Inequalities in SARS-CoV-2 case rates by ethnicity, religion, measures of socioeconomic position, English proficiency, and self-reported disability: cohort study of 39 million people in England during the alpha and delta waves

Study data

We linked national SARS-CoV-2 positive test results obtained through pillar 1 (swab testing in UK Health Security Agency laboratories and NHS hospitals for those with a clinical need, and health and care workers) and pillar 2 (swab testing for the wider population, as set out in government guidance) to the Office for National Statistics (ONS) Public Health Data Asset (PHDA) using NHS number.

The ONS PHDA is a linked data resource combining the 2011 census, death registrations, General Practice Extraction Service Data for Pandemic Planning and Research (GDPPR)22 and Hospital Episode Statistics.23 To obtain NHS numbers, we linked the 2011 census to the 2011–13 NHS patient registers using deterministic and probabilistic matching, with an overall linkage rate of 94.6%. The NHS numbers in national testing data were incomplete, with missing values for 21% of records. To retrieve additional NHS numbers, we linked the testing data to the NHS Personal Demographics Service using deterministic matching, achieving a linkage rate of 91.4%.

The study population consisted of all people aged ≥10 years living in England who were enumerated in the 2011 census, registered with a general practitioner (GP) surgery in November 2019, and alive on 1 September 2020 (figure 1). The cohort comprised 39 006 194 participants, 78.4% of the mid-year 2020 population estimate of people aged ≥10 years in England.

Request permission

We used national testing data up to 10 December 2021. Out of all test results, 83.0% were linked to the ONS PHDA. We could not calculate case rates and rate ratios for the first wave because mass testing was not available.

Characteristics and covariates

All individual level sociodemographic characteristics (sex, age, ethnic group, religious affiliation, disability status, educational attainment, National Statistics Socio-economic Classification (NS-SEC) of the household reference person, English language proficiency, country of birth) were obtained from the 2011 census. Place of residence variables (region within England and rural-urban classification24) and area based deprivation25) were derived based on postcodes held in GP records. Care home residence was retrieved from the 2019 NHS patient register. Pre-existing health conditions were derived from GDPPR data as in the QCOVID risk prediction model.3 We included the number of pre-existing conditions and a separate adjustment for learning disability because it could directly affect exposure to SARS-CoV-2.26 The number of pre-existing health conditions was included as a proxy for contact with the healthcare system, which might affect the risk of SARS-CoV-2 infection or lead to shielding. Contact with the healthcare system would also make the person more likely to be tested for SARS-CoV-2. We also adjusted for body mass index as a categorical variable with a category for missing values.

Missing data for 2011 census data were imputed using nearest neighbour donor imputation, the standard method used by the ONS to impute missing values.27 Because we do not have any information on which records were imputed, we could not perform multiple imputation. Therefore, the confidence intervals might not fully represent the level of uncertainty. However, the item non-response was less than 4% for all variables used in our analysis.28 Therefore, we would only expect this to have a minimal effect on the confidence intervals. Table S1 in supplemental file 1 lists all variables included in the analyses.

Outcome

The outcome was receiving a positive test result (polymerase chain reaction (PCR) or lateral flow device, including positive lateral flow device tests that were not confirmed by PCR) for SARS-CoV-2. We excluded any positive tests that occurred within 120 days of an initial positive test from the same person because these might have been part of the same infection episode.29 We classified tests from 1 September 2020 up to and including 22 May 2021 as having occurred in the second wave of the covid-19 pandemic, with tests from 23 May 2021 to 10 December 2021 classified as being in the third wave.17

Statistical analyses

We estimated age standardised SARS-CoV-2 case rates as the number of people who received a positive test result per 100 000 person weeks at risk, stratified by sociodemographic characteristics, and standardised to the 2013 European Standard Population30 using the approach described in the Association of Public Health Observatories’ third technical briefing.31 Rates were calculated separately for the second and third waves of the pandemic.

To explore differences in case rates by sociodemographic characteristics, for each factor, we compared rate ratios for testing positive for SARS-CoV-2 estimated from generalised linear regression models using a Poisson distribution, adjusted in a stepwise manner for three different sets of covariates: sex and age (model 1); sex, age, and geographical variables (region and rural-urban classification; model 2); and sex, age, geographical variables, sociodemographic characteristics (ethnicity, indices of deprivation as fifths, educational attainment, household tenure, and care home residence status), self-reported disability status, body mass index, and the number of pre-existing health conditions (model 3). Note that some of the variables in the covariate sets are considered as covariates and factor variables at different stages. Throughout the study, age is modelled using restricted natural cubic splines with 10 year age bands. The baseline rate ratios for each factor are therefore obtained under model 1, with the fully adjusted rate ratios given by model 3. This stepwise approach enables us to examine how much of the excess risk in certain groups can be accounted for by confounding factors. To account for the fact that some people died during the study period, the natural logarithm of time at risk (in days) was included in the model as an offset term.

