Ethnic minority groups in England and Wales - factors affecting the size and timing of elevated COVID-19 mortality: a retrospective cohort study linking Census and death records

Study design and data

In this observational, retrospective cohort study, the study population included all usual residents of England and Wales who were enumerated in private households (that is, not communal establishments) at the time of the 2011 Census (27 March 2011) and who were not known to have died before 2 March 2020. We excluded from the study population individuals who entered the UK in the year before the Census took place due to their high propensity to have left the UK prior to the study period, and those over 100 years of age at the time of the Census even if their death was not linked. Our study population therefore consisted of 47,872,412 individuals aged 9 years or over at 2 March 2020 (the youngest possible, given the nine-year lag from enumeration of the study population at the 2011 Census).

We used a unique, newly-created linked dataset that combines 2011 Census records and death registrations for England and Wales, created by linking the 2011 Census to NHS Patient Register records between 2011 and 2013, with a linkage rate of 94.6% (for further details, see the Supplementary Appendix on Census to Patient Register linkage rates by characteristics of interest).

Deaths were linked to the 2011 Census using NHS Number, with 89.9% of deaths occurring between 27 March 2011 and 1 March 2020 being linked to the 2011 Census. We linked deaths occurring between 2 March 2020 and 15 May 2020 that were registered by 29 May 2020 using NHS Number and a deterministic match key linkage method where NHS Number was unavailable, achieving a linkage rate of 90.8% of deaths. The unmatched deaths comprise people not present in the UK at the 2011 Census, people who arrived in the UK in the year before the Census (and were excluded from the study), and people who were present at Census but not enumerated.

The study dataset did not contain any information on whether individuals have left the country. To avoid biasing our denominators, we derived and applied weights reflecting the probability to have remained in the country between March 2011 and March 2020, based on data from the NHS Patient Register and the International Passenger Survey (IPS). For further information, see the Supplementary Appendix on the methodology for adjusting for emigration.

Despite being in the population at risk of COVID-19 death in March 2020, we did not replenish the sample with post-2011 births or immigrants; while the latter group could have been identified and in principle linked to our data, neither group are captured in the 2011 Census and therefore they have no ethnicity or covariate data recorded. Additionally, the younger population have been the least affected with COVID-19 related hospitalisation and mortality. For the same reason, individuals not enumerated at the 2011 Census (estimated to be 6.1% of the population of England and Wales [14]) were not included in our study population.

Outcome and exposure

In this study, the outcome of interest was death involving COVID-19 that occurred between 2 March 2020 and 15 May 2020, registered by 29 May 2020. Deaths involving COVID-19 included those with an underlying cause or any mention on the death certificate of the ICD-10 codes U07.1 (COVID-19, virus identified) or U07.2 (COVID-19, virus not identified). Time-at-risk started on 2 March 2020 and ran until 15 May 2020 or date of death, whichever was earlier.

Our exposure of interest was self-reported ethnic affiliation, as stated by respondents to the 2011 Census. Ethnic breakdown was based on the Census ethnicity question, with some groups collapsed to ensure outcome counts were large enough to permit reliable estimates of hazard ratios (see Supplementary Table for more detail on how the groups were aggregated). The White ethnic group was used as the reference category in our modelling analysis as it is the largest group. Ethnicity was imputed in 3.0% of 2011 Census returns due to item non-response, using the imputation methodology employed by the Office for National Statistics (ONS) across all 2011 Census variables [15].

Covariates

We controlled for demographic and structural socio-economic factors that may be associated with the risk of COVID-19 mortality (Table 1), through either the propensity to become infected or the propensity to die once infected (which could not be distinguished using the data available for this study). Such factors are likely to confound and/or mediate the relationship between ethnicity and COVID-19 mortality risk.

All models were adjusted for single year of age, included as a second-order polynomial to account for the non-linear relationship between age and the hazard of death involving COVID-19. Geographical variables included the local authority district of residence at the time of the 2011 Census and official estimates of mid-2018 population density for the Lower Super Output Area (LSOA) of residence.

To account for neighbourhood deprivation, we adjusted for the decile of the Index of Multiple Deprivation (IMD) 2015 and Welsh Index of Multiple Deprivation (WIMD) 2014 of the LSOA of residence at the time of the Census. The IMD and WIMD are overall measures of deprivation in the local area based on factors such as income, employment and health [16, 17].

