The Bacillus Calmette–Guérin (BCG) vaccine for tuberculosis (TB) is widely used globally. Many high-income countries discontinued nationwide vaccination policies starting in the 1980s as the TB prevalence decreased. However, there is continued scientific interest in whether the general childhood immunity boost conferred by the BCG vaccination impacts adult health and mortality in low-TB contexts (known as nonspecific effects). While recent studies have found evidence of an association between BCG vaccination and survival to ages 34–45, it is unclear whether these associations are causal or driven by the unobserved characteristics of those who chose to voluntarily vaccinate. We use the abrupt discontinuation of mandatory BCG vaccination in Sweden in 1975 as a natural experiment to estimate the causal nonspecific effect of the BCG vaccine on cohort survival to midlife. Applying two complementary study designs, we find no evidence that survival to age 40 was affected by the discontinuation of childhood BCG vaccination. The results are consistent among both males and females and are robust to several sensitivity tests. Overall, despite prior correlational studies suggesting large nonspecific effects, we do not find any population-level evidence for a nonspecific effect of the BCG vaccine discontinuation on survival to age 40 in Sweden.
Tuberculosis (TB) was a major cause of mortality in Europe in the eighteenth century and is estimated to have caused 1,000 deaths per 100,000 people (Murray 2001). The Bacillus Calmette–Guérin (BCG) vaccine was developed in the 1910s to address the global TB burden and was administered to the first child in 1921 (Dara et al. 2014). BCG vaccination was then scaled up widely in the 1940s across Europe. It became the main government strategy for containing TB, and many high-income countries introduced national policies making BCG vaccination compulsory for children (Zwerling et al. 2011). As the incidence of TB declined to negligible levels, several high-income countries discontinued compulsory vaccinations. However, there has been continued and significant scientific interest in the nonspecific effects of the BCG vaccine and whether it increases survival beyond its effects on TB.
The proposed mechanism behind nonspecific BCG effects is that early BCG vaccination improves general childhood immunity, which may confer a lasting health impact as individuals age (Cirovic et al. 2020; Dockrell and Smith 2017). This hypothesis is consistent with life course theories in demography and population health that link early-life conditions and events with later-life morbidity and mortality (Case and Paxson 2010; Cutler and Miller 2005; Elo and Preston 1992; Helgertz and Bengtsson 2019; Masters 2018; McEnry and Palloni 2010; Palloni and Souza 2013; van den Berg et al. 2011; Wen and Gu 2011; Zhao et al. 2021). Despite the growing scientific interest over the past two decades, especially among laboratory scientists studying cellular-level immune response to the vaccine (Dockrell and Smith 2017; Goodridge et al. 2016), it is still inconclusive whether the early childhood improvement in immune response induced by the BCG vaccine has a causal impact on health in later life.
There exists a body of literature analyzing the relationship between the BCG vaccine and morbidity and mortality among adults (Giamarellos-Bourboulis et al. 2020; Kölmel et al. 2005; Lankes et al. 2009; Pfahlberg et al. 2002; Rieckmann et al. 2019; Rieckmann, Villumsen, Jensen et al. 2017; Rieckmann, Villumsen, Sørup et al. 2017; Villumsen et al. 2009; Villumsen et al. 2013; Wardhana et al. 2011). While the nonspecific effects of the BCG vaccination may be most relevant for the elderly, a recent study suggested that children vaccinated with the BCG vaccine in Denmark had a 42% lower hazard of mortality to midlife (ages 34–45) (Rieckmann, Villumsen, Sørup et al. 2017). If this association was in fact evidence of a causal relationship, these results would have significant population policy implications and would suggest that reintroducing BCG vaccination in national vaccination plans, even in the absence of TB, may be a cost-effective and easy-to-implement way of improving survival to the midlife ages. However, these studies are based on small, nonrepresentative samples or limited to specific health outcomes rather than an overall measure like all-cause mortality, and most employ study designs that suffer from confounding. Thus, it is unclear whether the association between BCG and morbidity and mortality in adulthood is due to the causal effect of improved childhood immunity from vaccination or reflects the unobserved characteristics of the families and children that opted for the vaccine. To our knowledge, there is no causal evidence for the nonspecific effects of the vaccine on survival to midlife at the population level.
