key: cord-0878209-2i0o3k3u
authors: Jablonska, K.; Aballea, S.; Toumi, M.
title: Factors influencing the COVID-19 daily deaths peak across European countries
date: 2020-11-05
journal: nan
DOI: 10.1101/2020.11.04.20225656
sha: 0a62e8a1cd6e4d5b63db7ec752f96b44c5887cf2
doc_id: 878209
cord_uid: 2i0o3k3u

OBJECTIVES: The purpose of this study was to determine predictors of the height of COVID-19 daily deaths peak and time to the peak, in order to explain their variability across European countries. STUDY DESIGN: For 34 European countries, publicly available data were collected on daily numbers of COVID-19 deaths, population size, healthcare capacity, government restrictions and their timing, tourism and change in mobility during the pandemic. METHODS: Univariate and multivariate generalised linear models using different selection algorithms (forward, backward, stepwise and genetic algorithm) were analysed with height of COVID-19 daily deaths peak and time to the peak as dependent variables. RESULTS: The proportion of the population living in urban areas, mobility at the day of first reported death and number of infections when borders were closed were assessed as significant predictors of the height of COVID-19 daily deaths peak. Testing the model with variety of selection algorithms provided consistent results. Total hospital bed capacity, population size, number of foreign travellers and day of border closure, were found as significant predictors of time to COVID-19 daily deaths peak. CONCLUSIONS: Our analysis demonstrated that countries with higher proportions of the population living in urban areas, with lower reduction in mobility at the beginning of the pandemic, and countries which closed borders having more infected people experienced higher peak of COVID-19 deaths. Greater bed capacity, bigger population size and later border closure could result in delaying time to reach the deaths peak, whereas a high number of foreign travellers could accelerate it. Keywords: COVID-19, mortality, healthcare capacity, modelling.

The coronavirus infectious disease 2019 outbreak was announced as a pandemic 37 by the World Health Organization (WHO) on 11 March 2020. 1 By the end of March, Europe 38 exceeded Asia and became the region experiencing the highest percentage mortality from the 39 virus across the world, 2 until 10 June 2020, when the rapidly growing COVID-19 mortality rate 40

in the Americas exceeded all other continents. 3 According to the WHO report from 5 July 2020, 41 38% of the global mortality due to COVID-19 was from Europe. 4 42

Since incidence and mortality rates varied between countries, numerous studies have recently 43 been published investigating factors associated with COVID-19 infection and death rate 44 across countries. A variety of potential predictors have been assessed in the literature, such 45 as country-specific demographic and health characteristics, economic and social indicators, 5-46 8 mobility scores and social-distancing measures, 9, 10 as well as ecological and environmental 47 perspectives. 8, 11 To assess the relationship between covariates and COVID-19 incidence or 48 mortality, the most common approach used by authors was to analyse multivariate regression 49 models of total number of infections or deaths up to a given time point, using log-transformed 50 data or without any transformation, as well as daily data on infections or deaths as outcomes. 51

In this study, we used data on numbers of deaths, not infections, since the former has a much 52 higher degree of reliability than the latter, being better monitored and less dependent on the 53 number of tests done. Since all European countries seem to already reach the peak of deaths 54 by 3 June 2020 from the first wave of COVID-19 disease, our idea was to use height of daily 55 deaths peak as a primary outcome of interest, and time to the peak as a secondary outcome. 56

To the best of our knowledge, this perspective has not been investigated so far. When raw 57

and cumulative daily numbers of infections and deaths are subject to deviations from between-58 country differences in reporting and depend on the date up to which the analysis is performed, 59 maximum number of daily deaths can be assessed as an interesting new indicator of disease 60 mortality magnitude. In addition, analysis of the height of daily deaths peak enables us to 61 assess the overall capacity of healthcare systems. 62

