Regional excess mortality during the 2020 COVID-19 pandemic: a study of five European countries

The impact of the COVID-19 pandemic on excess mortality from all causes in 2020 varied across and within European countries. Using data for 2015-2019, we applied Bayesian spatio-temporal models to quantify the expected weekly deaths at the regional level had the pandemic not occurred in England, Greece, Italy, Spain, and Switzerland. With around 30%, Madrid, Castile-La Mancha, Castile-Leon (Spain) and Lombardia (Italy) were the regions with the highest excess mortality. In England, Greece and Switzerland, the regions most affected were Outer London and the West Midlands (England), Eastern, Western and Central Macedonia (Greece), and Ticino (Switzerland), with 15-20% excess mortality in 2020. Our study highlights the importance of the large transportation hubs for establishing community transmission in the first stages of the pandemic. Acting promptly to limit transmission around these hubs is essential to prevent spread to other regions and countries.

(95% CrI 6%-19%) in Spain, Fig. 2 and Supplementary Tables S2-6. For the country-level weekly trends see Sub-national level trends: NUTS2 regions. Excess mortality was evident for most regions, but with large intra-country variability, as shown in Fig. 1. Across the five countries, Madrid, Castile-La Mancha, Castile-Leon (Spain) and Lombardia (Italy) are the regions with the highest excess mortality in 2020, ranging from 28% (95% CrI 22%-34%) to 33% (95% CrI 27%-39%) for males and 25% (95% CrI 17%-38%) to 32% (95% CrI 23%-40%) for females ( Fig. 1 and Supplementary Table S4 -5). Ceuta (Spain) experienced a similar median excess for females (31%: 95% CrI 14%-54%), albeit associated with larger uncertainty, Fig. 1 and Supplementary   Table S5. For males, the regions most affected in England, Greece and Switzerland were Outer London and the West Midlands, Eastern, Western and Central Macedonia, and Ticino, with the median excess mortality varying from 15% (95% CrI 8%-24%) to 21% (95% CrI 15%-28%), Fig. 1  Finer sub-national level trends: NUTS3 regions. The higher resolution maps in Fig. 2 and 3 show the median relative excess mortality and the posterior probability of a positive excess, allowing appreciation of patterns missed by the lower resolution. In England, the high excess experienced by the West Midlands was driven by Birmingham, the largest urban area in the region, which recorded values above 20%. In Greece, the municipality of Drama, with an excess >20% was responsible for the high excess in Eastern Macedonia and Thrace. For Italy, outside the Northern regions, the model highlights localised excess in the provinces of Rimini, Pesaro-Urbino and Foggia, on the Eastern coast, while central Spain and Catalonia had the highest excess in Spain. In Switzerland, all regions had a relative excess of <20% but the French and Italian speaking regions experienced a posterior probability of a positive excess above 0.95. In contrast, the German-speaking regions were more diverse, Fig. 2 and 3. In Supplementary Fig. S7 to S16 we report the spatial trends by age and sex at the higher geographical resolution. The spatial trends differ depending on the age group, but are similar for men and women for age groups over 60 years.
Spatio-temporal trends at the regional level. Fig. 4 shows the relative excess mortality and the posterior probability that the excess mortality is greater than 0 for the different NUTS2 regions, for each week in 2020.
Across the five countries, relative excess larger than 200% is observed only during the first epidemic period 4 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 23, 2021. ; of 2020 (March-May 2020) in England (Greater London), Italy (Lombardia) and Spain (Madrid, Castille-La Mancha, Catalonia), Fig. 4. In Switzerland, during the first epidemic period, the geographical patterns of excess mortality were highly localised, with Ticino experiencing the highest excess mortality. In contrast, the geographical variability in Greece was more diffuse. During the second epidemic period (October-December 2020), the excess mortality in Italy and Switzerland was similar across the country, whereas in Greece it was highly localised, with Central Macedonia experiencing a relative excess mortality between 100-200% during November 2020, Fig. 4.

Discussion
To the best of our knowledge, this is the first multi-country study examining excess mortality in 2020 across five European countries at the sub-national level. We found that excess mortality in 2020 varied widely both between countries and within countries. Spain experienced the largest excess mortality among the five countries studied.
