s traits. Every model will be compared to real-
world data to see the appropriateness of each model and then to each other, to see overall if a
stochastic approach is better and if so, to what extent compared to the complexity increase
from standard systems. This will bring together knowledge about mathematical modelling (of
diseases) with that of stochastic models and processes. This application assumes that
stochastic properties can be applied to diseases that are usually modelled using standard
deterministic approaches which can help either represent the models better, in an alternative
way giving the same or similar results or give entirely different information which could be
used advantageously.
For the analysis of disease in this research, three diseases will be picked that can be
represented by a standard, deterministic mathematical model. The diseases were picked from
a list that can be modelling using standard techniques (CMMID 2014). Chickenpox, which
causes no (or few) deaths from the disease was picked to start with, followed by measles, a
disease with deaths and maternal immunity, and then the final disease, the H1N1 virus was
picked to allow for disease reoccurrence in the model and also to allow modelling and
comparisons of a disease which is usually almost non-existent at most times, but then
becomes very prevalent within populations when outbreaks occur.
10
2. Deterministic Modelling – an overview
“A model is a simplified version of something that is real.” (Schichl 2004 p.28). Taking this
and applying it to a mathematically specific interpretation gives that a mathematical model is
a simplified version of a real-world event that helps us to solve an existing problem.
Mathematical models of infectious diseases are those which help us see what a disease is
doing and how best to combat the infection. William Hamer and Ronald Ross were the
earliest known pioneers of infectious disease modelling, giving a basis to today’s current
models in the early part of the twentieth century. Within the next two decades, further
research, data and scientific knowledge had allowed them, along with other researchers to
develop their own epidemic model which showed a clear and logical relationship between
individuals in a population – they could be susceptible, infected or immune. This is the root of
the susceptible-infected-recovered (SIR) model (Weiss 2013). This simplistic model is
effectively just three states, where an individual fits into one and only one. Movement of
individuals between the states is described by deterministic parameters that help calculate the
rate of change of the sizes, or proportions, of these states. Analysing this gives an easily
interpretable system which shows the dynamics of individuals moving through the system,
which in turn, shows the basic properties that a disease exhibits to cause an outbreak.
This model can be applied to a wide variety of diseases, but mainly those which transfer when
humans, or any other animals, directly come into contact with each other via touch or close-
proximity sneezing and coughing, for example (Weiss 2013). So modelling this way does not
(easily) allow for diseases which are purely airborne or waterborne, for example, to be
simulated. Adaptations would be needed, making new types of systems to represent these
types of infections. Focusing just on those diseases which can transmit when the host in
question is coming into contact with another allows for much simpler analysis yet efficacy.
In the last century, these models have been developed to include many more compartments,
extra parameters that depend on disease traits and far more complicated systems of equations
11
due to this. Extra states have been added in certain systems such as ‘exposed’, where
individuals belong who have the disease but are not infectious yet, ‘carrier’, where individuals
belong who can pass on the disease but never themselves suffer from it (and likely keep the
disease for life) and ‘quarantine’, where individuals who have the disease are placed so as to
no longer pass on the disease but still need time to recover. Limitations on these states can
also be introduced, such as a space limit in quarantines so that once the state is ‘full’, no more
individuals can enter and must stay put in these rest of the system. Parameters involved in the
system normally depend on the compartments considered, but some are added to existing
models such as a vaccination rate or a separate birth rate that takes into account all individuals
born with the disease.
These representations can also be extended from a standard system of linear ordinary
differential equations with deterministic, explicit parameters, to more complex systems such
as partial differential equations which look at spatial variances and systems that have implicit
or time-based parameters. The more complicated the original disease is, the more complex the
system of equations must be to accurately represent the virus. However, analytical solutions
are only feasible on the simpler of models. Numerical methods must be used on the most
complex, which have their advantages and disadvantages. Some converge to solutions very
quickly but the sizes of errors during the calculations are larger than others. Methods with the
smallest of errors are the best possible approximation to the solutions of these systems, but
convergence is slow and the method, albeit doable, is intricate.
12
3. Deterministic Modelling – analysis
1.1. Model 1 – Chickenpox
To start with, modelling a disease than can be represented with a very simple model will give
a strong basis for the later, adapted model and comparison. Chickenpox, otherwise known
medically as varicella, which is a relatively basic and mild disease caused by the varicella-
zoster virus (VZV) spreads quickly and easily between individuals (NHS 2014a) but is rather
simplistic to model – it only very rarely (directly) causes deaths in the human population.
Symptoms of chickenpox are similar to that of influenza and often include sickness, high
temperatures, aching muscles, headaches, loss of appetite and most obviously, rashes of spots
which causes itching and irritation (NHS 2014b). These spots are the main way of the
infection spreading from person to person – and open spots are especially infective. Close
contact with individuals who have chickenpox themselves is the main cause of mass spread of
the disease, so reducing contact will lower the possibility that any individual will be able to
pass it on.
1.1.1. Compartmental model
Assuming the simplicity of this model, we can represent it with one of the simplest
compartmental models; that being, one that only includes one extra state besides the usual
three. This is the ‘susceptible-exposed-infectious-recovered’ model, or the SEIR model. This
is an adaptation of the ‘susceptible-infected-recovered’ model, or the SIR model, as it allows
for individuals who gain chickenpox to harbour the disease before being able to spread it
(exposed). Also, the ‘infected’ state is renamed ‘infectious’, to lessen the confusion as to what
individuals are capable of doing with the disease in each state. This fits with chickenpox, as
individuals will go through this stage before infecting others. After looking at how the
infection spreads, we can construct the first standard black-box model (Fig. 3.1. and Fig. 3.2.):
13
Figure 3.1.: Compartmental model for chickenpox
Or more mathematically:
Figure 3.2.: Compartmental model for chickenpox with mathematical symbols
1.1.2. Assumptions
The following assumptions can be made about this model with this disease. The aim of these
assumptions is to make the model viable and to enable a solution to be found, when at the
14
Natural Deaths
Recovered
Recovery rate
Natural Deaths
Infectious
Natural Deaths
Natural Deaths
Births
ExposedSusceptible
Transmission rate
Development rate
μ
Rγ
μ
I
μ μ
μ
ESβ σ
same time not taking away too much from the original traits of the disease and general
complexity of the disease transmission.
No individual is born with the virus (all individuals initially enter the susceptible
state).
The population has a life expectancy of μ, so that the natural mortality rate is 1μ
.
A closed population, so that the birth rate is also 1μ
.
The virus has a latency period of 1σ
.
A constant proportional population implies the individuals in the system add up to 1
(individuals either are susceptible, infected or have recovered, and nothing else):
S+E+ I +R=1 (1)
where S ( t ) , E (t ) , I ( t ) ,R ( t )≥ 0 , ∀ t and we simplify S (t )=S, E ( t )=E, I ( t )=I and R ( t )=R.
1.1.3. Representing the model
Using the above assumptions, with β, the transmission rate being the product of the contact
rate and probability of a successful transmission, γ , the recovery rate, and σ , the development
rate, we can construct the following ordinary differential equations (ODEs) to represent the
rate of change of each compartment in the model:
dSdt
=μ−βSI−μS (2)
dEdt
=βSI−σE−μE (3)
dIdt
=σE−γI−μI (4)
dRdt
=γI−μR (5)
15
where β ,σ , γ , μ>0.
1.1.4. Finding the steady states
At the steady states to this system of equations, there will be no change to the proportions
represented by S, I and R as time continues to pass. Therefore, the rates of change are equal
to zero. Hence:
dSdt
=0 ,dEdt
=0 ,dIdt
=0 ,dRdt
=0
Therefore:
μ−βSI−μS=0 (6)
βSI−σE−μE=0 (7)
σE−γI−μI=0 (8)
γI−μR=0 (9)
Rearranging all of (6) to (9) to make I the subject gives:
I=μ (1−S )
βS(10)
I=(σ+μ ) E
βS(11)
I= σEγ+μ
(12)
I=μRγ
(13)
If I=0 (no virus present), (10) ⟹S=1
16
(12) ⟹ E=0
(13) ⟹ R=0
Checking in (1), S=1, E=0, I=0 and R=0 clearly satisfy S+E+ I +R=1.
This gives the trivial steady state:
( S , E , I , R )=(1 , 0 , 0 ,0 ) (14)
If I ≠ 0 (virus present), then Maple confirms the following solution (along with the trivial
steady state):
>
Checking in (1), S=(σ+μ ) (γ+μ )
βσ, E=
μ ( βσ−(σ+μ)(γ+μ))βσ (σ+μ)
, I=μ ( βσ−(σ+μ)(γ +μ))
β (σ +μ)(γ+μ) and
R=γ ( βσ−(σ+μ)(γ+μ))
β (σ+μ)(γ+μ) satisfy S+E+ I +R=1 (manually checked).
This solution along with some manual working leads to the non-trivial endemic steady state:
17
( S¿ , E¿ , I ¿ , R¿)=(( σ+μ ) (γ+μ )
βσ,
μ ( βσ−(σ+μ)(γ+μ))βσ (σ+μ)
,
μ ( βσ−(σ+μ)(γ+μ))β(σ+μ)(γ +μ)
,
γ ( βσ−(σ+μ)(γ+μ))β (σ+μ)(γ +μ)
)(15)
1.1.5. Finding R0
For a closed population (which we have here), the critical value at which the virus becomes an
epidemic is when it exceeds 1S
. So, for this model:
R0=1S= βσ
(σ+μ ) ( γ+μ )
If R0 is above 1, the theory suggests that an epidemic starts:
R0=βσ
( σ+μ ) (γ+μ )>1
⟹ βσ>(σ+μ ) ( γ+μ ) (16)
Interpreting this inequality, by reducing the transmission rate or increasing the mortality rate
or recovery rate, we can cause a disease to die out in this instance. Increasing the mortality
rate is unethical within most populations, especially for those diseases affecting humans
(culling of other species can be introduced on a case by case basis), and the recovery rate is
generally unchangeable due to traits of the virus, along with the latency period due to the
properties of the disease. Hence reducing β, so in turn, reducing the contact rate is the best
way to prevent an epidemic in this model. Keeping those affected by the chickenpox away
from others is by far the most effective way of reducing the spread.
18
1.1.6. Analysing the data
19
Assuming a standard scenario for chickenpox (an average rate of spread), we take an estimate
for the basic reproduction number, R0, as equal to 3.83 for England and Wales (Nardone et al.
2007). It is stated that this is just an estimate, with a 95 % confidence interval (CI) given as
(3.32 – 4.49 ). By taking a mid-range value, we can model for the average spread in the United
Kingdom (UK). The latency period, 1σ
, is given as ten to fourteen days (Knott 2013), so we
take the average of twelve days. The recovery period, 1γ
, is also reported as ten to fourteen
days (Lamprecht 2012), so again, we take the average of twelve days. Average life
expectancy at birth in the UK is approximately 81 years (WHO 2012), so this is taken to be
the value of 1μ
. All parameters are rates, so all must be converted to the same units. Numerical
analysis will be taken at daily time periods with step size h=15
over the period of 365 days, so
all parameters are multiplied or divided as needed to give each value in terms of days. We can
substitute these values into R0=βσ
( σ+μ ) (γ+μ ) and rearrange, giving us a β value (transmission
rate) of 0.319426 (six significant figures). This system has been analysed using the 4th-order
Runge-Kutta method for a more robust approach, with an initial susceptible individual
proportion, S (0 ) of 0.999 and infected proportion, I (0 ) of 0.001 as a starting point (see
Appendix B1 for data).
20
0 50 100 150 200 250 300 350 4000.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
S
E
I
R
S*
E*
I*
R*
Time (days)
Pro
port
ion
of p
opul
atio
n
Figure 3.3.: Chickenpox 4th-order Runge-Kutta numerical analysis (with steady state
values)
As Fig. 3.3. shows, the chickenpox outbreak eventually subsides but does produce a worrying
epidemic initially. After around forty days, the exposed and infectious proportion starts to rise
rapidly and then both peaking just before day 100 at around 20 %, giving approximately 40 %
of the population having chickenpox at this point. This would most certainly be classified as
an epidemic if the outbreak spread across a national population, such as the entire UK. The
data again shows that a majority of people have the virus as early as day 100, as half the
population have entered the recovered state at this point. From the year-long graph, it show
the levels of all four compartments fluctuate somewhat, with S ever decreasing and R ever
increasing, and E and I showing the peak where the outbreak is at its worst before lowering
back down again. The S and R states eventually level out in the long-term, moving toward the
respective dotted lines shown, giving a steady proportion of the population who are present in
each state.
By substituting in the values for β, σ , γ and μ in (S¿ , E¿ , I ¿ , R¿) and using Maple to work them
out (see Appendix A1), we can find out the behaviour of the model analytically as t tends to
infinity (also shown in Appendix B1 as t → ∞):
(S¿ , E¿ , I ¿ , R¿)=(0.261097 , 0.000299789 , 0.000299667 , 0.738304)
So eventually, there is a constant 26.1 % of the population that are susceptible, 73.9 % that
have had chickenpox at some point and recovered, and a negligible amount suffering from
chickenpox at any time (¿0.06 %).
21
1.2. Model 2 – Measles
Now, a disease with an extra state is introduced to try to see if the adaptation later will work
with a different basis model. Measles, otherwise known as morbilli, English measles or
rubeola, and not to be confused with rubella (German measles), is a disease caused by a
paramyxovirus. This disease is a little more complex to model as individuals are more often
than not born with maternal immunity to measles which lasts sometime into the second year
of life (Nicoara et al. 1999) and can cause deaths directly from the disease in the human
population. Measles has a relatively high R0 value, meaning it is more likely to cause
epidemics but spread can be reduced significantly by use of vaccinations (CDC 2014).
Symptoms of measles are similar to that of a cold, with red eyes, high temperatures and spots
appearing in the mouth and throat (NHS 2013). These spots are the main cause of the spread
of measles; tiny droplets are ejected from the nose or mouth when sneezing or coughing
respectively (NHS 2013), and then infect others close by or land on surfaces or objects. In
turn, these are then touched the hands of others and through improper and irregular hand-
washing, this causes the infection to successfully transmit. Being in close proximity with
individuals who have measles is the main cause of mass infection, so removing contact and
introducing vaccinations will lower the possibility that any one individual will be infected.
1.2.1. Compartmental model
We can represent measles with a slightly more complex compartmental model; that being, one
that only includes one extra state besides the previously used four. This is the ‘maternal
immunity-susceptible-exposed-infectious-recovered’ model, or the MSEIR model. This is an
adaptation of the SIR model, as it allows for individuals who gain measles to harbour the
disease before spreading it (exposed). Again, the ‘infected’ state is renamed ‘infectious’, to
lessen the confusion. Also, a maternal immunity state is added. Individuals born enter this
category initially before be able to contract measles (being susceptible). This fits with
measles, as individuals will have immunity for a small time before losing it and entering a
standard SEIR part of the model. Vaccinations are also allowed in this model; individuals who
22
are vaccinated are only those in the susceptible state and will move into a recovered position
after the immunisation, preventing them catching the disease (all vaccinations are assumed to
be 100 % effective). After looking at how the infection spreads, we can construct the second
standard black-box model (Fig. 3.4. and Fig. 3.5.):
Figure 3.4.: Compartmental model for measles
Or more mathematically:
23
Gain of active immunity (vaccination)
Loss of passive
immunity
Development rate
Natural Deaths
Virus Deaths
Natural Deaths
Virus Deaths
Natural Deaths
Natural Deaths
Natural Deaths
Births (with passive
immunity)
Susceptible
Maternal Immunity
RecoveredInfectiousExposed
Transmission rate
Recovery rate
X
κ
δ σ
μ
χ
μ
χ
μμμ
μ
SM RIEβI γ
Figure 3.5.: Compartmental model for measles with mathematical symbols
When mathematically representing this model, we need to create a new state. A new death
state, X , is introduced to account for those who die due to the disease only.
