Interpreting Logistic Equation in Foot and Mouth Disease Epidemic
Purpose
This page offers a qualitative interpretation for the logistic equation's good fit to the
Foot and Mouth Disease epidemic that occurred in the United Kingdom in 2001. (See
here for more details).
The interpretation develops a connection between the logistic equation and
the more sophisticated epidemiological models used by Howard [HOW011100],
Ferguson, and the UK Chief Scientist and the Ministry of Agriculture, Fisheries
and Food (MAFF), in fighting the 2001 epidemic.
Caveat
This following is not peer-reviewed, is not guaranteed error-free and does not claim to be authoritative. Rather
it is meant as a contribution to discussions on utility of logistic curve modelling, particularly in the UK FMD
epidemic. Use for any other purpose at your own peril.
The interval between the
point of infection and the onset of infectiousness
(2)
Asymptomatic infectious
period
The interval between the
onset of infectiousness and the appearance of clinical
signs of disease
(3)
Incubation period (sum of
(1) and (2)
The interval between the
point of infection and the appearance of clinical signs
of disease
(4)
Force of Infection
The rate at which
susceptible herds acquire infection from infectious herds
(5)
Basic Reproduction Number
The average number of
secondary herds infected by the introduction of one
primary infected herd into a susceptible population.
"The value of the basic reproduction number at any time
is given by the length of the infectious period multiplied by the
the force of infection at that time" [HOW011100,p3]
So
BRN(t) = Infectious Period x Force of Infection(t) ...(1)
"For an epidemic to be sustained, the basic reproduction
number has to be greater than one"
ie.
BRN(t) > 1
Now,
"The rate of spread of infection between herds, by whatever
mechanism, is proportional to the number of susceptible herds multiplied by the
density of infectious herds at that time."
- [HOW011100,p191] ...(2)
Developing the Logistic Equation
Assume that the susceptible herds and density of infected herds are both
functions of time, then [HOW011100]'s definition (2) can be written:
d(Infected Herds)/dt = Ka x Susceptible Herds(t) x Density of Infected Herds(t)
... (3)
Let P = infected herds
S = susceptible herds
Ka = constant of proportionality
What does "density of infected herds" mean ?
In a highly controlled experimental situation, you might have enough data to be
able to represent the spatial distribution of herds. But that's a heck of a lot of work,
and if the data you have to use in real life is dodgy, you might be better to work
on aggregates or averages.
Let's make a simplification: an infected herd means one capable of passing on an infection.
Strictly, this is ignoring the latent period between the time when an animal in a herd acquires
an infection and when it is capable of passing it on. This simplification would be valid if
that doesn't matter too much, or can't be reliably measured.
Assume that:
Density of Infected Herds(t) = ka / Ka [P(t) - E(t)] ...(4)
where E(t) = total infected herds completely exterminated
(incapable of infecting, by any means)
dP/dt = ka S(t) [ P(t) - E(t) ] ...(5)
For the epidemic to end, we need the population of confirmed cases
(infected herds) to reach a limit - with no further change. The
objective function is then:
dP/dt = 0 ...(6)
We don't know what sort of function to use for dP/dt. Alfred
Lotka postulated that it might be some function of the population size:
dP/dt = f(P) = 0 ...(7)
The Taylor series expansion of any function f(P) gives:
dP/dt = bP + dP2 + gP3 + .... ...(8)
Using one term of the taylor series expansion for f(P) gives
an exponentially increasing population - and can't satisfy the
objective (6).
Taking the first two terms of the taylor series expansion is the
simplest approximation to f(P) that will permit a
non-trivial solution to (6).
dP/dt = bP + dP2 ...(9)
Solving (6) and (9) gives
0 = Plimit ( b + dPlimit )
hence
b + dPlimit= 0
Plimit = - b /d ...(9a)
(
Since
Plimit ≥ 0
we would expect that b ≥ 0 and d ≤ 0 in equation (9).
)
Now what is a susceptible herd (S(t))? We suspect there is some limit to the
number of infected herds - let's forget for the moment that we've just postulated
an expression for it in equation (9a).
Let's suppose that anything beyond the as yet unknown limit is not considered susceptible (so, for example,
herds in Australia would not be included as susceptible
when we were studying an outbreak in the United Kingdom).