Within Greece and Italy the northern regions were more affected than other regions. The temporal trends at the sub-national level showed patterns of localised excess mortality in England, Italy, Spain and Switzerland during the first wave, whereas in Greece the excess mortality was homogeneous. During the second wave, excess deaths were overall lower in magnitude and their distribution more homogeneous in England, Italy and Spain.
In contrast, in Greece and Switzerland, the second wave was more severe than the first one.
Our study has several strengths. It quantifies the short term, direct effect of the COVID-19 pandemic and indirect effects on mortality due to other life-threatening conditions, such as myocardial infarction [16], in five European countries. Reduced access to or uptake of medical care due to COVID-19 leading to delays in cancer screening, cancer diagnosis, rescheduling of surgery or cancellation of outpatient visits in patients with chronic conditions may have increased mortality during the pandemic, particularly in countries with weaker health systems [17,18,19]. Our modelling approach incorporated spatial and temporal mortality trends, factors such as temperature and public holidays, and the different population temporal trends across space, age and sex groups.
We carefully validated the model employing a cross-validation approach and found that it had high predictive accuracy. In contrast to previous studies, we stratified by age and sex, thus allowing the spatial and temporal mortality trends to vary across these groups. Weaknesses include the lack of detailed data on the causes of death, which would have allowed insights into the sources of the observed variation in excess deaths.
Several previous studies reported nationwide excess mortality for 2020. The Office for National Statistics in England reported a 17.9% increase in male mortality and 11.2% in females [11]. A recent study of 40 industrialised countries covered the period from February 2020 to February 2021 and found an excess mortality of 15% to 20% 5 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint in England and Wales, Spain and Italy [20]. Our estimates are lower but credibility intervals include the figures from both studies. Reasons for the discrepancies include the different periods used to train the model, different data sources, different prediction periods and the exclusion of Wales from our data [21]. Our results are in line with estimates from the Hellenic Statistical Authority, which reported a 7.3% increase in the relative excess in Greece during 2020 [10], a Swiss study reporting a 10.6% increase in excess mortality in males and a 7.2% increases in females relative to 2019 [22] and the estimates from the Italian National Institute of Statistics [23].
The latter reported a 15.6% excess for 2020 compared to the average number of deaths 2015 to 2019 [23]. In Spain, the relative excess mortality varied from 26.8% to 77.9% across the different age groups for the period March to May 2020 and from 10.0% to 18.9% during the period July to December 2020 [24]. At the regional level, our findings align with the Office for National Statistics in England, which reported a 20% increase in the relative excess mortality in London, the largest relative excess observed nationwide in 2020 [11]. Our estimates are also in line with the Hellenic Statistical Authority reports suggesting that Macedonia and Thrace experience the largest relative increase in excess deaths in Greece (14.9% in males and 12.9% in females) [10]. The observed north-to-south geographical gradient in the impact of COVID-19 in Italy is in line with previous studies [9,8]. The Italian National Institute of Statistics reported that the provinces with the highest excess of mortality in Northern Italy were Bergamo (51.5%), Cremona (47.5%), Lodi (39.9%) and Piacenza (35.7%) [23]. In central and southern Italy, Pesaro-Urbino (21.1%) and Foggia (16.1%) were the most affected provinces [23]. Several factors may have contributed to the differences in excess mortality we observed during the COVID-19 pandemic across countries and regions. Mortality depends on the probability of being infected and mortality among those infected. Both probabilities vary depending on the country's demographic and socio-economic characteristics, including age structure, ethnicity, level of deprivation, and environmental factors [14]. Further, the timeframe of non-pharmaceutical interventions in countries and regions and the resilience and capacity of health care systems have played a role [7]. The mobility of populations across borders and between regions and the timeliness of lockdowns have probably been the most important factors [25,26].