In this article, we address these three key limitations. First, rather than rely on small or potentially nonrepresentative samples, we use population data on births and deaths in Sweden for birth cohorts born between 1950 and 1981. Second, rather than focus on specific health conditions, we examine the effects of BCG vaccination on cohort survival to age 40 based on all-cause mortality. Lastly, we employ two causal inference study designs—the regression discontinuity design and the multiperiod difference-in-differences model—to address the potential issues of confounding seen in prior studies. In contrast to earlier studies, we do not find that the BCG vaccine had nonspecific effects on survival to age 40 in Sweden. This result is consistent across analysis approaches and robust to several sensitivity analyses and alternate specifications. While BCG vaccination in the absence of TB might still be important for certain specific subpopulations, our results do not provide evidence of nonspecific effects on survival to midlife in high-income countries with a low burden of TB.
Epidemiology of Tuberculosis
TB is an infectious disease caused by the Mycobacterium tuberculosis bacterium that mostly affects the lungs but can also spread to other organs and the bones (Raviglione 2018). TB is highly fatal if untreated but can be successfully treated with a six-month antibiotic drug regimen. It is transmitted through the air (e.g., from talking or coughing) and is mainly observed among adults, males, and those who are immunocompromised (e.g., living with HIV), live with diabetes, are undernourished, or use alcohol or tobacco (Lönnroth et al. 2009).
Mycobacterium tuberculosis has been around for thousands of years, but tuberculosis prevalence increased to an epidemic level only in the eighteenth century, reaching an estimated death rate of 1,000 per 100,000 inhabitants in 1800 (Barberis et al. 2017; Chisholm et al. 2016; Hershkovitz et al. 2015; Murray 2001). This dramatic increase in TB mortality is thought to be related to poor living standards and increasing urbanization and the close aggregation of individuals in cities (Lönnroth et al. 2009). In the nineteenth century, the TB prevalence declined in high-income countries (Wilson 1990). While TB remains the second highest cause of death from infectious diseases globally, after COVID-19, in 2020, this burden is concentrated in low- and middle-income countries, with high-income countries accounting for only 2% of all recorded TB cases (World Health Organization 2021).
Development and Introduction of the BCG Vaccine in Europe
The BCG vaccine is a live vaccine and was introduced in the 1920s (Dara et al. 2014). Since then, it has become one of the most widely used vaccines globally (Dara et al. 2014; Muhoza et al. 2021). Several strains of the BCG vaccine emerged from the original strain isolated by Albert Calmette and Camille Guérin in 1908 (Dara et al. 2014). There is conflicting evidence on the effectiveness of the vaccine and, to date, it is poorly understood how the BCG vaccine affects the immune system and why there is such a wide variation in the body's immunological response (Dockrell and Smith 2017). While evidence is mixed, it is generally believed that the BCG vaccine protects only against the development of active TB but not infection itself (World Health Organization 2021) and that the vaccine has the greatest benefits when administered after birth rather than to adolescents or adults (Abubakar et al. 2013). The BCG vaccine is safe for healthy children but poses a threat to immunodeficient children, who can develop a disseminated BCG infection (World Health Organization 2018). This fact is important for deciding which children should be given the BCG vaccination, a point we will return to later when discussing potential issues of confounding in prior research.
In Europe, the BCG vaccine was widely adopted in the 1940s (Dara et al. 2014), and in the large majority of countries it was administered right after birth (Zwerling et al. 2011). While some countries, such as Sweden and Norway, made the BCG vaccination compulsory, other countries adopted policies that targeted only high-risk groups, such as health care workers or children of parents born in high-incidence countries. In the decades after the introduction of the BCG vaccine, the TB incidence continued to decrease dramatically in Western Europe and other high-income countries (Daniel 2006). This decline led most of the governments of countries where vaccination was compulsory to discontinue their policies (Böttiger et al. 1982). In Sweden, our study country, TB mortality decreased from around 22 per 100,000 in 1950 to less than 4 per 100,000 in 1975 (Sakari Härö 1994; U.S. Senate Committee on Government Operations 1959). In Norway, which we use as a control in one of our analyses, mortality attributable to TB was 29 per 100,000 people in 1950 and decreased to 2 per 100,000 in 1975.1 Importantly, it remains unclear how much the BCG vaccine contributed to the decline in TB prevalence, or whether this decline was mainly due to improved living standards and nutrition (Wilson 1990).