1.2.2. Assumptions
The following assumptions can be made about this model with this disease. Again, the aim of
these assumptions is to make the model viable and to enable a solution to be found, when at
the same time not taking away too much from the original traits of the disease and general
complexity of the disease transmission.
No individual is born with the virus, but all are born with maternal immunity for some
length of time to it (all individuals initially enter the maternal immunity state).
The population has a life expectancy of μ, so that the natural mortality rate is 1μ
.
A closed population, so that the birth rate is also 1μ
. This only includes individuals
born that replace those who die naturally, so the birth rate relates to every state except
X (so in the first ODE, μ multiplies every state but X – see below).
The virus’ mortality rate is 1χ
, so individuals who die from the virus directly enter the
X state at this rate.
The virus has a latency period of 1σ
.
Vaccinations are only administered to susceptible individuals, not those who may have
the virus but are not showing symptoms or capable or passing the virus on (exposed)
nor infants with passive immunity from their mother (maternal immunity).
Vaccinations are assumed to be 100 % effective from the moment they are
administered (individuals who receive it cannot catch measles at any point in the
future).
24
Deaths caused by the virus have the same prevalence in both the exposed and
infectious individuals (virus mortality is the same whether you are exposed or
infectious).
To allow for the constant proportionality of the model, those who die of the disease
must be modelled to be allowed to ‘die again’ naturally. Although, this would not
happen in real life, it allows the model to maintain the same amount of individuals in
the system and those who do ‘die again’ from the virus death state will ‘re-enter’ as an
individual with maternal immunity. The low rate of virus and natural deaths within the
length of a long-term analysis makes this error relatively small, albeit still present.
A constant proportional population implies the individuals in the system add up to 1
(individuals either have maternal immunity, are susceptible, exposed, infectious, have
recovered or have died from measles, and nothing else):
M +S+E+ I +R+X=1 (17)
where M ( t ) , S (t ) , E (t ) , I ( t ) , R (t ) , X (t )≥ 0 ,∀ t and we simplify M ( t )=M , S ( t )=S, E ( t )=E,
I ( t )=I , R (t )=R and X ( t )=X .
1.2.3. Representing the model
Using the above assumptions, again with β, the transmission rate being the product of the
contact rate and probability of a successful transmission, γ , the recovery rate, and σ , the
development rate, but additionally δ , the rate at which individuals lose their (passive
maternal) immunity, and κ , the constant number of individuals immunised, we can construct
the following ODEs to represent the rate of change of each compartment in the model:
dMdt
=μ−δM−μM (18)
dSdt
=−βSI+δM−κS−μS (19)
dEdt
=βSI−σE−μE− χE (20)
25
dIdt
=σE−γI−μI− χI (21)
dRdt
=γI+κS−μR (22)
dXdt
= χE+ χI −μX (23)
where β ,σ , γ , δ , μ , χ ,κ>0.
1.2.4. Finding the steady states
Again, at the steady states to this system of equations, there will be no change to the
proportions represented by M , S, E, I , R and X as time continues to pass. Therefore, the rates
of change are equal to zero. Hence:
dMdt
=0 ,dSdt
=0 ,dEdt
=0 ,dIdt
=0 ,dRdt
=0 ,dXdt
=0
Therefore:
μ−δM−μM =0 (24)
−βSI +δM−κS−μS=0 (25)
βSI−σE−μE− χE=0 (26)
σE−γI−μI− χI=0 (27)
γI+κS−μR=0 (28)
χE+ χI−μX=0 (29)
Rearranging (24) to make M the subject gives:
26
M= μδ+μ
This shows that M is independent of any other states, and is only dependent on the parameters
given from the virus itself. More specifically, M is independent of I , whether directly or
indirectly through another state, so the M state will be the same for both steady states,
regardless whether an infection is present (I ≠ 0) or not (I=0).
Rearranging all of (25) to (29) to make I the subject and substituting M gives:
I=δμ−(δ +μ ) (κ+μ ) S
β (δ+μ ) S(30)
I=(σ+μ+ χ ) E
βS(31)
I= σEγ+μ+ χ
(32)
I=μR−κSγ
(33)
I=μX− χEχ
(34)
If I=0 (no virus present), (30) ⟹S= δμ(δ+μ ) (κ+μ )
(32) ⟹ E=0
(33) ⟹ R=κSμ
= δκ(δ +μ ) (κ+μ )
(34) ⟹ X= χEμ
=0
Checking in (17) (using Maple), M= μδ+μ
, S=δμ
(δ +μ ) (κ+μ ) , E=0, I=0, R= δκ( δ+μ ) (κ+μ ) and
X=0 satisfy M +S+E+ I +R+X=1:
27
>
>
>
>
>
>
>
This gives the trivial steady state:
( M , S , E , I , R , X )=( μδ+μ
,δμ
(δ +μ ) (κ+μ ),0 , 0 ,
δκ(δ +μ ) (κ+μ )
,0) (35)
This steady state can be simplified even further – no infection in the population makes
vaccinations unnecessary. So we can set κ=0, causing everyone to stay in the M or S states,
with no-one moving to the R state via the infection or vaccination. This makes the steady state
become:
( M , S , E , I , R , X )=( μδ+μ
,δ
δ +μ, 0 , 0 ,0 , 0)
If I ≠ 0 (virus present), then Maple confirms the following solution (along with the trivial
steady state) (solution omitted due to size, see Appendix A2):
>
28
Checking in (17), M=M ¿, S=S¿, E=E¿, I=I ¿, R=R ¿ and X=X¿ (see Appendix A2) satisfy
M +S+E+ I +R+X=1 (manually checked).
This solution along with some manual working leads to the non-trivial endemic steady state
(see Appendix A2):
( M , S , E , I , R , X )=( M ¿ , S¿ , E¿ , I ¿ , R¿ , X¿ ) (36)
1.2.5. Finding R0
For a closed population (which again, we have here), the critical value at which the virus
becomes an epidemic is when it exceeds 1S
. So, for this model:
R0=1S= βσ
σγ+ (μ+ χ ) (σ+γ+μ+ χ )
If R0 is above 1, the theory suggests that an epidemic starts:
R0=βσ
σγ+( μ+ χ ) (σ+γ+μ+ χ )>1
⟹ βσ>σγ +( μ+ χ ) (σ+γ+μ+ χ ) (37)
Interpreting this inequality, again, the simplest way of reducing the epidemic is to reduce the
transmission rate (as this only appears on the left hand side so reducing this quantity will not
affect the right hand side) or increasing the mortality rate or recovery rate (as these only
appear on the right hand, so similar to above). As before, increasing the mortality rate is
unethical and the recovery rate is generally unchangeable, along with the latency period in
this new model due to the properties of the disease. Hence reducing β, so in turn, reducing the
contact rate is the best way to prevent an epidemic in this model. Notice that the vaccination
number does not appear in the value of R0, so the rate at which we vaccinate people has no
effect of preventing the actual outbreak of an epidemic. It appears as if this value only helps
29
prevent the spread once an epidemic has begun. Keeping those affected by measles away from
other individuals in the population is by far the most effective way of reducing the spread.
1.2.6. Analysing the data
Assuming a standard scenario for measles (an average rate of spread), we take an estimate for
the basic reproduction number, R0, as between 12 and 18 for the United States (CDC 2014).
By taking the mid-range value of 15, we can model for the average spread in the United States
(US). The latency period, 1σ
, is given as seven to fourteen days (CDC 2009c), so we take the
average of 10.5 days. The recovery period, 1γ
, is reported as three to five days plus a few days
for the virus to completely subside (CDC 2009c), so again, we take the average of
approximately seven days. Average life expectancy at birth in the US is approximately 79
years (WHO 2012), so this is taken to be the value of 1μ
. Individuals lose maternal immunity
sometime between twelve and fifteen months (Nicoara 1999), so we take the average of 13.5
months. Individuals are vaccinated at approximately 91.9 % coverage per year (CDC 2013)
with a CI of (90.2 %−92.0 % ), so we can then work out a rate per day for κ . Approximately
one or two individuals per every thousand die directly due to measles (CDC 2009b) so we
take the average of 1.5 per thousand and then convert this for the parameter χ . All parameters
are rates, so all must be converted to the same units. Numerical analysis will be taken at daily
time periods with step size h=15
over the period of 365 days, so all parameters are multiplied
or divided as needed to give each value in terms of days. We can substitute these values into
R0=βσ
σγ+( μ+ χ ) (σ+γ+μ+ χ ) and rearrange, giving us a β value (transmission rate) of 2.14431
(six significant figures). Again, this system has been analysed using the 4th-order Runge-Kutta
method, with an initial maternally immune proportion, M (0 ) of 0.1, an initially susceptible
proportion, S (0 ) of 0.899 and infected proportion, I (0 ) of 0.001 as a starting point (see
Appendix B2 for data).
30
Figure 3.6.: Measles 4th-order Runge-Kutta numerical analysis (with steady state values)
As Fig. 3.6. shows, like chickenpox, the measles outbreak eventually subsides but again
produces a worrying epidemic initially, and this time, far worse than chickenpox. After
around only ten days, the exposed and infectious proportion starts to rise rapidly and then
both peaking just before day 25 or so at around 45 % and 20 % respectively, giving
approximately 65 % of the population having measles at this point. This would most certainly
be classified as another epidemic in the population. The data again shows that a majority of
people have the virus or have been vaccinated as early as day 40, as half the population have
entered the recovered state at this point. From the year-long graph, it shows the levels of all
compartments fluctuate somewhat, with M always decreasing, S mostly decreasing, R mostly
increasing and X always increasing but only slightly, and E and I showing the peak where the
outbreak is at its worst before lowering back down again. Like before and as predicted by the
steady state, the all states eventually level out in the long-term, moving toward the respective
dotted lines shown, giving a steady proportion of the population who are present in each state.
31
0 50 100 150 200 250 300 350 4000.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
M
S
E
I
R
X
M*
S*
E*
I*
R*
X*
Time (days)
Pro
port
ion
of p
opul
atio
n
By substituting in the values for β, σ , γ , δ , μ, χ and κ in (M ¿ , S¿ , E¿ , I ¿ , R¿ , X¿) and using
Maple to work them out (see Appendix A3), we can find out the behaviour of the model
analytically as t tends to infinity (also shown in Appendix B2 as t → ∞):
(M ¿ , S¿ , E¿ , I ¿ , R¿ , X¿)=(0.0111266 , 0.0666666 ,0.000159502 ,0.000106306 ,0.921910 ,0.0000314982)
So eventually, there is a constant 6.7 % of the population that are susceptible, 92.2 % that have
had chickenpox at some point and recovered or have been vaccinated, and a negligible
amount suffering from chickenpox at any time (¿0.026 %). Approximately 0.0031 % of the
population will have died from measles in a constant long-term population and about 1.1 %
will have maternal immunity at any time.
32
1.3. Model 3 – H1N1
Finally, a disease with a standard SEIR basis is proposed for analysis but with added loss of
natural immunity; that is, individuals can ‘loop’ in the model and gain the disease more than
once after contracting the virus or being vaccinated. H1N1, known fully as influenza A
(Haemagglutinin type 1-Neuraminidase type 1) or popularly (and incorrectly) as swine flu
(Flu.gov 2014), has exactly this property and exhibits a once-per-lifetime (on average) mass-
outbreak rate, which allows for easier analysis. The last two recorded major outbreaks were in
2009 and 1918 (known as Spanish flu). This disease is still complex to model as individuals
can gain H1N1 twice or more and like measles, it can directly cause deaths in the human
population. Symptoms of measles are that of seasonal flu, but much more extreme (Flu.gov
2014). The main contagious ways of passing the disease on is by coughing or sneezing, which
directly enters another person’s body or is transmitted via surfaces and improper hygiene
routines (CDC 2009d). Being in close proximity with individuals who have flu or related
symptoms is the main cause of spread, so removing contact and making sure individuals get
regular flu vaccinations will lower the chance that anyone will become sick.
1.3.1. Compartmental model
We can represent measles with the most complex compartmental model presented in this
research; that being, one that includes one extra state besides the standard three but allows for
‘looping’. This is the SEIR model again, but includes a rate that allows recovered individuals
can become susceptible again, making it a slightly adjusted model known as the SEIRS
model. This fits with H1N1, as individuals will have immunity from vaccinations or earned
immunity from catching the disease for a time; individuals who are vaccinated are only those
in the susceptible state and will move into a recovered position after the immunisation (again,
all vaccinations are assumed to be 100% effective at the time of immunisation). After looking
at how the infection spreads, we construct our final black-box model (Fig. 3.7. and Fig. 3.8.):
33
Figure 3.7.: Compartmental model for H1N1
Or more mathematically:
Figure 3.8.: Compartmental model for H1N1 with mathematical symbols
When mathematically representing this model, we need to create three new states. Firstly, just
as before, a new death state, X , appears to account for those who die due to the disease only.
Secondly, a ‘bin’ state, B, is created on the right of the model to account for individuals who
34
BD
X
δμμμ
χ
μ
R
μ
IES
χ
βI σ γ
κ
δ
Natural Deaths
Virus Deaths
Natural Deaths
Natural Deaths
Natural Deaths
Recovered
Births
InfectiousExposedSusceptible
Virus Deaths
Transmission rate
Development rate
Recovery rate
Gain of active immunity (vaccination)
Loss of active immunity
lose their active immunity and get H1N1 more than once. We could continue to represent this
with a loop (as in the previous diagram) but it would be impossible to analyse exactly how
many cases had occurred, let alone how many individuals got the disease at some point (at
least once). The individuals who enter the ‘bin’ state are replaced by another identical
individual in the system from a ‘dummy’ state, D, entering the susceptible state where they
are vulnerable to the disease again. The rate that individuals leave the R state to enter B is
exactly the same at which they leave the D state and enter S, and is modelled to only allow
each individual to leave D and enter S when the corresponding individual has left R to enter
B. The loop at the top is therefore removed, and the corresponding rate, δ , is placed on the
arrows in question. The dotted line along the bottom of the diagram represents that B affects
D but no individual actually travels between the two states. No deaths occur in these states (as
the individuals who die when in the ‘bin’ state are accounted for elsewhere in the system by
another state) and the two new states are not needed to be included in the sum for constant
proportionality (see final assumption in section 3.3.2.).