The first wave of the pandemic was mainly exogenous, with international airports and transport routes serving as main entry points. Thus, the highest number of excess deaths during the first wave was observed in 6 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 23, 2021. ; the areas affected first, i.e., big transit hubs like London, Madrid, Lombardia and Ticino, and Geneva. From the initial point of introduction, SARS-CoV-2 spread to nearby large urban areas where community transmission was established and increased exponentially, spreading to the entire country in the absence of mobility restrictions [27].
Furthermore, during the first wave stochastic super-spreader events like the Champion's League football game between Atalanta and Valencia on February 19, 2020 [28] played an important role in establishing community transmission [29]. The lockdowns in Italy, England and Spain were introduced after community transmission was established in the areas first affected. On the day of the national lockdown 1,797 new cases were reported in Italy, 1,159 in Spain and 2,349 in the UK. The lockdown reduced mobility, allowing some areas to maintain lower levels of community transmission and, for example, leading to the north-south divide in Italy. In Greece, a nationwide lockdown was imposed on March 13, 2020, before the country reached 100 reported cases per day, probably explaining the lack of a spatial gradient excess mortality during the first six months of 2020.
The spatial distribution of excess mortality during the second wave of the pandemic was more homogeneous, reflecting multiple routes of entry and transmission. Factors contributing to this situation included the relaxation of non-pharmaceutical interventions with the reopening of schools, retail and other activities, domestic and international travel, and the public's loosening of preventive behaviours [30]. The timeliness of the lockdowns and population mobility again played a crucial role. Lockdowns during the second wave were slower to be implemented and less rigorous [31]. In Italy and Switzerland, the geographical distribution of the excess deaths was equal nationwide, whereas it was more variable in England, Greece and Spain. In Greece, where community transmission was not established during the first epidemic wave, the patterns observed were highly localised, mimicking the patters observed in the other countries during the first epidemic wave. Central Macedonia (with the transit hub Thessaloniki), Eastern Macedonia and Thrace, and Western Macedonia which border on Bulgaria and Turkey, are the hardest-hit regions. On November 3, 2020, the Greek nationwide lockdown limited transmission in the rest of the country, resulting in lower excess mortality in areas in the south. In Switzerland, the area hit hardest during the second wave was the lake of Geneva region, potentially influenced by the French second wave [30].
In conclusion, this study provides the first comprehensive analysis of weekly sub-national excess mortality for 2020 across five countries, disaggregated by sex and age groups. Our findings highlight how excess mortality varied largely across countries, within countries and over time. They suggest that a timely lockdown led to reduced community transmissions and, subsequently, lower excess mortality. However, lockdowns have adverse short and long-term health, psychosocial and economic effects that need to be considered [7,32]. Community transmission was established in the transit hubs and nearby large metropolitan areas during the first stages of 7 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 23, 2021. ; https://doi.org/10.1101/2021.10.18.21264686 doi: medRxiv preprint the pandemic. Therefore, rapid action to limit transmission around these hubs is essential to prevent spread to other regions and countries.

Methods
All-cause mortality. We retrieved data for all-cause deaths and population counts from the Office for National  in particular NUTS3 (small regions for specific diagnoses) as the main spatial unit of our analysis [33]. We also show results at the NUTS2 level, which is defined to reflect basic regions for the application of regional policies (https://ec.europa.eu/eurostat/web/nuts/background/). The number of deaths from all-causes and the population denominator was available by sex, age, week and NUTS3 region defined as areas with a population varying from 150,000 to 800,000, for 2015-2020. We used the International Organization for Standardization (ISO) week calendar, i.e. the seven consecutive days beginning with a Monday and ending with a Sunday. We aggregated mortality and population data by age groups <40, 40-59, 60-69, 70-79 and 80 years and above to maintain consistency between countries and the literature [12].
Population at risk. Population estimates for the years 2014-2020 are available for Greece, Italy and Spain for the January 1 of every year, whereas for Switzerland for December 31, Supplementary Table S7. To obtain weekly 2020 population figures we performed a two-step linear interpolation. In a first step, using the years 2015-2020, we predicted population counts by age, sex and NUTS3 regions for January 1, 2021. In a second step, we calculated weekly 2020 population figures by linear interpolation of the estimates on January 1, 2020, and January 1, 2021, by age, sex and NUTS3 regions. For England, mid-year population figures were available, Supplementary Table S7, which for 2020 were affected by COVID-19 deaths during the first wave. We, therefore, used the data for 2015 to 2019 and estimated the midyear population of 2020 through linear interpolation for England. We estimated population numbers for January 1 2020, and then used linear interpolation to obtain the weekly population of 2019, which we used as a proxy for 2020.
