lloydchapman / seir

SEIR models with odin, dust and mcstate

seir

This package demonstrates use of the odin, dust and mcstate R packages for running stochastic SIR and SEIR models and fitting them to simulated data [1]. See also the SIR model odin.dust vignette and mcstate vignette.

Requirements

The package requires the following versions of the eigen1, odin, dust, odin.dust and mcstate packages to be installed:

eigen1 0.1.3
odin 1.5.10
dust 0.15.1
odin.dust 0.3.10 
mcstate 0.9.20

These can be installed in R with:

remotes::install_github(c(
  "mrc-ide/[email protected]",
  "mrc-ide/[email protected]",
  "mrc-ide/[email protected]",
  "mrc-ide/[email protected]",
  "mrc-ide/[email protected]"))

The vignettes use the coda package, which can be installed from CRAN with:

install.packages("coda")

Installation

Install the package with:

remotes::install_github("LloydChapman/seir", upgrade = FALSE, build_vignettes = TRUE)

References

1. FitzJohn RG, Knock ES, Whittles LK, Perez-Guzman PN, Bhatia S, Guntoro F, Watson OJ, Whittaker C, Ferguson NM, Cori A, Baguelin M, and Lees JA. Reproducible parallel inference and simulation of stochastic state space models using odin, dust, and mcstate. Wellcome Open Research, 5:288, 12 2021. doi:10.12688/wellcomeopenres.16466.2

seir's Issues

Particle filter trajectories all the same

In fit_seirdage.Rmd, the particle filter seems to generate identical trajectories for all particles in this line of code.

This lack of stochastic variation between particle trajectories is also apparent in the pMCMC, which seems to only capture parameter uncertainty, rather than the reconstructed trajectory from the filter:

I wondered if it might be something to do with seed=1 defined in the earlier model, but the later filter and pMCMC uses underlying seirdage, so this shouldn't be the problem?

The original odin model has random draws between transitions, so not clear to me why the simulations aren't different – but I'm probably missing something obvious!

Improving data/variable object readability

There are a few places in where could make data objects more user friendly, e.g. with the iterative simulation below. Maybe the array could have named dimensions to avoid having to rely on numeric indexing later on? Could also explore use of Map or similar to streamline these loops (e.g. as in {finalsize} vignettes).

# Create an array to contain outputs after looping the model.
# Array contains XX rows = Total S, E, I, R (4), and
# in each age compartment (XX) as well as the cumulative incidence (XX)
x <- array(NA, dim = c(seird_age_model$info()$len, 1, n_steps+1))

# For loop to run the model iteratively
x[ , ,1] <- seird_age_model$state()
for (t in seq_len(n_steps)) {
    x[ , ,t+1] <- seird_age_model$run(t)
}

It may depend a bit on how odin.dust is structured, but also wondered if it was possible to reduce reliance on numeric indexing in the state variables, e.g. in this line of code?