1.3.2. Assumptions
The following assumptions can be made about this model with this disease. Again, the aim of
these assumptions is to make the model viable and to enable a solution to be found, when at
the same time not taking away too much from the original traits of the disease and general
complexity of the disease transmission. This is the most complex of the three models (due to
the possibility of ‘cycling’ round the model via loss of active immunity) so the assumptions
below are comparatively more simple.
No individual is born with the virus (all individuals initially enter the susceptible
state).
The population has a life expectancy of μ, so that the natural mortality rate is 1μ
.
A closed population, so that the birth rate is also 1μ
.
35
The virus’ mortality rate is 1χ
.
The virus has a latency period of 1σ
.
Vaccinations are only administered to susceptible individuals, not those who may have
the virus but are not showing symptoms or capable or passing the virus on (exposed).
Vaccinations are assumed to be 100 % effective at the moment they are administered
(however individuals who receive it could lose immunity and catch H1N1 at another
point in the future).
Deaths caused by the virus have the same prevalence in both the exposed and
infectious individuals (virus mortality is the same whether you are exposed or
infectious).
As before, to allow for the constant proportionality of the model, those who die of the
disease must be modelled to be allowed to ‘die again’ naturally. Although, this would
not happen in real life, it allows the model to maintain the same amount of individuals
in the system and those who do ‘die again’ from the virus death state will ‘re-enter’ as
an individual with maternal immunity. The low rate of virus and natural deaths
compared to the overall length of a long-term analysis makes this error relatively
small, albeit still present.
A constant proportional population implies the individuals in the system add up to 1
(individuals either are susceptible, exposed, infectious, have recovered or have died
from H1N1, and nothing else, with those in the bin or dummy state being accounted
for somewhere else in one of the compartments):
S+E+ I +R+ X=1 (38)
where S ( t ) , E (t ) , I (t ) ,R (t ) , X (t ) , D (t ) , B (t )≥ 0 ,∀ t and we simplify S (t )=S, E ( t )=E, I ( t )=I ,
R ( t )=R, X (t )=X , D ( t )=D and B (t )=B.
1.3.3. Representing the model
Using these assumptions, again with β, the transmission rate being the product of the contact
rate and probability of a successful transmission, σ , the development rate, γ , the recovery rate,
36
δ , the rate at which individuals lose their (active) immunity, and κ , the constant number of
individuals immunised, we can construct the following ODEs to represent the rate of change
of each compartment in the model:
dSdt
=μ−βSI +δR−κS−μ S (39)
dEdt
=βSI−σE−μE− χE (40)
dIdt
=σE−γI−μI− χI (41)
dRdt
=γI−δR+κS−μR (42)
dXdt
= χE+ χI −μX (43)
dDdt
=−δR (44)
dBdt
=δR (45)
where β ,σ , γ , δ , μ , χ ,κ>0.
The final two equations here ((44) and (45)) are not included in the constant proportionality
equation (38) nor are they included in the solutions to the system (using Maple or otherwise).
This is because setting those equations to zero would imply the R state must be zero at any
trivial or endemic state found. Although this does make sense in long term as everyone would
eventually leave the R state one way or another, either by dying naturally and then being
replaced by a new individual in the system to account for the equal birth and death rates, or by
losing their active immunity and ‘looping’ back to enter the system as a susceptible
individual. As the terms in (44) and (45) are repeated from other equations, the system still
holds that the sum of equations (39) to (43) is one. The system would also be unsolvable for
D or B and these states do not appear explicitly in the system of ODEs (an infinite number of
solutions). These final two equations are only included when solving the system numerically
37
using the 4th-order Runge-Kutta method (see Appendix B3). The values of D and B will be a
fraction of R (multiplied by negative and positive delta respectively) and B specifically will
show the proportion of individuals having the possibility of getting the H1N1 virus more than
once.
1.3.4. Finding the steady states
Again, at the steady states to this system of equations, there will be no change to the
proportions represented by S, E, I , R and X as time continues to pass. Therefore, the rates of
change are equal to zero. Hence:
dSdt
=0 ,dEdt
=0 ,dIdt
=0 ,dRdt
=0 ,dXdt
=0
Therefore:
μ−βSI+δR−κS−μS=0 (46)
βSI−σE−μE− χE=0 (47)
σE−γI−μI− χI=0 (48)
γI−δR+κS−μR=0 (49)
χE+ χI−μX=0 (50)
Rearranging all of (46) to (50) to make I the subject gives:
I=μ+δR−(κ+μ ) S
βS(51)
I=(σ+μ+ χ ) E
βS(52)
38
I= σEγ+μ+ χ
(53)
I=(δ+μ ) R−κS
γ(54)
I=μX− χEχ
(55)
If I=0 (no virus present), (53) ⟹ E=0
(55) ⟹ X= χEμ
=0
Rearranging (54) to give R in terms of S and then substituting into (51) gives:
R= κSδ +μ ⟹0=
μ+δ( κSδ+μ )−( κ+μ ) S
βS
⟹S= δ+μδ+κ+μ
⟹ R= κδ +κ+μ
Checking in (38) (using Maple), S=δ +μ
δ +κ+μ, E=0, I=0, R= κ
δ +κ+μ and X=0 satisfy
S+E+ I +R+ X=1:
>
>
>
>
>
>
39
This gives the trivial steady state:
( S , E , I , R , X )=( δ +μδ+κ+μ
, 0 ,0 ,κ
δ+κ+μ,0) (56)
Again, this steady state can be simplified further – as before, no infection in the population
makes vaccinations unnecessary. So we can set κ=0, causing everyone to stay in the S state,
with no-one moving to the R state via the infection or vaccination. This makes the steady state
become:
( S , E , I , R , X )=(1 ,0 ,0 ,0 , 0 )
If I ≠ 0 (virus present), then Maple confirms the following solution (along with the trivial
steady state, solution omitted due to size, see Appendix A4):
>
Checking in (38), S=S¿, E=E¿, I=I ¿, R=R ¿ and X=X¿ satisfy S+E+ I +R+ X=1 (manually
checked).
This solution along with some manual working leads to the non-trivial endemic steady state
(see Appendix A4):
( S , E , I , R , X )=( S¿ , E¿ , I ¿ , R¿ , X¿ ) (57)
1.3.5. Finding R0
For a closed population (which again, we have here), the critical value at which the virus
becomes an epidemic is when it exceeds 1S
. So, for this model:
40
R0=1S= βσ
σγ+ (μ+ χ ) (σ+γ+μ+ χ )
If R0 is above 1, the theory suggests that an epidemic starts:
R0=βσ
σγ+( μ+ χ ) (σ+γ+μ+ χ )>1
⟹ βσ>σγ +( μ+ χ ) (σ+γ+μ+ χ ) (58)
This gives the exact same R0 value as the previous measles model, due to similar traits of the
diseases and models (vaccination and immunity loss). Interpreting this inequality, again, by
reducing the transmission rate or increasing the mortality rate or recovery rate, we can cause a
disease to die out in this instance. Just as before, increasing the mortality rate is unethical and
the recovery rate is generally unchangeable, along with the latency period in this new model
due to the properties of the disease. Hence reducing β, so in turn, reducing the contact rate is
the best way to prevent an epidemic in this model. Notice here that again the vaccination
number does not appear in the value of R0, so the rate at which we vaccinate people has no
effect of preventing the actual outbreak of an epidemic. It appears as if this value only helps
prevent the spread once an epidemic has begun. Keeping those individuals with the H1N1
virus away from others is clearly yet again the most effective way of reducing the spread of
any endemic that may occur.
1.3.6. Analysing the data
Assuming a standard scenario for H1N1 (an average rate of spread with one epidemic per
lifetime), we take an estimate for the basic reproduction number, R0, as 2.6 for the US (Barry
2009). The latency period, 1σ
, is given between one and four days (Balcan et al. 2009, CDC
2010b), so we take the average of 2.5 days. The recovery period, 1γ
, is reported as three to five
days (Asp 2009), so again, we take the average of approximately four days. Average life
expectancy at birth is the same as before in the US; 79 years (WHO 2012), so this is taken to
41
be the value of 1μ
. Individuals lose active immunity at around 5 % per year (Greenberg et al.
2009), so we take this for our value of δ . Individuals are vaccinated for H1N1 and flu at
approximately 91 million vaccinations per year (Drummond 2010), so accounting for a daily
rate and the population size of the US, we can then work out a rate per day for κ .
Approximately 12270 individuals died due to the 2009 outbreak (CDC 2010a), so again, we
can find the rate of death per day for the parameter χ . All parameters are rates, so all must be
converted to the same units. Again, numerical analysis will be taken at daily time periods with
step size h=15
over the period of 365 days, so all parameters are multiplied or divided as
needed to give each value in terms of days. We can substitute these values into
R0=βσ
σγ+( μ+ χ ) (σ+γ+μ+ χ ) and rearrange, giving us a β value (transmission rate) of
0.650147 (six significant figures). Again, this system has been analysed using the 4th-order
Runge-Kutta method, with an initial maternally immune proportion, S (0 ) of 0.999 and
infected proportion, I (0 ) of 0.001 as a starting point (see Appendix B3 for data). This analysis
includes B and B¿ to show the rate at which individuals loop round the system.
0 50 100 150 200 250 300 350 4000.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
S
E
I
R
X
B
S*
E*
I*
R*
X*
B*
Time (days)
Pro
port
ion
of p
opul
atio
n
Figure 3.9.: H1N1 4th-order Runge-Kutta numerical analysis (with steady state values)
42
As Fig. 3.9. shows, like the two diseases before, the H1N1 influenza outbreak eventually
subsides but again, a worrying epidemic ensues. After around twenty days, the exposed and
infectious proportion starts to rise rapidly and then both peaking at day 40 at around 12 % and
18 % respectively, giving approximately 20 % of the population having this flu at this point.
This would certainly be classified as another epidemic in the population. The data shows that
a majority of people have the virus or have been vaccinated as early as day 50, as half the
population have entered the recovered state at this point, however, the level of recovered
people quickly drops back below 50 % by about day 65. This is due to the fact that individuals
are constantly losing active immunity to H1N1. From the year-long graph, it shows the levels
of all compartments fluctuate somewhat but all have hit their respective steady level by about
day 130. This type of disease is a faster-hitting but quicker-settling one, due to the traits of the
virus and how it interacts within a population.
By substituting in the values for β, σ , γ , δ , μ, χ and κ in (S¿ , E¿ , I ¿ , R¿ , X¿) and using Maple
to work them out (see Appendix A5), we can find out the behaviour of the model analytically
as t tends to infinity (also shown in Appendix B3 as t → ∞):
(S¿ , E¿ , I ¿ , R¿ , X¿)=(0.384615 , 0.0574299 ,0.0918750 ,0.465119 , 0.000961296)
Therefore eventually, there is a constant 38.5 % of the population that are susceptible (through
not having the virus or having it at least once and losing their active immunity), 46.5 % that
have had H1N1 at some point and then either recovered or have been vaccinated, but this time
a significant amount of people suffering from it at any time (1.5 %). Approximately 0.096 %
of the population will have died from H1N1 in a constant long-term population, proving it to
be the most deadly of the three diseases in this research.
43
4. Overview of deterministic models
All three models show stable, long-term, analytical solutions. In comparison to the actual data
from real life outbreaks, two of the three models hold up particularly well (Table 7.1. and
Table 7.2.) and the latter two models predict some amount of death which can be used to help
discuss the extent at which vaccination or removal from the population needs to be
implemented. However, can these models be improved upon to provide better estimates? Or
are there other ways to represent the same information to give the same output? Are there any
advantages to doing this?
All dynamical systems that are represented in a similar way to those above are subject to the
‘Evolution Rule’. That is, if the state currently occupied in the system has pre-defined rate at
which the next state is entered, we define the system as deterministic. Whereas, if the change
in state is given as a probability, so that there is a chance that the subject in question can go to
one state or another, or even more than that, and not split off into both or any subset of them,
this is defined as a stochastic (or random) system (Meiss 2007). For the same initial starting
point, analysed at the same point in discrete time (whether a finite time period or an infinite
time steady state value), the standard deterministic model will always give the same output,
every time. This is due to the defined rates of the system, the parameters calculated before
analysis and the very same ones used in the analytical solution to the ODEs. The stochastic
model can, however, account for random processes which can occur during analysis. This is
done by calculating non-deterministic probabilities which are then applied as transition
probabilities between each state. Some of these may be one – forcing a state change in that
situation, whereas some could be zero – meaning there is no chance of moving to the state in
question. Other probabilities must between these two values, which is where the random
chance element comes into play. For example, if an individual is in a state which has two
exits; one to a state with probability of a third and the other with transition probability of two
thirds, it is unknown at the next step which state the individual will be in. Of an overall
population, though, the differences between analyses can be cancelled out to give a rather
(hopefully) accurate description of the overall behaviour of a system on a large scale.
44
One of the main problems with deterministic models is as stated before; the lack of
accountability for chance from the same starting point. During different time periods, diseases
can have different rates of infection and this makes it difficult to decide on exactly what the
rate involved will be (Yorke and London 1973). This also is a problem when analysing
diseases which have quiet periods and then sudden epidemics or those which have not
occurred in a long time. Current deterministic models of some populations exhibiting certain
diseases have problems when the population dynamics are not standard; for example, feral pig
population dynamics in Australia vary over time and in space due to the populations searching
for optimal feeding ground (Dexter 2003). This highly affects the outcomes of a deterministic
system, such as when this population is suffering from food and mouth disease (FMD) and a
model needs to account for exactly what point the dynamics are at and how they may change.
This can be represented by probabilistic elements, hence the application of a stochastic model.
However, applying these elements into the standard setup of deterministic ordinary
differential equations by Pech and Hone (1988) and Pech and McIlroy (1990) turned out to be
too complex – a stable equilibrium point was never reached for the model (Dexter 2003). This
was mainly due to the fact that the feral pig dynamics vary wildly depending on two other
factors – vegetation density and kangaroo density, alongside their own population’s density.
Simulation of these densities when random processes can affect the outcome left a constant
outbreak of FMD present in the system. Evaluation of the system leads to a suggestion that
the combination of both deterministic and probabilistic elements side-by-side causes complex
behaviour that is difficult to model and get meaningful and clear results from. Dexter agrees
that the untested assumptions of his model could substantially change the behaviour in the
dynamics – an intrinsic problem of all deterministic models. This is why I wish to test a
purely stochastic approach. This way, I will have two comparable models on each disease,
where any overlapping features that could cause common errors or deviations between the
models are eliminated as much as possible. This will allow a comparison of which method
seems better or easier, without underlying factors that are inherent in any sort of hybrid model
becoming significant.
Höhle et al. (2005) state that they wanted to take an SEIR model and extend it stochastically
to analyse previously undertaken disease transmission experiments in a more detailed way.
They do this by taking existing work on diseases involving Markov chains and applying it to
45
data from Belgian classical swine fever virus (CSFV). Like Dexter (2003), but this time with
a more standard SEIR model comparable to those presented earlier in this research, spatial
elements are considered to account for dynamics within the populations affected by CSFV.
