Introduction to

Mathematical Models of
Infectious Disease in Livestock
Lecture 1: Introduction to mathematical modelling
Andrea Doeschl-Wilson

Group leader, Genetics & Genomics, The Roslin Institute and Royal Dick
School of Veterinary Studies, University of Edinburgh, UK
[email protected]
Purpose of this lecture

• Get some understanding what mathematical models are &

what they can / cannot do

• Get acquainted with different types of mathematical models

• Learn the basic principles for building, analysing, testing and

using mathematical models
What is a mathematical model?

Model (Definition):
• A representation of a system that allows for investigation
of the properties of the system and, in some cases,
prediction of future outcomes.
• Always requires simplification
Mathematical model:
• Uses mathematical equations to describe a system
p
 p
  p 
var( f m )  E p  var( f m | p)   varp  E( f m | p) 
m 1  m 1   m1 
 E( p)var( f )  E ( f )var( p)
2

 p 2f  f 2 p2 .
Why do we need models?

• Models provide a framework for conceptualizing our ideas about the

behaviour of a particular system
• Models allow us to find structure in complex systems and to
investigate how different factors interact
• Models can play an important role in informing policies:
• By providing understanding of underlying causes for a complex phenomenon
• By predicting the future
• By predicting the impact of interventions
 Why mathematics?

• Mathematics is a precise language

• Forces us to formulate concrete ideas and assumptions in an unambiguous way
• Mathematics is a concise language
• One equation says more than 1000 words
• Mathematics is a universal language
• Same mathematical techniques can be applied over a range of scales
• Mathematics is an old but still trendy language
• The rich toolbox created by mathematicians over centuries is available at our
disposal
• Mathematics is the language that computers understand best
Mathematical models synthesize results from many
experiments
• Experimental studies concentrate on specific aspects of a
system

• Fragmented understanding
of the system
Mathematical models synthesize results from many
experiments
• Experimental studies concentrate on specific aspects of a
system

• Fragmented understanding
of the system

• Often hard to infer how the

system functions as a whole
Mathematical models can unravel the unobservable

Often the traits that we can measure are not the

most informative traits

From: The little Prince

Mathematical models can unravel the unobservable

Often the traits that we can measure are not the

most informative traits

From: The little Prince

Mathematical models are not bound by physical constraints

• Powerful tool to explore ‘what if scenarios’

• Extremely useful in the context of infectious disease where
experimental constraints are strong
What do we use mathematical models for?
• Combine fragmented information into a comprehensive framework (e.g.
combine results from in-vitro and in-vivo experiments)

• Determine the relationship between underlying biological traits and

observable traits

• Test hypotheses that are difficult to test in empirical studies

• Make predictions & generate new hypotheses for future testing

• Assist with decision making by exploring ‘what if’ scenarios

Limitations of mathematical models

1. Lack of quantifiable knowledge

• Models that encompass mechanisms (e.g. infection process) require quantitative
understanding of these mechanisms in order to make reliable predictions
2. Lack of available data / methods for estimating model parameters
• E.g. how to estimate e.g. individual susceptibility & infectivity?
• Much improvement to be expected over the next years due to recent advances in
statistical inference and data explosion
3. Inherent stochasticity of the biological system
• Infection is a stochastic process
• It is impossible to make accurate predictions for infection spread on the
individual level
Classification of mathematical models

• There are many different types of mathematical models

• Classifying them into broad categories can tell you much about their
purpose & scope and often require different mathematical techniques
• Typical distinctions:
• Empirical vs mechanistic
• Deterministic vs. stochastic All mathematical models are
• Systems vs molecular model composed of variables and a
• Static vs dynamic mathematical representation of
• Linear vs non-linear the relationship between them
• Discrete vs. continuous
Empirical vs mechanistic models We will use both types of
approaches to study host-
pathogen interactions.
Empirical Models (also called Statistical Models):
(See lecture 8)
• Data driven modelling approach
• Starting point: data obtained from empirical studies
• Aim: to determine patterns & relationships between data (model variables)
• Require no prior knowledge of the underlying biology

Mechanistic Models (also called Process Based Models):

• Hypothesis driven modelling approach
• Starting point: specific phenomena of interest – observed from data
• Aim: to provide understanding for underlying mechanisms of this phenomenon
• Require prior understanding of system
• Data are used to parameterise / validate the model
Deterministic vs stochastic models
Deterministic models
• Assume that the outcome is precisely determined by the model
inputs and relationships
• Ignore all random variation
• A given input always produces the same output
Stochastic models We will use both types of
• Incorporate inherent randomness approaches for modelling
• Use a range of values for the model variables in form of epidemics
probability distributions
• The same input produces an ensemble of outputs
Hybrid models
• include stochasticity on one scale (e.g. population)
• assume underlying deterministic processes (e.g. for individual)
Classification according to the scale of modelling

•National
Mechanistic models often combine 2 or
• Herd more adjacent levels of the hierarchy
• Individual Systems models combine several
• Organ levels of the hierarchy
• Cell See lecture on within host infection
• Molecules
dynamics: (molecules  cell  organ)
• Genes

The appropriate scale for modelling depends on the model objectives

What is a simulation model?

