Spatiotemporal pattern of COVID-19 mortality and its relationship with socioeconomic and environmental factors in England

Abstract

This paper investigated the spatiotemporal pattern of COVID-19 mortality and its socioeconomic and environmental determinants in the first and second wave of the pandemic in England. The COVID-19 mortality rates for middle super output areas from March 2020 to April 2021 were used in the analysis. SaTScan was used in the analysis of spatiotemporal pattern of COVID-19 mortality and geographically weighted Poisson regression (GWPR) was used to investigate the association with socioeconomic and environmental factors. The results show that there was significant spatiotemporal variation in hotspots of COVID-19 deaths with the hotspots moving from regions where the COVID-19 outbreak initiated and then spread to other parts of the country. The GWPR analysis revealed that age composition, ethnic composition, deprivation, care home and pollution were all related to COVID-19 mortality. Althoughthe relationship varied over space the association with these factors was fairly consistent over the first and second wave.

1. Introduction

The COVID-19 pandemic that is caused by a novel coronavirus (SARs-COV-2), has claimed over six million lives worldwide with an estimated case fatality of 0.6–1.2% (WHO, 2022; Sorensen et al., 2022). In the UK, one of the countries most affected, the first case was reported at the end of January 2020 and the first death was reported by the end of March 2020 (Mahase 2020). There have been six waves of the pandemic in the UK although there is no universally accepted definition for a wave (Department of Social and Health Care DSHC and Office for, Sutherland et al., 2021). The first wave started from the beginning of the pandemic, that was March of 2020 and lasted about 6 months to August of 2020. The second started from September 2020 till April 2021; the third wave lasted from May 2021 to October 2021. There were further three waves from November 2021 to August 2022 with peaks separately in December 2021, April and July 2022.

Similar to other countries, the UK government took many non-pharmaceutical measures to tackle the pandemic in the hope that these measures would reduce infection and mortality, while at the same time avoiding health care services becoming overwhelmed (Brown and Kirk-Wade, 2021). With initial outbreaks in London and southern parts of the country, then being spread to other regions, the first lock down officially started from 23rd of March. The peak of new infected cases and COVID-19 deaths occurred in April 2020. With the infection rate decreasing, the first lock down gradually eased from 13th of May, people were allowed to exercise outdoors more than once a day and interact with others by maintaining a safe social distance. From June to July, the COVID-19 restrictions were further eased with the infections dropping steadily. For instance, pubs, restaurants and hairdressers were allowed to reopen providing adherence to COVID-19 guidelines from 23rd of June. In September 2020 the infections started to rise up again, driven by new SARS-CoV-2 strains. As a result a second lockdown was announced on 5th November . Soon after the second wave peaked in mid-January 2021 with the infections declining after March 2021 and the second wave coming to an end in May 2021. In this period, the third lockdown was enforced from 6th January after a one-month break after the second lockdown . The third lockdown was again gradually eased following a roadmap and by July 2021 almost all restrictions were lifted. In the third wave and the subsequent waves, the situation was different from the first and second waves, although the infections were high especially in December 2021 when Omicron variants became dominant the death rate was relatively low thanks to the successful vaccination campaign. By early March 2022, the UK is amongst the top 10 countries in terms of COVID-19 death rates with over 160,000 deaths and a cumulative death rate of 237.6 per 100,000 (Larsen et al., 2021). Nevertheless, over the last year, mortality rates due to COVID-19 in England continued to decline. By August 2022 the age standardised mortality rate decreased to 30.5 deaths per 100,000 people in contrast to the highest rate of 626 per 100,000 in April 2020. On 18th March 2022 England became one of first of countries where all pandemic restrictions were lifted (Department of Social and Health Care (DSHC) and Office for National Statistics (ONS), 2022).

Research on spatio-temporal patterns of COVID-19 infection and death is important for understanding the transmission of the pandemic. In addition, research on geographical and environmental factors may play an important role in shedding light on understanding geographical inequalities of COVID-19 mortality. Evidence has been found that there are substantial geographical inequalities in COVID-19 mortality and these inequalities varied over time (Harris, 2020). Spatial analysis has played an important role in exploring spatial patterns and their changes over time since the beginning of the COVID-19 pandemic. Particularly local cluster analysis has been applied in identification of hotspots of COVID-19 and spatial regression has been used to analyse factors that determine the spatial disparities.

In previous COVID-19 studies, spatial patterns of COVID-19 incidence cases and death rates (Cavalcante and Abreu, 2020, Kim and Castro, 2020), and spatial epidemic dynamics of the virus (Dutta et al., 2021; Kang et al., 2020) have been explored using various cluster analysis methods. For example, space-time scan statistic (Kulldorff, 1997) has been used to analyse spatial clusters and risk areas of COVID-19 in a number of studies (Andrade et al., 2020; Cordes and Castro, 2020).