This model gets very complex very quickly, and two models are proposed – one that deals
with data with missing time elements and one that only looks at data with complete entries for
all fields. The first model has stochastic elements which allow for contact heterogeneity; that
is that not all contacts between pigs are the same and an element of probability is involved in
whether the disease transmits or not. The results are positive – the contact heterogeneity
model has a better statistical test outcome than the standard one proving that the added
complexity has provided better reliability. The second model also shows a promising
improvement on the standard deterministic-only model, but lacks the robustness of the first
model; some data had to be estimated or based off less applicable data than the first model.
Although the results show a good and efficient outcome in comparison to data, I argue that the
complexity of both models outweigh the effectiveness of the results. That is, is the extra data
and analysis required worth the relatively small improvements in reliability of data?
Obviously, there is a place for these models – anything that improves on something
previously existing is worth using, but sometimes, simplicity can be the best approach.
Following on from my previous work, I wish to use only probabilistic elements to model the
flow of an individual around a system. This will hopefully provide a similarly simple model
to the deterministic ones above, if not even simpler, yet still show the essence of the disease
that they are modelling.
46
5. Stochastic Processes – an overview
Stochastic processes are similar to traditional mathematical models, in the way that they
attempt to represent some real-world problem in a way that can be understood, adapted and
solved to provide a solution to the initial question. However, they differ in a major way – the
analysis of the system. Where traditional deterministic models allow the parameters in the
system to depend on time and or space as required, stochastic and probabilistic systems do not
rely on time intrinsically, instead conditional on just the probabilities of any one event
randomly occurring to an individual at any one time. Here, the time must be split up into
distinct, discrete chunks to allow for the analysis of the system at any given point in time. Any
system which must be or is chosen to be analysed using probabilistic theory in some way is
classified as stochastic (Nelson 1985), and many standard application revolve around
quantum theory; the underlying rules of physics which are based entirely on probabilistic
chance. This leads to a few definitions which will allow the implementation of this method to
analyse disease transmission:
A random walk (RW) is defined as (Urwin 2011b):
“A random process in discrete time steps { Xn:n=0,1,2 , …} such that it is only possible to
move forward, backward or remain in the same state, always with the same probability.”
A Markov chain (MC) is further defined as (Urwin 2011b):
“A stochastic process in discrete time steps { Xn:n=0,1,2 , …} with either a finite or infinite
state space, which has both the Markov property (MP) and the stationarity property (SP).”
The Markov property (MP) is that the currently occupied state is only dependant on the
immediately occupied previous state, and the stationarity (or homogeneity) property (SP) says
47
that the probability of transitioning from any one state to another (or to itself) stays the same
no matter which step the process is at (Urwin 2011b).
Therefore, any model consisting of a finite number of states, with constant probabilities and
state transitions only relating between the originating and target states in question can be
represented using a Markov chain. Looking at the original three compartmental models above,
if the natural and virus mortality arrows become targets for new states alongside if the
deterministic parameters that link each compartment are replaced with an assumed constant
probability (adding in self targetting probability arrows) for the transition between the two
states, then this model becomes a Markov chain. No state is needed for births as this method
looks at each individual separately, so all individuals can be assumed to be born and
susceptible to the disease from the outset.
Each compartmental model needs one compartment for each ‘alive’ state; M , S, E, I and R
are examples used in this research, and models not considered here that involved other
compartments would need extra (such as for the carrier state C in models which account for
individuals who can pass on an infection without ever suffering from it themselves).
Compartments are also needed for each ‘death’ state. These death states could be merged into
one as once an individual enters a death state, they cannot leave, but for easier reading of
results later, the death states are kept separate (except for the model presented in 6.3.). This is
so the results clearly show the percentage of individuals who died having had the disease at
one point, alongside rather than merged with the percentage who died never having caught it.
If there are m states, let pij be the probability of moving from state i to state j in a specified
time interval, where i , j∈ {1,2,…,m }. This implies that the total of the values given on all of
the arrows leaving any state must add exactly to one (Urwin 2011a):
∑j=1
m
pij=1 , ∀ i
Some states are known as absorbing states. These are states that once entered cannot be left.
This implies two things for an absorbing state i:
pii=1
and:
48
pij=0 , ∀ i≠ j
In the following models, all death states are clearly absorbing – once and individual enters
them, they will stay in them throughout the entirety of the rest of the analysis (Urwin 2011a).
Each compartmental model can be represented by something called a transition matrix, P.
This is a square matrix made up of m rows and m columns, representing each of the finite m
states. By definition, just as above, each row of P must add to one. Each element of the matrix
gives the probability of leaving a state and entering a new one (or staying in the same state) in
one discrete time step. That is, the element in the ith row and jth column of P is the probability
of leaving state i and entering state j (Urwin 2011a). The transition matrix for each model
will look like this:
P=(p11 p12 ⋯ p1 m
p21 p22 ⋯ p2 m
⋮ ⋮ ⋱ ⋮pm1 pm2 ⋯ pmm
)To measure the probability that any individual in this system is in state j after n steps given
that they started in state i can be calculated using the transition matrix. Matrix multiplication
allows us to find any probability we wish. If P is raised to the nth power, the following matrix
is produced:
Pn=(p11 p12 ⋯ p1 m
p21 p22 ⋯ p2 m
⋮ ⋮ ⋱ ⋮pm1 pm2 ⋯ pmm
)n
=(p11
(n ) p12(n ) ⋯ p1m
(n )
p21(n ) p22
(n ) ⋯ p2m(n )
⋮ ⋮ ⋱ ⋮pm1
(n ) pm2( n) ⋯ pmm
(n ) )Now, pij
( n) is not the same as pijn, the probability pij raised to the nth power. pij
( n) is denoted as the
probability of being in state j after n steps given that the individual started in state i. If we
wish to know what state someone will be in after 2 steps, for example, we calculate P2 and
then look at the appropriate element. Each of the following models will work out the
49
probability an individual has the disease in question after one year, or 365 days. Each of the
probabilities in P will be calculated based on a daily probability, so each transition matrix will
be raised to the 365th power to work out P365 (Urwin 2011b). Further multiplication will then
be carried out to find out the limiting distribution (Urwin 2011c).
The limiting distribution is the transition matrix P raised to the nth power, as n → ∞ (Urwin
2011c). This will produce a matrix where any further matrix multiplication by any power of P
will not change the answer you get from the previous step:
Pn∗Pa=Pn+a=Pn ,∀ a ≥1
This limiting distribution shows us when the disease will be steady in the population; where
people will catch the disease at exactly the same rate as others are recovering. Obviously, in
this model in this limiting matrix, each individual will end up dead in the long-term (end up in
an absorbing state). But depending on which dead state they end up in shows us exactly what
the probability of someone having the disease at some point was, and therefore, what the
long-term steady proportion of any given population having the disease in question was.
50
6. Stochastic Modelling – adaptations and analysis
1.4. Model 1 adaptation – Chickenpox
1.4.1. Stochastic model
The start of this method of infectious disease modelling includes a compartmental model,
much in a way similar to the original deterministic model given in 3.1.1. However, instead of
parameters given for each arrow, a probability of entering the state in question is given.
This gives an initial compartmental model for chickenpox (Fig. 6.1. and Fig. 6.2.):
Figure 6.1.: Stochastic model for chickenpox
51
Probability of natural death
Probability of natural death
Probability of natural death
Probability of natural death
Probability of staying dead
Probability of staying dead Probability of
staying deadProbability of staying dead
Probability of recovery
Probability of developmentProbability of
infection
Probability of staying recovered
Probability of staying infectious
Probability of staying exposedProbability of
staying susceptible
Recovered Deaths
Recovered
Infectious Deaths
Exposed Deaths
Susceptible Deaths
InfectiousExposedSusceptible
Or more mathematically:
Figure 6.2.: Stochastic model for chickenpox with mathematical symbols
Arrows that are missing (such as pSI) imply that the probability of leaving state i and entering
state j is zero (so pSI=0, for example). If at any time, an individual in state i has the chance of
staying in state i (pii>0), then this is represented by a circular arrow that has the state it left as
its target.
1.4.2. Transition matrix
This can be represented in a square matrix, where each element corresponds to the relative
probability of leaving the state and entering a new one, where the element in the ith row and jth
column is the probability of leaving state i and entering state j. Here, we need an eight-by-
eight (8 × 8) matrix as we have eight states, so this gives us the matrix P as follows:
52
pRZpEX p IYpSW
pZZpWW
pXX pYY
p IRpEIpSE
pRR
p IIpEE
pSS
Z
R
YXW
IES
P=(pSS pSE pSI pSR pSW pSX pSY pSZ
pES pEE pEI pER pEW pEX pEY pEZ
p IS p IE p II pIR p IW p IX p IY p IZ
pRS pℜ pRI pRR pRW pRX pRY pRZ
pWS pWE pWI pWR pWW pWX pWY pWZ
p XS pXE pXI pXR pXW pXX pXY p XZ
pYS pYE pYI pYR pYW pYX pYY pYZ
pZS pZE pZI pZR pZW pZX pZY pZZ
)Clearly, the probability of staying dead when already in a dead state is one, so:
pWW=pXX=pYY =pZZ=1
And the probability of moving out of a dead state to any other must be zero, so:
pWS=pWE=pWI=pW R=pWX=pWY=pWZ=0
pXS=pXE=pXI=pXR=pXW=pXY=pXZ=0
pYS=pYE=pYI=pYR=pYW=pYX=pYZ=0
pZS=pZE=pZI=pZR=pZW=pZX=pZY =0
There is no chance in this model that an individual can jump straight from being susceptible
to being infectious, being susceptible to having recovered or from being exposed to having
recovered, so:
pSI=pSR=pER=0
Individuals also cannot travel backwards in the model, so:
pES=p IS=p IE=pRS=pℜ=pRI=0
Each compartment that is not a dead state has its own respective dead state; individuals who
naturally die whilst susceptible go into the W state, those who naturally die whilst exposed go
into the X state, those who naturally die whilst infectious go into the Y state and those who
naturally die after having recovered go into the Z state. Therefore, for example, the
probability of a susceptible individual dying and going into states X , Y or Z is zero (similar
for states E, I and R), so:
53
pSX=pSY=pSZ=0
pEW=pEY=pEZ=0
p IW=p IX=p IZ=0
pRW=pRX=pRY=0
The reasons for having separate dead states is to make the analysis slightly easier later – those
who end up in states X , Y or Z must have had the infection at some point to end up in there,
whereas those who end up in state W have not.
The remaining eleven probabilities have values which are neither zero (impossible) nor one
(certain). These can be calculated (estimated) as follows (all probabilities are rounded only at
the final answer and are rounded to six significant figures):
The average life expectancy of an individual in the United Kingdom (UK) is currently 81
years (WHO 2012). As each step of this model accounts for one day, 81 years in days is 81
years ¿365 days per year ¿29565 days. Hence, the reciprocal of this gives the probability of
dying naturally each day during the analysis of this model. An assumption to be made here is
that the chances of dying naturally are equal regardless of whether the individual in question
is currently susceptible, exposed, infectious or has recovered from chickenpox, so:
pSW =pEX=p IY=pRZ=1
29565≅ 0.0000338238
Annually in the UK, there are approximately 57 cases of chickenpox per 10,000 people
presented to and recorded by doctors and practice nurses (Fleming et al. 2007 p.14). Although
this report is not taken from all hospitals, over 52 million patients were analysed (Fleming et
al. 2007 p.9) which is a high proportion of the UK population so it is representative.
Extrapolating this up to the entire population of the UK will give us the estimated number of
cases in one year. So there are approximately 57∗64100000
10000=365370 cases each year.
Dividing this number by 365 will give the number of cases per day, so 365370
365≅ 1001.01
54
cases per day. This number further divided by the population size will give the probability of
an individual contracting chickenpox on any given day, so:
pSE=1001.01
64100000≅ 0.0000156164
The latency period of chickenpox is given to be somewhere between ten and fourteen days
(Knott 2013), so taking the average of twelve days estimates the value of the reciprocal of pEI,
so:
pEI=1
12=0.083≅ 0.0833333
The recovery period of chickenpox once contagious is also given to be somewhere between
ten and fourteen days (Lamprecht 2012), so again, taking the average of twelve days estimates
the value of the reciprocal of p IR, so:
p IR=1
12=0.08 3≅ 0.0833333
Now, all rows of the matrix must add up to one, so the final stationary probabilities can all be
worked out using the values we already know (given to full decimal places):
pSS=1−0.0000156164−0.0000338238=0.9999505598
pEE=1−0.0833333−0.0000338238=0.9166328762
p II=1−0.0833333−0.0000338238=0.9166328762
pRR=1−0.0000338238=0.9999661762
Combining all of the above results give the following final transition matrix (given to
appropriate significant figures):
P=¿
55
(0.999951 1.56 ×10−5 0 0 3 .38 ×10−5 0 0 0
0 0.916633 0.0833 0 0 3 .38 ×10−5 0 00 0 0.916633 0.0833 0 0 3 .38 ×10−5 00 0 0 0.999966 0 0 0 3.38 × 10−5
0 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1
)1.4.3. n-step transition
The square of this matrix will give the probability that an individual will be in any given state
after one day. That is, the element in the ith row and jth column of P2, denoted pij(2 ), gives the
probability that an individual will be state j at the end of the second day, if they started in
state i. Following on from this, working out P365 will give all the probabilities that an
individual will end in any desired state, given any initial starting state, at the end of the 365th
day, or one year. Hence, using Maple (omitted due to size, see Appendix A6), this gives the
365-step matrix (given to appropriate significant figures):
P365=¿
(0.982116 0.000184080 0.000184114 0.00524607 0.0122353 2.22 ×10−6 2.14 ×10−6 3.04 × 10−5
0 1.6 ×10−14 5.28× 10−13 0.987730 0 0.000405721 0.000405556 0.01145870 0 1.6 × 10−14 0.987730 0 0 0.000405721 0.01186430 0 0 0.987730 0 0 0 0.01227000 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1
)Now, assuming we start with a susceptible individual, we can just concentrate on the first row
of P365. The chance of any susceptible individual having had chickenpox at any point in the
year of analysis is the sum of the following probabilities:
pSE(365)+ pSI
(365)+ pSR(365)+ pSX
(365)+ pSY(365)+ pSZ
(365)
56
The first three probabilities of these six represent those who were susceptible but have ended
day 365 being exposed, infectious or having recovered, respectively. The last three of these
probabilities represent those who were susceptible but have ended the year having passed
away naturally (not due to chickenpox directly) after being exposed, infectious or having
recovered, respectively. pSS(365) and pSW
(365) are not included in this sum as these represent those
who were susceptible and ended the year still being susceptible (not having had chickenpox)
and those who started susceptible but have died naturally before contracting the disease at any
point.
Hence, the probability that any individual in the UK contracts chickenpox in this year is:
pSE(365 )+ pSI
( 365 )+ pSR(365 )+ pSX
(365 )+ pSY(365 )+ pSZ
( 365 )=0.000184080+0.000184114+0.00524607+0.00000221723+0.00000214164+0.0000303784≅ 0.00564900
This shows that each individual has a 0.56 % probability of contracting chickenpox during
year one.
