GRFFmap

This project uses a Gaussian Process (GP)-based spatial-temporal model to estimate the prevalence of antimalarial resistance markers across Africa over time.

Overview

Resistance marker prevalence is modeled as a latent GP on a 2D spatial domain evolving over discrete yearly time steps. Observed data are binomial counts (mutant alleles out of total sequenced samples) linked to the latent field via a logistic transformation.

Key Components

Spatial representation

The spatial covariance follows a squared exponential kernel, approximated using Random Fourier Features (RFF) to reduce computational cost from O(N³) to a tractable linear operation. The latent logit-prevalence field is expressed as a linear combination of these features.

Temporal evolution

Temporal dynamics are modeled as a Gaussian random walk over the RFF coefficients, forming a linear state-space model. This keeps computation linear in the number of time points O(T).

Inference

Because the binomial likelihood is non-Gaussian, Pólya–Gamma data augmentation is used to convert observations into pseudo-Gaussian form, enabling exact Kalman filtering and RTS smoothing within an EM algorithm.

Output

Posterior prevalence surfaces are reconstructed via Monte Carlo sampling, yielding estimates, 95% credible intervals, and exceedance probabilities at each grid cell and year.

Installation

git clone https://github.com/IDEELResearch/GRFFmap.git
cd GRFFmap
open GRFFmap.Rproj

devtools::install_dev_deps() will install all required packages, as specified in the Imports in DESCRIPTION. (At a later date when analysis is finalised, renv can be used to create a reproducible R environment that anyone can use by calling renv::restore to set up package dependencies.)

Overview

The structure within analysis is as follows:

R/                            # Packaged R functions 

analysis/
    |
    ├── 01_xxxxx /            # analysis scripts used for generating figures
    |
    ├── data/               # data inputs for the model (shape files and prevalence data)

• Analysis scripts are to be run in the numbered order they are included. If there are shared numbers, then any order of these scripts works.

• Data that is read only, e.g. data shared from elsewhere and not generated using code in this repository, is stored in data

Compendium DOI:

Licenses

Code: MIT year: 2024, copyright holder: MRC Centre for Outbreak Analysis and Modelling

Data: CC-0 attribution requested in reuse