Majority of the earlier geospatial COVID-19 studies have been conducted at the global-continental or country scale (Melin et al., 2020; Moonsammy et al., 2021) or at the county/municipality level (Han et al., 2021; Martines et al., 2021; Sun et al., 2021; Travaglio et al., 2021). Less research has been carried out at the neighbourhood level (Cordes and Castro, 2020; DiMaggio et al., 2020, Harris, 2020, Kim and Castro, 2020). Studies along this line have been either carried out at large geographical scales such as local authorities that are heterogeneous over space for a country or at the neighbourhood level but focused on a particular region such as London (Harris, 2020, Sun et al., 2021; Siljander et al., 2022). Furthermore most research has been confined to the earlier stages of the pandemic instead of extending to the later stages when the spatial pattern may be very different. Only a limited number of studies have looked at COVID-19 mortality (Griffith et al., 2022) and infections (Morrissey et al., 2021; Harris and Brunsdon 2022) over a longer period. Thus this paper filled this gap by investigating the spatiotemporal pattern of COVID-19 mortality and its socioeconomic and environmental determinants in the first and second wave of the pandemic in England, based on 14-month small area data. We used spatial scan statistic (Kulldorff, 1997) to analyse the local clusters of COVID-19 deaths and use geographically weighted Poisson regression to examine the factors contributing to the spatial pattern taking into account of spatial heterogeneity.

1.1. Data

The COVID-19 death data was acquired from ONS (Office for National Statistics). (https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/bulletins/deathsinvolvingcovid19bylocalareasanddeprivation/). This data included the number of monthly deaths from March 2020 to April 2021. COVID-19 death referred to a death where COVID-19 was the underlying cause of death, with the date being based on date of occurrence. The geographical units were middle super output areas (MSOA, average population was 8200) in England. The age composition data was from the ONS mid-year population estimates in 2018 (https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/datasets/lowersuperoutputareamidyearpopulationestimates). The ethnicity and overcrowding data were from the 2011 census. Income deprivation data was from the index of multiple deprivation (IMD) statistics of 2019, provided by the Ministry of Housing, Communities & Local Government (MHCLG, 2019). Beds in care home were acquired from the Care Quality Commission (https://www.cqc.org.uk/).

We measured the percentage of the population from a minority group in each MSOA as the proportion that reported their ethnic group as separately: Black African, Black Caribbean, Indian, Pakistani and Bangladeshi in the 2011 Census. Overcrowding was measured by a proportion of people who lived in households where over 1 person shared a room from the 2011 census. Our measure of deprivation was the income deprivation domain from the IMD 2019 score. The income deprivation score measures the percentage of the population experiencing deprivation relating to low income, based on a non-overlapping count of people receiving welfare benefits for low-income. It includes those who are out-of-work and those who are in work but have low earnings (MHCLG, 2019). The population estimate data provided by the ONS were used to calculate percentages of the population aged 65 and over for MSOAs.

In order to adjust for viral exposure, we include days since the day when there were 10 confirmed COVID-19 cases in the local authority. The Public Health England (PHE) data (https://coronavirus.data.gov.uk/) was used to calculate the number of days since the day when 10 laboratory confirmed cases in a local authority were identified. Air pollutant PM_2.5 was included in the analysis which was obtained from DEFRA (Department for Environment, Food and Rural Affairs) modelled annual average pollutant concentrations at a 1 km by 1 km resolution from 2014 to 2018 (https://uk-air.defra.gov.uk/data/pcm-data). The five-year average of the pollutant concentrations was used to indicate long term exposure to air pollution. Concentration values were interpolated to MSOA level using the point-in-polygon method. For middle-layer Super Output Areas that did not have grid points falling within them, the concentration value from the nearest point of the air quality grid was assigned. PM2.5 has been explored in research on COVID-19 mortality (Travaglio et al., 2021; Wu et al., 2020). Last, as a health-related variable, obesity prevalence at the GP practice level during 2018–2019 from PHE was selected (Konstantinoudis et al., 2021).

