key: cord-0686814-cmqj8get
authors: Walach, H.; Hockertz, S.
title: What association do political interventions, environmental and health variables have with the number of Covid-19 cases and deaths? A linear modeling approach
date: 2020-06-22
journal: nan
DOI: 10.1101/2020.06.18.20135012
sha: ae68bb60957f2f5270c662bc37054448e9be1d93
doc_id: 686814
cord_uid: cmqj8get

Background: It is unclear which variables contribute to the variance in Covid-19 related deaths and Covid-19 cases. Method: We modelled the relationship of various predictors (health systems variables, population and population health indicators) together with variables indicating public health measures (school closures, border closures, country lockdown) in 40 European and other countries, using Generalized Linear Models and minimized information criteria to select the best fitting and most parsimonious models. Results: We fitted two models with log-linearly linked variables on gamma-distributed outome variables (CoV2 cases and Covid-19 related deaths, standardized on population). Population standardized cases were best predicted by number of tests, life-expectancy in a country, and border closure (negative predictor, i.e. preventive). Population standardized deaths were best predicted by time, the virus had been in the country, life expectancy, smoking (negative predictor, i.e. preventive), and school closures (positive predictor, i.e. accelerating). Model fit statistics and model adequacy were good. Discussion and Interpretation: Interestingly, none of the variables that code for the preparedness of the medical system, for health status or other population parameters were predictive. Of the public health variables only border closure had the potential of preventing cases and none were predictors for preventing deaths. School closures, likely as a proxy for social distancing, was associated with increased deaths. Conclusion: The pandemic seems to run its autonomous course and only border closure has the potential to prevent cases. None of the contributes to preventing deaths.

The novel Coronavirus SARS-Cov2 (CoV2) which surfaced in China in December 2019 for the first time created a world-wide pandemic 1, 2 and an associated disease, named , with respiratory stress, heart problems, kidney failures and immunological problems associated with it [3] [4] [5] [6] [7] . Countries closed down their borders, their schools, universities and cultural facilities and sometimes even their whole activities. This was due partially to its novelty and its largely unknown properties, but also, because it was soon clear that those infected by the virus could be asymptomatic for up to a week or longer, while still being infectious to others, and because high infectivity, virulence and mortality was assumed.

The spread of the virus was initially very quick following a seemingly exponential growth curve, but abated and the replication numbers went into decline. Currently it is highly debated what contributes to the variance that can be seen both in CoV2 cases, as well as in deaths attributed to . While most people assume that political measures have mitigated the spread of the virus 8 , others hold that the process is rather autonomous, that the virus recedes after having infected all those in a population susceptible to it and then the infection abates 9, 10 . Moreover, most modeling approaches that were used in early stages of the disease to inform political decision making did not take into account potential inhomogeneity of a population due to natural or specific immunity of a large part of the population 11, 12 . More recent models that take such inhomogeneity parameters into account, informed by novel data, estimated that after about 7 to 18% of a population have been infected herd immunity is reached, because the rest of the population might not be susceptible to the virus 13, 14 .

Since it is largely unclear what variables contribute to the variance in cases and deaths attributable to CoV2, we wanted to study this question by building linear models using various predictor variables to study their influence on the outcomes Covid-19 cases and deaths in various countries.

We collected data on Covid-19 cases and deaths as presented by the database of the European Center for Disease Prevention and Control on their website by 15th May 2020. We used European and OECD countries, because those data are most relevant to our question and are more validly accessible. We included the following 40 countries Covid-19 Cases and Deaths were summed for the total period covered by the ECDPdatabase and used as dependent variables (criterion). We standardized cases and deaths on 100.000 inhabitants, taken from the population size in the same data-base.

As predictors we collated data from publicly available sources (see Supplementary Material for a list and for sources) for population, health, health systems, and environmental indicators between May 15 th and 20 th 2020. The variables used as predictors are described in a protocol that was published on the server of the Open Science Framework (https://osf.io/x93np/). Briefly, we used population indicators (Life-expectancy, percent single households, city dwelling, age groups, population density), health systems indicators (number of doctors, hospital beds, ICU beds, PCR tests), health indicators (percentage of obese, diabetes, smoking and physically inactive persons), air pollution, and finally variables coding for political . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 actions: closure of borders, closure of schools, country lockdown, including the rapidity of implementation since the first case was noted (see Table 1 ).

