key: cord-0708101-5snkl0f0
authors: Aparicio, Ainoa; Grossbard, Shoshana
title: Are COVID fatalities in the US higher than in the EU, and if so, why?
date: 2021-01-16
journal: Rev Econ Househ
DOI: 10.1007/s11150-020-09532-9
sha: cb0b3eec81b6d63410226fc9d7adc0749978fea0
doc_id: 708101
cord_uid: 5snkl0f0

The COVID crisis has severely hit both the United States and Europe. We construct comparable measures of the death toll of the COVID crisis suffered by US states and 35 European countries: cumulative fatalities attributed to COVID at 100 days since the pandemic’s onset in a particular nation/state. When taking account of demographic, economic, and political factors (but not health-policy related factors) we find that, controlling for population size, cumulative deaths are between 100 and 130% higher in a US state than in a European country. We no longer find a US/EUROPE gap in fatalities from COVID after taking account of how each nation/state implemented social distance measures. This suggests that various types of social distance measures such as school closings and lockdowns, and how soon they were implemented, help explain the US/EUROPE gap in cumulative deaths measured 100 days after the pandemic’s onset in a state or country.

It is becoming increasingly common to compare Europe and the USA rather than the US and various European countries. For example, according to Richter (2020) "the trend of daily new COVID cases has taken completely different trajectories for the U.S. and the European Union." COVID fatalities are also routinely compared across the two sides of the Atlantic. For example, Drum (2020) charted 7-day averages of daily deaths in the two unions, letting the US lag the EU by 12 days (reproduced in Fig. 1 ). It shows weekly mortality in the US in June 2020 lying substantially above that of the EU.

There are at least three problems with such comparisons. First, they ignore the enormous variation in COVID outcomes within Europe and within the US (see Table 1 , showing cumulative deaths per million for 35 European countries and all 50 US states). To address this problem we analyze cumulative deaths from COVID in these 85 nations/states. In average population and a number of other characteristics, such as percent of the population aged 65 and older, US States tend to be similar to European countries (see Table 2 ). Second, a lag of 12 days between the average onset of COVID in the entire EU and its average onset in the entire US masks the great variation in onset dates among the 85 nations/states also reported in Table 1 . France was first to experience a death from COVID, on February 15, 2020 (we define time of onset as the day a first death was recorded). Wyoming was the last to experience its first COVID death on April 13, almost two months later. To address this second problem we use statistics on reported COVID deaths 50 or 100 days after the onset of COVID in that nation/state. 1 Looking at means we find a US/EUROPE gap of 207 more COVID-related deaths per million inhabitants 100 days after a nation/state's first death: the mean number of deaths per million is 407 in a US state and 200 in a European country (see Table 2 ). These averages include New York (the nation/state with most deaths per million inhabitants) and Belgium ranking 7th in the list of all nations/states. A number of other European countries rank among the 20 most affected, but most top 20 nations/states are part of the US. 2 The 5 nations/states with the best 100-days performance are all European countries (Malta, Greece, Latvia and Slovakia) except for Hawaii (see Table 1) .

A third problem with many previous comparisons of fatalities in the US and Europe is that they tend to be quick at assigning credit or blame to politicians, while overlooking other factors that may contribute to gaps in COVID deaths. We address this problem by taking account of differences in demographic, political, economic, and health-system characteristics. Demographic characteristics include proportion of the population aged 65 or older and proportion of young adults aged 18 to 34 who live with their parents. We also consider variation in the time that elapsed between onset of pandemic in France and its onset in each of the nations/states. After taking Fig. 1 Reproduced from K. Drum (2020) . The graph represents average mortality rates (deaths per million inhabitants) over 7 consecutive days. For comparability, US data is lagged by 12 days 1 In Wyoming 100 days from onset occurred on July 23, 2020.

2 California, Florida, and Arizona reached 100 days from onset respectively on June 12, June 14 and June 28. As of August 13 these states have experienced new increases in COVID infections. Cumulative deaths per capita in those states reached 816, 369 and 535, respectively. This implies that 160 days after its onset, Florida had fewer cumulative deaths from COVID per capita than France after 100 days. account of such factors, we find that US/Europe differences in cumulative deaths from COVID are considerably smaller than the gap in mean deaths per population shown in Table 2 and that the US/Europe gap is not related to whether a nation/ state's government is affiliated with the left or the right. Our main finding is that the large US/Europe gap in cumulative deaths becomes statistically insignificant in our models including various social distance measures and the timing of their implementation. Relative to US states, European countries were more likely to implement them and did so at a faster pace, and this appears to have saved lives.

We first estimate log-linear regressions of the log of cumulative number of deaths using a sample of 85 nations/states: 35 European countries and 50 US states. Logarithms allow us to interpret coefficients in percentage terms, which facilitates comparability across highly heterogeneous nations/states. 3 For example, we estimate Model 1 defined as:

where y is the log of cumulative COVID-caused deaths 100 days after the first death in nation/state r, U is a dummy for whether the nation/state is in the USA, POP stands for size of the population, and r indexes state or country. Epsilon is the error term. 4 Next, we add X 1r to this equation: it is a vector of demographic and economic characteristics including the following explanatory variables: intergenerational coresidence (measured as proportion of those aged 18 to 34 who live with their parents), percent of the population over 65, and percent urban, as well as economic variables (Gross Domestic or State Product per capita and rental prices). 5 This gives Eq. 2

Model 3 adds to model 2 by also including X 2r , a vector containing the following variables: number of days since first death in France, the square value of this number, 6 and whether a government is left-leaning or not. In the case of EU countries we defined 'left' as having a government that belongs to the Greens-European Free Alliance, European United Left-Nordic Green Left, or Progressive Alliance of Socialists and Democrats groups in the European parliament; in the case of US states 'left' is defined as presence of a governor belonging to the democratic party.

Regression Eq. 4 is similar to Eq. 4, except for the fact that it also includes X 3r , a vector of social distance measures specifying whether a state or country instituted a full or partial lockdown and number of days it took to implement the measure after the onset of the pandemic in each nation/state. The measures we consider are: full lockdown (all-day but could allow citizens to buy essential items), night curfew or other partial lockdown (could apply only to part of the population), closed schools, closed shops and closed social events. 7

We also estimate a model that is similar to Eq. 4, but in addition includes number of hospital beds per capita and number of per capita tests 14 days prior to the day cumulative deaths were measured. 8 All variables are defined in Table 2 . Sources are specified in Table 6 of the Appendix.

Parameter β U in all equations above estimates the difference in the conditional mean between US states and European countries. The predicted mean difference between the US (U) and European (E) death rates can then be written aŝ

where hats indicate predicted values and bars indicate means. This equation could be expanded if there are more than two vectors of explanatory variables. The question of 5 Rental prices may be a proxy for residential patterns and affect real income. 6 France recorded its first death on February 15, and this is the first death recorded in our sample. Time since onset in the West is specified in quadratic terms as we allowed for the possibility of a non-linear relationship with fatalities from COVID. interest to us is: what happens to b β u as more variables are included in the model? To the extent that these variables help explain the difference between European countries and US states the estimated value of β u is expected to decrease. Furthermore, the direction of the change is determined by the last terms in Eq. 5. If the mean value of a variable is greater (lower) in the US than in Europe, and it contributes to reducing b β u , the estimate of b β i is expected to be positive (negative). That is, some of the positive difference in the left-hand side that was captured by b β u is now redistributed to the last two terms. For example, to the extent that nations/states with higher income have more deaths, by including GDP per capita in the equation we expect that the coefficient of the US dummy will go down as some of the differential mortality is captured by differences in state/country income.

Regression results based on Eqs. 1-4 are reported in Table 3 (columns 1 to 4), where cumulative deaths are measured 100 days after onset in each country or state. The five regressions in Table 3 each include a dummy for US and population size, a central determinant of cumulative deaths. When these are the only variables taken into consideration (Model 1 in column 1) we find that the logarithm of cumulative deaths 100 days after onset is 1.3 higher in a US state than in a European country. This implies that cumulative deaths are 130% higher in a US state. On average, according to Table 2 cumulative deaths per capita were 169 in a European country. Multiplying this number by 1.3 gives 220 more cumulative deaths per capita for a US state according to model 1, which implies a total of 389 deaths per capita in a US state. Model 2 presented in Column 2 adds the following demographic and economic variables to the regression in Column 1 that was based on Eq. 1: share of young adults living with their parents, the proportion of the population aged 65 or older, percent urban, GDP or State Product per capita, rental price and dummies indicating that some of these variables have missing values (see Table 7 of Appendix for details about missing values). By adding these variables we see that the US/EUROPE differential in cumulative deaths shrinks to being 100% higher in a US state, which translates into a doubling in the number of deaths, on average from 169 for a European country to 338 for a US state.

Model 3 reported in Column 3 includes three additional variables: date of onset of the pandemic in a particular country, the square value of days since onset, and whether the government of a nation/state is left-leaning. Adding these variables is associated with a slight increase in the intercontinental differential: the coefficient of US in the regression rises, implying that the US/Europe differential increases to 110%.

The model shown in Col. 4 adds various types of social distance measures to Model 3. It corresponds to Eq. 4 above. By adding these measures the US/Europe differential in cumulative deaths shrinks considerably: from 110% based on column 3 to a value that is statistically insignificant and thus not different from zero. Finally, the model in column 5 indicates that by adding information on tests and beds to the model in column 4 the US/Europe differential continues to be statistically insignificant. Here we also add dummies when variables are missing for particular countries.

The differences in the coefficient of US state across the 5 models in Table 3 can be explained with the help of Eq. 5, interpreting X 1 and X 2 as different (vectors of) Table 3 The US-Europe differential in regressions of log of cumulative covid-19 deaths measured 100 days after onset Table 2 and Appendix Table. Robust standard errors in parentheses ***p < 0.01, **p < 0.05, *p < 0.1 explanatory variables added to the model. When comparing models 1 and 2 we see that the US/Europe differential in cumulative mortality shrinks, reflecting the addition of the following variables that favor the spread of the virus and that have a higher mean value in the US than in Europe: rental prices and percent urban. On average US states are slightly more urban and more urban states/countries have had more fatalities. In contrast, the US coefficient is expected to be larger in Model 2 that also includes share of multi-generational coresidence: the US has lower coresidence rates, and coresidence is associated with higher mortality. However, relative to European countries, the US suffers more mortality where intergenerational cohabitation is higher (as shown in Aparicio and Grossbard, 2020a). 9 From model 3 it can be seen that the later the pandemic started in a particular area the lower the number of cumulative deaths, as apparent from the coefficient of the squared value of 'Days since onset in France'. 10 As we add that variable to the model we see that the US/Europe differential rises slightly, given that, on average, COVID epidemics started later in US states than in European countries (the mean time that elapsed between first onset in the West and onset for a US state is 33.5 days; it is 30.7 for a European country). This suggests that the US was favored by the delay in experiencing the first Covid cases. It can be noticed that whether a government is left-leaning or not is not associated with differences in cumulative deaths once all the other variables are included in the regression models. This continues to be the case in the models reported in columns 4 and 5.

What could account for the substantial reduction in the coefficient of US state in column 4, after the addition of social distance measures? First, European countries took less time to close schools (on average, 12.1 days after onset, versus 24.9 days in the US) and to impose full lockdowns in case of full lockdown (on average 3.5 days from onset versus 7.7 days in the US). Second, 100 days after onset in 14 percent of European countries shops were closed (versus in 6 percent of US states) and in 9 percent of European countries there was a partial lockdown (versus zero percent in US states). Even though the results in Col. 4 do not indicate that any of these measures had statistically significant effects on cumulative death rates the presence of the extra vector of variables related to social distance measures does matter and other studies have shown that how quickly lockdowns were imposed was associated with fewer cases or fewer deaths (e.g. Pei et al 2020) . We don't expect the reduced coefficient of US state to be explained by the fact that on average US states closed shops faster (3 days after onset, versus 10 days after onset in European countries) and were faster at imposing a partial lockdown if it was imposed (3.8 days after onset, versus 5.2 days in Europe).

The results in column 5 suggest that little explanatory power is added by including information on hospital beds per capita and COVID tests performed 86 days after onset. Our results don't support or deny the possibility that lives were saved thanks to additional hospital beds. European countries had extra hospital beds (an average of 4.7 versus an average of 2.6 in US states). On average, more tests were given in US states than in European countries, however these tests differences do not seem to be at the origin of the extra deaths in the US.

We include dummies for missing values and it appears that the coefficients of these dummies are often statistically significant. Given that we only have a total of 85 countries or states and some data are missing for 6 European countries (no data are missing for US states) these dummies capture peculiarities unique to the countries missing that information. For instance, we miss information on proportion urban in Cyprus, Macedonia and Turkey. These three countries have fewer cumulative deaths for reasons we can't identify. Table 4 suggests that the US/Europe differential has grown over time. We reestimated the regressions presented in Table 3 , where cumulative deaths are measured 100 days past onset, and instead measured deaths at 50 days past onset. It can be seen that after 50 days the rough differential reported in Column 1 was smaller than after 100 days: cumulative deaths are 90 percent higher in the US, not 130 percent higher, when we only control for population size. Comparing the coefficient of US in Column 1 of Tables 3 and 4 suggests that the US/Europe gap in cumulative deaths has grown over time, as countries and states remain exposed to COVID for a longer time. Furthermore, at 50 days past onset, as soon as we add demographic and economic control variables the US/Europe differential becomes statistically insignificant (Column 2). The differential continues not to be significantly different from zero in the models presented in columns 3 to 5.

To test for the robustness of our results we also estimated regressions using deaths per million inhabitants as an alternative dependent variable (available upon request). Results support our findings that measures such as school closings and lockdowns, and how soon they were implemented, help explain the US/EUROPE gap in Covid deaths.

We also estimated regressions of the mortality rate, measured as number of deaths per COVID case. The same 5 models specified in Section 2 were estimated, but now with a different dependent variable. Results are reported in Table 5 . It can be seen that the coefficient of the US dummy is negative in all regressions and it grows in absolute value as we add an increasingly large number of explanatory variables. The negative coefficient indicates that given the number of cases identified 100 days after onset of COVID in a particular country or state fewer people died per case in a US state than in a European country. To explain the contrast with the US dummy coefficient in Table 3 , which was positive, we note that per capita there were, on average, more tests in US states than in European countries (Table 2) . Consequently, more cases were identified and the numerator is larger, on average, in a US state than in a European country. It is also possible that COVID has been less likely to lead to deaths in the US, conditional on number of cases.

Comparing model 3 in col. 3 (without controls for social distance measures) with the model in col. 4 (including social distance measures) we see that the US dummy rises in absolute value, from −45 to −65. This increase in coefficient is not statistically significant. In both columns 4 and 5 the coefficient of the US dummy is only significant at the 10% level; it was so at the 5% level in cols. 1 to 3. From Table 5 we can't derive the conclusion that US/Europe differentials in the use of social distance measures account for higher mortality in the US. Few variables have a statistically significant coefficient in Table 5 , an exception being a positive coefficient of GDP per capita, especially in col. 5 where we also control for hospital beds: richer countries and Table 4 The US-Europe differential in regressions of log of cumulative covid-19 deaths measured 50 days after onset Are COVID fatalities in the US higher than in the EU, and if so, why? ***p < 0.01, **p < 0.05, *p < 0.1 states may offer better medical care. We prefer our main results reported in Table 3 where the log of cumulative deaths is the dependent variable to the results in Table 5 which depend on both cumulative deaths and number of measured cases, in view of possible measurement errors in both cases and deaths, and the difficulty of establishing whether a variable affects number of cases, number of deaths per case, or both.

Using a sample of 50 US states and 35 European countries we find that 100 days after onset of the COVID pandemic in a particular state or country the US/Europe gap in cumulative deaths stands at 130% when we only control for population size. Given that on average a European country had 169 deaths this implies that on average a US state had 350 cumulative deaths. When we control for other demographic factors and some economic factors, the gap shrinks to 100%. Once we also control for national or state differences in social distancing measures related to COVID the US-Europe gap shrinks considerably and becomes statistically insignificant. This suggests that various types of social distance measures such as school closings and lockdowns, and how soon they were implemented, help explain the gap in cumulative deaths. Relative to US states, European countries were more likely to implement them and did so at a faster pace.

There is much left for further research to establish. We hope that our estimations will be computed with better statistics on deaths from COVID (such as comparisons of number of deaths before and after COVID), better health policy data, and based on a larger sample of countries. It would also be useful to further explore our findings at a more detailed level, such as the US counties, European provinces, or other subnational levels. There have been studies estimating determinants of fatalities using data for small geographic units in the US (e.g. Ahammer et al. 2020) or Europe (e.g. Arpino et al. 2020 , Belloc et al. 2020 , Laliotis and Minos 2020 . Insights could also be gained from combining such sub-national data from the US and Europe, but pooling large sets of data for small geographic units in the USA and Europe is a complex task that has not been undertaken yet. We also hope that further research will keep track of further changes in lockdown policy, beyond the measures taken in the first 100 days of the pandemic and covered in this study. 

Demographics Total population, and % over 65

Gross Domestic or State Product in dollars

Hospital beds

Mass gatherings contributed to early COVID-19 spread: Evidence from US sports

Intergenerational residence patterns and COVID-19 fatalities in the EU and the US

A comment on Landoni et al 's Is time our ultimate ally in defining the pandemic, Pathogens and Global Health

Are intergenerational relationships responsible for more COVID-19 cases? A cautionary tale of available empirical evidence

Cross-country correlation analysis for research on COVID-19. Vox-CEPR Policy Portal

A practitioner's guide to cluster-robust inference

The US Lags Way Behind Europe in COVID-19 Mortality. Kevin Drum blog

Reopening under COVID-19: what to watch for

Data From the COVID-19 Epidemic in Florida Suggest That Younger Cohorts Have Been Transmitting Their Infections to Less Socially Mobile Older Adults

Spreading the disease: the role of culture (20/12)

Is time our ultimate ally in defying the pandemic? Pathogens and Global Health

Differential effects of intervention timing on COVID-19 spread in the United States

COVID-19 PANDEMIC: The state of the unions

Acknowledgements We thank the coordinating editor (George Davis), two anonymous referees, Cynthia Bansak, Jeffrey E. Harris, Maurice Schiff, and Zev Shechtman for helpful comments.

Conflict of interest The authors declares that they have no conflict of interest.Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.