As further measures of socio-economic position, we used the level of household deprivation, a summary measure of disadvantage based on four selected household characteristics (employment, education, health and housing), the tenure of the household, and the National Statistics Socio-economic Classification (NS-SEC) of the household reference person. The NS-SEC is not defined for respondents aged 75 or over at the time of the Census, so we included individuals’ highest level of qualification as an additional indicator of socio-economic position.

Our measures of household composition and circumstances included the number of people in the household, the family type, and binary variables for living in a multigenerational household (defined as three generations living together) or with any children (aged 18 years or under). We included self-reported health status and a variable indicating whether the individual was activity restricted (disabled) at the 2011 Census.

Finally, in the fully adjusted model we incorporated measures of occupational exposure as recorded at the 2011 Census. We included a variable indicating if an individual was a key worker, and if so, what type. We also included a binary variable indicating if anyone in the household was a key worker. We accounted for exposure to diseases and contact with others using scores ranging from 0 (no exposure) to 100 (maximum exposure), derived using O*NET data based on US Standard Occupational Classification (SOC) codes mapped to UK SOC codes [18]. We assigned these scores based on individuals’ occupations and then derived the maximum value among all household members.

Statistical Analyses

Results

Risk of COVID-19 mortality by ethnicity

For both males and females, age-standardised mortality rates (ASMRs) of COVID-19 related death (see Supplementary Table 2) were greatest among individuals identifying as Black (250.6 and 116.9 deaths per 100,000 people, respectively) and lowest among those identifying as White (87.3 and 51.8 deaths per 100,000 people, respectively). The ASMRs indicate greater levels of absolute risk among all ethnic minority groups compared with the White population.

In terms of age-adjusted hazard ratios, males and females from all ethnic minority groups (except females of Chinese ethnicity) were at greater risk of death involving COVID-19 compared with those of White ethnicity (Figure 1, Model 1). The rate of death involving COVID-19 adjusted for age was 3.13 [95% confidence interval (CI): 2.93-3.34] times greater for Black males than White males, and 2.40 [2.20-2.61] times greater for Black females than White females. People of Bangladeshi/Pakistani, Indian, Mixed and Other ethnic backgrounds also had raised rates of death involving COVID-19 compared with those of White ethnicity.

• Download figure
• Open in new tab

Figure 1.

Hazard ratios for COVID-19 related death for ethnic minority groups compared to the White population, stratified by sex

Note: Results obtained from Cox regression models adjusted for age, population density, area and household deprivation, Socio-economic Status (SES), household composition, occupational exposure, self-reported health, with baseline hazards specific to local authority district. Errors bars represent limits of 95% confidence intervals of the hazard ratio. LAD BH denotes baseline hazard specific for each local authority district. Details of the covariates can be found in Table 1. Numerical results can be found in Supplementary Tables 3 and 4.

Including baseline hazards that are specific to each local authority district and controlling for population density substantially reduced the estimated differences in the rate of death involving COVID-19 across ethnic groups (Model 2). The attenuating effect of controlling for geography was greatest for Chinese males and Indian females, for whom the excess risk relative to the White group decreased by 89% and 84% respectively, relative to the age-adjusted model.

Further adjusting for deprivation and socio-economic position (Model 3), household composition (Model 4) and occupational exposure (Model 5) had a small impact on the hazard ratios, but not negligible for people of Bangladeshi and Pakistani background and Black males. Controlling for health and disability status (Model 6) at the time of the 2011 Census reduced the estimated hazard ratios for some groups, especially individuals of Bangladeshi/Pakistani or Indian ethnicity.

Taken together, the characteristics in the fully adjusted model for females statistically explained the differences in risk compared to the White population for all ethnic minority groups except those of Black ethnicity, as their hazard ratio remained at 1.29 [1.17-1.42] after adjustment.

Adjusting for these characteristics also reduced the estimated hazard ratios for males, but they remained larger than one for all groups other than Chinese and Mixed ethnic backgrounds. Compared to White males, the rate of death involving COVID-19 was 1.77 [1.64-1.91] times greater for Black males, 1.35 [1.22-1.50] times greater for Bangladeshi and Pakistani males, and 1.31 [1.19-1.43] times greater for Indian males.