We built two separate linear models to predict the influence of variables on populationstandardized CoV2-cases and Covid-19 associated deaths.

In order to investigate which variables might be potential predictors first order correlations of all relevant variables with the outcome variables were calculated, using nonparametric correlations, and their inter-correlation structure was studied. Only variables that contributed with an effect size r > .3 or with significant correlations were further considered for modeling.

As 40 cases offer enough stability to estimate about 4 parameters reliably 15 , we opted for small models to start with and included a further predictor only if it was theoretically meaningful, empirically supported (i.e. a significant predictor) and improved model fit. We included first all those relevant predictors from population, health systems, and environmental sets in separate steps that correlate significantly with the outcome and are not collinearly related with each other. We explored model fit to find the best subset for each small group of indicators with forced entry of not more than four variables at a time, retaining only significant predictors for the next step. In a final step we included potential predictors from the set of public health indicators to investigate whether there is any improvement in model fit and whether these variables were significant predictors. The rationale of this procedure is: If the public health measures contribute to preventing cases and deaths, then they would have to emerge as potential significant predictors with negative sign (as they were dummy coded with 1 coding for present and 0 coding for absend). In addition, the model fit of the enlarged model would have to improve.

As an indicator of improved model fit we used the difference of Akaike Information Criterion (AIC in its original and corrected version), the difference of Bayes Information Criterion (BIC) and the Chi 2 -Goodness of Fit test statistic divided by degrees of freedom conjointly to avoid over and underfitting. We always used the model that minimized all of them conjointly. To assess model adequacy, plots of predicted versus observed cases, residual distribution plots, and residuals vs. cases were visually analyzed and residual plots were screened for outliers (residuals vs. Chi 2 statistic).

In a sensitivity analysis the model was recalculated without outliers to see whether the model structure, i.e. the variables used as significant predictors would be the same and goodness of fit improved. For those sensitivity analyses, AIC and BIC were only used as a further criterion if the difference was large, as the efficiency of these information criteria change with number of cases/degrees of freedom and number of variables 15, 16 . We used Statistica Version 13.1 for all analyses.

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

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 The nonparametric correlations (Spearman's Rho) between the two predefined outcome variables, cases and deaths per 100.000 inhabitants, as well as the case-fatality rate (CFR), for illustration, are reported in Table 1 .

Insert Table 1 here Of the structural variables describing the health systems only the number of tests conducted correlated significantly with number of standardized cases (r = .32) and deaths (r = .46), as well as with case-fatality rate (r = .35) and with number of ICU beds (r = .39).

Of the variables describing political actions only border closure was negatively and significantly related with standardized cases (r = -.43), but only weakly and non-significantly with number of deaths (r = -.25): Cases tended to be higher in countries that did not close the border. But neither lockdown nor school closures were significantly and sizably related with number of cases or number of deaths. Only full closure of schools was slightly, but nonsignificantly related with number of cases (r = -.19) but not with number of deaths, indicating that cases were higher in countries that had not closed schools. However, as school closure was correlated positively, but non-significantly (r = .30) with CFR, it is necessary to clarify by modeling, which covariation might be influential.

The duration or length of border-or school closures was only marginally and nonsignificantly negatively correlated with number of deaths and cases. The rapidity with which countries reacted, i.e. the time difference between the registration of the first case and the initiation of political reactions was only slightly correlated with number of cases, and significantly correlated only with the case-fatality rate (r = .39), i.e. countries that were slower in initiating political actions had a higher case-fatality rate.