2. Methods

A space scan statistic method (Kulldorff et al., 1997) was used to analyse and identify where COVID-19 mortality clusters occurred. As COVID-19 deaths are expected to be proportional to population in the local area, a Poisson model was chosen to account for the population in each neighbourhood. The scanning window for space scan statistics is a circular window in a geographical base. The base is centred around one of the centroids positioned throughout the MSOAs, with the radius varying in size continuously from zero to a specified maximum value (Kulldorff et al., 1997). In this study, the spatial scan tool was repeatedly applied to monthly COVID-19 mortality data. The maximum spatial cluster size was set to 3% of the population at risk, in order to avoid large clusters (Kim and Castro, 2020). A likelihood ratio test was used to identify the spatial clusters of COVID-19 deaths. The likelihood was calculated for each circle to determine whether the observed number of mortality cases exceeded the expected number of cases, based on the number of cases and population size observed in the MSOAs over a month (Kulldorff, 1997). The ratio of observed to expected cases represents the risk within the window, and the relative risk represents the risk within the window compared to the risk outside the window. The statistical significance was evaluated using Monte Carlo simulation consisting of 999 random replications of the dataset (Kulldorff and Nagarwalla, 1995). The SatScan package (Kulldorff and Information Management Services, Inc., 2018) was used for scanning hotspots and all mappings were done in R. The data and R scripts are available upon request.

To investigate what factors determined COVID-19 deaths geographical weighted regression (GWR) was used. GWR takes spatial heterogeneity into account and as a consequence each area has its own relationship with the dependant variable. The Poisson distribution within the GWR framework is currently the most suitable for disease data, especially when observed counts of cases are low in specific areas (Nakaya et al., 2005). The dependant variable was specified within the geographically weighted Poisson regression (GWPR) as the observed number of COVID-19 deaths per MSOA and the offset variable was specified as the number of people per MSOA. The explanatory variables for the global and local Poisson regression models were the same variables. The centroids of each MSOA were used as input coordinates. The GWPR model then utilises a kernel and fits for each coordinate a regression equation where the coordinate in the centre of the kernel is the regression point. The data points inside the kernel are weighted from the centre of the kernel towards the edge of the kernel. Data points outside the kernel receive a weight of zero and are thus excluded in the regression equation.

The adaptive kernel method was chosen to account for differences in the density of MSOAs across England. The kernel varies with the size of the analysis window so as to incorporate the same number of MSOAs in each local estimate. An iterative approach identified 142 and 185 nearest neighbouring MSOAs as the optimal model bandwidth based on Akaike Information Criterion (AIC) for the first and second wave model separately (Nakaya et al., 2005). Statistical significance for each coefficient per MSOA was calculated using pseudo t-values. The computation of the GWPR was carried out using the GWR4 software (Nakaya, 2012).

3. Results

The total number of COVID-19 deaths were 45,795 and 70,783 respectively in the first and second waves from March to August 2020 and from September 2020 to April 2021. The total number of MSOAs included in our main analysis is 6971, of which 245 (3.5%) in the first wave and 40 (0.6%) in the second wave had not reported any COVID-19 deaths. The high death rates occurred mostly in local authorities like Hertsmere of East England, Harrow and Brent of London, Middlesbrough of Yorkshire and the Humber in the first wave and in Rother of Southeast England, Castle Point of East England, Burnley of Northwest, in the second (Fig. 1 ).

Fig. 1.

COVID-19 deaths per 100,000 in middle super output areas in England.

There was considerable variation in terms of demographic structure and income (Table 1 , Fig. 2 ). For examples, people aged 65 and over ranged from 1% to 52%. Percentage of black people in a local middle super output area varied from no black people at all to over half of the local population. Days from 10 confirmed cases were also markedly different in local authorities from 140 days to 180 days during the first wave and from 217 to 258 days in the second wave (Table 1).

Table 1.

Summary statistics of variables.

Variable	Mean	Std. Dev	Min	Max
# deaths, first wave	6.7	5.2	0	66
# deaths, second wave	10.5	6.7	0	70
% Age 65 and over	13.7	5.0	0.4	37.1
% Black African	1.7	3.5	0.0	38.1
% Black Caribbean	1.1	2.2	0.0	22.7
% Indian	2.4	5.4	0.0	77.2
% Pakistani	1.9	5.9	0.0	77.6
% Bangladeshi	0.8	3.1	0.0	71.6
Overcrowding	4.4	5.2	0.1	40.0
# Care home beds	67.5	73.5	0.0	784
% Low income	12.7	7.8	0.9	48.8
% People with obesity	17.8	4.3	3.6	38.9
PM2.5 (µg/m3)	9.8	2.1	5.0	14.6
Days from 10 confirmed cases in local authority
First wave	166.4	5.3	140	180
Second wave	252.2	5.4	217	258
Total population	8242.8	1936.3	2242	24,969

Data sources: ONS, PHE, MHCLG.

Fig. 2.

Hotspots of COVID-19 deaths by month, March 2020 – April 2021.

The hotspots of COVID-19 mortality by month identified by SatScan are showed in Fig. 2. The hotspots were clusters of high COVID-mortality relative to other neighbourhoods in a month. Therefore, there can be months when deaths are high all over the country but higher in some places (the hotspots).

