200-StochasticDynamics.Rmd

# Stochastic Dynamics and Extinctions {#stochastic}

## Overview and Learning Objectives
In this chapter, we take a look at stochasticity and how it impacts ID dynamics. A focus is on the potential of ID extinction. We discuss how one needs to account for stochasticity in models if one wants to study extinction. 

The learning objectives for this chapter are:

*    Understand the difference between stochastic and deterministic models
*    Know when one needs to make use of stochastic models
*    Understand the concept of critical community size
*   Understand how stochasticity affects ID dynamics and can lead to extinctions


## Introduction
So far, all the models we have used to explore different aspects of ID dynamics have been _deterministic_. For a deterministic model, once you choose the initial conditions and parameter values, you always get the same result, no matter how often you run the model. There is no randomness present. Biological systems are never deterministic. There is always some amount of randomness or noise present. Models that allow for such randomness are called stochastic models. We discuss here when it is important to consider stochastic models and some of the questions one can only study with those models.


## Deterministic systems
Such deterministic models are relatively easy to implement and to study. That is one of the reasons they are so commonly used. Such a modeling approach also often provides a good representation of the dynamics of real ID if numbers are large. That is due to the law of large numbers. While every individual host or pathogen undergoes somewhat random dynamics, these random parts get _averaged out_, and the dynamics of the whole population is fairly predictable and well described by such deterministic models. (This is the same concept we use in classical epidemiological studies, where we enroll sufficiently large numbers such that we can say something predictable about differences between groups, even if individuals are somewhat unpredictable.)


## Stochasticity
Biological systems are never deterministic. When numbers are small, randomness starts to matter. As an example, consider an outbreak of some ID. Let's say we have a group of 100 infected. At a given time, some of them will transmit to others, and after that, they will recover from the disease, while others will recover before infecting others. The overall outbreak dynamics can be well approximated by a deterministic model if we accurately specify the average rate of transmission and the average rate of recovery. In contrast, now assume that there is a single infected person. Now it makes a huge difference if this person transmits to someone else before recovering, or if they recover before transmission occurs. In the latter case, the outbreak is over, and the ID has gone extinct. Thus, at small numbers, randomness/unpredictability in the order of events can make a large difference. 


## Discrete Numbers and Extinctions
Another problem inherent in the deterministic differential equation based models we have looked at so far is that they treat individuals in each compartment as continuous. For instance, the models allow 142.7 infected hosts. Of course, in reality, those numbers have to be integers. If we are dealing with large numbers, the difference between 142.7 and 142 or 143 is minor and doesn't matter. However, once numbers get small, the fact that we are dealing with discrete numbers of hosts becomes significant. So once less than 1 host has an infection the ID is in reality gone, but a continuous model would instead allow the number of infected to drop below 1 but remain above 0. In this way, the ID in the continuous models never actually goes extinct, it only gets closer and closer to 0 (and at some point is so close to 0 that the computer can't distinguish it from 0 anymore).

## Modeling Stochasticity and Extinction
To capture both the randomness and discrete nature of real systems requires a slightly different modeling approach. We can still use compartmental models, i.e., we track total numbers of individuals in particular states (e.g., S-I-R). But now, instead of letting the numbers in each compartment change continuously through inflows and outflows, changes happen in discrete steps, determined by specific events that occur in a probabilistic manner. 

As an example, instead of new susceptible hosts entering the system at some continuous birth rate, we now model births occurring as discrete events. Each birth event leads to an increase in the number of susceptibles by 1. 
The timing of the birth events is somewhat random, with an overall probability determined by the model (e.g., a higher birth rate will mean births occur more frequently, but still randomly). 


## Factors Affecting Extinction
Extinction of ID is of interest because that is often our final goal. We would like to drive an ID to extinction, either locally or even better, globally. If and how an ID can be driven to extinction depends on different factors. For IDs in humans, the only real chance is to tackle those with no other hosts or reservoirs. As such, we will likely never be able to eradicate influenza. But measles eradication might be possible (though hard).

Factors related to the host population and distinct ID characteristics influence the likelihood of an ID to go extinct. Two significant population factors are the size of the host population and the speed at which new susceptible hosts are replenished. Important characteristics of the ID included the infectiousness, the duration of the latent and infectious periods, the ability of the ID to persist outside the host, and other.


## Critical Community Size
The minimum size of a population at which an ID can maintain itself in a population without extinction has been termed the Critical Community Size (CCS) (this also depends on the replenishment of new susceptibles, e.g., birth rates). Studies for measles suggested that the CCS was several hundred thousand [@keeling02], meaning that measles could have only become a human pathogen once human populations were large enough to maintain it. The CCS is the number of susceptibles in a population and does not include those who are immune. Through vaccinating it is possible to reduce the size of the susceptible population below the CCS. So it is not necessary to vaccinate everyone to get an ID to go extinct. We just need to reach enough individuals to get the population of susceptibles below the critical level. This task is already hard by itself, as evidenced by the fact that we have not been able to eradicate measles, despite having a good vaccine.


