A COVID-19 Model for Local Authorities of the United Kingdom

Swapnil Mishra, Jamie Scott, Harrison Zhu, Neil M. Ferguson, Samir Bhatt, Seth Flaxman, Axel GandymedRxiv 2020 · DOI

Abstract

We propose a new framework to model the COVID-19 epidemic of the United Kingdom at the level of local authorities. The model fits within a general framework for semi-mechanistic Bayesian models of the epidemic, with some important innovations: we model the proportion of infections that result in reported deaths and cases as random variables. This is in contrast to standard frameworks that model the latent infection as a deterministic function of time-varying reproduction number, $R_t$. The model is tailored and designed to be updated daily based on publicly available data. We envisage the model to be useful for now-casting and short-term projections of the epidemic as well as estimating historical trends. The model fits are available on a public website, and the model is currently being used by the Scottish government in their decisions on interventions.

Introduction

During the COVID-19 pandemic, national-level statistics often masked significant local variation. While the UK as a whole might have seen a decline in cases, specific cities—such as Leicester in mid-2020—experienced localized surges that required targeted interventions. Most existing models at the time provided nation- or region-specific estimates, which were often retrospective rather than preventive.

To address this, researchers from Imperial College London developed a semi-mechanistic Bayesian framework to track and project COVID-19 at the local authority level across the UK. By treating the relationship between infections, cases, and deaths as stochastic rather than deterministic, the model provides a robust tool for "now-casting" and identifying emerging hotspots before they overwhelm local healthcare systems.

Methodology

The framework builds upon established Bayesian transmission models but introduces several key innovations to handle the noise inherent in local-level data:

  1. Stochastic Infection Process: Unlike standard models that treat latent infections as a deterministic function of the reproduction number ($R_t$), this model treats the proportion of infections resulting in reported cases or deaths as random variables. This better accounts for the high variability seen in areas with low infection numbers.
  2. Three-Stage Top-Down Fitting: To ensure stability in local areas where raw data might be sparse, the model fits in three stages:
  3. National: Fitting to the whole of England to estimate the Infection Fatality Rate (IFR) and Infection Ascertainment Rate (IAR).
  4. Regional: Fitting to the 9 regions of England and the other UK nations using the national estimates as priors.
  5. Local: Fitting to individual local authorities, using the regional $R_t$ trends to stabilize the inference while allowing local data to drive the final 45 days of the trend.
  6. Weekly Random Walk: Instead of explicitly modeling every policy change, the model uses a weekly "random walk" on $R_t$ to capture shifts in transmission intensity.

Data Integration

The model integrates several high-quality public data sources, including reported cases and deaths from the UK Coronavirus dashboard, Public Health Wales, Health Protection Scotland, and the Northern Ireland Department of Health. Additionally, prevalence surveys from the Office for National Statistics (ONS) and the Imperial-led REACT study were used to calibrate "true" infection levels against reported cases.

Key Insights: Dynamic IFR and IAR

A critical feature of the model is its accounting for the fact that clinical outcomes and testing capabilities evolved over time. As medical treatments improved, the Infection Fatality Rate (IFR) decreased. Conversely, as testing capacity expanded, the Infection Ascertainment Rate (IAR) rose.

Trends in Infection Fatality Rate (IFR) and Infection Ascertainment Rate (IAR) throughout 2020.

As shown in the figure above, the model captures the steady decline in IFR and the corresponding increase in IAR as the pandemic progressed through 2020, allowing for more accurate estimates of the latent infection population.

The Hotspot Metric

The authors argue that looking at $R_t$ or case counts in isolation can be misleading. A low case count with a high $R_t$ represents a "ticking time bomb" of exponential growth, while a high case count with a low $R_t$ indicates a large but shrinking epidemic.

To provide actionable intelligence, the model combines these factors into a single "Hotspot" probability: the likelihood that an area will exceed a specific threshold (e.g., 100, 200, or 500 cases per 100,000 people) in the next two weeks.

Probability of local authorities becoming hotspots at various case thresholds (100, 200, 300, and 500 per 100,000).

This metric provides a 2-3 week lead time for local authorities to enhance testing or enforce social distancing measures.

Results and Policy Impact

The model's utility was demonstrated through its adoption by the Scottish government to inform their tiered intervention strategy. By November 2020, the model successfully identified significant regional variance, estimating $R_t$ was highest in the South East (1.24) and London (1.14), while it had fallen below 1.0 in Northern Ireland (0.79) and the North West (0.80). This granularity allowed for targeted responses rather than blunt national lockdowns.

Limitations and Future Work

Despite its successes, the model has several limitations. It treats each local authority as an independent "island," ignoring the impact of commuting and cross-border travel. It also assumes fixed delay distributions for symptom onset and death, which may vary by age. Furthermore, the current iteration lacks age-stratification, missing nuances in how the virus spreads differently among school children versus the elderly.

Future work aims to incorporate mobility data to link local authorities and add age-stratified compartments to better understand transmission risks in specific demographics.

Summary updated

Abstract

Abstract We propose a new framework to model the COVID-19 epidemic of the United Kingdom at the level of local authorities. The model fits within a general framework for semi-mechanistic Bayesian models of the epidemic, with some important innovations: we model the proportion of infections that result in reported deaths and cases as random variables. This is in contrast to standard frameworks that model the latent infection as a deterministic function of time varying reproduction number, R t . The model is tailored and designed to be updated daily based on publicly available data. We envisage the model to be useful for now-casting and short-term projections of the epidemic as well as estimating historical trends. The model fits are available on a public website, https://imperialcollegelondon.github.io/covid19local . The model is currently being used by the Scottish government in their decisions on interventions within Scotland [1, issue 24 to now].
BibTeX
@misc{mishra2020covid,
  author = {Swapnil Mishra and Jamie Scott and Harrison Zhu and Neil M. Ferguson and Samir Bhatt and Seth Flaxman and Axel Gandy},
  title = {{A COVID-19 Model for Local Authorities of the United Kingdom}},
  year = {2020},
  doi = {10.1101/2020.11.24.20236661},
  howpublished = {medRxiv},
}
Figure