In the beginning of the pandemic - March of 2020, the clusters were concentrated in Croydon, Ealing, Hillingdon, Harrow, Havering, Bromley, Barking and Dagenham of London and Dudley, Sandwell, and Birmingham of West Midlands, together they accounted for nearly 90% of MSOAs identified as hotspots. In April the pandemic spread to all over England with hotspots being in all regions. However, Yorkshire and The Humber, East Midlands and South West exhibited fewer hotspot MSOAs. In contrast, Birmingham was especially distinct with over 110 MSOAs being identified as hotspots. In May there was an outbreak in Yorkshire and the Humber. At the same time North East and North West continued to be the regions where hotspots concentrated. For example, County Durham and Cheshire East had respectively around 50 MSOAs as hotspots. In contrast London showed a considerable decline in COVID-19 mortality with only 10 MSOAs being identified as hotspots. The decline also occurred in West Midlands with Birmingham for example having only 4 MSOAs as the hotspots. In June, the hotspots remained in Yorkshire and the Humber, East Midlands, North West. London was free of hotspots for the rest of the first wave. Hotspots re-appeared in many places in West Midlands again but they were in areas outside of Birmingham. In July with the overall mortality rates declining South West and North East also dropped out as regions that contained hotspots. In August only 500 hotspot MSOAs remained in North West, South East and Yorkshire and the Humber, compared to the peak time of April 2020 when there were over three times of hotspots across the country. In general, the trend in the first wave was that the clusters started from London, West Midlands, spreading to the west part of England first and then reaching the east part of England. With the mortality rates declining, the hotspots turned localised in a limited number of areas in the North West and South East.

In September and October 2020 most hotspots remained in the three regions: North West, North East and Yorkshire and the Humber but expanded to East Midlands in September. In the peak period of the second wave, November and December 2020, hotspots spread to nearly all parts of England with County of Durham, Wakefield, Liverpool being top of the table in terms of hotspot concentration. The exception was South West where there were a very low number of hotspots. East of England and London also displayed a relatively low number of hotspots. In January of 2021 the high mortality areas moved to London, East of England, and South East. In contrast there was a sharp drop in mortality in Yorkshire and the Humber and North East. In February of 2021 there was an increase of hotspots in West Midlands around Walsall and Wolverhampton areas and a decrease in London while regions like North West and East of England remained stable. In March of 2021, Leeds, Wakefield, Birmingham, Sefton, Walsall, Leicester showed high concentration of hotspots. While areas in North East and London were free of hotspots with South West containing only one hotspot MSOA. In April 2021 the hotspot remained only in Bradford of Yorkshire and the Humber, and Nottingham, Earwash and Broxtowe of East Midlands. The general trend in the second wave is that hotspots in the north from the beginning moved to the south part of England. In March 2021 the clusters were still found in the north and central part of the country and the hotspots were localised in April 2021 when the overall level of COVID-19 mortality was low.

The results from the global Poisson regression models are shown in Table 2 . Two models separately refer to the first wave model and the second wave model. As expected from previous literature (Bambra and Smith, 2021, Harris, 2020), areas with a higher proportion of people aged over 65, Indian, Pakistani, low income people, and obesity were associated with higher risks of COVID-19 death for both periods. although in the first wave the association with Pakistani was significant at the 10% level. Similarly, a higher number of care home beds was also associated with higher risks of COVID-19 death for both periods. Areas where there was a longer duration in coronavirus infection was also linked to a higher risk of COVID-19 deaths. Long term exposure to PM2.5 is another predictor of COVID-19 mortality with higher risks being associated with higher exposure. The association between the proportion of Black Caribbean, Pakistani and Bangladeshi changed between the first wave and the second wave. In the first wave a higher proportion of Black Caribbean people were related to a higher risk of mortality in the local area but the association became negative in the second wave. This reverse in the relationship may happen due to the fact that Black Caribbeans disproportionately live in London where the pandemic hit harder in the first wave than in the second. For Bangladeshi the significantly positive association only appeared in the second wave, while it was positive but not significant in the first. Black Africans did not display a significant relationship with COVID-19 mortality in either of the two waves. The proportion of people in overcrowded households is consistent in its negative association with risk of mortality but only significant in the second wave.

Table 2.

Parameter estimates from Poisson regression.