• Simulation models are not specific types of mathematical models

• The term ‘simulation model’ refers to the process of implementing
mathematical model, i.e. via computer simulations
• Simulation models usually simulate the process of data generation
assuming the model was true
• E.g. simulate an epidemic or the within host infection process
• Simulate an experiment
 The 4 stages of modelling
2. Generate
predictions
& Analyse
3. Validate 4. Apply
Similar process as for conducting a
1. Build
biological experiment:
1. Design the experiment
2. Generate data
3. Analyse experimental data
4. Validate experimental findings
5. Apply results in practice
But modelling can be much more elaborate
 Stage 1: Building models

1. Define the model objectives

• Be clear about what you want your model to do
2. Determine the appropriate level & key model components
• What level of simplification is required?
• Apply the principle of Ockham’s razor (also known as the law of parsimony):
Stage 1: Building models (cont.)
3. Define your assumptions
• Assumptions reflect our beliefs how the system operates
• Remember: the model results are only as valid as the assumptions!
• Different assumptions can lead to fundamental differences:

Mass is
Mass is
variable
constant

Isaac Newton: founder of classical mechanics Albert Einstein: Founder of relativity theory
Example: A common assumption in population studies
“ A population grows at a rate that is proportional to its size”
𝑑𝑝
• Embedded in the deterministic model : = 𝑎𝑝,
𝑑𝑡
where p(t) is the population size at time t and a is a constant.
• The solution of this model is 𝑝 𝑡 = 𝑝(0)𝑒 𝑎𝑡 , i.e. population grows exponentially
• The model incorporates a number of other important assumptions:
A common assumption in population studies
“ A population grows at a rate that is proportional to its size”
𝑑𝑝
• Embedded in the deterministic model : = 𝑎𝑝,
𝑑𝑡
where p(t) is the population size at time t and a is a constant.
• The solution of this model is 𝑝 𝑡 = 𝑝(0)𝑒 𝑎𝑡 , i.e. population grows exponentially
• The model incorporates a number of other important assumptions:
1. There is no limiting factor that prevents the population to grow forever
2. Growth is a continuous process (embedded by the differential equation)
3. Growth follows a deterministic law
• Alternative stochastic approach: model birth and death events
• If any of these assumptions don’t hold, the model is wrong!
4. Produce a flow diagram

• Visual tool for formulating our beliefs and assumptions

• Describe the model components (variables) and their relationship
• Extremely important for complex models with many components and
relationships
• You will see many of these in this course

S I R
5. Write model equations

How to find the appropriate mathematical equations?

• Depends on the modelling approach:
• Statistical models are often represented by a single linear or non-linear function
• Deterministic mechanistic models of dynamical systems are usually represented
by systems of differential equations
• Stochastic models require expressions for the probability of events
• Start with equations from the literature
• You are likely not the first one to model a specific system. Start by exploring and
modifying existing models
• Explore your own data
• see e.g. ‘Woods model’ in within-host infection dynamics lecture
Stage 2: Generate model predictions & analyse
There are 2 ways of solving the model equations for given parameter values
1. Analytically (using mathematical principles)
• Ideal, provides exact solutions and hence a full insight of the model behaviour
• But usually only possible for very simple systems (e.g. one equation or system
of linear equations)
2. Numerically (using computers)
• Applies to most mathematical models
• Requires the use of numerical algorithms implemented in computational
routines (e.g. Euler method, Runge-Kutta, Monte-Carlo)
• Provides approximate solutions
• Use established code, avoid writing your own numerical solver!!!
Specifying appropriate model inputs & outputs

• Modeller’s dilemma: lack of physical constraints in the modelling

world implies that one can generate A LOT of data.
How to go about it in a systematic way?
1. Specify realistic value ranges for the model input parameters
2. Focus on relevant scenarios if the model involves simulations
3. Generate relevant outputs & summary statistics
Estimating model input parameter values

• Good estimates of the model input parameters are essential for

models with predictive power
• Apply principle of Ockham’s razor: favour the model with fewer parameters
• 2 sources for determining appropriate parameter values:
1. Use values reported in the literature
2. Fit your model to existing data (statistical inference)
• Note that it is often not possible to infer a unique value (with confidence interval) for
each model parameter from given data
• There are many different approaches of statistical inference; the right approach
depends on both the type of model & the data
See ‘Statistical Inference’
lectures on Thursday
Choosing relevant model scenarios & outputs

Criteria for choosing model scenarios:

• Realistic scenarios, to achieve your research objective
• Extreme scenarios, to determine the limitations of the model
Produce meaningful model outputs

