Excess mortality during the COVID-19 pandemic is the combination of deaths caused, or contributed to, by infection with SARS-CoV-2 plus deaths that resulted from the widespread behavioural, social and healthcare changes that accompanied national responses to the emergency^1,3-5. England has experienced one of the highest death tolls from COVID-19 in the industrialised world, far beyond what would be expected from its underlying health status and factors like obesity^1,2,6,7. COVID-19 may substantially widen existing national health inequalities as the direct and indirect impacts of the pandemic may disproportionately effect the population groups with the highest healthcare utilisation: the elderly, people with chronic health conditions, people from a minority ethnic background and people who live in more deprived areas^2,8.

Rates of diagnosed SARS-CoV-2 infections and deaths among people with confirmed infection vary substantially across England⁹. But neither local variations in all-cause mortality associated with the pandemic, nor their community determinants, are well understood; population density and urbanisation are often mentioned as major contributory factors in urbanised industrialised countries^10,11. Here, we analysed geocoded data on all-cause mortality at ages 40 years and over for 6,791 local communities (Middle Super Output Areas [MSOAs]; median population 7,956 in 2018, median area 3.04 km², Extended Data Table 1) in England to quantify local variations in excess mortality in the first wave of the pandemic, from 1 March to 31 May 2020, and to identify the community characteristics associated with these patterns.

From 1 March to 31 May 2020, 171,294 people at ages 40 years and over died in England, compared with a mean of 121,358 deaths in the same period in 2105-2019, equivalent to 49,936 excess deaths. Compared with 2015-2019, a greater proportion of the deaths in 2020 were in men, in care homes and a smaller proportion occurred in hospitals (Fig. 1).

Because communities (MSOAs) are small, we used a Bayesian spatial model to obtain stable estimates of excess death rates based on data for each community and those of its neighbours to reduce uncertainty (Methods). The spatial model included terms for potential community determinants of mortality: percent population on income support as a marker of area poverty; population density; percent who are non-White; and percent population living in overcrowded homes. We also included air pollution, namely annual average nitrogen dioxide (NO₂) and fine particulate matter (PM_2.5) and number of care homes per 1,000 population. Data sources and definitions of these variables are detailed in Methods. Each variable was divided into quintiles of the distribution to allow for non-linear relationships (Extended Data Table 2).

All but 360 communities in men and 668 in women had higher mortality in 2020 than expected based on prior years (Fig. 2) with a posterior probability of increased mortality of at least 90% in 4,240 (62.4%) communities in men and 3,166 (46.7%) in women. Of these, mortality more than doubled in 697 (10.3%) and more than tripled in 18 (0.26%) communities in men and in 498 (7.3%) and 15 (0.22%) communities respectively in women (Extended Data Table 3).

The communities with an increase in mortality were spread across the country with the lowest increases in remote rural areas (Fig. 3). The largest increases in mortality were concentrated in London¹², especially for men; for women high excess mortality also occurred in suburban areas. In men, 30.8% of variation in excess mortality was explained by local clustering but only 18.4% in women, suggesting greater correlation in excess mortality between neighbouring areas for men than women. On average, communities with large increases in mortality tended to have greater social and environmental deprivation than those with small increases (Fig. 2).

The combination of a large relative increase in mortality and a high baseline (i.e. pre-pandemic) death rate meant that men in 2,240 communities and women in 1,544 communities experienced 250 or more excess deaths per 100,000 people aged 40 years and over compared with the prior years; in 333 communities for men and 305 for women, the excess mortality burden was at least 500 per 100,000 people. The large variation in excess death rates meant that 25% of all excess deaths during the pandemic occurred in only 9.4% of communities for men and 9.1% for women, and one half of excess deaths occurred in 24.3% and 24.2% of communities, respectively. Excess deaths per 100,000 people were only moderately correlated between men and women (Extended Data Fig. 1 and 2).