1.4.4. Limiting distribution
If we take P to higher and higher powers, we can get towards the limiting distribution of this
matrix system. P365 gives the probability of an individual, hence the percentage of a
population, that will get the disease in one year. Raising this new matrix to the power n will
give the percentage of a constant population that will contract the disease at some point during
n years. As n → ∞, the matrix ( P365 )n=P365∗n → Pn. So raising the original matrix to
exceptionally high powers will give the steady proportion of a population (based on UK data)
that will contract chickenpox.
57
After some manual checking (using integer powers of 10 for n), both P365∗100000=P36500000 and
P365∗1000000¿ P365000000 give the same matrix (using Maple, omitted due to size, see Appendix
A7). This implies that the limiting distribution, Pn, is (to six significant figures):
Pn=(0 0 0 0 0.684136 0.000128153 0.000128101 0.3156080 0 0 0 0 0.000405721 0.000405556 0.9991890 0 0 0 0 0 0.000405721 0.9995940 0 0 0 0 0 0 10 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1
)Some rounding errors have occurred during the calculation process due to the sheer size of the
power the matrix P has been raised to. However, rounding to significant figures shows that all
rows of Pn still add up to one.
From this, the result here shows that given enough time, all individuals will pass away (all
elements in the first four columns are zero so no-one ends up in the S, E, I or R states), which
is obvious. It also shows a steady percentage of pSW( n) =68.4 % of the population will have died
directly from the S state, so 68.4 % of the population will never contract chickenpox. The sum
of the other three non-zero values in this row gives the percentage of people who will get
chickenpox (this can also be worked out using 1−pSW(n) as all rows must add to one):
pSX( n) + pSY
(n ) +pSZ(n )=0.000128153+0.000128101+0.315608=0.315864254
So 31.6 % of the population will contract chickenpox in a steady population that stays
constant for many years.
58
1.5. Model 2 adaptation – Measles
1.5.1. Stochastic model
Just like the previous adaptation, the start of this model includes a compartmental model,
much in a way similar to the original deterministic model given in 3.2.1., with a probability of
entering the state in question is given instead of deterministic parameters.
This gives an initial compartmental model for measles (Fig. 6.3. and Fig. 6.4.):
Figure 6.3.: Stochastic model for measles
59
Exposed Virus
Deaths
Exposed Natural Deaths
Infectious Natural Deaths
Susceptible Deaths
Maternal Deaths
Recovered Deaths
Infectious Virus
Deaths
Probability of gain of active immunity (vaccination)
Probability of loss of passive
immunityProbability of development
SusceptibleMaternal Immunity
RecoveredInfectiousExposed
Probability of infection
Probability of recovery
Prob. of virus death
Prob. of natural death
Prob. of virus death
Prob. of natural death
Prob. of natural death
Prob. of natural death
Prob. of natural death
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying immune Probability of
staying recovered
Probability of staying infectious
Probability of staying exposed
Probability of staying susceptible
Or more mathematically:
Figure 6.4.: Stochastic model for measles with mathematical symbols
Again, arrows that are missing (such as pME) imply that the probability of leaving state i and
entering state j is zero (so pME=0, for example). If at any time, an individual in state i has the
chance of staying in state i (pii>0), then as before, this has been added and is represented by a
circular arrow that has the state it left as its target.
1.5.2. Transition matrix
This can be represented in a square matrix, where each element corresponds to the relative
probability of leaving the state and entering a new one, where the element in the ith row and jth
column is the probability of leaving state i and entering state j. Here, we need a twelve-by-
twelve (12 ×12) matrix as we have twelve states (five ‘alive’ states, five ‘natural death’ states
and two ‘virus death’ states), so this gives us the matrix P as follows:
60
TX YWV ZU
pSR
pMS pEI
SM RIEpSE
p IR
pET p IY p IUpEX pRZpSWpMV
pUUpYYpTTpXXpWW
pZZpVV
pMM pRR
p IIpEEpSS
P=(pMM pMS pME pMI pMR pMV pMW pMX pMT pMY pMU pMZ
pSM pSS pSE pSI pSR pSV pSW pSX pST pSY pSU pSZ
pEM pES pEE pEI pER pEV pEW pEX pET pEY pEU pEZ
pℑ p IS p IE p II p IR pIV p IW p IX p¿ p IY p IU pIZ
pRM pRS pℜ pRI pRR pRV pRW pRX pRT pRY pRU pRZ
pVM pVS pVE pVI pVR pVV pVW pVX pVT pVY pVU pVZ
pWM pWS pWE pWI pWR pWV pWW pWX pWT pWY pWU pWZ
pXM pXS pXE pXI pXR pXV pXW p XX pXT pXY pXU pXZ
pTM pTS pTE pTI pTR pTV pTW pTX pTT pTY pTU pTZ
pYM pYS pYE pYI pYR pYV pYW pYX pYT pYY pYU pYZ
pUM pUS pUE pUI pUR pUV pUW pUX pUT pUY pUU pUZ
pZM pZS pZE pZI pZR pZV pZW pZX pZT pZY pZU pZZ
)Like before, each probability needs to be worked out. Clearly, the probability of staying dead
when already in a dead state is one, so: pVV =pWW=pXX=pTT=pYY=pUU=pZZ=1
And the probability of moving out of a dead state to any other must be zero, so:
pVM=pVS=pVE=pVI=pVR=pVW=pVX=pVT=pVY=pVU=pVZ=0
pWM=pWS=pWE=pWI=pWR=pWV=pWX=pWT=pWY=pWU=pWZ=0
pXM=pXS=pXE=pXI=pXR=pXV=pXW=pXT=pXY=pXU =pXZ=0
pTM=pTS=pTE=pTI=pTR=pTV=pTW=pTX=pTY=pTU=pTZ=0
pYM=pYS=pYE=pYI=pYR=pYV=pYW=pYX=pYT=pYU=pYZ=0
pUM=pUS=pUE=pUI=pUR=pUV =pUW=pUX=pUT=pUY=pUZ=0
pZM=pZS=pZE=pZI=pZR=pZV=pZW=pZX=pZT=pZY =pZU=0
There is no chance in this model that an individual can jump straight from having maternal
immunity to being exposed, infectious or having recovered, from being susceptible to being
infectious, or from being exposed to having recovered, so:
pME=pM I=pMR=pSI=pER=0
Individuals also cannot travel backwards in the model, so:
61
pSM=pEM=pES=pℑ=p IS=pIE=pRM=pRS=pℜ=pRI=0
Each compartment that is not a dead state has its own respective dead state; similar to before,
individuals who naturally die whilst susceptible go into the W state, those who naturally die
whilst exposed go into the X state, those who naturally die whilst infectious go into the Y
state and those who naturally die after having recovered go into the Z state, but also,
individuals who naturally die whilst having passive immunity go into the V state. Individuals
can only die from the disease whilst exposed or infectious, so the according states after a virus
death can only being reached from these states too. Therefore, for example, the probability of
an immune individual dying and going into states W , X , T , Y , U or Z is zero (similar for
states S, E, I and R), so:
pMW=pMX=pMT=pMY=pMU=pMZ=0
pSV =pSX=pST=pSY=pSU=pSZ=0
pEV=pEW=pEY=pEU=pEZ=0
p IV=p IW=p IX=p¿=p IZ=0
pRV=pRW =pRX=pRT=pRY=pRU=0
Again, the reasons for having separate dead states is to make the analysis slightly easier later
– those who end up in states X , Y or Z must have had the infection at some point to end up in
there, whereas those who end up in states V or W have not.
The remaining seventeen probabilities have values which are neither zero nor one. These can
be calculated as follows (as before, all probabilities are rounded only at the final answer and
are rounded to six significant figures):
The average life expectancy of an individual in the United States (US) is currently 79 years
(WHO 2012). Similar to before, 79 years in days is 79 years ¿365 days per year ¿28835 days.
The reciprocal of this gives the probability of dying naturally each day. An assumption
continuing to be made here is that the chances of dying naturally are equal regardless of
62
whether the individual in question is currently young with maternal immunity, susceptible,
exposed, infectious or has recovered from measles, so:
pMV=pSW=pEX=p IY=pRZ=1
28835≅ 0.0000346801
Annually in the US, there are only approximately 55 cases of measles (CDC 2009a). Simply
dividing by the US population will give the probability of contracting measles each year, and
then further dividing by 365 will give the chance per day:
pSE=55
316148990∗365≅ 0.000000000476626
The chance of dying from measles if an individual has caught it is approximately one or two
in a thousand, for children over the entire length of the disease (17.5 days) (CDC 2009b). As
most adult individuals in the US will have already been vaccinated, it is safe to assume that
any individual catching the disease is relatively young. This allows the child virus mortality
rate to be used in this model (taking the average of one and a half out of one thousand and
dividing by the length of having the disease). An assumption that dying due to measles
whether the individual is exposed or infectious is equivalent in the two states is made here:
pET=p IU=1.5
1000∗17.5≅ 0.0000857143
Again, looking specifically at the child vaccination rate as most individuals in the US will be
immunised at a young age, the data shows that a mean of 91.9 % of children across the US are
immunised per year. The ages in question are well after children are estimated to have lost
their maternal immunity so it is safe to assume that all individuals immunised in this model
are in the susceptible state, and not the maternal immunity state. So, dividing the value by 365
will give the probability per day:
pSR=0.919365
≅ 0.00251781
The latency period of measles is somewhere between seven and fourteen days (CDC 2009c),
so taking the average of ten and a half days gives the value of the reciprocal of pEI, so:
pEI=1
10.5≅ 0.0953381
63
The recovery period of measles after becoming contagious is approximately seven days; about
four days for a rash to form and another three for it to be completely gone and have fully
recovered (CDC 2009c), so again, taking this value and finding the reciprocal gives p IR, so:
p IR=17=0. 142857≅ 0.142857
Maternal immunity is basically fully lost by the age of twelve to fifteen months (Nicoara
1999). So, if an individual has a 100 % chance after these 13.5 months (taking the average),
converting this into years (1.125 years), and dividing by the 365 days per year gives the
probability of losing the conferred immunity:
pMS=1
1.125∗365≅ 0.00243531
All rows of the matrix must add up to one, so the final stationary probabilities can all be
worked out using the values we already know (given to full decimal places):
pMM=1−0.00243531−0.0000346801=0.9975300099
pSS=¿
1−0.000000000476626−0.00251781−0.0000346801=0.997447509423374
pEE=1−0.0953381−0.0000346801−0.0000857143=0.9045415056
p II=1−0.142857−0.0000346801−0.0000857143=0.8570226056
pRR=1−0.0000346801=0.9999653199
Combining all of the above results give the following final transition matrix (given to
appropriate significant figures):
P=¿
64
(0.997530 0.00243531 0 0 0 3.47 ×10−5 0 0 0 0 0 0
0 0.997448 4.77 ×10−10 0 0.00251781 0 3.47 ×10−5 0 0 0 0 00 0 0.904542 0.0953381 0 0 0 3.47 × 10−5 8 .57 × 10−5 0 0 00 0 0 0.857023 0.142857 0 0 0 0 3.47 × 10−5 8.57 × 10−5 00 0 0 0 0.999965 0 0 0 0 0 0 3.47 ×10−5
0 0 0 0 0 1 0 0 0 0 0 00 0 0 0 0 0 1 0 0 0 0 00 0 0 0 0 0 0 1 0 0 0 00 0 0 0 0 0 0 0 1 0 0 00 0 0 0 0 0 0 0 0 1 0 00 0 0 0 0 0 0 0 0 0 1 00 0 0 0 0 0 0 0 0 0 0 1
)1.5.3. n-step transition
As before, working out P365 will give all the probabilities that an individual will end in any
desired state, given any initial starting state, at the end of one year. Hence, using Maple
(omitted due to size, see Appendix A8), this gives the 365-step matrix (given to appropriate
significant figures):
P365=¿
(0.405489 0.355942 1.77 ×10−9 1.18× 10−9 0.225990 0.00834728 0.00312797 1.50 × 10−11 3.70×10−11 9.70 ×10−12 2.40× 10−11 0.00110342
0 0.393431 2.02 ×10−9 1.37×10−9 0.593991 0 0.00824132 4.04 ×10−11 9.99×10−11 2.66 ×10−11 6.58× 10−11 0.004337360 0 1.25 × 10−16 2.50 × 10−16 0.985943 0 0 0.000363300 0.000897922 0.000242251 0.000598739 0.01195500 0 0 3.48 × 10−25 0.986829 0 0 0 0 0.000242557 0.000599495 0.01232870 0 0 0 0.987421 0 0 0 0 0 0 0.01257870 0 0 0 0 1 0 0 0 0 0 00 0 0 0 0 0 1 0 0 0 0 00 0 0 0 0 0 0 1 0 0 0 00 0 0 0 0 0 0 0 1 0 0 00 0 0 0 0 0 0 0 0 1 0 00 0 0 0 0 0 0 0 0 0 1 00 0 0 0 0 0 0 0 0 0 0 1
)Now, differing from the previous disease, if we assume we start with a maternally immune
individual (as this model is based around children catching measles and being vaccinated at
an early age), we can just concentrate on the first row of P365. The chance of any individual
65
having had measles at any point in the first year of analysis (hence in the first year of their
life) is the sum of the following probabilities:
pME(365)+ pMI
(365)+ pMR(365)+ pMX
(365)+ pMT(365)+ pMY
(365)+ pMU(365)+pMZ
(365)
The first three probabilities of these eight represent those who were immune to start with but
have ended day 365 being exposed, infectious or having recovered, respectively. The last five
of these probabilities represent those who were immune but have ended the year having
passed away naturally after being exposed, infectious or having recovered, or died due to
measles whilst they were exposed or infectious.
Hence, the probability that any individual in the US contracts measles in this year is:
pME(365)+ pMI
(365)+ pMR(365)+ pMX
(365)+ pMT(365)+ pMY
(365)+ pMU(365)+pMZ
(365)=0.00000000177157+0.00000000117831+0.225990+0.0000000000149744+0.0000000000370103+0.00000000000969920+0.0000000000239723+0.00110342≅ 0.227093
This shows that each initially immune (i.e. someone who has just been born) individual has a
22.7 % probability of contracting measles during year one. However, due to vaccinations, not
all of these 22.7% of the population will have caught measles. Some will have become
immune due to vaccination – approximately 91.9 % in this year. So to account for these, we
need to take away 91.9 % from 100 % to find the proportion that were not vaccinated, and
then multiply the two resulting percentages together:
(1−0.919 )∗0.227093=0.081∗0.227093=0.018394533
Hence, an individual with maternal immunity will have a 1.84 % probability of catching
measles in the first year of life.
66
However, if we assume we start with a susceptible individual (someone who has already lost
their maternal immunity but has not yet contracted any form of measles), we can look at the
second row of P365. The chance of any of these individuals having had measles at any point in
the first year of analysis is the sum of the following probabilities:
pSE(365)+ pSI
(365)+ pSR(365)+ pSX
(365)+ pST(365)+ pSY
(365)+ pSU(365)+pSZ
(365)
Again, the first three probabilities of these eight represent those who were susceptible to start
with but have ended day 365 being exposed, infectious or having recovered, respectively. The
last five of these probabilities similarly represent those who were susceptible but have ended
the year having passed away naturally after being exposed, infectious or having recovered, or
died due to measles whilst they were exposed or infectious.