Higher and significant correlations were visible with descriptors of populations and health status. There were more cases in countries that had a higher life-expectancy at birth (r = .53), and there were more deaths (r = .43) in such countries. There were more cases (r = .47), as well as more deaths (r = .36) in countries with a higher percentage living in cities. However, neither population density of a country, nor percentage living in single households emerged as a potential predictor. Potentially interesting correlations emerged between case-rate and death rate with percentage of population taking lipid lowering drugs (r = .42) and with amount of mercury used in the alkaline-chlorine industry (cases: r = .42, deaths: r = .40). But since we were unable to find enough data for all countries of interest, these variables could not be used for modeling. None of the other variables describing the health status of a population (obesity rate, insufficient physical activity, sleep problems, vaccination rate, percent of diabetes patients in a population) emerged as potential predictors.

Paradoxically, there were more cases (r = -.38), as well as deaths (r = -.33) in countries that had a lower percentage of smokers in the population. As this correlation was even higher for male smokers, likely because smoking is predominantly a male phenomenon, the percentage of male smokers was used for further modeling. The same paradoxical relationship can be seen with variables that code for air-pollution, especially with very small particles (PM2 -particulate matter of 2 micron size per m 3 air), where we see significant negative correlations of r = -.52 with cases and r = -.39 with deaths. Although the correlation with PM10 was somewhat higher, we used PM2 for modeling, because PM2 and PM10 are highly intercorrelated (r = .75) and because we had more cases with data for PM2, most notably USA.

Following our protocol, we constructed a model to account for the covariance structure of the variables. Since the outcome variables, cases and deaths standardized on 100.000 inhabitants per country, showed adequate fit to a gamma-distribution (see e- Figure 1 ), we calculated a generalized linear model with a log-link-function on gamma-distributed outcomevariables:

where Γ refers to the Gamma distribution with shape ki and shape si, i 1, … , .

We used the Gamma distribution as it maximizes entropy. Although the (overdispersed) Poisson distribution may have been a choice, we opted for the Gamma distribution because by modelling standardized cases, we are effectively modelling a continuous variable, thereby excluding the Poisson distribution (which models discrete events). We also considered logtransforming the outcome variables to approximate a normal distribution, but the fit was not adequate and sensitivity analyses using linear regression on a log-transformed outcome variable yielded essentially the same results, but with inadequate fit.

The models that best predicted standardized cases are presented in Table 2 . The first model describes the best fitting model for all countries predicting cases. The variables entering the model are life-expectancy, number of tests and smoking. Parameter estimates are positive for life expectancy and number of tests, and negative for smoking. In a second step variables coding for political decisions (country lockdown, border closure, school closures) were entered. The best fitting model emerged with border closure as a negative predictor, with smoking removed. The model fit statistics show improved model fit over the first model (Akaike Information Criterion -AIC 462,62 vs. 465,69; Bayes Information Criterion 470,94 vs. 474; Chi 2 /degrees of freedom 0,43 vs. 0,46). We inspected the Chi 2 vs. prognosis plot to spot outliers. There was only one clear outlier, Belgium (eFigure 3). Removing this outlier improved model fit considerably (AIC 379, 21; BIC 388,19; Chi 2 /degrees of freedom 0,24), with air-pollution PM2 added to the model as a negative predictor. Table 2 here

The model predicting Covid-19 related deaths is presented in Table 3 : Here the duration the infection had been in the country is a significant positive predictor, and so is life expectancy. Smoking is a negative predictor. When entering the public health variables only school closures emerged as a significant positive predictor that improved model fit. Excluding . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 Belgium, the only serious outlier, improved model fit. The same variables remain in the model as significant predictors with nearly the same regression coefficients including their sign. Table 3 here Inspection of residuals show that the linearity assumption is warranted. The model can predict the cases reasonably well. A plot of countries vs. residuals is presented in e- Figure 4 .

The major findings of this modeling study using population data for 40 countries are clear, if surprising: Life-expectancy emerges as a stable positive predictor both for standardized cases of CoV2 infections, as well as for Covid-19 related deaths. Surprisingly, smoking emerges as a stable negative predictor, i.e. protective factor. Of the public health or political variables only border closure emerges as a strong negative predictor for cases. But school closures is a strong positive predictor for deaths, i.e. is associated with more deaths. The parameter for number of tests conducted in a country emerges as a strongly significant positive predictor.

