Time varying association between deprivation, ethnicity and SARS-CoV-2 infections in England: a space-time study

Introduction

The differential impact of COVID-19 on minorities has reportedly varied over the course of the pandemic. An early review of the literature by Public Health England (1) concluded there is evidence of higher test positivity rates, defined as the percentage of processed SARS-CoV-2 swab tests that are positive, for people identifying as Black. Later studies reported an increased test positivity rate and mortality for the Black, Asian and Minority Ethnic (BAME) population in the first wave of the pandemic in England (2) and Wales (3). Mathur and colleagues additionally found that between September and December 2020 the increased risk of testing positive disappeared and there was a reduced risk of death for Black and mixed ethnicity groups but not for South Asians (2). Morales and Ali (2021) suggested that this change might be due to interventions that occurred between the two waves (e.g. better access to testing, targeted public communication and reduction of occupational risks), but stressed the importance of additional studies to better disentangle the effect of ethnicity from that of other variables (4).

Socio-economic deprivation is another key factor that contributes to health inequalities. People living in the most deprived areas in England and Wales were reportedly twice as likely to die from COVID-19 in the first wave (5). A Europe-wide study tested a range of socio-demographic characteristics at country level and found that income was the best predictor of COVID-19 cases between December 2019 and April 2020 (6). As socio-economic deprivation correlates with ethnicity¹, it is necessary to consider both dimensions simultaneously to accurately assess their impact on COVID-19 burden. Socio-economic and ethnic inequalities in England have become more pronounced during the pandemic and have contributed to unequal death tolls across communities (7). Rose and colleagues found an association between socio-economic deprivation and mortality in England at Upper Tier Local Authority level in the first wave, after adjusting for ethnicity and other socio-demographic factors such as population density and age (8). A recent study by Lo et al. reported that ethnic minorities were at a higher risk of testing positive in the UK and the US in 2020 and some of this increase could be explained by these communities living in more deprived areas (9).

A full detailed analysis of whether and how the role of social and ethnic inequalities on SARS-CoV-2 infections varied over time is still missing from the literature. This is especially true for understanding their combined effect. Most literature to date covers March to December 2020 and has focused either on mortality or test positivity rates, with the latter indicator being affected not only by the actual number of infections but also by testing capacity and strategies. In this study we focus on SARS-CoV-2 infections in England and evaluate how the effects of area levell socio-economic deprivation and ethnic composition evolved over the course of the pandemic (from 1^st June 2020 to 19^th September 2021). A key novelty of our analysis is that we consider test positivity rate as well as a new derived measure of prevalence proportion which corrects for the ascertainment bias of the tested population, enabling us to assess the impact of inequalities both on observed disease surveillance indicators and on the estimated underlying infection rate.

Methods

We examine the effects of inequalities on two outcome metrics: test positivity rate and unbiased prevalence proportion. We analyse outcomes as weekly summaries at Lower Tier Local Authority (LTLA) level, the finest spatial resolution at which the UK government monitors epidemic indicators and implements local interventions against COVID-19. Due to data sparsity we remove Isles of Scilly, and, in accordance with government reporting of COVID-19 cases, we combine the Boroughs of Hackney and City of London, leaving 312 LTLAs in the analysis. Aggregating data by week mitigates day-of-week effects in administering, processing and reporting the outcomes of interest. Our analysis covers the period from the 1st of June 2020 to the 19^th of September 2021. We exclude data from the first wave of the pandemic (March - May 2020) since community testing was not widely available. Data is stratified by age (5-12, 13-17, 18-24, 25-34, 35-44, 45-54, 55-64, 65+ years old); we do not consider 0-4 years old, as prevalence estimates are not available for this age group.

Exposure variables

We consider two community-level characteristics: (i) social deprivation, measured through the 2019 Index of Multiple Deprivation (IMD)⁴ and (ii) ethnic diversity, defined as the proportion of BAME (Black, Asian and Minority Ethnic) population in each geographical area as reported in the 2011 Census. The distribution of these two covariates by LTLA is presented in Figure 1 of Supplementary material. We also repeat the analysis disaggregating BAME into Black, South Asian (those identifying as Pakistani, Indian or Bangladeshi) and Other BAME (people not identifying as White, Black or South Asian), and include all three as covariates. Figure 2 in Supplementary material presents the proportion of people from a South Asian, Black and Other BAME background in each LTLA and shows the strongly spatially clustered nature of this variable.

• Download figure
• Open in new tab

Figure 1:

Spatial distributions of the median test positivity (top panel) and debiased prevalence (bottom panel) over the entire study period outputted by the model, presented on the logit scale. The overall spatial distribution (left column) is disaggregated into two components, representing what is explained by social deprivation, ethnic composition and confounders (centre column) and the unobserved spatial residuals (right column).