Hence, the probability that any susceptible individual in the US contracts the disease in this
year is:
pSE(365)+ pSI
(365)+ pSR(365)+ pSX
(365)+ pST(365)+ pSY
(365)+ pSU(365)+pSZ
(365)=0.00000000201838+0.00000000137033+0.593991+0.0000000000404158+0.0000000000998904+0.0000000000266171+0.0000000000657859+0.00433736≅ 0.598328
This shows that each initially susceptible individual has a 59.8 % probability of contracting
measles during the year. Again, however, due to vaccinations, not all of these 59.8 % of the
population will have caught measles. Some will have become immune due to vaccination –
91.9 % in this year. To account for these, we again need to multiply 8.1% by the other
percentage:
(1−0.919 )∗0.598328=0.081∗0.598328≅ 0.0484646
Hence, an unvaccinated but susceptible individual will have a 4.85 % probability of catching
measles.
67
1.5.4. Limiting distribution
Like before, if we take P to higher and higher powers, we can get towards the limiting
distribution of this matrix system. P365 gives the probability of an individual, hence the
percentage of a population, that will get the disease in one year. Raising this new matrix to the
power n will give the percentage of a constant population that will contract the disease at
some point during n years, just as we had with the previous disease. As n → ∞, the matrix
( P365 )n=P365∗n → Pn. So raising the original matrix to exceptionally high powers will give the
steady proportion of a population (based on US data) that will contract measles.
After some manual checking (using integer powers of 10 for n), just like before both
P365∗100000=P36500000 and P365∗1000000¿ P365000000 give the same matrix (using Maple, omitted due
to size, see Appendix A9). This implies that the limiting distribution, Pn, is (to appropriate
significant figures):
Pn=¿
(0 0 0 0 0 0.0140406 0.0133960 6.69 × 10−11 1.65 ×10−10 4.46×10−11 1.10× 10−10 0.9725630 0 0 0 0 0 0.0135868 6.78 × 10−11 1.68 ×10−10 4.52×10−11 1.12×10−10 0.9864130 0 0 0 0 0 0 0.000363300 0.000897922 0.000242251 0.000598739 0.9978980 0 0 0 0 0 0 0 0 0.000242557 0.000599495 0.9991580 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 1 0 0 0 0 0 00 0 0 0 0 0 1 0 0 0 0 00 0 0 0 0 0 0 1 0 0 0 00 0 0 0 0 0 0 0 1 0 0 00 0 0 0 0 0 0 0 0 1 0 00 0 0 0 0 0 0 0 0 0 1 00 0 0 0 0 0 0 0 0 0 0 1
)As before, some rounding errors have occurred during the calculation process due to the sheer
size of the power the matrix P has been raised to. However, rounding to significant figures
shows that all rows of Pn still add up to one.
68
From this, the result here shows that given enough time, all individuals will pass away (all
elements in the first five columns are zero so no-one ends up in the M , S, E, I or R states),
which is obvious. By adding pMV( n) + pMW
(n ) , it also shows a steady percentage of the population
who will have died directly without ever getting the disease or being vaccinated (assuming
immune start), so:
pMV( n) + pMW
(n ) =0.0140406+0.0133960=0.0274366
So 2.74 % of the population will never contract measles even when not vaccinated. By taking
pMZ( n) and multiplying by 0.919 to find out the percentage of those who are vaccinated
(assuming the rate of 91.9 % per year is kept steady), we get:
pMZ( n) ∗0.919=0.972563∗0.919=0.893785397
So 89.4 % of the population will have been vaccinated and be safe from measles at any point.
This seems counter-intuitive if a constant 91.9 % are being vaccinated, however, this value
accounts for the fact that individuals who have been vaccinated can die as others are being
immunised, and others can be born and lose their maternal immunity. The difference between
pMZ( n) and this answer gives the percentage of people who do contract measles but eventually
recover from it:
0.972563−0.893785397=0.078777603
So 7.88 % of those who wind up in the recovered state (or associated death state) will have
contracted measles and lived (with no prior vaccination). The sum of this value and the other
four non-zero values in this row gives the percentage of people who will get measles at some
point in their life (whether they survive from it or not):
pMX( n) + pMT
( n) + pMY( n) + pMU
(n ) +0.078777603=0.0787776033870348
69
So 7.88 % of the population will contract measles in a steady population that stays constant
for many years. As this value is very similar to our previous value, most people who contract
measles will live. The difference will give the overall exact percentage who will die:
0.0787776033870348−0.078777603=0.0000000003870348
So 0 .000000039 % of the long-term population will die from measles.
70
1.6. Model 3 adaptation – H1N1
1.6.1. Stochastic model
Again, much in a way similar to the original deterministic model given in 3.3.1., the
probabilistic model starts with a compartmental model, with a probability of entering the state
in question is given instead of deterministic parameters.
This gives an initial compartmental model for H1N1 (Fig. 6.5. and Fig. 6.6.):
Figure 6.5.: Stochastic model for H1N1
71
Probability of loss of active immunity
Exposed Virus
Deaths
Exposed Natural Deaths
Infectious Natural Deaths
Susceptible Deaths
Recovered Deaths
Infectious Virus
Deaths
Probability of gain of active immunity (vaccination)
Probability of development
Susceptible RecoveredInfectiousExposed
Probability of infection
Probability of recovery
Probability of virus death
Probability of natural
death
Probability of virus death
Probability of natural death
Probability of natural death
Probability of natural death
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying dead
Probability of staying recovered
Probability of staying infectious
Probability of staying exposed
Probability of staying susceptible
Or more mathematically:
Figure 6.6.: Stochastic model for H1N1 with mathematical symbols
Here, we are trying a simplification on this final model by merging together the exposed and
infectious death states but keeping natural and virus deaths separate. This will make the
calculations simpler later as we have two fewer states, hence two fewer rows and columns in
the transition matrix P. This is considered as this model loops and cannot be simplified
further in any other way. No ‘dummy’ or ‘bin’ states are added here – a dummy state would
be pointless as you can only leave it and never enter it (when dealing with each individual
separately as per this model, we are only interested in the states that can be entered), and the
bin state would be an absorbing state, so the individual cannot leave. However, as opposed to
the traditional model for H1N1 where anyone entering the bin would be ‘replaced’ by another
72
individual from the dummy state, this would not occur here. The model here, though, is
accepted as a valid comparison to that of above.
Again, arrows that are missing (such as pSI) imply that the probability of leaving state i and
entering state j is zero (so pSI=0, for example). If at any time, an individual in state i has the
chance of staying in state i (pii>0), then as before, this has been added and is represented by a
circular arrow that has the state it left as its target.
1.6.2. Transition matrix
This can be represented in a square matrix, where each element corresponds to the relative
probability of leaving the state and entering a new one, where the element in the ith row and jth
column is the probability of leaving state i and entering state j. Here, we need an eight-by-
eight (8 × 8) matrix as we have eight states (four ‘alive’ states, three ‘natural death’ states and
one ‘virus death’ state), so this gives us the matrix P as follows:
P=(pSS pSE pSI pSR pSW pSX pSY pSZ
pES pEE pEI pER pEW pEX pEY pEZ
p IS p IE p II pIR p IW p IX p IY p IZ
pRS pℜ pRI pRR pRW pRX pRY pRZ
pWS pWE pWI pWR pWW pWX pWY pWZ
p XS pXE pXI pXR pXW pXX pXY p XZ
pYS pYE pYI pYR pYW pYX pYY pYZ
pZS pZE pZI pZR pZW pZX pZY pZZ
)Again, the probability of staying dead when already in a dead state is one, so:
pWW=pXX=pYY =pZZ=1
And the probability of moving out of a dead state to any other must be zero, so:
pWS=pWE=pWI=pWR=pWX=pWY=pWZ=0
73
pXS=pXE=pXI=pXR=pXW=pXY=pXZ=0
pYS=pYE=pYI=pYR=pYW=pYX=pYZ=0
pZS=pZE=pZI=pZR=pZW=pZX=pZY =0
There is no chance in this model that an individual can jump straight from being susceptible
to being infectious or from being exposed to having recovered, so:
pSI=pER=0
Individuals also cannot travel backwards in the model in most cases, so:
pES=p IS=p IE=pℜ=pRI=0
Each compartment that is not a dead state has its own respective dead state (with exposed and
infectious merged); individuals who naturally die whilst susceptible go into the W state, those
who naturally die whilst exposed or infectious go into the Y state and those who naturally die
after having recovered go into the Z state. Similarly, those who die due to H1N1 whilst
exposed or infectious go into the X state. Therefore, for example, the probability of a
susceptible individual dying and going into states X , Y or Z is zero (similar for states E, I and
R), so:
pSX=pSY=pSZ=0
pEW=pEZ=0
p IW=p IZ=0
pRW=pRX=pRY=0
As before, the reasons for having separate dead states is to make the analysis slightly easier
later – those who end up in states X , Y or Z must have had the infection at some point to end
up in there, whereas those who end up in state W have not. However, it is valid to merge the
death states for exposed and infectious individuals as when summing up whether an
individual had died whilst affected by H1N1 (whether due to the virus or not), these states
would be added together anyway.
74
The remaining fifteen probabilities have values which are neither zero nor one. These can be
worked out as follows (all probabilities are rounded only at the final answer and are rounded
to six significant figures):
As already stated, the average life expectancy of an individual in the United States (US) is
currently 79 years (WHO 2012). So, 79 years in days is 79 years ¿365 days per year ¿28835
days. The reciprocal of this gives the probability of dying naturally each day. An assumption
continuing to be made here is that the chances of dying naturally are equal regardless of
whether the individual in question is susceptible, exposed, infectious or has recovered from
H1N1, so:
pSW =pEY=p IY=pRZ=1
28835≅ 0.0000346801
In the US during the 2009 H1N1 epidemic, there were approximately 60000000 (sixty
million) cases of H1N1 influenza (CDC 2010a). So the chance of contracting it during this
year is one in sixty million, so dividing by 365 will give the chance per day:
pSE=1
60000000∗365≅ 0.0000000000456621
Again, during the 2009 epidemic, the US had 12270 deaths in the year (CDC 2010a). So, the
reciprocal of this quantity further divided by 365 will give the chance of dying from H1N1 if
you have the virus on any given day. An assumption that dying due to H1N1 whether the
individual is exposed or infectious is equivalent regardless of the state they are in is made
here:
pEX=p IX=1
12270∗365≅ 0.000000223287
The US government administered a figure of 91000000 (ninety-one million) (Drummond
2010) vaccinations during the year of the outbreak. That works out to approximately 249315
vaccines performed each day. The reciprocal of this gives the chance of being vaccinated each
day:
pSR=1
249315=0.000004 010989≅ 0.00000401099
75
The latency period of the virus is approximately 2.5 days (Balcan et al. 2009, CDC 2010b),
and taking the reciprocal of this gives pEI, so:
pEI=1
2.5=0.4
The recovery period after becoming contagious is approximately four days on average (Asp
2009), so again, taking this value and finding the reciprocal gives p IR, so:
p IR=14=0.25
The final probability we need to calculate from data is the chance of someone who has
recovered (after either having H1N1 or having the vaccination) loses their active immunity
and becomes susceptible again. Approximately 5 % of individuals who recover or are
vaccinated lose their immunity at some point in the year or do not gain immunity from the
vaccination in the first place (Greenberg et al. 2009). This implies our value for pRS is 5 %
divided by 365:
pRS=0.05365
≅ 0.000136986
Just as before, like the other stochastic models, all rows of the matrix must add up to one, so
the final stationary probabilities can all be worked out using the values we already know
(given to full decimal places):
pSS=1−0.0000000000456621−0.00000401099−0.0000346801=0.9999613088643379
pEE=1−0.4−0.000000223287−0.0000346801=0.599965096613
p II=1−0.25−0.000000223287−0.0000346801=0.749965096613
pRR=1−0.000136986−0.0000346801=0.9998283339
Combining all of the above results give the following final transition matrix (given to
appropriate significant figures):
76
P=¿
(0.999961 4.57×10−11 0 4.01×10−6 3.47×10−5 0 0 0
0 0.599965 0.4 0 0 2.23×10−7 3.47× 10−5 00 0 0.749965 0.25 0 2.23×10−7 3.47× 10−5 0
0.000136986 0 0 0.999828 0 0 0 3.47× 10−5
0 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1
)1.6.3. n-step transition
As before, P365 gives the probabilities that an individual will end in any desired state, given
any initial starting state, at the end of one year. Hence, using Maple (omitted due to size, see
Appendix A10), this gives the 365-step matrix (given to appropriate significant figures):
P365=¿
(0.986012 1.13 ×10−10 1.80 ×10−10 0.00140919 0.0125697 2.37 × 10−14 3.68 ×10−12 9.01×10−6
0.0472912 5.36 ×10−12 8.48 ×10−12 0.940129 0.000296913 1.45 ×10−6 0.000225382 0.01205640.0476128 5.40 ×10−12 8.54 ×10−12 0.939808 0.000301039 8.93 ×10−7 0.000138701 0.01213890.0481270 5.46 ×10−12 8.64 ×10−12 0.939294 0.000307686 5.72× 10−16 8.89 ×10−14 0.0122710
0 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1
)Like the first stochastic model, if we assume all individuals start off susceptible, we can just
concentrate on the first row of P365. Approximately 5 % of all individuals who are in the
recovered state lose their immunity and if 91 million people per year are immunised, this
gives a vaccination proportion in the first year of 91000000
316148990≅ 0.287857. This allows us to
work out the proportion of pSR(365) and pSZ
(365) that actually have had the virus at some point and
77
are not in those states because of vaccination and to account for those who loop round the
system:
pSE(365)+ pSI
(365)+(1−0.287857+0.05 )∗pSR(365 )+ pSX
( 365 )+ pSY(365 )+ (1−0.287857+0.05 )∗pSZ
( 365 )=pSE(365 )+ pSI
( 365 )+0.762143∗pSR( 365 )+ pSX
(365 )+ pSY(365 )+0.762143∗pSZ
(365)
The first three probabilities of these six represent those who were susceptible to start with but
have ended day 365 being exposed, infectious or having recovered, respectively. The
coefficient of pSR(365) accounts for looping and loss of immunity, as not all individuals in the
recovered state after one year will have necessarily had H1N1. The last three represent those
who were susceptible but have ended the year having passed away naturally after being
exposed, infectious or having recovered, or died due to the virus whilst they were exposed or
infectious. Again, the coefficient of pSZ(365) acts similarly to above.
Hence, the probability that any individual in the US contracts H1N1 in this year is:
pSE(365 )+ pSI
( 365 )+0.762143∗pSR( 365 )+ pSX
(365 )+pSY(365 )+0.762143∗pSZ
(365 )=0.000000000112559+0.000000000180098+0.762143∗0.00140919+0.0000000000000236918+0.00000000000367973+0.762143∗0.00000900923≅ 0.00108087
Hence, an individual in this population will have a 0.108 % probability of contracting H1N1
in this year of life.