Because of the considerable overlap between ethnicity and religion, when considering religion as our main factor of interest, we excluded ethnicity from the third covariate set. To examine the relation between ethnicity and religion in our data and their impact on rate ratios, we ran additional models using an interaction term between ethnicity and religion, adding back ethnicity to the third covariate set alongside religion as our factor. Similarly, in a separate model we investigated the interaction between ethnicity and English language proficiency (self-defined from the 2011 census), which could act as a proxy for a range of factors from cultural upbringing to the length of time a person had been in England before the 2011 census. These models are included in the online supplemental file 1.

We explored how differences in the risk of testing positive for SARS-CoV-2 changed over the course of the pandemic by fitting separate models for the second and third waves. We also fitted separate models for those aged <65 years and ≥65 years.

All analyses were conducted using R version 3.5.1 (in Cloudera Data Science Workbench) using Spark base engine 8,32 and the packages sparklyr33 and dplyr.34

Patient and public involvement

We did not directly involve patients and the public in the design and conception of the study because of the pace at which this study was conducted to inform the UK government’s response to the covid-19 pandemic. The use of deidentified data precludes direct dissemination to participants. For the purpose of open access, the authors have applied a Creative Commons Attribution (CC BY) licence to any author accepted manuscript version arising. Results will also be disseminated by all coauthors through their home institutions.

Main findings

Our analysis using population level linked data in England shows that there were major inequalities in covid-19 case rates in people aged ≥10 years during the second and third waves for several sociodemographic characteristics, most notably by ethnic group, religious affiliation, and rural-urban classification. During the second wave, case rates were highest among Bangladeshi and Pakistani ethnic groups, with adjustments for geographical variables, socioeconomic factors, and pre-existing health conditions accounting for 27.2% and 31.7% of the excess risk, respectively. For religious affiliation, those who identified as Muslim or Sikh had the highest rates, with adjustments only accounting for 27.2% and 15.8% of the excess risk, respectively. While some differences were found by deprivation and other sociodemographic factors, these were less pronounced than for ethnicity or religious affiliation. However, there is considerable overlap between ethnicity and religion; 93.4% of people from the Pakistani and Bangladeshi ethnic groups within the study self-identified as Muslim. The highest rates were seen among people from the most deprived areas, even in model 3. Those who do not speak English well or at all were at greater risk of having a positive test than those with English as their main language, with adjustments for geographical variables, socioeconomic factors, and pre-existing health conditions accounting for 79.2% of the excess risk.

For the third wave, corresponding to the emergence of the delta variant, we observed a different pattern for several factors. The white British ethnic group had the highest case rates and rate ratios, while those who self-identified as Christian had the highest rates among religious affiliations. Case rates also became highest among people born in the UK and whose main language was English. A potential reason is that levels of population immunity were higher for the groups that had the highest case rates in the first and second waves, even considering the potential for reinfection.35

Changes in the rate ratios observed in wave three compared with wave two could also be due to changes in testing behaviours in response to rollout of vaccination, changes in the perceived risk of infection or reinfection, and policy changes related to isolation periods and compensation after testing positive for SARS-CoV-2. Rates of access to sick pay in England and Wales were lower among South Asian workers than white British workers36 and it was more difficult for ethnic minority groups to access Test and Trace services,37 which probably had an impact on case rates among these groups. Interestingly, when stratifying these models by broad age groups (<65 years v ≥65 years) as an exploratory analysis, we found that the rate ratios for all ethnic minority groups were higher in the model restricted to people aged ≥65 years compared with the unrestricted model and the model restricted to those aged <65 years. These results could indicate the presence of further factors affecting the underlying risk of infection and the likelihood of being tested, such as living in multigenerational and overcrowded households. This finding is consistent with the continued increased risk of mortality during the third wave for ethnic minority groups compared with the white British population.7 38

Comparison with other studies

Our findings are consistent with results from the Coronavirus Infection Survey, which found that between September 2020 and May 2021, people living in urban areas and deprived areas, and of a younger age were most likely to test positive in the UK.16 Studies using UK covid-19 surveillance data have also suggested that black and South Asian ethnic groups were more likely to test positive than white British people in England.6 39 In addition, our results support previous analyses using UK administrative data that have shown higher age standardised case rates among ethnic minority groups until June 2021, when rates increased among the white population.38 Similar patterns of increased infection in the most deprived areas and among minority ethnic groups have been observed worldwide.10 40

Studies have shown that covid-19 vaccinations significantly reduced the risk of SARS-CoV-2 infection.18 From December 2020 onwards, unadjusted vaccination uptake rates were lower among adults from ethnic minority groups, people living in the most deprived areas, those self-reporting as having a disability, people younger in age, those who did not speak English as their first language, and people who belonged to a lower socioeconomic group.38 41 These data are consistent with our findings when adjusting for age and sex only during the second wave, suggesting that lower vaccine uptake rates for certain groups and younger people might contribute to case rate inequalities. Although vaccination rates were lower for the Bangladeshi and Pakistani groups than the white British population, the lowest rates were found in black African and black Caribbean groups.