• Download figure
• Open in new tab

Figure 2:

Test positivity (left) and debiased prevalence (right) for profiles of ethnicity (all BAME) and deprivation. Each tile represents the average weekly outcome (test positivity or debiased prevalence) over the entire period of analysis, obtained as output of the model including ethnicity, IMD, confounders and the spatio-temporal correlation structure. In parenthesis we report the relative change in the outcome between each profile and the reference, defined as low deprivation and low BAME population.

To illustrate the combined effect of ethnicity and deprivation, we define profiles based on the distribution of the IMD score and the percentage of BAME population across LTLAs. We consider as a representative low value the first quartile, that is, the value that divides the bottom 25% from the remainder of the distribution of the variable. Conversely, we consider a representative high value to be the one dividing the bottom 75% from the remaining 25%. Finally we consider the median as a representative medium value.

Results

Overall spatial distribution and temporal trends

The median of logit-transformed test positivity and debiased prevalence outputted by the model is shown in Figure 1 (left column) and depicts the spatial contrasts for the whole period. As shown in the central panel of Figure 1, social deprivation, ethnic composition and the confounders partially explain the higher values of both outcomes in large cities, like London, Birmingham, and the area around Manchester, Leeds and Sheffield. Altogether, exposures and confounders account for approximately 24% of the overall spatial distribution of test positivity and 20% of debiased prevalence, reaffirming the need for including spatially structured residuals in the model (Figure 1, right panel).

The median of the temporal component in the model captures the overall time trend for the two outcomes, shown in Figure 3 in Supplementary material. Both test positivity rate and debiased prevalence show three peaks during the period under analysis. The first two peaks, in November 2020 and December 2020, were curbed by the introduction of a national lockdown (shown in grey). The third, highest, peak is in July-August 2021. When including LFT in the analysis the summer peak is lower than the two winter ones (Figure 4 in Supplementary material, top panel). This might potentially reflect the change in testing policy: LFT became more widespread as businesses were encouraged to sign up for a testing scheme for their staff, increasing the denominator of the positivity rate. The summer peak is also somewhat less pronounced when removing vaccine uptake from the confounders; in this case the temporal random effect implicitly accounts for the fact that higher vaccination rates lead to a reduction in infections (Figure 4 in Supplementary material, central panel). Finally, the time pattern does not change when we include a lockdown indicator, suggesting that the temporal component already captures much of this effect (Figure 4 in Supplementary material, bottom panel).

• Download figure
• Open in new tab

Figure 3:

Time-varying test positivity (top) and debiased prevalence (bottom) for profiles of ethnicity (all BAME) and deprivation. Each line represents the monthly median odds ratio for the profile relative to a population low in deprivation and BAME, while the shaded area around it represents the corresponding 95% credible interval. The median odds ratio and its uncertainty are outputs of the model after adjusting for confounders and the spatio-temporal correlation structure. Periods of national lockdown are marked in grey.

• Download figure
• Open in new tab

Figure 4:

Time-varying test positivity (top) and debiased prevalence (bottom) for profiles of ethnicity disaggregated by BAME subgroups (South Asian, Other). Each line represents the monthly median odds ratio for that profile relative to a population with a low percentage of all BAME subgroups, while the shaded area around it represents the corresponding 95% credible interval. The median odds ratio and its uncertainty are outputs of the model after adjusting for IMD, confounders and the spatio-temporal correlation structure. Periods of national lockdown are marked in grey.

Discussion

Using a flexible spatio-temporal framework, we investigated the effects of social and ethnic inequalities on two LTLA-level measures of SARS-CoV-2 infection load. Overall, we found that area-level percentage of BAME population and area-level deprivation (IMD) are positively associated both with test positivity and with prevalence. When disentangling the effects of ethnicity, we found specifically that areas with a high percentage of South Asian and Other BAME populations were more adversely affected, but not areas with a high percentage of individuals identifying as Black. Moreover, we found that the impact of the two covariates varies over time: deprivation had a stronger effect during the first period of the analysis (until October 2020) whereas the effect of ethnicity became more pronounced during part of the second wave, in December-January (Figure 3). From the spring of 2021 onwards there is a less clear distinction between the impact of BAME and IMD on either outcome. The consistency of results across the two outcome measures leads us to conclude that disease exposure or susceptibility, rather than testing access and habits, differs between areas with those community-level characteristics.

This is the first study to evaluate the impact of social and ethnic inequalities on infections over the whole course of the pandemic so far (excluding the first few months when testing was not widely available at the community level). It is also the first study to model how these effects have evolved over time. Additionally, as we consider test positivity rate and an unbiased estimate of prevalence proportions, we can evaluate the impact of deprivation and ethnicity on both the monitoring and spread of the pandemic, robustly and free from potential confounding by testing ascertainment bias. Lastly, our flexible space time random effects formulation captures the constantly evolving dynamics of the disease arising from emergence of new variants or interventions such as lockdowns, allowing our analyses to automatically adjust for these potential confounders.