Individuals who die directly due to H1N1 in this year can be estimated as follows. All
individuals who die due to the virus must end up in the X state. Once in there, they cannot
leave, so the probability of dying from the virus is simpler to see than the cases of H1N1:
pSX(365 )=0.0000000000000236918
1.6.4. Limiting distribution
Like before, if we take P to higher and higher powers, we can get towards the limiting
distribution of this matrix system. P365 gives the probability of an individual, hence the
78
percentage of a population, that will get the disease in one year. Raising this new matrix to the
power n will give the percentage of a constant population that will contract the disease at
some point during n years, just as we had with the previous disease. As n → ∞, the matrix
( P365 )n=P365∗n → Pn. So raising the original matrix to exceptionally high powers will give the
steady proportion of a population (based on US data) that will contract H1N1.
After some manual checking (using integer powers of 10 for n), just like before both
P365∗100000=P36500000 and P365∗1000000¿ P365000000 give the same matrix (using Maple, omitted due
to size, see Appendix A11). This implies that the limiting distribution, Pn, is (to appropriate
significant figures):
Pn=¿
(0 0 0 0 0.977168 0.00000000000186701 0.000000000289977 0.02283190 0 0 0 0.779583 0.00000145112 0.000225382 0.2201900 0 0 0 0.779651 0.000000893025 0.000138701 0.2202090 0 0 0 0.779760 0.00000000000148983 0.000000000231396 0.2202400 0 0 0 1 0 0 00 0 0 0 0 1 0 00 0 0 0 0 0 1 00 0 0 0 0 0 0 1
)From this, the result here shows that given enough time, all individuals will pass away (all
elements in the first four columns are zero so no-one ends up in the S, E, I or R states), which
again, is obvious. By adding pSX( n) + pSY
(n ) , it shows a steady percentage of the population who
will have caught H1N1 and died whilst infected (naturally or due to the virus). Then adding a
proportion of those who ended up dying whilst susceptible or recovered will give us the total
number of cases. This proportion is worked out as follows:
As before, 0.762143∗pSZ(n ) would give the correct proportion for this compartment. We need to
work a similar one out for the pSW( n) element. Now, (1−0.762143 )∗pSW
( n) would give those who
79
have been round the system before, so working these two values out and then adding to pSX( n)
and pSY( n) gives:
pSX( n) + pSY
(n ) +0.762143∗pSZ(n )+ (1−0.762143 )∗pSW
(n ) =0.00000000000186701+0.000000000289977+0.762143∗0.0228319+0.237857∗0.977168≅ 0.249904
So 25.0 % of the population will contract H1N1 at some point during their life, assuming a
long-term outbreak is sustained.
Deaths from the virus are much simpler to work out, it is just the value of pSX( n) :
pSX( n) =0.00000000000186701
So 0.000000000187 % of the population will contract H1N1 and go on to die directly due to
the virus, in a steady population that stays constant for many years. As this value is very
small, we can assume a low death rate from H1N1 if an outbreak or epidemic is sustained.
80
2. Results and model comparisons
Using the results above from all six models, presented here is an overview and comparison of
how accurate each model was. Although comparisons are made between each model and the
number of cases, the main comparison is that of how much the models differ in their
predictions in the long-term (i.e. non-trivial endemic steady states against limiting
distributions). Note that for chickenpox and measles, actual cases and actual deaths are
multiplied by the life expectancy in the relevant country as these diseases are constantly
around and never truly flare up or outbreak under normal circumstances. However, for H1N1,
there is approximately one outbreak in each lifetime (based on the space between the previous
two epidemics) so actual cases per lifetime are just based off a single outbreak. The UK
population is taken to be 64.1 million (ONS 2014) and the US population as 316148990 (U.S.
Census Bureau 2013). These errors are shown in Table 7.1., below.
As the maximum positive error in Table 7.1. is theoretically infinite, yet the maximum
negative error is only −100 % (as the lowest prediction is clearly 0, so
0actual cases /deaths
−1=−100 % for all cases or death values), it could be seen that all
under-predictions are closer to the actual values than any over-predictions. To counter this
unwanted effect, taking the logarithm (base 10) of each of the error calculations plus one (i.e.
just the fractional part) will allow simpler and more visual comparisons, and permit each error
to be seen as to how many orders of magnitude difference they are away from the real number
and each other’s predictions. These calculations are also shown in Table 7.2., below.
81
Disease Chickenpox (UK) Measles (US) H1N1 pandemic (US)
Deterministic
model cases
prediction per
lifetime:
0.7389033941048∗64100000≅ 473637080.07497197730196∗316148990≅ 237023150.1967780042823∗(1+0.02325592285 )∗316148990≅ 63657945
Stochastic model
cases prediction
per lifetime:
0.315864254∗64100000≅ 202468990.0787776033870348∗316148990≅ 249054590.249904∗316148990≅ 79006897
Actual number of
cases per
lifetime:
365370∗81=29594970 55∗79=4345 60000000 per outbreak∗app . 1 outbreak per
lifetime=60000000
Deterministic
model deaths
prediction per
lifetime:
N/A 0.00003149821716∗316148990≅ 99580.0009612958123∗316148990≅ 303913
Stochastic model
deaths prediction
per lifetime:
N/A 0.0000000003870348∗316148990≅ 0.1223610.00000000000186701∗316148990≅ 0.000590253
Actual number of
deaths per
lifetime:
N/A 55∗79∗1.51000
≅ 712270 per outbreak∗app .1 outbreak per
lifetime=12270
Deterministic
model case error:
4736370829594970
−1≅+60.0 %error23702315
4345−1≅+5454 %error
6365794560000000
−1≅+0.061 % error
Stochastic model
case error:
2024689929594970
−1≅−31.6 %error24905459
4345−1≅+5731 %error
7900689760000000
−1≅+31.7 % error
Deterministic
model death error:
N/A 99587
−1≅+1422 % error30391312270
−1≅+23.8 % error
Stochastic model
death error:
N/A 0.1223617
−1≅−98.3 %error0.000590253
12270−1≅−100 %error
Table 7.1.: Comparison of models to actual cases and deaths plus errors in predictions
Disease Chickenpox (UK) Measles (US) H1N1 pandemic (US)
82
Deterministic
model case error:log10( 47363708
29594970 )≅ 0.2042log10( 237023154345 )≅ 3.7368 log10( 63657945
60000000 )≅ 0.0257
Stochastic model
case error:log10( 20246899
29594970 )≅−0.1649log10( 249054594345 )≅ 3.7583 log10( 79006897
60000000 )≅ 0.1195
Deterministic
model death
error:
N/Alog10( 9958
7 )≅ 3.1531 log10( 30391312270 )≅ 1.3939
Stochastic model
death error:
N/Alog10( 0.122361
7 )≅−1.7575log10( 0.00059025312270 )≅−7.3178
Table 7.2.: Logarithmic model errors in predictions
2.1. Model 1 comparisons
As can be seen from the tables above, the results for each disease vary, especially when
comparing each model with the same disease. Taking chickenpox for the first comparison, it
can be seen that the traditional model over-predicts the number of cases by approximately
60%, whereas the stochastic model under-predicts them by just over 30%. Looking at the
logarithmic figures directly, both predictions are less than an order-of-magnitude difference
away from the real-world case number (within one either side of zero), but more importantly,
both the predictions themselves are less than an order-of-magnitude difference away from
each other (|0.2042−(−0.1649 )|=0.3691<1). This implies that the basis for the adaptation to
construct the stochastic model for chickenpox was good and provided a good yet novel model.
2.2. Model 2 comparisons
For measles, the results are not as reliable. Both the standard model and the stochastic one
predict case values well above that of real life, giving huge over-predictions of well into the
83
order of the five-thousand percentage range. Looking at the logarithmic figures, both
predictions are huge, being an order-of-magnitude of almost four away. However, like for
chickenpox, the adapted model does follow the original model quite well; both the predictions
themselves are again, less than an order-of-magnitude difference away from each other (
|3.7368−3.7583|=0.0215≪1). This error is likely due to the inclusion of a maternally
immune state. In most models, individuals born enter the susceptible state but here, they do
not. This had added huge complexity to the models, such as how many individuals should
start maternally immune during numerical analysis and which row or combination of rows of
the limiting distribution matrix should be used from the stochastic approach. Deaths from the
disease are included in this model analysis too, yet both are wildly inaccurate. The standard
model over-predicts by well over one-thousand percent and the adapted model under-predicts
virtually no deaths (less than one person would die). The logarithmic values do differ here, as
they are an order-of-magnitude of almost five away from each other (
|3.1531−(−1.7575 )|=4.9106). This implies that the mortality number predictions for both
models are flawed, not only in comparison to the actual deaths from measles but in
comparison to each other.
2.3. Model 3 comparisons
Finally, for the H1N1 virus, the results are better. Both the standard model and the stochastic
one predict case values only a little more than that of real life, giving slight over-predictions
in both approaches. The errors are similar, with the standard model being closer at a lowly
error of 0.061% and the stochastic model not far behind at 31.7 %. Although percentage point
wise, these look vastly different, looking at the logarithmic figures, both predictions are very
small, and are well within an order-of-magnitude difference away from the real-life cases.
Compared to each other, they give the closest comparison of all parts of all accurate model
adaptations, having a tiny difference in relative errors (|0.0257−0.1195|=0.0938≪1). Like
measles, deaths from H1N1 are included in this model analysis too, however just as before,
both are inaccurate. The standard model does well and only over-predicts by approximately
five times the actual value. The stochastic adaptation does far worse, having an even smaller
value for mortality numbers than measles did, even though more individuals would die from
84
H1N1 in real life. The logarithmic values differ hugely, as they are an order-of-magnitude of
almost nine away from each other (|1.3939−(−7.3178 )|=8.7117), giving our worst prediction
for any part of any model, contrasting the accuracy of the case prediction for this disease. This
implies that the mortality number predictions for both the standard and stochastic models are
inconsistent yet again, in comparison to the actual deaths in the real world and to each other.
85
3. Conclusions
3.1. Conclusion and critical evaluation
In conclusion, the stochastic approaches attempted in this paper worked well only when the
original deterministic model was relatively accurate. The models for chickenpox and the
H1N1 virus both estimated very accurate case numbers, as seen in Table 7.2., and the
following analysis showed that both models were accurate to each other as the difference
between their respective logarithmic values were well within one. The models for measles,
both deterministic and probabilistic, were not accurate to real life, well over-predicting the
actual number of cases. This was due to the complexity of the model and the fact that not all
individuals start out maternally immune in real life. That is, starting analysis at any one point,
a certain (low) proportion of the population would be maternally immune (the long-term
Runge-Kutta analysis suggests this as approximately 1.1 %). However, even when starting
with M (0 ) at around that point, the analysis was still not close to approaching a real life case
value.
The mortality modelling parts for the latter two models which did include it was also not
accurate for both types of models. This implies that the setup of the mortality sections was
flawed from the start; the deterministic model did not have the correct structure for modelling
deaths due to the disease, so when adapted and changed into the stochastic method, this
suffered the same fate. However, the stochastic model predicts very few deaths (even as low
as a fraction of a person), whereas the deterministic models well over-predicted the mortality
from disease. This makes sense, however, as the rate in the deterministic model would slowly
‘build up’ the deaths and they will never fall, where the stochastic model has a constant an
unchanging probability each day due to the fact it has been modelled as a simple Markov
chain.
86
Comparing the models directly to each other shows the main point of this research; stochastic
modelling using Markov chains is feasible, relatively simple and provides results as accurate
as those of standard deterministic modelling using systems of ordinary differential equations.
The initial models are built on the same state-space within each disease, and the same
characteristics of each disease are used between the deterministic and stochastic models. But
the steps each model takes in the analysis differ based upon fixed rates and fixed probabilities,
respectively. This allows for the most reliable comparison between two completely different
methods, negating any effects besides the direct comparisons intended to be made here. To
improve the robustness of each model, more data over longer period of time (i.e. longer than
one year) could be used; however, this is unlikely to be accurate or available for most diseases
where a lifetime comparison is to be completed (as data would need to stretch back
approximately 80 years).
Furthermore to the issues encountered whilst deciding on reliability of the results,
timekeeping and planning ahead during the entire dissertation was difficult. I had originally
planned to get all of the analysis of all of the models finished with over a month to go, thus
allowing me time to read many more articles that I did and take my time during the write-up
and critical evaluation of other people’s work and my own. However, unsurprisingly, my
plans had to change and adapt accordingly after my models became complicated from the
outset. This meant my final model was only finished with two weeks to go, halving the
amount of time I had available. Obviously, the write-up and evaluations had been going on
along the way and had been started well before this, but some elements could not be fully
included until my own analysis was complete. My time schedule was shorter than anticipated
and involved more work than originally planned, but I feel I coped with the demand and
workload well over the period since I started this dissertation.
Overall, although the stochastic approach did not reduce the reliability or precision of the
cases predicted, it did not improve on them either. That way, presented here is not a
progression of the original model into a better one, but more of an alternative approach which
can be used in the right circumstances if the researcher believes it to be more beneficial. This
87
then poses a new question; is the novel stochastic method worth implementing in place of an
already established method that is used by mathematicians and biologists worldwide?
3.2. Further work
Models that use differing rates and probabilities of infection for each time period could be
used. These models would, however, be far more complex, possibly even too complex to
apply to a disease analysis. Constant rates and probabilities were used throughout this
research; firstly to keep the simplicity of each model and remove over-complicating the
problem, and secondly to allow for the use of a Markov chain in the adaptations. Markov
chains have a property that requires constant probabilities for every step of the analysis,
regardless of the point the analysis is at. If the probability was variable over time, the limiting
distribution could not be found and hence, this research would not be possible. More work
could be undertaken where probability changes. This would either require a completely
different stochastic process to be used or Markov chains with a theoretically infinite state-
space where the infectious state is split up into states representing each day of a disease. The
states would have probabilities associated with them that the individual could either pass on to
the next day (an average case) or recover and leave the infectious part of the model. These
states could be limited in number (say N ) but this would force any individual in the system to
recover within N stages (for example, N days) and would not realistically and truly give an
appropriate view of the disease’s traits. Besides these points, a transition matrix of order N
would need to be used; where small errors in this method would be enhanced to unreliable
levels after hundreds of iterations of matrix multiplication. If another stochastic process was
used besides Markov chains, probability density functions could be introduced, but as stated
before, the complexity of this system would shoot up and make it unviable for use instead of a
deterministic model. It should be noted that over-fitting of the data to new models can occur if
too many factors are allowed for and including in working out the probabilities associated.
This would present a perfect model for the problem at hand but later use of the model will
likely be questionable.
88
Further work would need to be attempted to model deaths more accurately from diseases,
possibly even as far to say as developing specific and separate ‘mortality model’ for the
disease in question where the entire population is assumed to be infected and we are
effectively ‘zooming in’ on just the virus deaths part to each model. Demography in this
model would not represent births and natural deaths, but more like the contraction and
recovery from the disease. Constant proportionality could be used in this model if the
assumption is made that the long-term steady state has already been reached so I (t ) will be
unchanged over time.