		First wave		Second wave
Variable	Estimate (S.E)	Z	Exp (Est)	Estimate (S.E)	Z	Exp (Est)
Intercept	−7.166(0.005)	−1455.57	0.001	−6.714(0.004)	−1713.37	0.001
% Aged 65 and up	0.133(0.007)	19.787	1.142	0.237(0.005)	43.536	1.267
% Black African	0.001(0.008)	0.125	1.001	−0.005(0.007)	−0.614	0.995
% Black Caribbean	0.049(0.006)	7.818	1.050	−0.011(0.006)	−1.959	0.989
% Indian	0.065(0.005)	13.684	1.067	0.064(0.004)	16.245	1.066
% Pakistani	0.01(0.006)	1.800	1.010	0.066(0.004)	15.050	1.068
% Bangladeshi	0.000(0.006)	0.004	1.000	0.041(0.005)	8.766	1.042
% Low income	0.074(0.007)	11.031	1.077	0.164(0.005)	31.053	1.178
Days since 10 infection cases	0.159(0.006)	28.114	1.172	0.031(0.004)	7.172	1.031
% People in crowded household	−0.017(0.011)	−1.460	0.983	−0.105(0.01)	−10.431	0.900
PM2.5	0.036(0.007)	5.334	1.037	0.168(0.006)	30.283	1.183
Care home beds	0.232(0.004)	63.184	1.261	0.145(0.003)	45.277	1.156
% People of obesity	0.079(0.006)	13.661	1.082	0.106(0.005)	23.227	1.112

In the first wave, the number of care home beds was the most important predictor, with a standard deviation change leading to a 23% higher COVID-19 mortality risk, followed by duration of COVID-19 infection (17%) and proportion of people aged 65 and over (14%). In the second wave, the proportion of people 65 and over (27%) became the most important predictor followed by income deprivation (18%) and air pollution represented by PM2.5 (18%).

Based on the AIC goodness-of-fit statistic for comparing models, the model with the lower statistic is the one with the better model fit. Using this criterion to compare the models, the GWPR model had the better fit (first wave mode: AIC=13,255; second wave model AIC=12,969) in comparison with global models (first wave model: AIC=18,858, second wave model AIC=18,502). It is important to stress that both first wave and second wave GWPR models showed evidence of non-stationarity of all the regression coefficients. This is evidenced by the fact that the interquartile ranges of the local regression coefficients were all larger than twice the standard errors of the regression coefficients of the global Poisson GWR model (Table 3 ). This implies that the regression coefficients of each of the variables included in the local GWR models were not constant but changed across MSOAs in England. The implication of this is that, the strength of the associations between COVID-19 mortality risk and each of the explanatory variables vary depending on the spatial location.

Table 3.

Parameter estimates from GWPR.

First wave	Mean	Lower Quartile	Median	Upper Quartile	Interquartile	Global error X 2
Intercept	−7.213	−7.527	−7.120	−6.773	0.754	0.010
% Aged 65 and up	0.307	0.180	0.305	0.428	0.248	0.014
% Black African	−0.177	−0.333	−0.055	0.153	0.486	0.016
% Black Caribbean	−0.162	−0.281	0.012	0.282	0.562	0.012
% Indian	−0.070	−0.149	0.047	0.249	0.398	0.010
% Pakistani	0.476	−0.196	0.005	0.278	0.474	0.012
% Bangladeshi	−0.189	−0.410	−0.049	0.132	0.542	0.012
% Low income	0.147	0.037	0.138	0.252	0.215	0.014
Days since 10 infection cases	0.091	−0.027	0.093	0.209	0.236	0.012
% People in crowded household	−0.026	−0.218	0.009	0.227	0.445	0.022
PM2.5	0.113	−0.089	0.108	0.312	0.401	0.014
Care home beds	0.303	0.230	0.292	0.369	0.138	0.008
% People of obesity	0.032	−0.056	0.024	0.113	0.169	0.012
Second wave	Mean	Lower Quartile	Median	Upper Quartile	Interquartile	Global error X 2
Intercept	−6.747	−6.935	−6.683	−6.399	0.535	0.008
% Aged 65 and up	0.286	0.193	0.286	0.381	0.187	0.010
% Black African	−0.196	−0.318	−0.109	0.005	0.323	0.010
% Black Caribbean	−0.045	−0.202	0.021	0.119	0.320	0.008
% Indian	0.053	−0.066	0.033	0.137	0.203	0.020
% Pakistani	0.041	−0.037	0.055	0.202	0.239	0.012
% Bangladeshi	0.083	−0.022	0.044	0.181	0.204	0.006
% Low income	0.216	0.108	0.200	0.307	0.199	0.010
Days since 10 infection cases	0.023	−0.076	0.029	0.126	0.202	0.014
% People in crowded household	−0.062	−0.191	−0.042	0.130	0.320	0.012
PM2.5	0.095	−0.056	0.114	0.241	0.297	0.008
Care home beds	0.160	0.106	0.147	0.197	0.091	0.008
% People of obesity	0.043	−0.010	0.047	0.102	0.112	0.010