Each of the community characteristics considered was individually (i.e. in univariate analysis) associated with excess deaths during the pandemic in graded fashion across quintiles (Fig. 4, Extended Data Table 4). However, there were strong inter-correlations between some variables; for example, Kendall’s Tau was 0.67 and 0.55 between percent non-White population and levels of NO₂ and PM_2.5 respectively (Extended Data Table 5). When the community characteristics were considered jointly in multivariable analyses, air pollution and population density were no longer associated with excess deaths, contrary to reports in some studies that air pollution is a contributory factor for COVID-19 deaths (Extended Data Table 6)^13-16. Relationships with income support, percent non-White population and overcrowded homes persisted, although were attenuated – with a ∼10% higher rate across quintiles for men, and somewhat weaker associations for women. The relationship with care home density, even after accounting for the other variables, remained strong, with a ∼22% higher excess death rate for men and ∼27% for women in communities with the highest compared to lowest density of care homes; many of these deaths were not assigned to COVID-19¹⁷. Overall, the community variables accounted for 17.6% of the variation in mortality at community level in men and 14.9% in women (Extended Data Table 6). Sensitivity analyses with different smoothing parameters, excluding deaths in care homes, and combining data for men and women, did not materially alter our findings (Extended Data Table 7 a-e)

Our study has strengths and limitations. We included excess mortality from all causes, not just deaths coded to COVID-19. This should give the most complete picture of the effects of the pandemic on mortality and is comparable across geographies, as it is not dependent on availability of testing or diagnostic facilities nor variations in national or local coding practices. Not only could COVID-19 deaths have been wrongly ascribed to other causes but deaths from other causes may have been affected by the switching of healthcare resources to deal with the pandemic^6,8,18-20. Although we accounted for population changes in the communities during the study period, population estimates at this scale were only available to 2018 and were extrapolated for the subsequent years. We used mortality data for the same three months of the year (March to May) over the previous five years to estimate the expected numbers of deaths in those months during 2020. But factors like temperature may modify the number of deaths observed. In addition, we used data from the last national census in 2011 to obtain information on sociodemographic characteristics of communities. To the extent that there have been demographic changes in the nine years since then, this may have led to misclassification of areas with respect to their community characteristics.

Our finding on the importance of care home density as a predictor of local excess mortality is consistent with the policy in the National Health Service to discharge up to 15,000 medically fit inpatients to avoid hospitals becoming overwhelmed²¹. It is likely that many of the elderly individuals discharged in this way will have needed support from social care services (including care homes) on discharge²² and may have not been tested for the SARS-CoV-2 virus prior to discharge²³. Our study also underlines the associations between excess mortality and poverty, non-White ethnicity^24,25 and overcrowded housing^26,27at the community-level. Those living in poor communities and overcrowded homes have fewer opportunities for adopting measures that reduce transmission²⁸, have higher exposure to infection at work or may be more restricted in terms of accessing healthcare for COVID-19 and other conditions²⁹. Recent and ongoing research indicates that higher risks associated with ethnicity may at least in part reflect higher levels of overcrowding and poverty (adjusted for in our analysis), higher representation in frontline jobs in the health and care sector^12,30, slower access to and utilisation of healthcare^26,29,31,32, and possibly higher rates of co-morbidities such as diabetes and obesity^2,33,34. Further research to understand the pathways underpinning these associations is needed to inform long-term strategy to tackle the social and environmental drivers of inequality that may have contributed to differential mortality during the pandemic.

In the short-term, as industrialised countries in Europe and elsewhere confront the pandemic’s second wave, in addition to more attention to protecting care home residents and workers³⁵, the response from many governments has been either a national lockdown or a tiered lockdown applied primarily to cities³⁶. England has now entered its second national lockdown. Lockdowns in the first wave were highly effective at driving down the rates of new infection, but they are not sustainable ^4,37-39. Therefore, the immediate priority is to continue to strengthen public health systems to ensure they have the capacity, in real-time, to test and diagnose newly infected individuals; identify their contacts; provide self-isolation and quarantine advice; and undertake national surveillance to inform the evolving policy response⁴⁰. In parallel, economic interventions that support job security and provide financial compensation to low-paid workers required to self-isolate are essential to support population-level compliance with public health advice^41,42.

In summary, in one of the worst affected industrialised countries in the first wave of the COVID-19 pandemic, we found substantial community-level variation in excess mortality, ranging from small declines to tripling in mortality in some areas. Although at first glance the high increases are more evident in cities, population density itself does not appear to be a driver of excess mortality; rather excess mortality risks are related to poverty, overcrowded homes and non-White ethnicity, parallel to large impacts in communities where care homes are located. While we found that these community factors and geographical clustering contributed independently to patterns of excess mortality, a large proportion of the variance remained unexplained. This underlines the importance of using real-time surveillance to identify local outbreaks and target public health resource according to need. A robust public health response is essential if England and other industrialised countries are to control the transmission of SARS-CoV-2 and avoid further widening of inequalities in mortality patterns during the second wave.