Spatial elements would be the next logical step from this research; using systems of partial
differential equations to represent a disease flow within a population and show how exactly
the disease spreads. The space in the model could represent ‘hotspots’ where a disease
initially flares up, so that certain individuals in the model are technically immune until the
individuals with the disease spread to areas nearby. However, without extensive
simplifications, the stochastic method using matrices may struggle to keep up with the
relatively ease of use and application of the traditional modelling methods. Another type of
‘grouping’ could be established, such as specific age ranges being more susceptible to certain
diseases or individuals becoming more or less susceptible at certain times of the year or
period of analysis. The easiest way to analyse age ranges may be to have separate models for
each age group and then combine the results at the end, using real life age distributions to
fully evaluate the data. Different parameters or probabilities could then be used to account for
certain ages being more likely to catch a disease. The seasonal effects of a disease could be
modelled similarly; possibly four models representing each season or even twelve
representing each month. The initial values of one model would have to be the final values in
the analysis of the previous one, meaning the differing parameters in the next model could try
to accurately account for temperature or other environmental factors.
89
4. References
Adeleye, E.S. (2014) Modelling The Impact of BCG Vaccines on Tuberculosis Epidemics.
Journal of Mathematical Modelling and Application, 1(9), pp.49-55.
Asp, K. (2009) Recovering From H1N1: What to expect after you've been struck with the
swine flu [online]. Everyday Health. [Accessed 14 January 2014]. Available at:
<http://www.everydayhealth.com/the-swine-flu/recovering-from-h1n1.aspx>
Balcan, D., Hu, H., Goncalves, B., Bajardi, P., Poletto, C., Ramasco, J.J., Paolotti, D., Perra,
N., Tizzoni, M., Van den Broeck, W., Colizza, V. & Vespignani, A. (2009) Seasonal
transmission potential and activity peaks of the new influenza A(H1N1): a Monte Carlo
likelihood analysis based on human mobility. BMC Medicine, 7(45).
Barry, J.M. (2009) White paper on novel H1N1 [online]. Massachusetts Institute of
Technology Engineering Systems Division. [Accessed 14 January 2014]. Available at:
<http://esd.mit.edu/wps/2009/esd-wp-2009-07-072709.pdf>
Centre for Disease Control and Prevention (CDC) (2009a) Measles (Rubeola): Measles
Vaccination [online]. CDC. [Accessed 29 August 2014]. Available at:
<http://www.cdc.gov/measles/vaccination.html>
Centre for Disease Control and Prevention (CDC) (2009b) Measles (Rubeola): Overview of
Measles Disease [online]. CDC. [Accessed 29 August 2014]. Available at:
<http://www.cdc.gov/measles/about/overview.html>
Centre for Disease Control and Prevention (CDC) (2009c) Measles (Rubeola): Signs and
Symptoms [online]. CDC. [Accessed 29 August 2014]. Available at:
<http://www.cdc.gov/measles/about/signs-symptoms.html>
Centres for Disease Control and Prevention (CDC) (2009d) Novel H1N1 Flu: Background on
the Situation [online]. CDC. [Accessed 13 January 2014]. Available at:
<http://www.cdc.gov/h1n1flu/background.htm>
90
Centres for Disease Control and Prevention (CDC) (2010a) CDC Estimates of 2009 H1N1
Influenza Cases, Hospitalizations and Deaths in the United States, April 2009 – March 13,
2010 [online]. CDC. [Accessed 15 January 2014]. Available at:
<http://www.cdc.gov/h1n1flu/estimates/April_March_13.htm>
Centres for Disease Control and Prevention (CDC) (2010b) Interim Guidance on Infection
Control Measures for 2009 H1N1 Influenza in Healthcare Settings, Including Protection of
Healthcare Personnel [online]. CDC. [Accessed 15 January 2014]. Available at:
<http://www.cdc.gov/h1n1flu/guidelines_infection_control.htm>
Centre for Disease Control and Prevention (CDC) (2013) Estimated Vaccination Coverage,
with Individual Vaccines and Selected Vaccination Series Among Children Aged 19-35
Months by State and Selected Area -- National Immunization Survey, United States, 2013
[online]. CDC. [Accessed 29 August 2014]. Available at: <http://www.cdc.gov/vaccines/imz-
managers/coverage/nis/child/tables/13/tab02_antigen_iap_2013.pdf>
Centre for Disease Control and Prevention (CDC) (2014) Smallpox: Disease, Prevention, and
Intervention [online]. CDC. [Accessed 29 August 2014]. Available at:
<http://www.bt.cdc.gov/agent/smallpox/training/overview/pdf/eradicationhistory.pdf>
Centre for the Mathematical Modelling of Infectious Diseases (CMMID) (2014) The Centre
for the Mathematical Modelling of Infectious Diseases [online]. London School of Hygiene &
Tropical Medicine (LSHTM). [Accessed 31 July 2014]. Available at:
<http://cmmid.lshtm.ac.uk/>
Dexter, N. (2003) Stochastic models of foot and mouth disease in feral pigs in the Australian
semi-arid rangelands. Journal of Applied Ecology, 40(2), pp.293-306.
Drummond, K. (2010) Once Scarce, H1N1 Vaccines Now Getting Dumped [online]. AOL
News. [Accessed 15 January 2014]. Available at:
<http://web.archive.org/web/20100402224921/http://www.aolnews.com/health/article/once-
hard-to-get-h1n1-vaccines-now-getting-dumped/19422863>
91
Fleming, D.M., Elliot, A.J., Miles, J., Barley, M.A. and Grant, S. (2007) Weekly Returns
Service Annual Prevalence Report 2007 [online]. Royal College of General Practitioners
(RCGP). [Accessed 15 September 2014]. Available at: <http://www.rcgp.org.uk/clinical-and-
research/~/media/Files/CIRC/CIRC-76-80/BRU_Annual_prevalence_report_2007.ashx>
Flu.gov (2014) H1N1 (originally referred to as Swine Flu) [online]. Flu.gov. [Accessed 26
August 2014]. Available at: <http://www.flu.gov/about_the_flu/h1n1/>
Greenberg, M.E., Lai, M.H., Hartel, G.F., Wichems, C.H., Gittleson, G., Bennet, J., Dawson,
G., Hu, W., Leggio, C., Washington, D. & Basser, R.L. (2009) Response to a Monovalent
2009 Influenza A (H1N1) Vaccine. The New England Journal of Medicine, 361, pp.2405-
2413.
Hille, P. (2014) Ebola outbreak in DRC: same virus, but different [online]. Deutsche Welle
(DW). [Accessed 3 September 2014]. Available at: <http://www.dw.de/ebola-outbreak-in-
drc-same-virus-but-different/a-17875656>
Höhle, M., Jørgensen, E. and O'Neill, P.D. (2005) Inference in disease transmission
experiments by using stochastic epidemic models. Journal of the Royal Statistical Society:
Series C (Applied Statistics), 54(2), pp.349-266.
Inaba, Y. (2014) An Example of Statistical Modeling for Count Data Analysis in Secondary
Education. Journal of Mathematical Modelling and Application, 1(9), pp.14-21.
Knott, L. (2013) Chickenpox [online]. Patient.co.uk. [Accessed 12 September 2014].
Available at: <http://www.patient.co.uk/doctor/chickenpox-pro>
Lamprecht, C.L. (2012) Chickenpox [online]. KidsHealth. [Accessed 12 September 2014].
Available at: <http://kidshealth.org/kid/ill_injure/sick/chicken_pox.html>
Meiss, J. (2007) Dynamical systems. Scholarpedia, 2(2), p.1629.
92
Nardone, A., de Ory, F., Carton, M., Cohen, D., van Damme, P., Davidkin, I., Rota, M.C., de
Melker, H., Mossong, J., Slacikova, M., Tischer, A., Andrews, N., Berbers, G., Gabutti, G.,
Gay, N., Jones, L., Jokinen, S., Kafatos, G., Martínez de Aragón, M.V., Schneider, F.,
Smetana, Z., Vargova, B., Vranckx, R. and Miller, E. (2007) The comparative sero-
epidemiology of varicella zoster virus in 11 countries in the European region. Vaccine,
25(45), pp.7866-7872.
National Health Service (NHS) (2013) Measles: Introduction [online]. NHS Choices.
[Accessed 29 August 2014]. Available at:
<http://www.nhs.uk/conditions/measles/Pages/Introduction.aspx>
National Health Service (NHS) (2014a) Chickenpox: Introduction [online]. NHS Choices.
[Accessed 11 September 2014]. Available at:
<http://www.nhs.uk/conditions/chickenpox/Pages/Introduction.aspx>
National Health Service (NHS) (2014b) Chickenpox: Symptoms of chickenpox [online]. NHS
Choices. [Accessed 11 September 2014]. Available at:
<http://www.nhs.uk/Conditions/Chickenpox/Pages/Symptoms.aspx>
Nelson, E. (1985) Quantum Fluctuations. Princeton: Princeton University Press.
Neves, R.G. and Teodoro, V.D. (2010) Enhancing Science and Mathematics Education with
Computational Modelling. Journal of Mathematical Modelling and Application, 1(2), pp.2-
15.
Nicoara, C., Zäch, K., Trachsel, D., Germann, D. and Matter, L. (1999) Decay of Passively
Acquired Maternal Antibodies against Measles, Mumps, and Rubella Viruses [online].
Clinical and Diagnostic Laboratory Immunology, 6(6), pp.868-871. [Accessed 29 August
2014]. Available at: <http://www.ncbi.nlm.nih.gov/pmc/articles/PMC95790/>
Office for National Statistics (ONS) (2014) Population Estimates for UK, England and
Wales, Scotland and Northern Ireland, Mid-2013 [online]. ONS. [Accessed 10 September
2014]. Available at: <http://www.ons.gov.uk/ons/rel/pop-estimate/population-estimates-for-
uk--england-and-wales--scotland-and-northern-ireland/2013/index.html>
93
Pande, H. (2012) Mathematical modeling of the dependence of pattern of occurrence of
graphemes in different texts with reference to the occurrence in various words in intra text
perspective. Journal of Mathematical Modelling and Application, 1(8), pp.84-90.
Pech, R.P. and Hone, J. (1988) A model of the dynamics and control of an outbreak of foot
and mouth disease in feral pigs in Australia. Journal of Applied Ecology, 25, pp.63-78.
Pech, R.P. and McIlroy, J.C. (1990) A model of the velocity of the advance of foot and mouth
disease in feral pigs. Journal of Applied Ecology, 27, pp.635-650.
Public Health Action Support Team (PHAST) (2011) The Uses of Mathematical Modelling
Techniques in Health Service Planning [online]. HealthKnowledge. [Accessed 31 July 2014].
Available at: <http://www.healthknowledge.org.uk/public-health-textbook/health-
information/3c-applications/mathematical-modelling>
Schichl, H. (2004) Models and the History of Modelling. in Kallrath, J. (ed.) Modeling
Languages in Mathematical Optimization. Massachusetts: Kluwer Academic Publishers, p.28.
Spencer, S.E.F. (2007) Stochastic Epidemic Models for Emerging Diseases. Ph.D. Thesis,
University of Nottingham.
Stohlmann, M.S., Moore, T.J. and Cramer, K. (2012) Preservice Elementary Teachers'
Mathematical Content Knowledge from an Integrated STEM Modelling Activity. Journal of
Mathematical Modelling and Application, 1(8), pp.18-31.
United States Census Bureau (U.S. Census Bureau) (2013) U.S. and World Population Clock
(July 4, 2013) [online]. U.S. Census Bureau. [Accessed 3 September 2014]. Available at:
<http://www.census.gov/popclock>
Urwin, H. (2011b) MAT-20006: Stochastic Processes Lecture at Keele University, 8 February
2011.
Urwin, H. (2011b) MAT-20006: Stochastic Processes Lecture at Keele University, 10
February 2011.
94
Urwin, H. (2011c) MAT-20006: Stochastic Processes Lecture at Keele University, 15
February 2011.
Weiss, H. (2013) The SIR model and the Foundations of Public Health. MATerials
MATemàtics (MAT2), 2013(3).
World Health Organization (WHO) (2012) Global Health Observatory Data Repository
[online]. WHO. [Accessed 12 September 2014]. Available at:
<http://apps.who.int/gho/data/node.main.688?lang=en>
Yorke, J.A. and London, W.P. (1973) Recurrent outbreaks of measles, chickenpox and
mumps: II. Systematic differences in contact rates and stochastic effects. American Journal of
Epidemiology, 98(6), pp.469-482.
95
5. Bibliography
Breathing earth
Age differences H1N1 http://www.eurosurveillance.org/ViewArticle.aspx?ArticleId=19344
96
6. Appendices
6.1. Appendix A
Omitted Maple outputs
1.1.1. Appendix A1
Endemic steady state values for deterministic model 1 – Chickenpox
>
>
>
>
>
>
>
>
>
>
>
97
>
1.1.2. Appendix A2
Endemic steady state for
deterministic model 2 – Measles
(with trivial steady state shown
first)
98
1.1.3. Appendix A3
Endemic steady state values for deterministic model 2 – Measles
>
>
>
>
>
99
>
>
>
>
>
>
>
>
>
>
>
>
>
>
100
1.1.4. Appendix A4
Endemic steady state for deterministic model 3 – H1N1 (with trivial steady state shown
second)
101
1.1.5. Appendix A5
Endemic steady state values for deterministic model 3 – H1N1
>
>
>
>
>
>
>
>
>
>
>
>
102
>
>
>
>
>
103
1.1.6. Appendix A6
365-step matrix for stochastic model 1 – Chickenpox
>
>
104
1.1.7. Appendix A7
Limiting distribution (both matrices equal) for stochastic model 1 – Chickenpox
>
>
105
1.1.8. Appendix A8
365-step matrix for stochastic model 2 – Measles
>
>
106
1.1.9. Appendix A9
Limiting distribution (both matrices equal) for stochastic model 2 – Measles
>
>
107
1.1.10. Appendix A10
365-step matrix for stochastic model 3 – H1N1
>
>
108
1.1.11. Appendix A11
Limiting distribution (both matrices equal) for stochastic model 3 – H1N1
>
>
109
1.2. Appendix B
4th-order Runge-Kutta numerical analysis using Microsoft Excel
1.2.1. Appendix B1
4th-order Runge-Kutta numerical analysis of deterministic model 1 – Chickenpox (first and
last pages of analysis only – see Appendix C for full data)
110
111
112
1.2.2. Appendix B2
4th-order Runge-Kutta numerical analysis of deterministic model 2 – Measles (first and last
pages of analysis only – see Appendix C for full data)
113
114
115
1.2.3. Appendix B3
4th-order Runge-Kutta numerical analysis of deterministic model 3 – H1N1 (first, stage when
the steady state values are reached and last pages of analysis only – see Appendix C for full
data)
116
117
118
119
1.3. Appendix C
Compact disc (CD) containing full data for all model analysis (4th-order Runge-Kutta
numerical analysis using Microsoft Excel) and full electronic copy of dissertation
120