kenya-covid-three-waves

Simulation and inference code for SARS-CoV-2 transmission within and between socio-economic groups in Kenya over the first three waves of COVID-19 in Kenya.

This code repository contains the underlying model code for the paper COVID-19 Transmission Dynamics Underlying Epidemic Waves in Kenya.

Prerequisites and recommended background knowledge:

• Basic familiarity with the Julia programming language.
• Solutions of the infection process are generated using the performant, and well documented, package SciML/DifferentialEquations.jl. Familiarisation with this package is desirable.
• Hamiltonian MCMC (HMC) is implemented using the dynamicHMC.jl package. The log-likelihood function for parameters is directly defined in KenyaCoVSD, and log-likelihood gradients (necessary for HMC) are calculated using forward-mode automatic differentiation. The combination of ODE solutions and log-likelihood function gradients in code was inspired by DiffEqBayes.jl. A good conceptual introduction to HMC can be found here.
• MCMC posterior draws for parameters are stored as a Chains struct from the MCMCChains package. The MCMC draws for each county are available from the JLD2 objects stored in \modelfits, e.g.

using JLD2,MCMCChains
import KenyaCoVSD
@load("modelfits/Nairobi_model.jld2") #Loads an object called model into Main scope which contains information about Nairobi county
model.MCMC_results.chain #Summary information of MCMC posterior draws

Data sets in this repository

/opendatacsvsfolder contains .csv datafiles (see supplementary information for the main manuscript to read data file captions). This data is also present on the repository in the form of .jld2 datafiles, which are directly used in the tutorial notebooks (see below).

Supplementary Information

Please find the supplementary information containing further details on the data analysis, and mathematical/statistical reasoning and assumptions behind the model in /supplementary_info as a Word file.

Getting started

The source code for the module KenyaCoVSD is located in /src. KenyaCoVSD is not part of the Julia package ecosystem general directory, therefore, to use this module:

1. Clone this project using e.g. git clone.
2. Activate the Julia REPL.
3. cd to the directory containing the Project.toml and Manifest.toml files for the cloned repository.
4. Open the Package manager using ], activate and instantiate your copy of KenyaCoVSD by

(v1.6) pkg> activate .
(KenyaCoVSD) pkg> instantiate

This will download and precompile the dependencies for KenyaCoVSD.

Model selection and sensitivity analysis

Code and saved fits for alternative model structures and sensitivity analysis are in sensitivity_analysis. Considered alternative model structures are one-group and three-group models. Sensitivity analysis includes changing assumptions about rate of loss of immunity to reinfection with SARS-CoV-2, the relative susceptibility of individuals after a first episode, compared to naive individuals, and, the relative transmissibility of individuals in subsequent episodes of infection. The script sensitivity_analysis/model_selection.jl runs and presents model selection. The scripts:

• sensitivity_analysis/waning_immunity_sensitivity.jl runs MCMC inference under a range of waning immunity assumptions.
• sensitivity_analysis/infectiousness_sensitivity.jl runs MCMC inference under a range of subsequent episode infectiousness assumptions (these require an alternative formulation of the transmission model).
• sensitivity_analysis/sensitivity_analysis.jl combines saved results from the two scripts above and prints MCMC posteriors for all parameters under each assumption as plots to plots/sensitivity_analysis_plots.

County specific plots

We fitted a transmission to each of the 47 Kenyan counties. In the main paper analysis was presented in terms of aggregated pan-Kenyan totals. Please find county-specific plots of infections, daily deaths, daily PCR positive, overall population exposure and SES group specific R(t) in /plots/county_plots.