Reducing parameters required for final size estimation
Source:vignettes/finalsize_comparison.Rmd
finalsize_comparison.RmdNew to epidemics? It may help to read the “Get started” vignette first!
Additionally, read how to quickly calculate the final size of an epidemic - with fewer epidemiological inputs than epidemics required - the finalsize package, which uses an analytical solution to avoid simulating the full epidemic dynamics.
The total number of individuals expected to have been infected over an epidemic is called the epidemic final size (Miller 2012). The R package finalsize can help to calculate the epidemic final sizes using the final size equation (Miller 2012), while accounting for heterogeneity in population contacts as well as susceptibility to infection, such as due to vaccination.
epidemics can also be used to calculate the final size of an outbreak while accounting for heterogeneity in contacts and the rollout of vaccination campaigns.
This vignette lays out which of the two is more suitable for specific tasks, and how to convert between when seeking more insight into the effect of policy decisions made during epidemic response.
Code
Different use cases of finalsize and epidemics
While both finalsize and epidemics can be used to calculate the epidemic final size, they are aimed at different use cases.
finalsize assumes that both infection and population characteristics are fixed to their initial conditions. For the infection, this includes properties such as the transmission rate, while for the population, it includes social contacts between demographic groups, and the proportion of demographic groups that have specific levels of susceptibility to infection.
An advantage is that finalsize only requires that we define the initial transmission rate of the infection and susceptibility of the population, rather than all the time-dependent processes that drive the epidemic shape, such as the duration of infectiousness or latent period of the infection. For questions relating to epidemic shape, rather than shape, this results in a much simpler set of inputs.
However, it also means that finalsize cannot be used to model temporal dynamics of epidemic response, or be used to answer policy questions with a temporal component, such as when to implement interventions.
epidemics includes a number of scenario models, each with its own assumptions. Most models allow for some modification of the initial characteristics of the outbreak due to a range of events. For the infection, this includes interventions (such as masking or treatments) that reduce the number of forward transmissions or deaths, but also seasonal effects which may increase or decrease the transmission rate. For the population, initial conditions of social contacts can be influenced by interventions as well.
This makes it much easier to model the temporal dynamics of public-health policy decisions which are taken during epidemic response as epidemics has many more features that allow for such modelling.
However, it requires more inputs to be defined, including time-dependent infections processes. It can also be more difficult to model scenarios in which more complicated susceptibility structure is required, such as when some demographic groups have underlying immunity to infection due to past exposure or vaccination. Thus epidemics is likely to be especially useful for outbreaks of novel pathogens (such as the Covid-19 pandemic) where there is little population immunity to infection.
It is easier to configure finalsize out-of-the-box for scenarios with complex demographic patterns of underlying susceptibility to (or immunity against) infection, potentially due to a history of previous outbreaks and the policy responses (such as vaccination). Thus, while it cannot model temporal dynamics, it can quickly provide useful initial estimates of the final size of outbreaks, without having to write compartmental models which implement multiple policy decisions.
Converting scenarios between finalsize and epidemics
Here, we show an example in which we show how to model a similar scenario using both finalsize and epidemics. For example, we might want to study the full epidemic dynamics in future, but start with a simpler final size estimate in the meantime. Here we show how to build equivalent models, while allowing extensions to model epidemic temporal dynamics (using epidemics) later.
As an illustration, we use epidemics to model the effect of vaccination on the trajectory of an epidemic. While finalsize does not allow for the implementation of a dynamic vaccination calendar, we can model reduced susceptibility to infection in the population, such as due to prior vaccination.
The two methods are comparable if we model vaccination in epidemics as occurring before the main wave of the epidemic, as this sets up underlying susceptibility.
Prepare population and model parameters
We first prepare the population (modelled on the U.K.), initial conditions, and model parameters, before passing them to epidemics::model_default().
This example does not include any interventions or vaccination regimes.
Code to prepare model inputs is folded below for brevity. See the “Get started” vignette for an explanation of how to prepare these inputs.