Strengths and limitations

The primary strength of the study is using nationwide linked population level data that combine a diverse set of demographic and socioeconomic factors from the 2011 census with timely data on national SARS-CoV-2 testing. Unlike studies based solely on electronic health records, our study is based on self-identified ethnicity, limiting the potential for factor misclassification bias. We also have information on a wide range of sociodemographic factors not typically available in electronic health records, such as religion, main language, and educational attainment. Another strength is the size of the dataset, comprising 78.4% of people aged ≥10 years living in England in 2020. Therefore, this study is sufficiently powered to detect small differences in the relative risk of testing positive for SARS-CoV-2 by detailed characteristics after adjusting for confounding factors and interactions with age.

An important limitation is that the PHDA only contains information on people who were enumerated in the 2011 census. Therefore, it excludes people living in England in 2011 who did not participate in the 2011 census (estimated to be approximately 5% of the population at the time); respondents who could not be linked to the 2011-13 NHS patient registers (5.4% of census respondents); people who have immigrated since 2011; children <10 years old in 2021; and people not registered with a GP surgery or who had opted out of GDPPR. Additionally, the NHS patient register is known to have coverage issues,42 with undercoverage of specific groups such as migrants and recent returnees to the UK, armed forces and dependants, prisoners, and people registered only with private practices. Therefore, because our study population is based on the PHDA, specific groups might not be adequately covered,43 which could result in biased estimates of relative risks for some groups. However, the coverage is high and the biases are probably small.

A further limitation is that many of the sociodemographic variables were derived from the 2011 census. Some of these characteristics (for example, disability status, English language proficiency, and NS-SEC) might have changed since the 2011 census and might not accurately reflect peoples’ circumstances during the pandemic. Some unaccounted factors might also exists that could contribute to the inequalities in case rates observed across ethnicities, such as current occupation or household size, with Pakistani and Bangladeshi groups being most likely to work in occupations which carry greater risk of infection37 and live in overcrowded households with poor ventilation.44 45 Because our occupation data are from the 2011 census, we have used the NS-SEC of the household reference person to give wider coverage of age groups. Using this as a proxy for occupation means people could have changed NS-SEC categories since the 2011 census, particularly those who are not the household reference person and have moved out.

National SARS-CoV-2 testing data do not provide a representative measure of infections because people are more likely to get a test for covid-19 if they have symptoms, as they are advised to do, and because there might also be other biases in the choice to get a test. About 40% of people who tested positive in the Coronavirus Infection Survey did not develop symptoms within 35 days of testing positive.29 Therefore, these figures are likely to under-represent the number of people without symptoms and so might not be generalisable to all infections in the population. Additionally, people in certain occupations and school children are required to undergo regular testing, and so might be more likely to test positive for covid-19 as a result of higher testing rates. Adherence to testing has been shown to be lower among men and boys, those of younger age, and people of lower socioeconomic status,46 meaning inequalities in case rates are likely to be underestimated.

We were not able to account for the impact of lockdown measures on relative risks because these varied over time throughout the waves and differed by geographical areas. These policies were also not consistent across occupations and so varying rates of sociodemographic characteristics across regions and occupations could lead to differential risks which are not accounted for in this study.

Different diagnostic tests have been used for identifying SARS-CoV-2 infection, with the gold standard being reverse transcription PCR testing, a technique based on amplifying genetic material present in a sample to confirm the presence of the virus. All test types have been found to have high specificity, meaning that false positives are rare, while the test sensitivities have been found to differ across type of tests.47 With the accuracy of tests being affected by the timing and the conditions of the test, and in people with symptoms the ability and willingness to identify their symptoms and seek a test,48 49 the case rates reported in this study are probably underestimates. A large scale population study would be valuable to understand the differences in test seeking behaviours and estimate the probability of being tested for SARS-CoV-2 according to sociodemographics.

Conclusion

SARS-CoV-2 case rates were found to vary considerably across different sociodemographic groups, particularly ethnicity and religion, in the second and third waves of the covid-19 pandemic. Further research is needed to understand why these inequalities exist and how they can best be addressed through policy interventions. Continued surveillance is essential to ensure that changes in the patterns of infection are identified early to inform public health interventions.