• Models produce predictions for every variable over time

• Model variables are not always measurable  comparison to data difficult
• Produce also model outputs that can be directly compared to data
• essential for model validation
• Apply similar statistical analysis as for experimental data (frequency
distributions, means, variance etc.)
• Assess relationships between observable and underlying biological
traits
• useful for gaining new insights
Analysing the model

• The aim is to obtain a thorough understanding what your

model can / cannot do
• Comprises both qualitative & quantitative analysis:
• What types of response patterns does the model generate?
• How realistic are these?
• What mechanisms / parameter values produce the diverse patterns?
• Which inputs correspond to which outputs?
• How sensitive are the model output to changes in the input parameter
values?
• How stable are the model predictions to small changes in starting
values / assumptions?
• Very elaborate step and often results in rebuilding the model
Analysis techniques:
Distinguish between short- and long-term behaviour
1. Asymptotic behaviour
• Does the system eventually settle to a steady state?
• E.g. will the infection eventually clear or persist?
• How many steady states (long-term outcomes) are there?
• Under what conditions will a particular steady state be
reached
• Use mathematical stability analysis, bifurcation theory
2. Initial phase behaviour
• E.g. will the infection kick off after introduction of 1
infectious agent?
• How does the initial behaviour depend on the starting point?
Analysis techniques:
Sensitivity analysis & Uncertainty analysis
• Uncertainty analysis: assess variability in model outputs that arise from
uncertainty in model inputs
• How confident are we about the model predictions?
• Sensitivity analysis: quantifies the influence of each parameter or modelled
process on the model outputs
• How sensitive are the model predictions to changes in the input parameter
values or modelled processes?
Sensitivity analysis & Uncertainty analysis cont.

• Essential components of model analysis, especially when parameter values are

unknown
• Complex tasks, given that there are usually complex interactions between
parameter values
• Typical approaches:
• Change one / few parameters at a time, keeping the others fixed
• Adopt partial factorial designs, e.g. Latin Hypercube Sampling
Stage 3: Validating the model

• Ideally (but not necessarily!) involves comparison of model predictions to

experimental data
• Important to use independent data to those used for parameter estimation
• If independent data don’t exist, split your data into training and validation set
• Useful summary statistics for comparing model predictions (Pi) to observations (Oi):
1 𝑛
Bias (B) =
𝑛 𝑖=1(𝑃𝑖 − 𝑂𝑖 )
1 𝑛
Standard deviation (SD) = 𝑖=1(𝑃𝑖 − 𝑂𝑖 − 𝐵)2
𝑛
1 𝑛
Prediction mean square error (MSE) = 𝑖=1(𝑃𝑖 − 𝑂𝑖 )2
𝑛
What if model predictions are different to the
observations?
Identify potential reasons for imperfect predictions:
1. Natural variability in the real system and environment
• Equates to experimental measurement errors
• Obtain confidence intervals directly from the data; if model predictions fall within
these limits, don’t worry
2. Mis-specifications in the model
• Wrong parameter values  extend parameter range, use fitting algorithms
• Errors in the choice of model equations
• Restrict the scope of the model or look for better equations and start again
3. Effects of factors ignored in the model
• Increase model complexity and start again
Comparing alternative models

Independent models:
• Subjective choice: no objective model selection criterion available
• Balance between generality, flexibility, predictive ability, computing requirements

See Foot & Mouth Disease

models example later today
Related (e.g. nested) models:
• For models with likelihood (L), k parameters and n available independent data
points, use information criteria (IC) such as
• AIC (Akaike IC): -2log(L) + 2k; defined for nested models
• BIC (Bayesian IC): - 2log(L) + k log(n); penalizes models with more parameters
Stage 4: Applying the model

• Mathematical models can be a valuable decision support tool

• For risk assessment – particularly important in infectious disease context
• To predict consequences of various (disease) control strategies

• It requires trust that the model predictions are valid

• It is crucial to keep the purpose of the model and the end user of the model
in mind at all modelling stages
• The user should have a thorough understanding of the model assumptions,
model predictions (with uncertainty estimates) and limitations

See Foot & Mouth Disease

models example later today
What is a good model?

Key attributes of a good model:

1. Fit for purpose
• As simple as possible, but sufficiently complex to adequately represent the
real system without obstructing understanding
• Appropriate balance between accuracy, transparency and flexibility
2. For predictive models: Parameterizable from available data

Keep in mind that no model is perfect!

Further reading

• Otto, Sarah P., and Troy Day. A biologist's guide to mathematical modeling in ecology and
evolution. Vol. 13. Princeton University Press, 2007.
• A nice introduction to mathematical modelling with plenty of applications from ecology and
evolutionary systems.
• Renshaw, Eric. Modelling biological populations in space and time. Vol. 11. Cambridge University
Press, 1993.
• A good and not too mathematical introduction to deterministic and stochastic models of
biological systems.
• Cross, Mark, and Alfredo O. Moscardini. Learning the art of mathematical modelling. John Wiley
& Sons, Inc., 1985.
• A readable, non-technical book on how to start modelling and how to teach others.