Code
# load contact and population data from socialmixr::polymod
polymod <- polymod
contact_data <- contact_matrix(
polymod,
countries = "United Kingdom",
age.limits = c(0, 20, 40),
symmetric = TRUE
)
# prepare contact matrix
contact_matrix <- t(contact_data$matrix)
# prepare the demography vector
demography_vector <- contact_data$demography$population
names(demography_vector) <- rownames(contact_matrix)
# initial conditions: one in every 1 million is infected
initial_i <- 1e-6
initial_conditions <- c(
S = 1 - initial_i, E = 0, I = initial_i, R = 0, V = 0
)
# build for all age groups
initial_conditions <- rbind(
initial_conditions,
initial_conditions,
initial_conditions
)
rownames(initial_conditions) <- rownames(contact_matrix)Code
# prepare the population to model as affected by the epidemic
# the contact_matrix, demography_vector, and initial_conditions
# have been prepared in the folded code above
uk_population <- population(
name = "UK",
contact_matrix = contact_matrix,
demography_vector = demography_vector,
initial_conditions = initial_conditions
)
# view the population
uk_population
#>
#> Population name:
#> Demography
#> [0,20): 14,799,290 (20%)
#> [20,40): 16,526,302 (30%)
#> 40+: 28,961,159 (50%)
#>
#> Contact matrix
#>
#> contact.age.group [0,20) [20,40) 40+
#> [0,20) 7.883663 2.794154 1.565665
#> [20,40) 3.120220 4.854839 2.624868
#> 40+ 3.063895 4.599893 5.005571We model the spread of influenza with pandemic potential, assuming $R_0 = $ 1.5, a pre-infectious period of 3 days, and an infectious period of 7 days.
This leads to the following model parameters for transmission rate \(\beta\), the rate of transition from exposed to infectious \(\sigma\), and the recovery rate \(\gamma\).
Code
# simulate pandemic parameters
transmission_rate <- 1.5 / 7
infectiousness_rate <- 1 / 3
recovery_rate <- 1 / 7Implementing vaccination in epidemics
For simplicity, we implement a vaccination regime in which vaccines are delivered to individuals over the age of 40 over the course of roughly six months (150 days) before the main epidemic peak. We assume that 1 in every 1000 individuals in this age group is vaccinated every day; this translates to approximately 28,900 individuals per day.
We assume that the vaccination is non-leaky, protecting vaccinated individuals from infection, and this due to the model structure of the ‘default’ SEIR-V model in epidemics that we use here.
New to modelling vaccination using epidemics? It may help to read the “Modelling a vaccination campaign” vignette first.
We first create a <vaccination> object to represent this vaccination regime.
Code
We then pass this vaccination regime to the epidemic function in the optional vaccination argument, and model 600 days of the epidemic.
We obtain the ‘final’ epidemic size using epidemics::epidemic_size().
Code
# model epidemic with vaccination prior to the main epidemic wave
output <- model_default(
population = uk_population,
transmission_rate = transmission_rate,
infectiousness_rate = infectiousness_rate,
recovery_rate = recovery_rate,
vaccination = vaccinate_elders,
time_end = 600, increment = 1.0
)
# Calculate the epidemic size using the helper function
finalsize_dat <- tibble(
demography_group = names(demography_vector),
value = epidemic_size(output) / demography_vector
)
# View the data
finalsize_dat
#> # A tibble: 3 × 2
#> demography_group value
#> <chr> <dbl>
#> 1 [0,20) 0.617
#> 2 [20,40) 0.525
#> 3 40+ 0.343Calculating individuals vaccinated in epidemic model
To compare the results of the ODE model against the analytical method, it is necessary to set up the population’s susceptibility and demography-in-susceptibility matrices to reflect the proportion of individuals in each age group vaccinated before the outbreak.
To do this, we need to calculate the total proportion of individuals in each age group vaccinated at the end of the epidemic model run.
Implementing vaccination in finalsize
We use the proportion of individuals vaccinated in each age group to set up matrices to pass to the susceptibility and p_susceptibility arguments of finalsize::final_size().
New to final size estimation with heterogeneity in susceptibility to infection? It may help to read the “Modelling heterogeneous susceptibility” vignette, and the “Guide to constructing susceptibility matrices” for finalsize first.
We create the susceptibility matrix to have two groups, ‘unvaccinated’, with full susceptibility, and ‘vaccinated’, who are immune to infection.