The fact that life expectancy is the most consistent positive predictor -the longer the life expectancy in a country the more cases and deaths -is easy to understand. The disease affects most aggressively elderly and multimorbid patients. Life expectancy is a complex variable, incorporating social and medical progress in a country as well as economic indicators, and hence it denotes the number of the elderly in a population, as well as the intensity of medical care. Only the number of CoV2-PCR-tests, out of all health systems variables, enters the model as a significant positive predictor. Other quality indicators of the medical system (number of doctors per 10.000 inhabitants, number of ICU or hospital beds) do not enter the model. That number of tests should be related to the number of cases is evident: the more tests are conducted in a country the more cases can be potentially registered.

The absence of other indicators from the set of medical system variables shows that the development both of infections and deaths is rather independent of the preparedness of the medical system. Although some interesting first order correlations indicated that health status variables might be interesting to explore, none of them emerged as a predictor, except smoking as a somewhat protective variable. This might have to do with the fact that smokers may have a hyperactive system to combat airborne noxes and hence might have a small advantage against this particular disease 17,18 . A large cohort study has documented a similar counterintuitive effect 19 , and an argument could be made that this might have to do with the fact that smokers express fewer ACE2 receptors 20 , which are the main entry gate of CoV2 into the lungs. 2 However, the correlation of smoking with life expectancy is negative (r = -.50 for men), and hence smoking might confer other risks that shorten lives. Only in one model, excluding the outlier Belgium, does air pollution play a role as a potential negative predictor, i.e. as a potential preventive factor. This is rather counterintuitive. Either it could be understood along the same lines: lungs prepared to deal with small noxes might be better prepared to fight a virus. Or else, airborne viruses might be captured by small airborne particles and might fall to the ground earlier. Again, this could be an accidental effect that should not encourage air pollution, as this has detrimental effects elsewhere.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 The closing of borders is a significant negative predictor, denoting a protective effect, in the model including all countries and predicting number of cases, but not for predicting number of deaths.

In all models predicting death the duration of the infection is a positive predictor. As deaths develop with a delay of perhaps 3-4 weeks after the first contact with the virus 21-24 , this relationship reflects a quite independent temporal dynamic of the infection. It is interesting to observe that closure of schools emerges as a strong positive predictor for the number of deaths, i.e. school closures are associated with more deaths. This could be an indicator for strong social distancing rules in a country which might be counterproductive in preventing deaths, as social distance for very ill, and presumably also very old patients, might enhance anxiety and stress and could then become a nocebo 25, 26 . It could also reflect the fact that countries which saw a rising tendency of deaths closed schools as an emergency measure, and hence school closure is an indicator of fear in a country. But considering the prevention of deaths, none of the public health measures studied are associated with the prevention of deaths. Border closure might be an exception in that it is associated with reduction in the number of cases and hence, indirectly, the number of deaths. But in the full model predicting deaths it is not a significant predictor. This seems to contradict new modeling data using time series models 27,28 that report clear evidence for the effectiveness of non-pharmaceutical interventions. We actually doubt the validity of these findings. The major shortfall of these models is that they ignore the most likely reason why we find the data we find: immunity in the population and neglecting the strength of natural immunity (see below). Thus, a new reliability study of such models shows that they are crucially dependent on assumptions, parameters assumed and the time point at which they capture data 29 . If the wrong assumption about a potential resistance against an infection in a population is made, the results are far off from true values.

Rapidity of reaction can be a positive predictor in some models, but reliably leaves the equation, as soon as the duration of infection is taken into account. This signals in our view, the fact that the dynamics of the infection develops quite independent of political actions, or rather that political actions are mostly too late.

The duration of an infection in a country was only a significant predictor for standardized deaths. A model including this variable to predict standardized cases is not significant and does not improve model fit.

One might argue that a perhaps more conventional way of modelling would have been to log-transform the outcome variables and use standard linear regression approaches. We tried this as a sensitivity analysis but did see essentially similar results with a residual distribution that signaled model inadequacy, and hence we doubt that such a model would have helped with understanding the data.

As data on number of tests were not available for China one might argue that our model is inadequate, as it excludes an important country. While this is true, we fitted models without the number of tests as a predictor which did not lead to better fitting or more meaningful models. We also used population standardized tests as an alternative to raw number of tests, but found that the model fit was much worse.