Risk of COVID-19 mortality before and after lockdown

The smoothed Schoenfeld residuals by ethnic group (see Supplementary Supplementary Figure and 2) indicated a non-constant hazard ratio of deaths involving COVID-19 over the course of the pandemic for some groups, with the hazard ratios for these groups decreasing with time. For the Bangladeshi/Pakistani and Chinese groups for males and the Black group for females, the log hazard ratios tended to zero midway through the outcome period, indicating diminished differences in risk between ethnic minority groups and the White population.

• Download figure
• Open in new tab

Figure 2.

Fully adjusted hazard ratios for COVID-19 related death for ethnic minority groups compared to the White population, before and after lockdown plus 21 days, stratified by sex

Note: Results obtained from a Cox regression model adjusted for age, population density, deprivation, SES, household composition, occupational exposure, self-reported health, with baseline hazards specific to local authority district. Errors bars represent limits of 95% confidence intervals of the hazard ratio. Estimates for each ethnic group are stratified by a time variable indicating pre- and post-lockdown periods.

The ASMRs decreased after lockdown for males of all ethnic minority groups, most notably by 61.8 deaths per 100,000 of the Black population, but increased by 12.1 deaths per 100,000 of the White population (Supplementary Table 5). Similar patterns were observed for females, with ASMRs after lockdown decreasing by 42.4 deaths per 100,000 of the Black population but increasing by 13.3 deaths per 100,000 of the White population. The increase in ASMRs for the White population may be due to deaths in care homes peaking later than throughout the rest of the community (See Supplementary Figure 3). The White population was on older average than other ethnic groups (Table 2), and therefore more likely to live in care settings.

When modelling the effect of lockdown in our fully adjusted Cox regression models, all the estimated hazard ratios decreased after the lockdown for both males and females (Figure 2). This finding is consistent with the results from the smoothed Schoenfeld residuals and indicates the elevated risk of COVID-19 mortality among ethnic minority groups relative to the White population was more pronounced before the lockdown restrictions were announced.

For females, the rate of mortality involving COVID-19 was elevated for the Bangladeshi/Pakistani and Black groups compared with the White population over the pre-lockdown period, even after accounting for individual and household characteristics, with hazard ratios of 1.22 [1.01-1.47] and 1.72 [1.53-1.93] respectively. After the lockdown, the hazard ratios decreased to 0.87 [0.71-1.07] for the Bangladeshi/Pakistani population and 0.83 [0.70-0.97] for the Black population.

The adjusted differences between pre- and post-lockdown hazard ratios were generally less pronounced for males than females. The Black, Indian and Other ethnic minority groups continued to experience a greater rate of COVID-19 mortality than the White population, but with reduced hazard ratios of 1.18 [1.04-1.34], 1.51 [1.34-1.71], 1.22 [1.06-1.39] and 1.30 [1.10-1.52] respectively.

Results obtained using shorter and longer lag durations for capturing post-lockdown deaths (14 and 28 days instead of 21 days) were similar to those in the main analysis, although less precisely estimated when using a lag of 28 days (see Supplementary Figures 4 and 5).

Principal findings

Our paper has three novel findings. Firstly, using population-level data we observed a substantially elevated risk of COVID-19 related death among ethnic minority groups. Secondly, this elevated risk was largely – but not entirely – mediated by location, living circumstances, socio-economic factors, occupational exposure, and self-reported health status. Thirdly, differences in COVID-19 mortality rates between ethnic groups were considerably reduced following the introduction of lockdown measures in the UK.

Comparison with related studies

Our finding that ethnic minority groups were at greater risk of COVID-19 related death compared with the White population is broadly consistent with other UK-based research published to date. An analysis of laboratory confirmed cases of COVID-19 in England found that some, but not all, ethnic minority groups were at greater risk of death after adjusting for demographics and deprivation, with Black and Asian (particularly Bangladeshi) groups having the highest mortality rates [12]. However, the tested sample is unlikely to be representative of the general population, which could bias the estimates of the relationship between ethnicity and COVID-19 mortality [23]. A retrospective study of electronic primary care records covering approximately 40% of patients in England found elevated COVID-19 mortality risk for Black and South Asian individuals after controlling for deprivation and clinical risk factors [3]. Despite its relatively large size and richness of clinical information, the data may not be fully representative of the full population of England and do not include socio-economic covariates other than area deprivation.

