The BCG vaccine does not protect against COVID-19: Applying an economist’s toolkit to a medical question
As COVID-19 has spread across the globe, so has an intense search for treatments and vaccines, with numerous trials running in multiple countries. Several researchers (e.g. Hegarty et al. 2020, Miller et al. 2020, Ozdemir et al. 2020)1 and prominent news outlets2 have noticed that countries still administering an old vaccine against tuberculosis had fewer coronavirus cases and fewer deaths per capita in the outbreak’s early stages. But is that correlation evidence that the Bacillus Calmette-Guérin (BCG) vaccine provides some defence against COVID-19? In this column, we describe the findings of our recent work (Bluhm and Pinkovskiy 2020) looking at the incidence of coronavirus cases along the former border between East and West Germany, using modern econometric techniques to investigate whether historical differences in vaccination policies account for the lower level of infection in the East.
Tuberculosis has been largely eradicated from the developed world, and many rich countries have stopped using the BCG vaccine. Spain, which ended mandatory BCG vaccination in 1985, experienced more than 580 coronavirus deaths per million of its population and is one of the hardest-hit countries in per capita terms. Portugal, Spain’s neighbour, continues to perform mandatory BCG vaccination to this day and has experienced only 135 coronavirus deaths per million (all numbers as of 29 May 2020). More generally, studies have documented that countries with mandatory BCG vaccination tend to have substantially fewer coronavirus cases and fewer deaths per capita than countries without mandatory vaccination, and that the intensity of the epidemic is lower for countries that began vaccinating earlier (e.g. Hegarty et al. 2020, Miller et al. 2020, Ozdemir et al. 2020). Such cross-country correlations do not imply causation (Belloc et al. 2020) and sceptics have rushed to suggest that the passive immunisation from the BCG vaccine would at best be short-lived (Faust et al. 2020). The WHO now cautions that there is currently no evidence that the vaccine protects against the novel coronavirus (WHO 2020).
Because they are often unable to perform randomised experiments, economists have developed a suite of methods to exploit “natural experiments” that occur by chance as human action unfolds. We use one of these tools—a geographic regression discontinuity analysis (Dell 2010)—to determine whether coronavirus prevalence changes markedly at the Cold War-era border between East and West Germany, whose vaccination policies differed prior to their reunification in 1990. East Germany had mandatory BCG vaccination until reunification, while West Germany stopped voluntary but de facto universal vaccination in 1975.
Our setting presents a puzzle of its own. Many observers in Germany have noticed a much lower coronavirus prevalence in the former East than in the West but found few convincing explanations for the difference. Moreover, low mortality in Germany as a whole has been the subject of widespread speculation. We provide formal evidence that there is indeed a sizable discontinuity in COVID-19 cases at the border, flexibly controlling for everything continuous there. Other important variables—such as disposable income and average age—also jump at the border, but do not explain the discontinuity in COVID-19 cases. However, the difference in coronavirus prevalence is uniform across age groups rather than being dependent on whether individuals were born before or after the cessation of (near) universal vaccination in each part of the country. This fact cannot be explained by a broader immunity acquired via the BCG vaccine.
COVID-19 prevalence is much higher in former West than in East Germany
We start our analysis with a map of coronavirus cases by county as of 26 April, using data from the Robert Koch Institute. The former border between East and West Germany is outlined in red. The darker the shading of a county, the more coronavirus cases it has per million inhabitants. Several counties with high concentrations stand out (such as Heinsberg, bordering the Netherlands, and much of Bavaria – both places where the epidemic was first recorded). There is clearly a greater density of coronavirus cases around major cities (Berlin, Hamburg, and Stuttgart), as in the US. More importantly, we immediately see that counties just to the west of the former border are a much darker shade of blue than counties just to the east.
Figure 1 COVID-19 cases in Germany, 26 April 2020, log (1 + cases/million)
Notes: Illustration of the spatial distribution of COVID-19 cases in Germany as of 26 April 2020. The map shows the log (1 + cases/million people) in each county using population data from 2018.
To formalise the intuition of this map, we use a regression discontinuity design in which we nonparametrically estimate coronavirus prevalence as a function of the distance to the border and compare estimates from the eastern and western sides. We present a graphical representation of this exercise in the chart below, where the dependent variable is the logarithm of coronavirus cases per million inhabitants—a gauge of the exponential spread of the virus. The jump at the border is about 0.7 log points, which implies that there are half as many cases per capita in a former East German county than in a West German county just across the border. This halving of cases dominates the variation in coronavirus prevalence among counties in the East (where it is uniformly low) and is sizable relative to the average prevalence in the West. Far-away counties in Bavaria or near the border with France have the highest cases, but a smooth spread from there would imply a continuous decline, and cannot explain the precipitous drop at the border.