Code
We then create the demography-in-susceptibility matrix to reflect that only some individuals in the >40 age group are vaccinated. Since vaccination has not been implemented for other age groups, we assume that all individuals in those groups are fully susceptible.
Code
# Second column holds the vaccinated (who are protected fully)
susceptibility[, "vaccinated"] <- 0
# Assume susceptibility varies within age groups
p_susceptibility <- matrix(
data = 1.0,
nrow = length(demography_vector),
ncol = 2,
dimnames = list(
names(demography_vector),
c("unvaccinated", "vaccinated")
)
)
p_susceptibility[, "vaccinated"] <- p_vacc
p_susceptibility[, "unvaccinated"] <- 1 - p_vaccWe then calculate the expected final size using the modified susceptibility matrices, and combine this data with the output of the ODE model as in the previous example.
We need to scale the contact matrix appropriately.
Code
Code
# Calculate the proportion of individuals infected in each age group
dat1 <- final_size(
r0 = 1.5,
contact_matrix = contact_matrix,
demography_vector = demography_vector,
susceptibility = susceptibility,
p_susceptibility = p_susceptibility
)
# Final size returns the proportion infected in each susceptibility group
# (i.e. non vaccinated and vaccinated)
# here we calculate the proportion of infected for the age group as a whole
dat3 <- dat1 %>%
select(
-susceptibility,
final_size = p_infected
) %>%
filter(susc_grp == "unvaccinated") %>%
select(-susc_grp)
fs <- dat3$unvaccinated * p_susceptibility[, "unvaccinated"]
finalsize_dat <- finalsize_dat %>%
select(demo_grp = demography_group, seir_v = value) %>%
left_join(dat3)We can print the data to compare values, which are similar.
Code
finalsize_dat
#> # A tibble: 3 × 3
#> demo_grp seir_v final_size
#> <chr> <dbl> <dbl>
#> 1 [0,20) 0.617 0.660
#> 2 [20,40) 0.525 0.542
#> 3 40+ 0.343 0.272However, note that while epidemics::epidemic_size() can be used for any time point in the epidemic, finalsize::final_size() only returns the size at the end of the epidemic, showing how epidemics is more suitable to examine temporal dynamics.
Code
# note that epidemic_size() returns the absolute values
# epidemic size after 10% of the epidemic model run
epidemic_size(data = output, stage = 0.1)
#> [1] 574.4348 488.3563 585.5760Code
# epidemic size at 50%
epidemic_size(data = output, stage = 0.5)
#> [1] 5565772 4890178 5254726Consideration of computational speed
An important reason to use finalsize for final size calculations may be speed — finalsize::final_size() is much faster than epidemics::epidemic_size() applied to epidemics::model_default().
This makes it much more suitable for high-performance applications such as fitting to data.
The benchmarking shows that the total time taken for 1,000 runs using both methods is very low, and both methods could be used to scan across multiple values of the input parameters, such as when dealing with parameter uncertainty.
Code
# run benchmarks using {bench}
benchmark <- mark(
analytical_method = final_size(
r0 = 1.5,
contact_matrix = contact_matrix,
demography_vector = demography_vector,
susceptibility = susceptibility,
p_susceptibility = p_susceptibility
),
ode_model = epidemic_size(
model_default(
population = uk_population,
transmission_rate = transmission_rate,
infectiousness_rate = infectiousness_rate,
recovery_rate = recovery_rate,
vaccination = vaccinate_elders,
time_end = 600, increment = 1.0
)
),
iterations = 1000,
time_unit = "s",
check = FALSE
)
# view the total time for 1000 runs in seconds
select(as_tibble(benchmark), expression, total_time)
#> # A tibble: 2 × 2
#> expression total_time
#> <bch:expr> <dbl>
#> 1 analytical_method 0.470
#> 2 ode_model 10.7Note that some model runs using epidemics that implement more complex compartmental structure, or multiple interventions and vaccination regimes are likely to be slower than the example shown here.
However, users should choose finalsize only when the assumptions underlying a final size calculation are met. For example, in cases where vaccinations are concurrent with a large number of infections, or when interventions are applied to reduce transmission, the final size assumptions are not met. In these cases, users are advised to use a dynamical model such as those in epidemics instead.