Covariates. As ambient temperature influences death rates [34], we retrieved data on temperature from the ERA5 reanalysis data set of the Copernicus climate data [35]. Using data from global in situ and satellite measurements, ERA5 provides hourly estimates of a large number of atmospheric, land and oceanic climate 8 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 23, 2021. ; https://doi.org/10.1101/2021.10.18.21264686 doi: medRxiv preprint variables, spatially and temporally compatible with our analysis [35]. For each centroid of the grid cells (at 0.25 • × 0.25 • resolution) that fall into the NUTS3 regions, we calculated the daily mean temperature during 2015-2020 and then the weekly mean, align temperature and mortality data. Additionally, as mortality from all causes can be different during national holidays, we also included a binary variable taking the value 1 if the week contains a public holiday and 0 otherwise.
Statistical methods. We used Bayesian hierarchical models to predict deaths in 2020, under the scenario of absence of the pandemic. Let y jtsk be the number of all-cause deaths, P jtsk be the population at risk and r jtsk the risk in the j-th week of the t-th year (t = 1, . . . , 5 with year 1 corresponding to 2015), for the s-th spatial unit (s = 1, . . . , S) and k-th age-sex group (k = 1, . . . , 10) (male-female and <40, 40-59, 60-69, 70-79, ≥ 80).
The models in the main analysis excluded the age group below 40 years, based on the cross-validation exercise.
We assume a Poisson distribution for the number of deaths y jtsk and modeled the risk r jtsk using the following specification: where β 0t is the year specific intercept given by β 0t = β 0 + t , with β 0 being the global intercept and t ∼ Normal(0, τ −1 ) an unstructured random effect representing the deviation of each year from the global intercept, with τ denoting the precision of t . The term β 1 represents the effect of public holidays (i.e. Z j = 1 if week j contains a public holiday and 0 otherwise). The linear predictor includes also a non-linear effect f (·) of the average weekly temperature in each area, x jts ; in particular, we assume the following second-order random walk (RW2) model: with τ x denoting the precision.
The term b s is a spatial field defined as an extension of the Besag-York-Mollié model given by the sum of an unstructured random effect, v s ∼ Normal(0, τ −1 v ), and a spatially structured effect u s [36,37,38]. In particular b s is defined as follows: where u s and v s are standardised version of u s and v s to have variance equal to 1 [39]. The term 0 ≤ φ ≤ 1 is a mixing parameter which measures the proportion of the marginal variance explained by the structured effect.

9
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint To account for seasonality, we included in the linear predictor a non-linear weekly effect w j , common to all the areas, with a first order random walk (RW1) structure: where τ w is the precision of w j .
For the spatial field hyperparameters φ and τ b we adopted priors that tend to regularise inference while not providing too strong information, the so-called penalize complexity (PC) priors introduced in [39]. In particular, for the standard deviation σ b = 1/τ b we selected a prior so that Pr(σ b > 1) = 0.01, implying that it is unlikely to have a spatial relative risk higher than exp(2) based solely on spatial or temporal variation. For φ we set Pr(φ < 0.5) = 0.5 reflecting our lack of knowledge about which spatial component, the unstructured or structured, should dominate the field b. Finally, PC priors are also adopted for all the standard deviations σ = 1/τ , σ x = 1/τ x and σ w = 1/τ w such that for each hyperparameter Pr(σ > 1) = 0.01.
We train the model using the years 2015-2019 and predict area level weekly mortality for 2020 assuming that the pandemic did not take place. To summarise the results we retrieve samples by age, sex, week and NUTS3 regions for 2020 from the posterior predictive distribution: where θ θ θ is the vector of the model parameters and D the observed data. We report the weekly observed number of deaths in 2020 together with p(y jsk6 | D) at the weekly resolution. We also compare p(y jsk6 | D) with the observed number of deaths in 2020 and retrieve the posterior of the relative (percent) increase in mortality (i.e. relative to what is expected had the pandemic not occurred). We summarise the above posterior reporting medians, 95% credible intervals and posterior probability that the relative excess mortality is larger than 0.