## Host Extinction
Another way an ID can go extinct is if its hosts go extinct - either due to mortality from the ID or other causes. 
For most human diseases, such host extinction is fortunately not very common - though highly lethal ID such as Ebola can lead to marked reduction in the number of hosts in a given population. For non-human disease, extinctions of hosts due to disease is a more significant issue. Here, conversation efforts might try to achieve non-extinction of the host and - ideally - extinction of the ID. 

The host extinction approach is also often considered and used for vector-borne diseases. Here, the idea is to drive one of the hosts (commonly called the vector) to extinction. The obvious reasoning is that if there are fewer vectors (e.g., mosquitoes), the chances for humans to get infected are lower. Widespread use of DDT or insecticide-coated bed-nets are examples of this. A similar but more controversial approach is the attempted control of rabies in humans by killing its primary host, vampire-bats [@stoner-duncan14].


## Disease Emergence
The flip-side of extinction is the emergence of a new disease. During extinction, infected/pathogen numbers move from a level that can be decently approximated by a deterministic model to numbers so small that it requires a stochastic analysis approach to allow the possibility of extinction. During emergence, the new disease starts at zero, then is introduced in modest numbers (possibly only a single introduction) into a new population, and "bounces around" for a while in small numbers. If conditions are right (i.e., local reproductive number greater than 1), the disease might take off and spread. It could however also go extinct without taking off. 

In contrast to deterministic models, a reproductive number that is larger than 1 is necessary, but not sufficient, for the pathogen to produce an outbreak. An initial introduction can by chance be followed by extinction (i.e. an infected person recovers before they can infect someone else), even if R~0~>1. Thus, the concept derived from deterministic models, that there is no outbreak for R~0~<1 and one gets an outbreak for R~0~>1 needs to be modified for stochastic (arguably closer to the real world) settings. An R~0~<1 still means no outbreak, but now for R~0~>1, an outbreak is not guaranteed. Instead, there is a certain probability that an outbreak happens. We won't go into details here (see e.g. [@keeling08]) but one finds that for an SIR-type stochastic model, the probability that an outbreak occurs, _P_, if started by a single infected is given by _P=1-1/R~0~_. Thus, as expected, the larger _R~0~_, the more likely an outbreak is to occur, but it's not guaranteed. Similarly, one can show that if there are initially _N_ infected individuals, the probability for an outbreak is _P=1-1/(R~0~)^N^_. Again, this makes intuitive sense, namely more infected individuals make it more likely that an outbreak occurs. For instance if the pathogen has an R~0~=2 and there are 10 initially infectious individuals, the probability for an outbreak is 99.9%. That means as soon as the first few individuals are infected, it is almost certain that the disease takes off.

Once several 10s-100s of individuals are infected, a deterministic approximation which focuses on the mean dynamics is often a reasonable approximation. Often, during the emergence process, the pathogen evolves which makes it potentially easier to establish itself and eventually take off. Both emerging infectious diseases and evolutionary dynamics are an important topics and thus get their own dedicated chapters.




## Are stochastic models better?
After having gone through this chapter and learned that stochastic models allow you to address questions that deterministic models cannot address, you might wonder why we should still use deterministic models at all. The main reasons are that deterministic models are generally easier to build, easier to analyze, faster to run on a computer, and importantly, much easier to fit to data. Often this simplicity is worth the loss in some level of realism. But of course, it depends on the specific scenario and question. Recall the example of different maps in a chapter \@ref(idstates). The right map depends on the question. Similarly, the most appropriate model depends on your question. For some questions, stochastic models are needed. For others, deterministic models might still be the better choice.


## Summary and Cartoon
This module provides a brief discussion of stochasticity in ID dynamics, and its relation to disease extinctions.



```{r smallpox,  fig.cap='Smallpox is _almost_ extinct.', echo=FALSE}
knitr::include_graphics("./images/smallpox.png")
```




## Exercises
* The _Stochastic dynamics_ app in the DSAIDE package provides hands-on computer exercises for this chapter.
* Read the article "Measles Endemicity in Insular Populations: Critical Community Size and Its Evolutionary Implication" by Black [@black66]. The paper discusses critical community size for measles. Find 2 papers/analyses that discuss the CCS for other IDs. Read, summarize and critique the papers.


## Further Resources
* Some further analysis of critical community size and measles can be found in [@bartlett57; @bartlett60; @keeling97; @keeling97a].
* Some further discussion of stochastic epidemic models can be found in [@black12; @britton10].
* A broad discussion of invasion and persistence of ID and the role of stochasticity is given in [@lloyd-smith05b].
* An interesting discussion of critical community size and its dependence on birth dynamics, especially birth pulses, is provided in [@peel14].

## References