Figure 2 Discontinuity in log (1 + cases/million) at former border
Notes: Illustration of the discontinuity log (1 + cases/million people) across the former border between West and East Germany. The figure shows non-parametric local polynomial estimates for bins of the dependent variable, where each bin is 20 km wide; 95% confidence intervals are shaded in grey.
BCG vaccination is not driving this discontinuity
Our key piece of evidence makes the BCG hypothesis appear very unlikely. We exploit the fact that East Germany had mandatory vaccination from 1953 to 1990, while West Germany recommended vaccinating everyone from 1961 to 1975. If the vaccine protects against the virus, then we would expect that discontinuities in detected cases among people aged 15-34 (most of whom did not get the vaccine anywhere) would be small or non-existent. On the other hand, discontinuities among people aged 35-59 (all of whom were vaccinated in the East, as opposed to only those above age 45 in the West) should be larger. The chart below shows that the discontinuities in coronavirus prevalence for both age groups are stark and roughly identical in size, contradicting what we would expect to see if the BCG hypothesis were true. Moreover, for early April, we have more fine-grained data on cases for individual ages, allowing us to form age groups that directly correspond to the policy shift in BCG vaccination. They also show discontinuities that are equal or even smaller in the older cohort.
Figure 3 Age-specific discontinuity in log (1 + cases/million) at former border
Notes: Illustration of the discontinuity log (1 + cases/million people) across the former border between West and East Germany for different age groups. The left panel shows results for ages 15-34 and the right panel shows results for ages 35-59. Both figures show non-parametric local polynomial estimates for bins of the dependent variable, where each bin is 20 km wide; 95% confidence intervals are shaded in grey.
Alternative explanations: Commuter flows and the geography of the early outbreak
If the BCG vaccine does not explain the East-West differential in coronavirus cases, then what does? Historically, pandemics have been spread by the movement of people and trade from one area to another (Voth 2020), so in that spirit we investigate Germany’s commuting patterns. If those who live in the West work in the West and, in spite of a large wave of migration post-reunification, most former East Germans still work in the East, it may be the case that travel flows have not readjusted completely since reunification. In other words, Western counties along the former border may remain more disconnected from their Eastern neighbours than they would if a national border had never divided them. As the epidemic started in the West, it may have had a harder time spreading eastward because relatively fewer people commute between East and West than commute across comparable distances within the West or within the East. The eastward spread of the virus was then further interrupted by the nation-wide lockdown instituted on 22 March.
We simulate a canonical SIR model of the coronavirus epidemic in each German county, allowing infections to spread along commuting patterns starting from the distribution of coronavirus cases on 29 February (Bjørnstad and Grenfell 2008, Wesolowski et al. 2017). We find that in the simulated data, the number of cases also discontinuously declines as one crosses from West to East over the former border, with a decline smaller but close in magnitude to the decline observed in the actual data. Our methodology cannot exclude alternative explanations, and officially registered commuter flows likely do not represent person-to-person movement across Germany perfectly. However, our simulation constructs a situation that is consistent with the data on coronavirus prevalence and explains the drop in prevalence at the former East German border without any reference to the BCG hypothesis.
While it is disappointing to find evidence against a partial remedy, we believe that negative results are necessary to redeploy resources in the right direction. The BCG vaccine is already in low supply (Namkoong et al. 2020) and is an important tool in the fight against Tuberculosis—a lethal disease that killed 1.5 million people in 2018 according to the WHO. Our results also help guard against a false sense of security in countries with a current BCG vaccination policy, and show how the econometric toolkit can help address identification problems in current debates.
The views expressed in this post are those of the author and do not necessarily reflect the position of the Federal Reserve Bank of New York or the Federal Reserve System. Any errors or omissions are the responsibility of the author. A version of this post originally appeared in the Federal Reserve Bank of New York’s Liberty Street Economics blog as Richard Bluhm and Maxim Pinkovskiy, “Does the BCG Vaccine Protect Against Coronavirus? Applying an Economist’s Toolkit to a Medical Question”, Federal Reserve Bank of New York Liberty Street Economics, May 11, 2020.
Wesolowski, A, E zu Erbach-Schoenberg, A J Tatem, C Lourenço, C Viboud, V Charu and N Eagle et al. (2017), “Multinational patterns of seasonal asymmetry in human movement influence infectious disease dynamics”, Nature Communications 8(1): 1-9.