In Sweden, compulsory vaccination of neonates was introduced in 1940 (Public Health Agency of Sweden 2020) and then reversed in April 1975 because of the persistently low TB prevalence (Bergström et al. 2001; Zwerling et al. 2011). Thus, children born in 1976 were the first birth cohort in which no newborn received the routine BCG vaccine. This shift was abrupt: vaccine coverage dropped from 95% to 2% in the years following its discontinuation (Romanus 2006). Coverage then gradually increased to 13.1% in the cohort born in 1989 and focused on at-risk groups such as children born to parents from countries with high TB incidence (Bergström et al. 2001). In Norway, BCG vaccinations became compulsory in 1947 for students leaving elementary school, usually at age 13 (Waaler et al. 1971). Mass vaccination campaigns started in 1947 (Tverdal and Funnemark 1988). However, unlike Sweden, mandatory vaccination in Norway remained in place until 2005, when the policy changed to a voluntary scheme; in 2009, the vaccination recommendation was limited to at-risk groups (Norwegian Institute of Public Health 2008, 2016). BCG vaccine coverage in Norway increased from 88.6% among adolescents aged 13 in 1959 to 92.2% in 1973; coverage has been above 97% since 1975 (Norwegian Institute of Public Health n.d.; Tverdal and Funnemark 1988).
Nonspecific Effects of the BCG Vaccine
Nonspecific effects are health effects of a vaccine that are unrelated to the targeted disease. The first nonspecific effects were reported for the vaccinia vaccine against smallpox, as the vaccine also seemed to protect against measles and other infectious diseases (Mayr 2004). These findings led to the assumption that other live vaccines, such as the BCG vaccine, might have positive nonspecific effects (Higgins et al. 2016). It is not entirely clear through which channels the BCG vaccine could provide protection against other diseases. There is evidence that the BCG vaccine causes an adaptation of the innate, or nonspecific, immunity, which is the fraction of the immune system humans have from birth on and is, thus, not induced in response to specific diseases through infection or vaccination (Alberts et al. 2015). This trained immunity, that is, the change in the innate immunity following BCG vaccination, is one of the potential channels for related nonspecific effects (Moorlag et al. 2019). This physiological mechanism is consistent with a broader body of evidence from demography and population health that finds a robust link between early childhood health and changes to the immune system and what happens in later life (Bonanni 1999; Drummond et al. 2007; Karafillakis et al. 2015; Mercer 1985; Michiels et al. 2011).
That the BCG vaccine's nonspecific effects improve health through altering the trained immunity is also in line with previous research on the hygiene or “old friends” hypotheses (Cepon-Robins and Gildner 2020; Garn et al. 2021; Lambrecht and Hammad 2017; Renz et al. 2017; Rook 2012; Rook et al. 2017; Stiemsma et al. 2015). These hypotheses postulate that microbial exposure in early life regulates inflammation and, thus, improves the immune system's response to infection over the life course. This leads to better protection against allergies, autoimmune diseases (such as rheumatoid arthritis, inflammatory bowel disease, and diabetes mellitus type 1), and even psychiatric disorders (Rook 2012; Rook et al. 2013). However, there is also extensive evidence that stimulation of the trained immunity increases inflammation, which in turn leads to a higher risk of cardiovascular disease, cancer, and chronic obstructive pulmonary disease (Accardi and Caruso 2018; Badii et al. 2022; Emerging Risk Factors Collaboration 2010; Furman et al. 2019; Gan et al. 2004; Hajishengallis et al. 2019; Netea et al. 2020; Reuter et al. 2010). Inflammation is also associated with a higher all-cause mortality (Maluf et al. 2020; Morrisette-Thomas et al. 2014; Proctor et al. 2015; Suemoto et al. 2017). Today, the BCG vaccine is widely applied as a treatment for bladder cancer, and research on its effectiveness in the treatment of other cancers is ongoing (Lamm and Morales 2021).