A limitation of our study is that the demographic data considered may not fully represent the structure of the population at the time of analysis. For local population sizes, we used mid-year population estimates from 2020 as reference (the most up-to-date available at the time of the study), hence neglecting any changes occurring during the time of the analysis. More critically, we retrieved the proportion of BAME population from the 2011 Census (the most up to date source available at the time of the study). We work under the implicit assumption that the ethnic composition of LTLAs has not changed drastically, but we cannot verify this. Additionally, this study shows an area-level association between inequalities and SARS-CoV-2 infection burden. Given that we consider aggregated data, the analysis may suffer from ecological bias (17) (18).

Our ecological framework is similar to (8) but we focus on SARS-CoV-2 infections rather than mortality and, crucially, adjusts for spatiotemporal dependence in the residuals, a necessary adjustment for obtaining correct estimates of the precision of the effect of deprivation and ethnicity (19). Additionally, while Rose et al. focus on the first wave of the pandemic (8), we consider a longer time period. Our results are consistent with other findings of higher test positivity for BAME groups, but the dynamic nature of our analysis allows us to evaluate how this effect changes during the different phases of the pandemic. Additionally, by considering ethnic composition in conjunction with deprivation, we are able to assess the results for combinations of these two variables, disentangling which variable has a greater impact on test positivity and prevalence at different times. For instance, we show how the nature of the association changed over time, with deprivation being more important in discriminating COVID-19 burden during the start of the second wave of the pandemic and, particularly, in the third national lockdown. This could be a manifestation of financial hardship and/or pandemic fatigue leading to lower adherence to restrictions.

Conversely, in the last few months, we saw the effect of BAME become more prominent, perhaps due to factors like employment in essential sectors and usage of public transport. We found that the effect was not uniform across all BAME groups: the proportion of Black population within an area was not associated with the outcomes as strongly as the proportion of the South Asian and the Other BAME subgroups. Previous reports (1) highlighting an increased test positivity rate for those identifying as Black are limited to the first wave of the pandemic (February - August 2020). Mathur et al. reported that in September - December 2020 the increased test positivity rate in the Black group disappeared (2). Our results are consistent with their findings, even though we only consider Pillar 2 PCR tests while they considered all Pillar 1⁸ and 2 tests combined. In addition, we observe an increased test positivity for South Asians and Other BAME groups for most of the period under analysis prior to July 2021, thus replicating and extending findings reported by Mathur et al. in the first part of the second wave (2).

To conclude, this nationwide population-based study adds to evidence on the impact of social and ethnic inequalities on SARS-CoV-2 infections. Additionally, it shows that, as the pandemic evolved, with the introduction of new variants as well as new policy responses, so did the susceptibility of different communities. We believe that with the constantly changing epidemic dynamics, it is important to continually monitor how different communities are responding, in order to inform relevant policies aimed at eliminating social inequality in COVID-19 burden.

Ethics statement

This research has been approved by the Alan Turing Institute’s Ethics Advisory Group, nb C2103-105.

Funding statement

BL was supported by the UK Engineering and Physical Sciences Research Council through the Bayes4Health programme [Grant number EP/R018561/1] and gratefully acknowledges funding from Jesus College, Oxford. SR is supported by MRC programme grant MC\_UU\_00002/10; The Alan Turing Institute grant: TU/B/000092; EPSRC Bayes4Health programme grant: EP/R018561/1. GN and CH acknowledge support from the Medical Research Council Programme Leaders award MC_UP_A390_1107. CH acknowledges support from The Alan Turing Institute, Health Data Research, U.K., and the U.K.Engineering and Physical Sciences Research Council through the Bayes4Health programme grant. MB acknowledges partial support from the MRC Centre for Environment and Health, which is currently funded by the Medical Research Council (MR/S019669/1). MB and TP were partially supported by the National Institutes of Health, grant: R01HD092580-01A1. Infrastructure support for the Department of Epidemiology and Biostatistics is also provided by the NIHR Imperial BRC. Authors in The Alan Turing Institute and Royal Statistical Society Statistical Modelling and Machine Learning Laboratory gratefully acknowledge funding from Data, Analytics and Surveillance Group, a part of the UKHSA. This work was funded by The Department for Health and Social Care with in-kind support from The Alan Turing Institute and The Royal Statistical Society.

Authors contributions

TP: conceptualisation, methodology, formal analysis, visualisation, writing - original draft

RJ: methodology, formal analysis, writing - original draft

BJ: methodology, formal analysis, writing - review & editing

PD, SR, CH: conceptualisation, methodology, funding acquisition, writing - review and editing

GN: conceptualisation, methodology, formal analysis, writing - review and editing

AM, RK: data curation

MB: conceptualisation, methodology, visualisation, supervision, writing - original draft

Declaration of interests

All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work.