Studies based on COVID-19 surveillance data suggest that Black and South Asian individuals were more likely than White British people to test positive for COVID-19 after controlling for demographic, socio-economic, behavioural and health-related covariates [24, 25]. These findings shed some light on our own results, as observed inequality in the risk of dying with COVID-19 may partly be a result of heterogeneity in the risk of infection. Our results could also be driven by differences in prognosis. Studies examining differences in outcomes among hospitalised patients found that patients from ethnic minority groups were more likely to be admitted to critical care [26, 11] and had higher in-hospital mortality rates among COVID-19 cases in England [11].

Beyond the UK, a systematic review of global literature on the association between ethnicity and COVID-19 infection and outcomes concluded that BAME individuals were likely to be at elevated risk [13]. However, published evidence available at the time of the review was found to be limited, with much of the emerging data originating from grey literature and preprint articles.

Whilst there is evidence that lockdown measures reduced infection [27], little is known on whether these measures affected ethnic groups differently.

Potential mechanisms

We found that area of residence and population density together accounted for over half of the excess hazard of COVID-19 mortality across all ethnic minority groups. These groups tend to be concentrated in large urban deprived areas in the UK, which have been more severely affected by the COVID-19 pandemic in terms of both infection [28] and mortality rates [7].

The elevated risk of COVID-19 mortality for ethnic minority groups compared to the White population was also statistically attenuated by the inclusion of socio-economic variables in our models, suggesting that these factors collectively mediate much of the observed risk. The existence of a social gradient in general health status and life expectancy is long-established [29], and recent official data have shown that White individuals in England are less likely to live in deprived neighbourhoods than non-White individuals [30]. These findings emphasise the importance of policy interventions aimed at reducing socio-economic inequalities as a means of reducing heterogeneity in COVID-19 (and broader) mortality risk, particularly among ethnic minority groups who are observed to accrue smaller gains from socio-economic improvements than the White population [31], and highlights the need to study the health of ethnic minority groups within the context of their broader socioeconomic disadvantage.

After full adjustment we found no difference in COVID-19 mortality for females from all ethnic minority groups except those of Black ethnicity, but elevated risks remained for males from most ethnic minority groups. The lockdown measures attenuated the disparity in COVID-19 mortality across ethnic groups, suggesting that the differences are largely due to an increased risk of infection rather than a worse prognosis. However, males from most ethnic minority groups were still at higher risk of COVID-19 mortality.

The residual differences may in part be due to unobserved factors, including pre-existing conditions and other clinical risk factors not fully captured by self-reported health status. People from Black and Asian backgrounds tend to have a higher prevalence of obesity and associated conditions of diabetes, hypertension and cardiovascular disease compared to the White population [32]. These conditions are known risk factors for adverse COVID-19 outcomes and could mediate the observed differences in mortality rates between ethnic groups; however, they may also be partly attributable to the same socio-economic inequalities already captured in our models, or to specific features of the disadvantage suffered by ethnic minority populations such as structural discrimination.

Socio-economic inequality is a complex, multi-faceted, partly ecological concept. The statistical measures of socio-economic status and deprivation included in our models are unlikely to capture fully this conceptual complexity [33], nor the fact that disadvantage accumulated over the life-course may be partly responsible for observed health inequalities [34]. Given these measurement issues, we cannot rule out the possibility of structural socio-economic factors explaining more of the heterogeneity in outcomes than estimated in our study, rather than the unexplained variation necessarily being attributable to other influences.

Strengths and limitations

The primary strength of our study lies in the use of a unique, newly linked population-level dataset based on the 2011 Census, comprising a whole population level cohort and a broad range of demographic, socio-economic and self-reported health variables. We report results disaggregated by groupings of self-reported ethnic affiliation, across the full range of ethnicities. To our knowledge, our study is the first to demonstrate an association between COVID-19 lockdown measures and ethnicity-specific mortality.

The main limitation of our study dataset is the nine-year lag between Census day and the start of the pandemic. The covariates included in our models reflect the socio-economic situations of individuals as they were in 2011, not necessarily those at the start of the COVID-19 pandemic. In addition, the study population did not include people who immigrated or were born between 2011 and 2020, and therefore did not fully represent the population at risk.