Evidence on Nonspecific Effects of the BCG Vaccine
A number of studies, most of them observational, estimate nonspecific effects of the BCG vaccine on child survival (Biering-Sørensen et al. 2018; Breiman et al. 2004; de Castro et al. 2015; Elguero et al. 2005; Garly et al. 2003; Lehmann et al. 2005; Nankabirwa et al. 2015; Roth et al. 2005; Schaltz-Buchholzer et al. 2021; Schaltz-Buchholzer et al. 2022; Steenhuis et al. 2008; Thysen et al. 2020; Vaugelade et al. 2004). The majority were conducted in low- and middle-income countries and suggest a positive association between BCG vaccine and child survival. Evidence on the vaccine's nonspecific effect on adult morbidity and mortality has important limitations (Giamarellos-Bourboulis et al. 2020; Kölmel et al. 2005; Lankes et al. 2009; Pfahlberg et al. 2002; Rieckmann et al. 2019; Rieckmann, Villumsen, Jensen et al. 2017; Rieckmann, Villumsen, Sørup et al. 2017; Villumsen et al. 2013; Villumsen et al. 2009; Wardhana et al. 2011). All studies but one (Rieckmann, Villumsen, Sørup et al. 2017) are based on small, nonrepresentative samples or are limited to specific health outcomes. Most importantly, all studies except for one clinical trial (Giamarellos-Bourboulis et al. 2020) do not sufficiently account for potential confounding and can thus only estimate associations but not establish causal relationships.
For example, Pfahlberg et al. (2002) studied the relationship between BCG vaccination and malignant melanoma among 1,230 individuals from 11 health centers in seven European countries between 1994 and 1997. Using a case–control design, they compared those who did and did not receive the BCG vaccine and concluded that the vaccine might reduce the risk of malignant melanoma. In contrast, Rieckmann et al. (2019) linked data on 5,090 adults from the Copenhagen School Health Records to the Danish Cancer Registry and found no association between BCG vaccination and malignant melanoma. Focusing on lymphoma and leukemia, Villumsen et al. (2009) also used data from the Copenhagen School Health Records and found that individuals who had received the BCG vaccine had half the risk of developing lymphoma than those who had not. There was no difference in the risk of developing leukemia.
Two observational studies analyzed the association between BCG vaccination and adult survival. Kölmel et al. (2005) examined the association of the vaccinia and BCG vaccines with survival among individuals with malignant melanoma in six European countries and Israel. The authors compared participants with the vaccine to those without and found no evidence of an association between the BCG vaccine and survival among individuals with malignant melanoma. The only study analyzing the relationship between the BCG vaccination and survival to midlife used data from the Copenhagen School Health Records. Rieckmann, Villumsen, Sørup et al. (2017) linked these data to the Civil Registration System and the Danish Register of Causes of Death databases, which focus on death by natural causes (excluding tuberculosis or smallpox). In comparing those who did receive the BCG vaccine with those who did not, they found that vaccinated individuals had approximately half the risk of dying from natural causes as unvaccinated individuals.
While some of these studies suggest that the BCG vaccine has nonspecific effects on morbidity and mortality, none of the study designs can establish a causal relationship2 because they do not sufficiently account for confounding due to factors such as the healthy vaccinee bias (Schaltz-Buchholzer et al. 2022). The healthy vaccinee bias occurs if healthy children, and thus children with a higher overall probability of survival, are more likely to receive the vaccine than unhealthy children. The decision to not administer the vaccine can be based on concerns regarding side effects or because parents have lower health service uptake in general. Other factors that may influence parents' decision to vaccinate their children that are also linked to children's health over the life course are parental education and socioeconomic status. However, the relationship between these and vaccine hesitancy is unclear and could generate bias in both directions (Larson et al. 2014). Furthermore, access to health services may confound the association. In Sweden, universal health coverage was introduced in 1955, and over the following decades, health services were increasingly provided exclusively by public providers (Savedoff and Smith 2011). Thus, children living in underserved areas during this transition might have had less access to the BCG vaccine and other health services. If the vaccine was not offered free of charge, this might have discouraged parents from vaccinating their children (Larson et al. 2014).
Comparing vaccinated to unvaccinated individuals without considering systematic differences in characteristics that are correlated with the decision to vaccinate can result in an over- or underestimation of the vaccine's true nonspecific effects. We address this knowledge gap by applying quasi-experimental methods that take advantage of the sudden discontinuation of compulsory vaccination to estimate the causal effect of BCG vaccination on survival to midlife.
Data and Methods
We screened all available information on the BCG Atlas to identify countries that would be eligible for inclusion in our study (Zwerling et al. 2011). First, we identified all countries that had a change in vaccination policy, moving from compulsory to optional vaccination, as potential natural experiments. Next, we screened the documentation provided on each of these countries in the Human Mortality Database (HMD) as of January 19, 2022, to identify which of these countries have available data for all years in our study period (HMD 2023). We identified seven potential countries with a change in vaccination policy that also had data available in the HMD (Austria, Denmark, Finland, France, Great Britain, Norway, and Sweden). We chose Sweden as most suitable for the following reasons: (1) there was a sharp discontinuation of compulsory vaccinations in 1975, and BCG vaccine coverage dropped from 95% before 1975 to 2% in the years following the discontinuation (Romanus 2006); and (2) Sweden was the first country to abolish a nationwide policy and discontinue the BCG vaccine. Thus, Sweden allows for the maximum observation period in terms of survival to midlife.