Thus, the image that emerges from the data and the attempt to understand their relationship through modeling is that of a largely autonomous development. It affects mainly the elderly. Smoking is somewhat protective and border closures is associated with a lower number of cases. But other measures -closing of schools and lockdown of whole countries -. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 do not contribute to a reduced number of cases or deaths. This may have to do with the fact that the virus travels extremely quickly. Even the shutdown of Wuhan airport delayed the spread across China only by 2,8 days 30,31 , and as Chinese airports remained open the spread of the virus across the world was guaranteed and could not be stopped by gross measures such as border closures, as these came too late. An Italian seroprevalence study estimated that even at the very beginning of the pandemic in Italy there were 2.7% of the population in Milan that had had contact with the virus. 32 The examples of Taiwan 33 and Hongkong 34 show that containment is possible, if reactions come very quickly and if cases can be traced close to 100%. But already the presence of 5 cases in a population increases the likelihood of a pandemic by 50%. 35, 36 Once infections are in the vulnerable segments of a populations, like in hospitals or homes for the elderly, political actions like school closures or country lockdowns do not prevent deaths. If anything, social distancing seems to be harmful. What might be useful but cannot be seen in our coarsegrained data are special protective measures geared to protect these vulnerable populations, such as protective masks for personnel and visitors in hospitals and old people's homes, or the wearing of face masks in places with bad ventilation and close proximity of people.

Why, then, have infections subsided and deaths receded since we gathered our data on May 15 th 2020? Most people would say this was due to the public health measures 1 , and recent modelling studies seem to support this 8, 27, 28 . However, we have pointed out that the peak of the cases had been reached in Wuhan already on January 26 th , only 3 days after the city lockdown. 37 This was surely too short to be an effect of public health measures, as cases manifest with a delay of at least 5, rather more days. And a careful analysis shows that, if one uses realistic retrodiction of cases, then effects of public health measures cannot be seen 38 . Thus, our modelling supports the view that the public health measures of school closures and country lockdown, with the exception of the closure of borders to reduce cases, were likely ineffective in influencing cases and deaths. If anything, social distancing might even be harmful for seriously ill patients.

Very likely, scientists and governments overestimated the danger this virus presented and underestimated the immunological resistance in the population. While there is no doubt that those really falling seriously ill from this infection suffered a lot more and were in much greater danger than comparable patients suffering from flu or other respiratory infections 22 , there can also be little doubt that basic immunological insights were neglected from the outset. Both specific [39] [40] [41] and non-specific immunity [42] [43] [44] seems to have been much greater in the population than initially assumed. This is likely the case because the difference of CoV2 from other Corona-viruses is not as great as initially thought. Thus, a considerable percentage of any population would have been immune through specific cross-immunity against other coronaviruses, apart from the fact that non-specific immunity has been neglected in the discussion nearly completely. This is the reason why more recent models that account for this fact and introduce inhomogeneity parameters reach the conclusion that it is sufficient if 7%-18% of a population have had contact with CoV2 to reach herd immunity 13 , and that further waves are unlikely given immunity 45 .

Our data are not foolproof but first important hints. We were unable to code more countries, due to restrictions in time, resources and availability of data. This reduces the stability of estimates and to some degree also variance, although for those variables of interest variance was large enough to estimate stable models. 46 For the chosen models the goodness of fits test signals good fit, and the relative improvement of AIC and BIC values is obvious and . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 well below conventional thresholds of differences between source model and improved model. [47] [48] [49] We opted for completeness of data as much as possible rather than for a large number of countries, as modeling depends on the completeness of case-wise data. Some interesting and potentially useful predictors we were unable to gather: a more fine-grained resolution of different social distancing rules in different countries, availability and wearing of face masks for medical personnel and the public, for instance. Every model is wrong 50 , but given the data our models have a comparatively good fit. This can also be seen indirectly, as excluding outliers improved model fit, but did not change the predictors and their overall structure. Obviously, in a population-based study we have to rely on the validity of the data provided by other sources, which may be of variable, even doubtful quality. This limitation has to be borne in mind.