The spatial distribution of the GWPR regression coefficients of the explanatory variables common to models first wave and second wave are shown in Fig. 3 . Only the coefficients that are significant at the 5% level based on the pseudo t value were mapped in colour while areas with non-significant coefficients were displayed in light grey. The local estimates are predominantly positive which suggests that as the proportion of old people in an area increases, COVID-19 mortality tends to increase in both waves. There were a few exceptions where the negative coefficients indicating an increase in old people was associated with a decrease in COVID-19 mortality. For example, in terms of number of MSOAs that exhibited negative coefficients there are only 2 MSOAs with an absolute t-value over 1.96 and none with a t-value over 2.58 from the first wave model showing these areas looked very much outliers in the whole country. In the areas that are urban or close to urban areas the impact of people of older age appear to have stronger impact on COVID-19 mortality risk. This reflects that COVID-19 infection opportunities and the subsequent death risks are usually higher in urban areas where people are more likely to interact with others and thus to be exposed to the infection risk.

Fig. 3.

Estimated coefficients from GWPR (a) first wave model (b) second wave model.

For the proportion of the Black African group, the coefficients range from negative and positive with a higher number of areas with a negative coefficient than those with a positive coefficient. The number of areas which displayed potentially significant positive association with COVID-19 mortality (t > 2.58) was 128 while the equivalent number of MSOAs with a negative association (t<−2.58) was 402 in the first wave model. For Black Caribbeans more areas show a positive association than those showing a negative association in both waves. For example 1078 MSOAs showed a significantly positive (t >1.96) relationship while only 369 MSOAs showed significantly negative coefficients in the first wave. The association with the proportion of Indian, Pakistani and Bangladeshi also varied considerably. For Indians a larger number of areas showing positive association (second wave, N = 1168, t>1.96) compared to areas showing negative association (151, t<−1.96) in both waves. In contrast a similar number of MSOAs showing either positive or negative associations for Pakistani. For Bangladeshi more areas showed positive association than those with negative association in the second wave while it was opposite in the first wave. There was no clear pattern along the urban rural spectrum although in urban areas the association was more likely positive in contrast to rural areas where the association was more likely to be negative.

Association with income deprivation was relatively consistent over space with only a minor part of areas displaying negative association. For example only 80 and 3 MSOAs with negative coefficients that has a t-value < −2.58 and 1203 and 3346 MSOAs that have a t value over 2.58 in the first wave and second wave model separately. Overcrowding was negatively related to mortality risk in the global model but the local estimation showed that some areas had positive relationship while others negative. If we looked at the coefficients with a t value over the absolute value of 2.58, about 259 showing negative while 490 a positive association in the model of first wave. However, in the second wave model, there were 457 MSOAs displaying negative coefficients with a t value less than −2.58 while only 157 displaying positive coefficients with a t value over 2.58.

The most consistent association was for care home capacity with all MSOAs displaying positive association with the stronger association appeared to be in West Midlands, part of East Midlands and North East. Furthermore, the number of care home beds is the most consistent variable with over 95% of areas displaying a high t-value over 2.58 indicative of significance at the 1% significance level. It is surprising that PM2.5 also had positive and negative association with COVID-19 mortality risk. Some 233 areas showed a negative relationship while 955 had a positive relationship on a 99% confidence level measured by pseudo-t value in the first wave model. Obesity has a positive relationship with mortality risk. Again, the relationship varied over space with more areas displaying positive relationship with COVID-19 mortality risk. For example, some 101 areas had negative coefficients while 1035 areas with positive coefficients at the 99% confidence level measured by pseudo t value. For the viral spread variable, again a fairly dominant number of areas showing positive relationship but there are about 190 for example displaying negative relationship with mortality risk. This shows that the viral spread is a pre-requisite condition but not sufficient condition in leading to covid-19 mortality.

The varying relationships between these predictors and COVID-19 mortality risk are fairly consistent across waves. Most relationships remain in the directions and similar magnitudes. Some exceptions exist. There are more areas displaying negative relationships with Bangladeshi in the first wave while there are fewer areas in the second wave model. For overcrowding there are more areas with positive relationships in the first wave but more areas with negative relationships in the second. The largest changes appeared amongst the variables of proportion of Black African people and Pakistani people. For example for North East the average coefficients for Black African people changed from positive to negative from the first wave to the second wave. In contrast for Yorkshire and the Humber, the relationship with Pakistani has changed from strong positive to weak negative.

4. Discussion

Using data from various sources including small area COVID-19 death registrations, census, and COVID-19 cases from public health bodies of England this study examined the spatial and temporal pattern of COVID-19 mortality and its association with a number of socioeconomic and environmental factors in the two waves of the pandemic in England. We found the distribution of hotspots varied considerably month-to-month over two waves. There were more hotspots across the whole country when the pandemic reached peak times and fewer localised ones when the pandemic was in off-peaks times. We found statistically significant associations between various socioeconomic and environmental factors and COVID-19 mortality risk. In addition, these associations varied over space but relatively consistent over two waves.