One of our study designs, the multiperiod difference-in-differences (DiD), additionally requires selecting a control country that followed a similar survival trajectory prior to the discontinuation of BCG vaccination in Sweden but that did not discontinue the vaccine at the same or similar time as Sweden. From this criterion, we selected Norway as the most suitable counterfactual country for the DiD analyses.3 For both Sweden and Norway, we accessed data on the number of births and age-specific death rates for each birth cohort born from 1846 to 1981 from the HMD. To minimize confounding from the aftermath of World War II, we included only birth cohorts born in 1950 or later. Thus, our observation period covers cohorts born from 1950 to 1981, which allows us to study the impact of BCG vaccination on the cohort probability of survival to age 40 (for those born in the final 1981 birth cohort, we have data on their survival up to the year 2021, when cohort members turned 40).
Unfortunately, we were not able to use Swedish register data, which is commonly used in demographic analyses of the Swedish population (Barclay and Kolk 2018; Bucher-Koenen et al. 2020; Helgertz and Bengtsson 2019; Lazuka 2019; Oudin Åström et al. 2016), for the following reasons: (1) our analysis requires survival data for cohorts born several years before the BCG discontinuation in 1975, but the Demographic Data Base includes data only from 1973 and later; and (2) it is not currently possible to link the Demographic Data Base to BCG vaccination data.
Our outcome of interest is the probability of survival to age 40 in each birth cohort. Constructing this outcome requires information on the number of individuals who survived to age 40 in each cohort. The HMD data include cohort age-specific death rates but not the total number of deaths that occurred in a birth cohort by age. Therefore, we had to convert the age-specific death rates to death counts. We did this through the following procedure. First, we converted the cohort age-specific mortality rates to the age-specific probabilities of survival using the standard life table approach (Preston et al. 2001). Second, we generated a dataset with the number of observations equal to the number of births in a cohort. We then generated a variable for “alive at age 40” and used the cohort life table probability of survival from birth to age 40 to determine what share of this initial birth cohort survived to age 40, and correspondingly assigned values to the “alive at age 40” variable. For example, if among the 115,414 births for the 1950 birth cohort, the cohort–life table probability of survival to age 40 was 90%, we would assign values of 1 for “alive at age 40” to 103,873 observations (.90 × 115,414). This procedure was repeated for each birth cohort, resulting in a final dataset of 3,456,602 observations, which equals the total number of births in Sweden from 1950 to 1981. We repeated this approach to generate individual-level datasets for Norway as well as separately for males and females in each country.
We used two complementary quasi-experimental designs to assess the effect of the sharp BCG discontinuation in Sweden on cohort survival to age 40: the regression discontinuity design (RDD) and the multiperiod DiD approach. The RDD estimates the causal effect of an intervention or policy in situations where exposure to the intervention or policy is based on whether individuals fall on one side or another of some type of cutoff. This cutoff point is usually referred to as the “threshold,” and the continuous variable on which the cutoff is based is known as the “running variable.” For example, individuals may be eligible for free health care (the policy) depending on whether their income (the running variable) is below the poverty line (the threshold). In such circumstances, the RDD design estimates the causal effect of the policy under the assumption that in the absence of the treatment, the relationship between the running variable and outcome would have evolved continuously. Heuristically, this can be expressed in the following way: those just below the threshold are, on average, identical in all characteristics to those just above the threshold. The only difference being that those above the threshold receive the treatment (in our study, this would be not receiving BCG vaccination). Importantly, the RDD estimates the local average treatment effect (LATE) at the threshold value.