An exploratory modeling approach like ours is always open to critique. We decided to build parsimonious models that are theoretically guided and conceptually informed 15 , starting with health systems, structural and population indicator variables and entering political public health variables in a last step, then adapt the model to find the best model fit. We followed a predefined, published protocol which guarded us against aimless fishing, and strove for parsimonious models that could explain the data with a minimum of predictors and good model fit. We avoided computer guided step-down and step-up procedures as they are inefficient or prone to overfitting. 15 Thus, we are quite confident that we did not overlook an important contribution of political actions to an explanatory model: they are not visible in our data except for those we report.

In conclusion: In our data-set of 40 countries, only border closure had the potential to prevent cases. Other public health measures were not associated with reduced CoV2-cases or Covid-19 associated deaths. Rather, the pandemic seems to take its own course. Since being elderly is a risk factor for many diseases, and eventually death, and cannot be changed, political actions in future pandemics would likely need to focus on protecting these members of society first. Apparently, closing schools and locking down countries is not the right method to prevent deaths. Perhaps the most sensible measures against pandemics are high alertness and an early warning system that initiates rapid actions that can prevent pandemics from developing.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 

We are grateful to Sebastian Sauer for advice on modeling and for counter-checking the validity of our modelling with R-routines and for critically commenting on an earlier draft. We thank the students of the master class "Quantitative Research Methods" of the MSc course "Health Promotion" who gathered the data for this study and participated in discussing and initiating this project.

Not applicable for secondary data-analysis.

No external funding and no external influence.

None of the authors has a conflict of interest.

Data will be made publicly available after publication and for peer review and qualified requests beforehand.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 16 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020 . . https://doi.org/10.1101 Residuals for full models per country; the last country is USA; China is missing, because there were no data for tests conducted in China . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020 . . https://doi.org/10.1101 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020 . . https://doi.org/10.1101 Figure 4e -Model diagnostic: histogram of raw residuals -predicting cases . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020 . . https://doi.org/10.1101 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020 . . https://doi.org/10.1101 

Data were extracted from the the Database of the European Center for Disease Prevention and Control, as of 15 th May 2020. This yields day-wise cases, deaths and population numbers for each country. The data were parsed (cases and deaths summed, date of first case registration and populations numbers extracted) using the Statistica Data Reporting Tool and the countries of interest included. These were all European countries including Switzerland, as well as other countries of the OECD, including China, Iran, Russia, India, Japan and Brazil to represent all countries that were at the beginning of the crisis, as well as other large countries in the world. We excluded Africa, other South and Middle American and Asian countries because of a lack of resources, time and because we were not sure we would be able to find sufficient data. This yielded the 40 countries described in our protocol. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 Data for border closure of Iran suggested that other countries closed their borders against Iran, but not Iran against other countries. Hence this was marked as open.

We wanted to include the availability of face masks and their number, but those data were unavailable for most countries, so we excluded this variable.

Were taken from the UNESCO data base which updated them on a daily basis; we extracted the data on the 15 th of May:

UNESCO. COVID-19 Educational Disruption and Response. Access on 18.05.2020. https://en.unesco.org/covid19/educationresponse

We extracted data for vaccination rates from https://de.statista.com/statistik/daten/studie/1034782/umfrage/laender-mit-d,er-hoechstenimpfquote/ This is a representative survey and covers all different vaccinations. As this covers only Europe, we searched other sources. For China, India, Iran, Japan data came from https://ourworldindata.org/grapher/immunization-coverage-against-diphtheria-tetanus-andpertussis-dtp3-vs-gdp-per-capita For Brazil from https://academic.oup.com/jtm/article/25/1/tay100/5127106 For Canada from https://www.canada.ca/en/services/health/publications/vaccinesimmunization/vaccine-uptake-canadian-children-preliminary-results-2017-childhood-nationalimmunization-coverage-survey.html USA from https://www.epa.gov/outdoor-air-quality-data/air-quality-statistics-report CSV formatted for all counties for 2017; imported into new spreadsheet and calculated median across PM2.5 weighted 24 h average, because there were a few counties with very high . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 doi:https://doi.org/10.1016/j.atmosenv.2019.05.003) did not contain data. Prof. Kümmerer, an expert in environmental toxicology did not have any information, neither did other experts asked. Hence we dropped this variable in the final analysis, as the data from the UN-EN report on mercury in the alkaline industry was too patchy.