This study aims to detect significant drivers of COVID-19 mortality with the use of multivariate 63 generalised linear models (GLM) and distinct selection algorithms, to explain the variability of 64 height of and time to the deaths peak among European countries. Due to relatively low sample size, a risk of bias could appear for models with too high number 140 of variables included. To overcome the problem and limit the number of covariates, sensitivity 141 analyses were performed using variety of selection algorithms: stepwise, backward, forward 142 and the genetic algorithm. For the backward, forward and stepwise algorithms, a criterion of 143 having p-value lower than 0.1 was applied for each variable to stay in a model (backward and 144 stepwise) and to enter a model (forward and stepwise). For the genetic algorithm, the best 145 model was fitted based on the value of Akaike's Information Criterion (AIC) corrected for small 146 sample sizes (AICC). Models with only main effects were considered. Additional sensitivity 147 analysis was performed removing countries for which imputation of government restriction 148 dates was needed. 149 Similar methods were used to analyse time to deaths peak but without any transformation 150 since it seemed to follow a normal distribution. GLM models with a normal distribution function 151 and identity link function were analysed. 152

For all analyses, a p-value lower than 0.05 was considered as statistically significant. For each 153 GLM model, fit statistics such as AIC and its equivalent AICC were produced with lower values 154 indicating better fit. 155

Analyses were performed using SAS 9.4 software. R 3.6.2 software was used to apply the 156 genetic algorithm with the package glmulti. 22

Descriptive analysis 160 Characteristics of the countries included in the study are presented in Table 1 , and a histogram 161 of the height of COVID-19 deaths peak is depicted in Figure 1 . The graph presenting deaths 162 peak height by country is provided in Figure 2 . 163

Median height of the peak per 1 million inhabitants per day was equal to 3.48 deaths ([lower  164 quartile; upper quartile] = [1.68; 12.78]), with Belgium reaching outstandingly higher peak than 165 other countries. As can be seen from these statistics and the histogram, the height of the 166 deaths peak does not follow a normal distribution, but it seems it can be well approximated 167 using a log-normal distribution. areas was found significant (p<0.001) and the mobility score at the day of first reported death 184 was very close to reaching significance (p=0.052). Results are presented in Table 2 . 185

Considering small sample size (N=34), selection algorithms were applied to the base case 186 model to limit the number of covariates, increase model precision and improve the model fit. 187

All selection algorithms were consistent and selected the model with three significant 188 parameters ("final" model), indicating its best fit. The proportion of the population living in urban 189 areas (6.848, p<0.001), mobility score at the day of first reported death (0.049 p<0.001) and 190 number of infections when borders were closed per 1 million inhabitants (0.0002, p=0.016) 191

were all significantly associated with the deaths peak height ( preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted November 5, 2020. found significantly related with time to COVID-19 deaths peak. Results are presented in Table  208 3. Considering the small sample size (N=34), selection algorithms were applied to the above 209 multivariate model. The "final" model had four covariates and was selected by the backward 210 and the genetic algorithm (details in Supplementary Materials 

This study provides some evidence about factors associated with COVID-19 mortality peak 221 and time to the peak. One of the strongest predictors of the COVID-19 mortality peak identified 222 in our analysis was the proportion of population living in urban areas. The relationship between 223 urbanisation and COVID-19 infections ratio was earlier outlined by the United Nations 224

Association, 23 estimating that 90% of all reported COVID-19 cases (by July 2020) came from 225 urban areas, becoming the epicentre of the pandemic. High population density increases the 226 propensity of viruses spread by increasing the contact rates of individuals. However, some 227 authors 24 warn readers against putting too much weight on urban density, arguing that large 228 cities just faced the coronavirus earlier (due to the higher number of incoming people) and that 229 the timing of epidemy start was of bigger interest than population density itself. This 230 observation seems to be on the contrary to our study, since neither proportion of population 231 living in urban areas, nor proportion living in metropolitan cities were related with the time to 232 deaths peak. 233

Number of infections when borders were closed was found to be another important factor 234 associated with the COVID-19 deaths peak height, whereas a positive association between 235 borders closure day and time to reach the peak was observed in this study. It shows that 236 stopping arrivals to the country at the earlier stage of epidemy can be crucial in reducing the 237 peak height, stopping the increase of daily number of deaths earlier and, consequently, to 238 flatten the deaths curve. These findings are on the contrary with Chaudhry who showed no 239 association between rapid border closures and COVID-19 mortality per million people, using 240 cumulative data available as of 1 April 2020. 25 However, the peak height and mortality may 241 not be strongly correlated. For example some countries experience a point peak and dropping 242 fast, (France, United Kingdom, Spain, Italy) while other countries may experience a flattened 243 peak with a mortality staying high over a long period of time (US, Brazil, Mexico). 12 244 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.20225656 doi: medRxiv preprint