The spatial pattern revealed by the spatio-temporal variation of hotspots manifests the transmission of the coronavirus. For example in March 2020 when the pandemic started in England it was London that was first in experiencing the outbreak. Birmingham is also a large city, second only to London. Both cities have high volume of connections with external world, especially London (Harris, 2020). The hotspot of COVID-19 mortality moved up north and down south to all over the country when the first wave peaked in April and in May. By the end of first wave hotspots were localised and shrank to a limited number of local areas.

The diffusion in relation to COVID-19 from the initial hotspots to other regions is not uniform. This was demonstrated in the case of Yorkshire and the Humber which avoided the April 2020 outbreak like in North West. However, the hotspots appeared later in May and June 2020 displaying a lag in its outbreak compared with other regions. The simultaneous spread of COVID-19 to most of regions showed the diffusion was likely to be a process of both expansive and relocation process (Haggett, 2000). In contrast to the first wave when hotspots were concentrated in south and expanded to north, the diffusion in the second wave was opposite, with hotspots firstly in the north part of the country and then moving to South.

The spatial scan statistics revealed several consistent, important clusters or hot-spots. The most consistent areas of clustering throughout the two waves were in the Northwest region accounting for 21% and 25% of total hotspots respectively in the first and second wave. The considerable change occurred in London which accounted for 20% of hotspots in the first wave but only accounted for 6% in the second.

The North-South divide manifested in COVID-19 mortality was identified by the hotspot analysis (Table A1), which confirmed what was suggested in previous reports based on the first wave ( Harris, 2020). In the first wave the number of hotspots in North was twice those in South. The divide increased in the second wave, with the number of MSOAs in North tripled that of the South. Part of the reason in the North-South divide is due to population composition, industrial difference and behaviour disparity (Nicodemo et al., 2020). For example, people in the South may be more likely to work at home because the large fraction of employees were in service industry. In contrast people in the North were more likely to work in manufacturing industry which required on-site working and commuting to work. Socioeconomic conditions also underpinned the North-South divide (Harris, 2020). This was consistent with the findings that areas in the North of England were more vulnerable compared with their counterparts in the South (Nicodemo et al., 2020).

The disparity across urban-rural spectrum is more evident (Table A2). There is a high concentration of hotspot MSOAs in urban areas, especially urban conurbation. In the first wave, over one in every five of MSOAs in urban conurbation were identified as hotspots. Urban City and Town, and Urban City and Town in a Sparse Setting displayed lower concentration of hotspots with proportions of 13% and 3% separately. In comparison to urban areas, rural areas had lower risks of being a hotspot. Less than 10% of MSOAs were identified as hotspots with the exception of Rural Town and Fringe where over 10% of MSOAs as hotspots. The pattern is similar in the second wave showing a stronger urban –rural divide of the pandemic. One noticeable difference is that there was an increase in the number of areas in both Urban city and town and Rural town and village in a sparse setting showing the further diffusion from urban to remote areas in the lengthy second wave.

The spatial modelling exercises using GWPR revealed that local socioeconomic and behavioural characteristics played a significant role in determining the COVID-19 mortality. The consistent and positive effect was observed for income, old people, care home beds, underlying medical conditions. The analysis adds to the evidence showing that age, deprivation, underlying health condition are key risk factors associated with higher mortality rates from Covid-19. with the UK having one of the highest per capita Covid-19 death rates in Europe (Burn-Murdoch and Giles, 2020).

Whilst many of those hot spots changed over the first and the second wave spanning over 14 months, reflecting the spatial diffusion of the disease, the demographic and social composition of neighbourhoods continue to be predictive, with a greater number of deaths associated with more from older populations, lower average income, pollution, and underlying medical conditions. In spite of the positive relationship with black Caribbean people were found in the first wave, the non-significant relationship found with Bangladeshi in the first wave and black Africans in both waves in the global morel is different from previous studies (Aldridge et al., 2020; Raleigh 2022). Evidence showed that the rate of death involving COVID-19 for the Black African group was 3.7 times greater than for the White British group for males, and 2.6 greater for females (Aldridge et al., 2020). The elevated risks remained in the second wave but to a lower level. One potential explanation is that ecological analysis does not conform with individual level analysis thanks to the ecological fallacy. Other possibilities are that the modelling exercise did not adjust for potentially important factors about social interaction which might confound the relationship. As a result, the model failed to reveal the positive association with black African and Bangladeshi at the MSOA level.

