weave

weave turns noisy, gappy spatio-temporal count data — such as malaria cases reported by health facilities — into smooth estimates of the underlying rate. It fills in missing weeks and attaches an honest prediction interval to every estimate.

How it works

weave treats the counts as arising from a smooth latent surface that varies across space and time, modelled as a separable Gaussian process — a spatial kernel combined with a temporal kernel (periodic season × long-term drift). In three steps it:

1. builds a quick plug-in field from the counts (per-site standardised log(1 + y));
2. estimates the kernel hyperparameters — how fast counts decorrelate across space, within the season, and over the long term — by maximising the Gaussian-process marginal likelihood, kept fast by the kernel’s Kronecker structure; and
3. predicts the latent rate at every site and week, conditioning on the observed cells (so gaps are genuinely interpolated rather than guessed) and combining rate uncertainty with Negative-Binomial count noise to form a 95% prediction interval.

Installation

You can install the development version of weave from GitHub with:

# install.packages("pak")
pak::pak("mrc-ide/weave")

Getting started

weave needs two inputs: obs_data, a data frame with columns id (site), t (week) and y_obs (the count, NA where missing); and coordinates, giving each site’s lon/lat. The workflow is two calls — estimate the kernels, then predict:

library(weave)

# 1. estimate the spatial + temporal kernel length-scales
est <- infer_kernel_params(obs_data, coordinates, nt = nt, period = 52)

# 2. predict the rate, fill the gaps, and attach a 95% prediction interval
pred <- gp_predict(obs_data, coordinates, hyperparameters = est,
                   nt = nt, period = 52, n_draws = 100)

pred has one row per site-week, with the posterior rate and a lower/upper prediction interval.

Learn more

• Gaussian processes: a gentle introduction — what a Gaussian process is, the role of the kernel, and conditioning on data.
• The weave walkthrough — the full method worked end to end, from estimating the kernels to predicting rates with honest uncertainty.