The difference in average mobility before the pandemic and at the day when first death was 245 reported in a country was found to be associated with deaths peak height, which suggests that 246 rapid reduction in mobility was important in terms of reducing overall mortality. These findings 247 are consistent with other authors opinion that changes in social-distancing covariates can 248 flatten the curve by changing the peak death rate. 9, 10 249 Tourism indicators were found not to be associated with deaths peak height, but to be 250 significant drivers of time to reach the peak. It suggests that the magnitude of inbound tourism 251

can have an impact on accelerating the moment of reaching the peak. However, Aldibasi 5 and 252

Garcia de Alcaniz 8 found tourism to be a significant predictor of COVID-19 mortality and 253 infection rates. Ostig and Askin 26 also found a significant positive relationship between number 254 of airline passengers and number of COVID-19 infections. 255

Another conclusion resulting from the study is that timing of government restrictions, especially 256 border closure, can be found as an important factor in terms of COVID-19 mortality. What can 257 be observed for many countries is that closing borders (and other government restrictions as 258 well) took place several days before any death was reported in a country. Therefore, increase 259 or decrease in mortality can be seen as a result of undertaking preventive action by the 260 government. 261

There is no general conclusion on the association between COVID-19 fatality and hospital 262 beds capacity in literature. Our study showed lack of association between beds capacity and 263 deaths peak height, but higher beds capacity was related with longer time to peak. Garcia de 264

Alcaniz showed no association between hospital beds density and both the number of 265 infections and number of deaths at any moment of the pandemic. 8 preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted November 5, 2020. ; https://doi.org/10.1101 https://doi.org/10. /2020 preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in 

All rights reserved. No reuse allowed without permission.

perpetuity.

preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in

The copyright holder for this this version posted November 5, 2020. ; https://doi.org/10.1101 https://doi.org/10. /2020 

GLM univariate and multivariate models with normal distribution and logit link function were used to explore factors associated with COVID-19 deaths peak height as of 3 rd June 2020. Each model 380 was run using 34 observations. Variables significant in univariate models were included into the multivariate base case model, avoiding highly correlated pairs. The final multivariate model was 381 selected based on the use of selection algorithms (backward, forward, stepwise and the genetic algorithm). 382 383 384 385 

Abbreviations: AIC, Akaike's Information Criterion; AICC, AIC corrected for small sample sizes; GLM, generalised linear models; ICU, intensive care unit; mln, million. 389 GLM univariate and multivariate models with normal distribution and identity link function were used to explore factors associated with time to COVID-19 deaths peak (starting from the day when the 390 first death was reported in a given country), as of 3 rd June 2020. Each model was run using 34 observations. Variables significant in univariate models were included into the multivariate base case 391 model, avoiding highly correlated pairs. The final multivariate model was selected based on the use of selection algorithms (backward, forward, stepwise and the genetic algorithm). perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted November 5, 2020. ; https://doi.org/10.1101/2020.11.04.20225656 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted November 5, 2020. ; https://doi.org/10. 1101 

World Health Organisation

World Health Organisation. Coronavirus disease (COVID-19) Situation Report -142

Covid-19 coronavirus pandemic

National responses to the COVID-19 pandemic

Google. COVID-19 Community Mobility Reports

What is loess regression? SAS Blogs: SAS

Eurostat. Eurostat Database

Air traffic increases, finances in crisis

Air passenger traffic in Turkish airports rose 8.8 percent in 2018

Record-breaking 20.5 million passengers at Ukrainian airports in 2018

Statistics corner: A guide to appropriate use of correlation coefficient in medical 356 research

An R Package for Easy Automated Model Selection with 358 ( Generalized ) Linear Models

COVID-19 in an Urban World

Are Crowded Cities the Reason for the COVID-19 Pandemic? Placing too 362 much blame on urban density is a mistake