Methods

Data sources

The study uses data held by the UK Small Area Health Statistics Unit (SAHSU), obtained from the Office for National Statistics (ONS). The ONS individual mortality data included date of death, date of registration of death, place of residence of the deceased, place of death (e.g. hospital, hospice, care home, at home), and International Classification of Diseases tenth revision (ICD10) codes of the underlying cause of death. Annual population was from ONS mid-year population estimates by age and sex for communities (MSOAs) in England, 2015 to 2018. 2019 MSOA populations were estimated by distributing 2019 local authority district populations according to MSOA shares in 2018. No 2020 population data are yet available and 2019 estimates were used instead.

Characteristics of MSOAs (communities)

To investigate the association of community characteristics with excess mortality we included the following data at MSOA level:

• Income deprivation: Proportion of the population (adults and children, including asylum seekers) on government assistance due to low income and unemployment⁴³.
• Population density: Number of people per square kilometre from 2019 mid-year population estimates, as described above.
• Ethnicity: Percentage of the population of ethnic origin other than White from 2011 census data⁴⁴.
• Housing: Percentage of overcrowded households defined as households with at least one fewer bedroom as required based on the number of household members and their relationship to each other, from 2011 census data⁴⁴.
• Air pollution: Annual average concentrations of nitrogen dioxide (NO₂) and fine particulate matter (PM_2.5) for 2018 at 1km x 1km grids, modelled to MSOA level using 2011 postcode headcount information⁴⁵.
• Location of care homes: Care homes per 1,000 population using data from the Care Quality Commission via Geolytix⁴⁶.

All covariates were divided into quintiles, giving approximately 1,360 MSOAs in each quintile.

Statistical methods

All analyses were carried out for males and females separately. We split age into four groups: 40-59 years; 60-69 years; 70-79 years; 80+ years.

We used a two-stage approach in order that the pandemic and comparison periods were treated as independent and distinct. First, we obtained estimates of the death rates in each MSOA for the comparison period of 1 March to 31 May 2015-2019 using a model that incorporated spatial and age terms to obtain stable estimates of death rates in each age group. Then in a second stage, we modelled the death rates from 1 March to 31 May 2020, relative to the death rates estimated for the comparison period. We estimated excess mortality for each MSOA by comparing death rates for these three months between 2020 and 2015-2019 by sex and age-group. In the second stage we included spatial and age terms as well as community variables to assess their effect on excess mortality. The spatial terms in both stages allowed for local smoothing across communities as well as global smoothing across England and were shared across all age groups.

In the first stage we adjusted for age and smoothed over space to obtain stable estimates of the death rates for the comparison period. We assumed that the number of deaths y_itkfor the i-th MSOA (i=1,…6,791), the t-th year (t=2015,…,2019) and k-th age group (k=40-59, 60-69, 70-79, 80+) arose from a Poisson distribution: with the log-transformed death rates modelled as a sum of space, age and time terms: The common intercept for log-transformed death rates is represented by α₀, with β_0k the age effect for thek-th age group. We modelled MSOA-level intercepts using a Besag, York and Mollie spatial model ⁴⁷; this includes spatially unstructured, independent and identically distributed Gaussian random effects (V_0i) and spatially structured random effects (U_0i). The latter were modelled with an intrinsic conditional autoregressive prior, which allows for death rates to be more similar across neighbouring MSOAs than those that are far away. This spatial model provides both local and global smoothing on the underlying death rate λ_ikt1.

We obtained the posterior distributions of the death rates in each MSOA, age group and year, λ_ikt1, and averaged over March to May for the 5 years of the comparison period (2015-2019) to obtain λ_ik1, the expected death rate for the i-th MSOA and k-th age group during March-May 2020 had there been an absence of the pandemic.

In the second stage we estimated the ratio between death rates in March to May 2020 and the death rates we would have expected had there been no pandemic, using data for the same three months for 2015-2019. We estimated the effect of community variables on this ratio. For the number of deaths in the i-th MSOA and k-th age group in 2020, we specified the following model: where ρ_ikrepresents the age-specific ratio between death rates in 2020 and the comparison period (λ_ik1).