The varying relationship between COVID-19 mortality and a number of predictors may be due to omission of some predictors such as compliance of lock down policy and mobility tendencies. There were three lock-downs in the two waves. The first started from 23 March 2020, the second from 5 November 2020 and the third from 6 January 2021. Although there were surveys that include questions on compliance of lock-down policies no data was available on the local level. Google mobility data provided changes in mobility frequencies compared with the baseline before the pandemic but the data was only available for local authority levels. In England the vaccination campaign started from late December 2020 which may have some impact on the COVID-19 mortality rate in the late stage of the second wave. Future research can investigate to what extent local uptake of vaccine impact has on COVID-19 mortality.

The strengths of this analysis include examination of the pandemic across two waves. The majority of research so far has focused on the first wave in England (Harris, 2020, Sun et al., 2021). Our analysis utilised COVID-19 death data up to the end of April 2021 allowing us to capture almost the entire course of the two phases of the pandemic in England and hence much more fully than the previous studies which have examined data up to only the first wave or the initial phase of the pandemic. So far only one paper has used the same mortality data looking at persistent inequalities relating to COVID-19 mortality (Griffith et al., 2022). In addition, the analysis examined a number of socioeconomic, demographic confounders and the local regression approach allowing investigation of variation of spatial relationship between demographic, socioeconomic and environmental factors. Viral exposure is also adjusted for the analysis. Use of small areas is also an advantage because it allows for identification ofsmall hotspot at neighbourhood level. Large administrative areas like local authority could be heterogeneous and include both hotspots and coldspots. Also, potential socioeconomic and environmental determinants may vary considerably within the large areas used in other studies (Larotz et al., 2021; Sun et al., 2021; Konstantinoudis et al., 2021; Travaglio et al. (2021) which may smooth out some important relationships. A few studies analysed both COVID-19 infection and/or deaths at super output area level in England but their analytical period was limited to the first wave (Sartorius et al., 2021). The research by Harris (2020) again used MSOA level data but focused on London only.

There are limitations to this ecological study. The association with the proportion of black Africans could be due to ecological fallacy which may also apply to associations with other variables. A few variables were not up to date but from the 2011 census which may lead to some measurement errors. The study was observational and therefore any causal interpretation needs to be taken with caution. For example, the relationship with PM2.5 may not be causative because it could be proxy for London and Southeast England. However, there are other studies suggestive of the positive association with COVID-19 deaths as well and therefore further research may be required (Konstantinoudis et al., 2021). In the GWPR analysis a single optimal bandwidth was used to define the kernel for the whole of England. However, the optimal bandwidth may vary over space. In the future, multi-scale GWPR can be used to allow multiple bandwidths being used and improve the goodness of the model fit.

5. Conclusions

To our knowledge, this study is amongst the first that provides results of spatiotemporal patterns for the two waves from a comprehensive multivariable analysis of MSOA-level predictors of COVID-19 death rates in England. Our findings, based on the data of 14 months to the end of April 2021, have significant implications for COVID-19 providing insights on potential intervention in future pandemics. The identification of those areas at excess risk can help guide local policy to develop interventions in protecting vulnerable populations. The assessment of demographic, socio-economic and environmental factors predictive of COVID-19 mortality risk at high spatial resolution is useful to enhance the understanding of the transmission and severity of the COVID-19 on local level.

This analysis provides further evidence that the pandemic is syndemic and determined by both biological process in transmission of coronavirus and socioeconomic process that different population groups tend to be exposed to the risk at different levels (Bambra and Smith, 2021). Health inequalities in terms of COVID-19 mortality are evident in the analysis showing evidence that interventions should be provided to protect those who are at high risk of COVID-19 death. In addition the spatiotemporal pattern of hotspots provide evidence on how the transmission and severity COVID-19 varied over space and time.

Declaration of Competing Interest

None.

Appendix

Tables A1 and A2

Table A1.

Summary statistics of COVID-19 mortality hotspots by North and South.

	First wave		Second wave
	#MSOA	%	#MSOA	%
Central	1152	17.6	2069	17.6
South	1967	11.2	2690	8.5
North	2209	22.6	4411	25.1

Table A2.

Summary statistics of COVID-19 mortality hotspots by urban rural classification.

	First wave		Second wave
Urban rural type	N MSOA	%	N MSOA	%
Urban major conurbation	2478	20.7	3943	18.3
Urban minor conurbation	282	22.7	488	21.8
Urban city and town	1961	13.3	3547	13.4
Urban city and town in a sparse setting	2	3.1	10	8.5
Rural town and fringe	355	12.1	686	13.0
Rural town and fringe in a sparse setting	5	5.0	22	12.2
Rural village and dispersed	240	8.9	437	9.0
Rural village and dispersed in a sparse setting	5	2.2	37	9.1

Fig. A1, Fig. A2

Fig. A1.

Distribution of socioeconomic and environmental predictors.

Fig. A2.

Estimated coefficients from GWPR and the pseudo t values, (a) first wave model (b) second wave model.

Data availability

• Data will be made available on request.