At each step new partial data-sets were created and added to the master database with the Statistica-data "merge" command with variable merging according to country names and each set was saved under a new name.

Since some variables were created as functions of others and indexed by the original variable number, the final database contained all variables, even those that were never further used except for orienting correlations.

Outcome variables were log-transformed to check whether they would then be normally distributed to allow for a normal linear regression. As the result was less than satisfactory we decided to go for a generalized model and regress on a gamma-distributed variable, as the variables were clearly gamma-distributed.

Date variables were calculated from starting date to the 15 th May for border closure, school closure, and finally for the rapidity of reaction as the difference. This was defined as the date when the first measure, either border closure or school closure was registered and the days to first case registration was calculated.

Distribution analyses of the dependent variable showed that it was gamma distributed and that a log transformation cannot rectify this. Hence it was decided that a regression on gamma-distributed variables should be calculated.

First non-parametric correlations were correlated to see which variables correlate at all with the outcome, and as defined in the protocol, only variables with r > .3 and/or significantly correlating variables were considered further in regression models. The regression models used the functionality Generalized Linear Models, stipulating a gamma distributed outcome variable with a log-link function. The parametrization method was overparametrized, as there was no sigma restricted coding in our data.

All potentially included variables were inspected for their descriptive parameters and to see, whether they contribute any variance. If several similar variables (.e.g. air pollution variables) were available we used those that had the least missing data in order to not lose power.

All modeling approaches included first health serviced, population and health parameters in a model, trying to fit the most parsimonious model with only significant predictors in the equation. This was done by calculating forced entry models, excluding non-. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 significant predictors and recalculating the model. Highly intercorrelated variables were never used together, but separate models were calculated and the model with the best model fit was selected.

After that variables representing political actions (country lockdown, school closure) were entered in an additional model and used if significant as predictors.

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

The copyright holder for this preprint this version posted June 22, 2020. . https://doi.org /10.1101 /10. /2020 

Association of Public Health Interventions With the Epidemiology of the COVID-19 Outbreak in Wuhan

On the origin and continuing evolution of SARS-CoV-2. National Science Review

SARS-CoV-2 infects T lymphocytes through its spike protein-mediated membrane fusion

SARS-CoV-2 invades host cells via a novel route: CD147-spike protein

Functional exhaustion of antiviral lymphocytes in COVID-19 patients

Baseline Characteristics and Outcomes of 1591 Patients Infected With SARS-CoV-2 Admitted to ICUs of the Lombardy Region

Epidemiology and transmission of COVID-19 in 391 cases and 1286 of their close contacts in Shenzhen, China: a retrospective cohort study. The Lancet Infectious Diseases

Inferring change points in the spread of COVID-19 reveals the effectiveness of interventions

The end of exponential growth: The decline in the spread of the coronavirus

Von der fehlenden wissenschaftlichen Begründung der Corona-Maßnahmen

Impact of non-pharmaceutical interventions (NPIs) to reduce COVID19 mortality and healthcare demand. London: Imperial College

Individual variation in susceptibility or exposure to SARS-CoV-2 lowers the herd immunity threshold

Why herd immunity to COVID-19 is reached much earlier than thought

Model Selection and Inference: A Practical Information-Theoretic Approach

Regression and time-series model selection

Feasibility of controlling COVID-19 outbreaks by isolation of cases and contacts. The Lancet Global Health

Early dynamics of transmission and control of COVID-19: a mathematical modelling study. The Lancet Infectious Diseases

Wuhan Covid19 data -more questions than answers

Inferring change points in the spread of COVID-19 reveals the effectiveness of interventions). advance Social Sciences and Humanities Preprint

Asymptomatic Seroconversion of Immunoglobulins to SARS-CoV-2 in a Pediatric Dialysis Unit