Model validation. We perform a cross-validation like procedure to examine the validity of our predictions.
Using the years 2015-2019, we fit the proposed model multiple times, leaving out one year at a time and predicting the weekly number of deaths by NUTS3 regions for the year left out. We repeat for the different age and sex groups, and different countries. We assess the agreement between the predicted and observed deaths at the year t. We use the following metrics: a) the correlation between the predicted and observed deaths and b) the 95% coverage, defined as the probability that the observed deaths lie within the 95% interval estimated from the model.

10
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Data availability statement
We provide at https://github.com/gkonstantinoudis/ExcessDeathsCOVID the final version of the datasets for Italy and Switzerland as their mortality data is available online.
Raw mortality data files for Switzerland are provided at https://www.bfs.admin.ch/bfs/en/home/statistics/ population/births-deaths/deaths.assetdetail.19184461.html and https://www.bfs.admin.ch/bfs/en/ home/statistics/population/births-deaths/deaths.assetdetail.13187299.html, whereas for Italy at https://www.istat.it/it/archivio/240401. Access to mortality data for Greece, England and Spain is subject to requests. For England, the data were obtained from the Small Area Health Statistics Unit (SAHSU), which does not have permission to supply data to third parties. The data can be requested through the Office for National Statistics (https://www.ons.gov.uk/). For Greece, mortality data can be requested from ELSTAT (https://www.statistics.gr) and for Spain from the National Centre of Epidemiology at the Carlos III Health Institute (https://eng.isciii.es/eng.isciii.es/Paginas/Inicio.html).
Population data is available at the following locations: England: https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/ datasets/populationestimatesforukenglandandwalesscotlandandnorthernireland Greece: The selected aggregation by age, sex and NUTS3 regions is subject to a request at: https://www.

Code availability
All models were fitted using the Integrated Nested Laplace Approximation (INLA) using its R software interface [40]. To ensure reproducibility and transparency to our results and approach the code for running the analysis is available at https://github.com/gkonstantinoudis/ExcessDeathsCOVID. Results are also provided in a Shiny app (http://atlasmortalidad.uclm.es/excess/), to facilitate communication with the general public and stakeholders.

11
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 23, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 23, 2021. ; Fig. 1: Posterior distribution of relative excess deaths (%) across the different countries by NUTS2 region and sex in 2020. The first panel shows the posterior of relative excess deaths in the different NUTS2 regions in England, the second in Greece, the third in Italy, the fourth in Spain and the last in Switzerland. The red line highlight the 0% relative excess deaths, which mean no observed difference in the 2020 mortality compared to the counterfactual scenario that the pandemic did not occur. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Figures
The copyright holder for this this version posted October 23, 2021. ; https://doi.org/10.1101/2021.10.18.21264686 doi: medRxiv preprint Fig. 2: Median relative excess deaths (%) across the different countries by NUTS3 region in 2020. The different panels show the median relative excess deaths in (clockwise) England, Greece, Italy, Spain and Switzerland in categories. Areas in blue indicate areas that observed less deaths than expected had the pandemic not occurred, whereas the different shades of red indicate the higher relative excess mortality. The black solid lines correspond to the NUTS2 region borders. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Spain
The copyright holder for this this version posted October 23, 2021. ; https://doi.org/10.1101/2021.10.18.21264686 doi: medRxiv preprint Fig. 3: Probability that the relative excess deaths is higher than 0% across the different countries by NUTS3 region in 2020. The different panels show the probability that the relative excess deaths is higher than 0% in (clockwise) England, Greece, Italy, Spain and Switzerland in categories. Areas in blue indicate weak evidence of an increased relative excess, areas in white insufficient evidence, whereas areas in red strong evidence. The black solid lines correspond to the NUTS2 region borders. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint