Did World War II Deaths Help Prevent Deaths from COVID-19?

Abstract

ABSTRACTThe paper documents and tries to explain a striking negative correlation between COVID-19 mortality across countries and deaths during World War II. The correlation persists with various controls for observables and allowing for latent omitted variables, using the pre-war distribution of the Jewish population for identification. The correlation also survives influence and falsification tests, measurement-error adjustments, and tests for spatial autocorrelation, which can generate spurious historical dependence. We suggest a theoretical explanation whereby large shocks promote institutions and cooperative behavioral norms – interpretable as civic capital – that initially help attenuate losses from future large shocks, though with fading impact over time.KEYWORDS: COVID-19Europerare eventsshocksWorld War IIJEL CLASSIFICATION: D74I12N10 Disclosure StatementNo potential conflict of interest was reported by the author(s).FindingThis paper’s findings, interpretations, and conclusions are entirely those of the authors and do not necessarily represent the views of their employers including the World Bank, its Executive Directors, or the countries they represent. The authors thank Branko Milanovic for discussions, Toan Do, Ivan Torre and Dominique van de Walle for their comments, and two anonymous referees for their constructive comments and suggestions.Notes1. For a recent survey see Cioni et al. (Citation2021) and Dupraz and Ferrara (Citation2023).2. All but the last is reasonably well known; on the last see Kelly (Citation2020). The historical literature on the development of social policies points to a degree of spatial correlation (Ferrera Citation2005).3. See, for example, the discussion in Arthi and Parman (Citation2021), with reference to the 2020/21 pandemic and its potential future impacts.4. The literature on social capital and health has pointed to such a distinction between “cognitive” and “structural” social capital (Murayama, Yoshinori, and Kawachi Citation2012).5. The index appears to rank countries differently from rankings based on similar perception-based measures (i.e., WGI). For example, El Salvador has Civil Capital index of 0.08 while that index for France is −0.59 and 0.00 for Finland. Such differences might arise from differences in country/culture-specific subjective scales respondent use when answering these questions (Ravallion and Lokshin Citation2001)6. Also see Egorov (Citation2020) on Russia’s success in rapidly containing a smallpox outbreak around 1960. The success would not have been possible without widespread public acceptance.7. Surveys data for the U.S. indicate a strong association between the acceptance of social norms for cooperative health behaviors and actual personal preventative actions during the pandemic (Goldberg et al. Citation2020). Also, for the U.S., Barrios et al. (Citation2021) find greater use of face masks in counties with higher measures of civic capital.8. WWII is often mentioned as a turning point in social policy making in Europe; see, for example, the discussion in Ferrera (Citation2005).9. Wu (Citation2020) provides a more complete review of the sociological literature on the role of social capital in success in dealing with the COVID-19 pandemic.10. When we add to our empirical specification the government effectiveness indicator (Kaufmann, Kraay, and Mastruzzi Citation2006) used in two studies, its coefficient is not statistically significant, while the sign and significance of the coefficient on WWII total losses do not change. Likewise, we find no significant results regressing government effectiveness indicator on WWII total losses. We re-estimate our model with an alternative definition of state capacity from the International Country Risk Guide’s (ICRG) Bureaucratic Quality rating (Howell Citation2011) and obtain similar results. Therefore, we find no evidence of state capacity affecting the relationship between the WWII losses and COVID mortality.11. Educationalists have often emphasized the importance of direct experience to knowledge, separately to formal education; see, for example, the discussions in Boud et al. (Citation1993).12. The latter assumptions can be rationalized by imagining the special case in which u(τ,0) = u ̃(0)-c(τ) where c(τ) is an increasing convex cost function, although we do not need this separable structure.13. The second-order conditions are satisfied given that expected welfare across shock prospects is concave in τ.14. Our other assumptions so far cannot rule out a non-stationary process, implying that successive big shocks have larger and larger welfare effects, alternating positively and negatively. That can be considered an empirical question.15. We are not aware of a unified data source on the country-level losses during WW2. For example, one of the most reputable sources, the Human Mortality Database (Citation2021), produced by Max Plank Institute for Demographic Research, University of California Berkley, or Uppsala Conflict Data Program (UCDP Citation2021) lack mortality data for the WW2 period. We address the issue of precision of the WW2 loss estimates in Section 3.16. For example, losses for Balkan countries are imputed based on the losses of Yugoslavia.17. For post-Soviet countries, we used 1939 and 1937 USSR population census, respectively.18. While the human losses from WWII were huge, the GDP losses appear to have been modest for victors and are thought to have dissipated over 15–20 years for losers (Organski and Kugler Citation1977).19. There are other measures available in this source, though they tend to be highly correlated.20. The “Axis powers” formally took the name after the Tripartite Pact was signed by Germany, Italy, and Japan on 27 September 1940, in Berlin. The pact was subsequently joined by Hungary, Romania, and Bulgaria (Hill Citation2003).21. The Council for Mutual Economic Assistance (COMECON) was an economic organization from 1949 to 1991 under the leadership of the Soviet Union that comprised, among other countries, Albania, Bulgaria, Czechoslovakia, Hungary, Poland, Romania, and the Soviet Union (Kaser Citation1967).22. The moral distance between country c and country j is the mean over all dimensions d of the squared difference between the countries’ moral difference value I for dimension i weighted by the variance V of that dimension i as in MDcj≡∑i=1dIij−Iic2/(Vid).23. While a nonlinear relationship is suggested by Figure 1, we chose a more parsimonious linear regression. We did two tests on functional form. First, we included the squared value of WW2 mortality, but its coefficient was not significantly different from zero. Second, we tested a specification with the inverse hyperbolic sine transformation of deaths per million as a dependent variable, with the same transformation applied to WW2 deaths. This gave qualitatively similar results.24. Estimations of the cumulative COVID-19 infection rates on the same set of covariates produce no significant results. We also estimated specifications with other governance indicators from WGI dataset: Political Stability and Absence of Violence, Government Effectiveness, Regulatory Quality, Rule of Law, and Control of Corruption. None of these variables show significant coefficients in estimations. These results are available from the authors.25. Two countries have missing data, namely Greenland and Montenegro.26. Two countries have missing data, Greenland, and Tajikistan.27. The 1939 population shares gave similar results but were slightly less significant in the first-stage regressions.28. For example, with the set of controls in row (3) of Table 2 the Jewish population share had a coefficient of 1.17 with a robust standard error of 0.27; the overall F statistic was 11.56 (prob. <0.0000). The instrument on its own had a coefficient of 1.54 with s.e. = 0.34 and F = 26.32.29. Here we are invoking well-known results in econometrics; see, for example, Wooldridge (Citation2002, Section 6.2.1).30. Using OLS without controls the regression coefficients are −12.37 (S.E. = 4.47) and −12.80 (9.62) for civilian and military deaths respectively. The (robust) standard error for the difference is 10.03.31. An emerging literature that studies heroic actions and altruism during the war finds that the majority of heroic acts happened when the combatants defended their land (vs. being on the attack), e.g., Franco et al. (Citation2011). A possible explanation of this phenomenon could be that it is based on the evolutionary mechanism of protecting of close kin (Rusch and Stormer Citation2015).32. The added countries are Australia, Burundi, Brazil, Canada, China, Egypt, Ethiopia, Indonesia, India, Iran, Japan, Cambodia South Korea, Laos, Mexico, Myanmar, Mongolia, Malaysia, Nepal, New Zealand, Philippines, Papua New Guinea, Rwanda, Singapore, Thailand, United States of America, Vietnam, South Africa.33. For example, during WW2, Royal Nepalese Army fought on the Burmese front, and, at the same time, Nepalese soldiers fought as a part of British army (Cross and Gurung Citation2002).34. We define wave 1 of the pandemic as a period between February 1, 2020 until August 1, 2020. Wave 2 is a period between August 1, 2020 and December 1, 2020.35. There could be other confounding factors that affect the relationship we see in Figure 1. Drozdzewski et al. (Citation2019) suggest that wars might lead to a more collectivist culture, and Gorodnichenko and Roland (Citation2011) show that individualism led to better economic outcomes and innovations. There is also literature on the relationship between economic freedoms, liberal institutions and pandemics (e.g., Geloso, Hyde, and Murtazashvili Citation2022; Troesken Citation2015). These studies provide important directions for future research.36. This test identifies the sensitivity of our results to outliers but does not test the causality of the relation between the WWII and COVID total deaths.37. Our criterion for selecting observations is that: DFBETAi>2/√N, where N is the sample size.38. We did not include the health system efficiency control or the moral distance control since three countries were lost.39. The regressions based on error-prone regressors will yield inconsistent estimators not only for the variable measured with error but also for all model parameters (see, for example, Buonaccorsi Citation2010).40. The errors-in-variables regression is implemented in Stata eivreg routine (Lockwood and McCaffrey Citation2020).Additional informationNotes on contributorsMichael LokshinMichael Lokshin is a Lead Economist in the Office of Chief Economist, Europe and Central Asia Region of the World Bank. His research focuses on the areas of poverty and inequality measurement, labor economics, and applied econometrics. Michael holds a Masters in Physics from Moscow Institute of Physics and Technology and a Ph.D. in Economics from the University of North Carolina at Chapel Hill.Martin RavallionMartin Ravallion holds the inaugural Edmond D. Villani Chair of Economics at Georgetown University, prior to which he was the Director of the World Bank's research department. He has advised numerous governments and international agencies on poverty and policies for fighting it, and he has written extensively on this and other subjects in economics, including four books and 200 papers in scholarly journals and edited volumes. He is a past President of the Society for the Study of Economic Inequality, a Senior Fellow of the Bureau for Research in Economic Analysis of Development, a Research Associate of the National Bureau of Economic Research, USA, and a non-resident Fellow of the Center for Global Development.Vladimir KolchinVladimir Kolchin is an economist with the Poverty and Equity Practice, Europe and Central Asia Region of the World Bank. His research focuses on the areas of labor economics and applied econometrics. Vladimir holds a Ph.D. in Economics from Rutgers University, New Brunswick.