Let
klimit = P(t=¥)
then:
S(t) = klimit - P(t) ...(10)
where klimit = a positive constant > 0
Now substitute (10) into (5):
dP/dt = ka [ klimit - P(t) ] [ P(t) - E(t) ] ...(11)
dP/dt = ka [ klimit P(t) - P(t)2 + P(t) E(t) - klimit E(t) ]
then
let b = kaklimit d = ka
dP/dt = b P(t) - d P(t)2 +d P(t) E(t) - b E(t) ...(12)
If you exterminated no infected herds, or were so slow in doing it that the virus was free
to spread in a way that made no difference compared to zero extermination
(for example, by the time you culled, all the
infectivity had gone), then E(t) = 0
At the time of writing there is no clear evidence as to how much infection occurs
after reporting, case confirmation, or culling (see especially Ferguson's
difficulty with estimating his rI factor). It's then reasonable to
explore the possibility:
E(t) = 0 ...(13)
then substituting in (12):
dP/dt = b P(t) - d P(t)2 ...(14)
which is just the logistic equation and dP/dt = 0 means end of
the epidemic.
The goodness of fit of such a logistic equation to the data is
evidence that the suppositions are correct: namely the rate of
spread of infection is determined primarily by what happens
before an infected herd ceases to exist (I'm being careful here
to allow that carcasses might still exist from an infection
standpoint); and closeness to an infected doesn't particularly
matter as far as population aggregates are concerned. This is simplistic, but if reality is like that - e.g.
wind or other transport mechanisms are more important than
assumed to date - then the logistic fit will be good, or at least
better than alternative models.
But there might be a better qualitative explanation.
The number of exterminated cases will be proportional to
the infected cases, but with perhaps a loss factor and some lag due to the
lead time of detecting, then organising slaughter and disposal:
kvulnerable is a loss factor that
represents the proportion of the infected population
that will be vulernable to treatment - i.e. detected, and culled. Surveillance
and detection will never be 100% (e.g. asymptomatic carriers),
and some detected farms will manage to infect others after slaughtering
(sloppy infection control by slaughtermen etc.,.).
td is the time lag between diagnosis and slaughter &
disposal,
and (15) means you will are trying to exterminate all the cases infected
td days previous.
but if detection and slaughter is prompt (
as pre-emptive neighbour culling might cause) then, td=0
E(t) = kvulnerable * P(t) ...(16)
This is saying you are exterminating a high proportion of the infected cases
on the day they get infected. It's more than saying a high
proportion of confirmed cases is exterminated.
More generally, you could formulate the problem as a pair of differential
equations relating E(t) and P(t). And then, you could think of more complicated
ways to define the terms, and bring in more variables and more equations. So you
might represent the distinction between infected and confirmed cases, as Ferguson
et al did. Then you could go on to add even more variables than Ferguson (e.g. for
species differences). Where
do you stop ? At some point, there won't be sufficient data to make any meaningful
correlation with the data. Ferguson's inability to estimate rI suggests
he went too far. The suggestion here is that you stop when the model you get fits
good enough, and that (16) might be good enough.
substituting in (11):
dP/dt = ka [ klimit - P(t) ] [ P(t) - kvulnerable P(t) ] ...(17)
= ka [ klimit - P(t)] P(t) [ 1 - kvulnerable ]
= ka [1-kvulnerable] klimit P(t) - ka [1-kvulnerable] P(t)2
so
dP/dt = b' P(t) - d' P(t)2
which is the logistic equation
where
d' = ka [1-kvulnerable]
b' = ka [1-kvulnerable] klimit
= d' x klimit
with b'/d' = klimit
Note b'/d' is just the expression for the asymptote for P(t) (from
the solution to the differential equation),
and hence matches the interpretation put on klimit.
In using MAFF web-published statistics, we have only confirmed cases data to go by. Provided the confirmed cases are
proportional to the infected cases, the analysis still holds.
Conclusions
If the size of the exterminated infected population E(t) is
either effectively zero , or proportional to the size of the infected
population, you get a logistic equation.
Anecdotal evidence suggests that delays to slaughter were significant in the 2001 UK
outbreak. But it's
not clear whether the pre-emptive nature of the 3 km cull might still have left
an effective td of zero.
To decide conclusively which of the two explanations is correct,
you would need more data and research that was able to discriminate what amount of infection occurred after case confirmation,
and quality evidence of the slaughter delays.
In the context of the UK FMD epidemic, the pre-emptive nature of 3 km culling may have helped exterminate FMD in the
quicker manner predicted by the logistic equation.
Execution failures with the pre-emptive cull may have led to FMD lasting longer than
the logistic equation projections suggested.
Further research, assuming data can be obtained, might help decide whether there was another reason for
the overly pessimistic projections from the complex models used by MAFF. Or whether a performance like
the earlier logistic fits to the data could have been
achieved with more aggressive predation on the virus.
[KINGS010382] - "The Refractory Model: The Logistic Curve
and the History of Population Ecology", S Kingsland, The Quarterly Review of Biology, Vol 57,
P 29-52, March 1982.
First published 13th May, 2001. Last Revised
Last Revision: vdeck modification