Human coronavirus reinfection dynamics: lessons for SARS-CoV-2. medRxiv

The infection fatality rate of COVID-19 inferred from seroprevalence data. medRxiv

Epigenetic landscape during coronavirus infection

SARS coronavirus pathogenesis: host innate immune responses and viral antagonism of interferon. Current Opinion in Virology

Recognition of virus infection and innate host responses to viral gene therapy vectors

Second waves, social distancing, and the spread of COVID-19 across America

Interpreting and understanding logits, probits, and other nonlinear probability models

Model selection and model averaging in behavioural ecology: the utility of the IT-AIC framework

AIC and BIC: Comparisons of assumptions and performance

AIC model selection using Akaike weights

Model Selection and Model Averaging

values, median was taken, else it would have been 103 instead of 40. Only PM2 and PM10

Data including US air-pollution data Covid10-mastertabelle10.sta Some countries missing on that variable

Worldwide trends in insufficient physical activity from 2001 to 2016: a pooled analysis of 358 population-based surveys with 1·9 million participants

World health statistics 2020: monitoring health for the SDGs, sustainable development goals. Geneva: World Health Organization; 2020. Licence: CC BY-NC-SA 3.0 IGO. Other Sources on Health and Population Data

Japan und Deutschland im Zahlenvergleich (2): Bevölkerung. Accessed 15

Türkei: Mehr Einpersonenhaushalte und kleinere Haushaltsgrößen. Accessed 15

IDF MENA Members. Accessed 15

south-and-central-america/members.html International Diabetes Federation. (2020). IDF SEA members. Accessed 15

IDF Western Pacific members. Accessed 15

Bevölkerungsdichte nach Ländern. Accessed 15

Do more people live in urban or rural areas? Accessed 15

Anteil der Einpersonenhaushalte an allen Privathaushalten in den Ländern der EU im Jahr

China: Altersstruktur von

Großbritannien: Altersstruktur von

Japan: Altersstruktur von

Norwegen: Altersstruktur von

Prävalenz von Diabetes bei zwischen 20-und 79-Jährigen in ausgewählten Ländern weltweit im Jahr

Russland: Altersstruktur von

Türkei: Altersstruktur von

USA: Altersstruktur von

Altersstruktur der ständigen Wohnbevölkerung in der Schweiz von 2009 bis 2019. Accessed 15

Brasilien: Altersstruktur von

Europäische Union: Altersstruktur in den Mitgliedsstaaten im Jahr 2019. Accessed 15

Europäische Union: Bevölkerungsdichte in den Mitgliedsstaaten im

Indien: Altersstruktur von

Iran: Altersstruktur von

Kanada: Altersstruktur von

substantial-investment-needed-to-avert-mental-health-crisis Prevalence of insomnia from various review sources: Canada

Ohayon & Sagales (2010) Incidende of insomnia

European guideline for the diagnosis and treatment of insomnia

Sleep dissatisfaction from Canada

Switzerland: Statista "Leiden Sie unter Schlafstörungen

And prevalence of sleep problems from Van de Straat & Bracke

Then a new hyper-variable was constructed in which the mean rank of those variables that were available per country was deposited. Finally, this new mean rank-variable was again ranked to yield the rank order of countries with sleep problems. This was used for further analysis. Health Services Data: We exctracted number of doctors, standardized per 1.000 inhabitants, number of hospital beds and number of ICU beds, standardized. Hospital Beds (per 1.000 inhabitants) WHO: Global Health Observatory data repository

beyond-containment-health-systems-responses-to-covid-19-in-the-oecd-6ab740c0/ The following countries are not contained in this list and data for these countries come from the according sources: China: Phua

Drugs We attempted to get data for lipid-lowering drug consumption, but as these data were sparse and not systematically comparable, we desisted from further attempts. Mercury Country wide mercury consumption is not easily available. UN-EN reports give tons of consumption for those countries that use mercury in the chlorine-alkaline industry but this is only about half of the countries. For the rest only regional data were available. We contacted Dr. Steenhuisen and Prof. Kümmerer in the hope to get help. Dr. Steenhuisen did not provide data and the publication