We modelled the ratio ρ_ikin a similar way to stage one using terms to account for both space and age: Community variables were incorporated into this second stage log-linear model to evaluate their effect on the mortality rate ratio. For univariable effects we added the term δX_i where X_iis the quintile of the variable in the i-th MSOA and δ is the associated effect. Similarly for the full multivariable model evaluating the joint effect of all variables we added ∑_j δ_jX_ij withj=1,…6.

To ensure uncertainty in the estimation of λ_ik1in stage 1 is expressed in stage 2, we drew 50 samples from the posterior distribution of each λ_ik1and ran a stage 2 analysis fixing λ_ik1 to each of these values in turn. For each of these 50 analyses we sampled 100 values from the posterior distribution of each λ_ik2. In this way we fully expressed the uncertainty resulting from the two stages of our analysis. For the neighbourhood variable effects, we report posterior mean and 95% credible intervals (2.5th to 97.5th percentiles) based on the 5,000 sampled values (50 x 100). In addition, for each MSOA we report both excess deaths per 100,000 people and the percentage change in deaths, as described below.

Excess deaths per 100,000 people

We obtained the posterior distribution of the estimated number of deaths across ages for March – May 2020, calculated as ŷ _i.2020. = ∑ _kλ_ik1 × ρ_ik× Pop_i.2020.k, and subtracted the corresponding number for 2015-2019, calculated as ŷ_{i.2015–2019}. = ∑ _kλ_ik1 × Pop_{i.2015–2019.k}. This difference was then divided by the 2020 population over 40 years old in that MSOA and multiplied by 100,000. Fig. 3(A) shows the excess deaths in map form: the colour key on the maps is categorical such that all MSOAs with excess deaths above 500 per 100,000 are coloured darkest red.

Percentage change in deaths

We obtained the posterior distribution of the estimated number of deaths across ages for March – May 2020 as summed over the age groups, as ŷ_i.2020. = ∑ _kλ_ik1 × ρ_ik× Pop_i.2020.k and divided by the corresponding number for 2015-2019 as ŷ_{i.2015–2019}. = ∑ _kλ_ik1 × Pop_{i.2015–2019.k}. We then subtracted 1 and multiplied by 100

We fitted the models using the Integrated Nested Laplace Approximation (INLA), through the R-INLA software package (http://www.r-inla.org/)^48,49. We specified a minimally informative prior, logGamma(1, 0.1), on the hyperparameters τ_V and τ_U.

As sensitivity analyses we re-ran the model using alternative priors for the hyperparameters τ_V and τ_U firstly using logGamma(0.5, 0.05), and secondly using the penalised complexity prior, as described by Moraga⁵⁰. The results from these sensitivity analyses (Extended Data Table 7a and 7b) show little difference to the main model.

We carried out an analysis in which deaths in care homes were excluded (Extended Data Table 7c and 7d). Finally, we carried out an analysis in which deaths from both sexes were combined (Extended Data Table 7e).

Ethics and governance

The study was covered by national research ethics approval from the London-South East Research Ethics Committee (Reference 17/LO/0846). Data access was covered by the Health Research Authority Confidentiality Advisory Group under section 251 of the National Health Service Act 2006 and the Health Service (Control of Patient Information) Regulations 2002 (Reference 20/CAG/0008).

Data availability

• No identifiable information will be shared with any other organisation. SAHSU does not have permission to supply data to third parties. Individual mortality data can be requested through the Office for National Statistics (https://www.ons.gov.uk/).
• Mid-year population estimates can be downloaded from https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/datasets/middlesuperoutputareamidyearpopulationestimates.
• English Index of Multiple Deprivation data can be downloaded from https://www.gov.uk/government/statistics/english-indices-of-deprivation-2019.
• 2011 Census data can be downloaded from https://www.ons.gov.uk/census/2011census/2011censusdata.
• Modelled air pollution data (NO₂ & PM_2.5) can be downloaded from https://uk-air.defra.gov.uk/data/pcm-data.
• Locations data of care homes can be downloaded from https://covid19.esriuk.com/datasets/e4ffa672880a4facaab717dea3cdc404_0.

Code availability

The computer code written in R⁵¹ for the two stages of Bayesian models used in this work is available on request.