For our analysis, we followed the continuity-based RDD approach described by Cattaneo et al. (2019). The first assumption of an RDD is that the probability of being treated changes discontinuously at the specified threshold. In our analysis, the running variable is year of birth and the threshold is 1976, the birth year of the first cohort for which the BCG vaccine was no longer mandatory. There are no yearly data on BCG coverage for the years around the threshold. However, Romanus et al. (2006) reported that BCG vaccine coverage dropped from 95% to 2% in the five years following the discontinuation, which shows that the probability of receiving the BCG vaccine decreased sharply at the threshold. A second assumption is that the outcome would be continuous at the threshold in the absence of the treatment. In other words, there was no substantial change in the probability of survival to age 40 induced by an event other than the BCG vaccination discontinuation. To our knowledge, there have been no major shifts in policies or national-level events, such as natural disasters, that could have affected the probability of survival of the cohorts born just before the discontinuation from those born just after. Finally, for the RDD approach to be valid, individuals should not systematically sort around the threshold with the aim to obtain or avoid the treatment. We assess this risk to be negligible in the context of our study, as we consider it unlikely that parents accurately anticipated and based their family planning on the future BCG policy change. This is supported by the data, which shows no clustering of births either above or below the threshold (see Figure S1 in the online appendix).
We followed the current best practice for RDD estimation (Cattaneo et al. 2019). This includes using a local linear model (to prevent overfitting the data) and triangular kernel weights (to give more weight to observations closer to the threshold). The bandwidth was calculated in a data-driven way for identifying the mean square error (MSE) optimal bandwidth around the threshold (to remove arbitrary bandwidth selection and potential manipulation of the bandwidth size by researchers) (Imbens and Kalyanaraman 2012). From our data, the estimated bandwidths were ±4 birth cohorts for the total population and for the male and female subpopulations. For inference, we used robust bias-corrected standard errors and confidence intervals (Cattaneo et al. 2019). Our estimation model is
Here Yi is the probability of survival to age 40 for newborn i, Xi is a continuous measure of year of birth (the running variable), and Di is a binary indicator that takes the value of 1 if a birth cohort was born after the discontinuation of mandatory BCG vaccination (Cunningham 2021). is the causal effect of the discontinuation of the BCG vaccine on survival to age 40 for the cohort born in the threshold year, that is, 1976.
While the RDD estimates the effect of the BCG discontinuation for the cohort born in the threshold year, it could be the case that the effect emerges only after several birth cohorts (e.g., if it took a few birth cohorts to fully discontinue the vaccine). Although unlikely, it could also be the case that another shock unrelated to BCG discontinuation occurred over the same period of the discontinuation. To investigate the effect of BCG vaccination in the case of these two scenarios, we conducted a multiperiod DiD model (sometimes referred to as a leads and lags model) with Norway as the comparison country.
The multiperiod DiD estimates the effect of the vaccine discontinuation by comparing the probability of survival in Sweden and in Norway in each birth cohort, accounting for temporal trends (i.e., the general positive trend in the probability of survival), as well as for the baseline difference between the two countries. The central identifying assumption of the multiperiod DiD is that in the absence of the BCG vaccination discontinuation, the cohort probability of survival to age 40 in Sweden would follow the same trend as the cohort probability of survival to age 40 in Norway (the so-called parallel trends assumption). This assumption is commonly assessed by comparing the trends in the outcome variable in the pretreatment period. While we formally assess this assumption later, the visual evidence presented in Figure 1 shows that among several potential comparison countries, the cohort mortality trends in Norway most closely mirrored those in Sweden. By estimating the effect of BCG discontinuation on survival for multiple birth cohorts born after the discontinuation—and not just the 1976 birth cohort as in the RDD—this approach reveals whether the effect wanes or intensifies over time or whether there is a lag in the effect and only cohorts born some years after the discontinuation are affected.
The multiperiod DiD regression model is set up as
with i indexing individuals, c indexing countries, and t indexing years (Cunningham 2021). µc is a country fixed effect that accounts for the baseline difference in the probability of survival between the treatment (Sweden) and control (Norway) countries. λt is a vector of year fixed effects that accounts for the time trend common to both countries (1975, the year of the BCG vaccine discontinuation, is the reference year). µcλt is a vector of the interaction between the country and year dummy variables, and its coefficients can be interpreted as the treatment effect for each of the birth cohorts. Importantly, the coefficients on the country–year interaction terms for cohorts born before 1975 provide a test of the parallel trends assumption (these coefficients should be close to 0). Lastly, rather than estimate treatment effects for each birth cohort born after the BCG discontinuation, we can also estimate a common treatment effect for all birth cohorts by interacting the country dummy variable with a post dummy variable (=1 for all cohorts born after the year of discontinuation) and including year fixed effects as individual dummy variables (Angrist and Pischke 2008):
The final data include one observation per person born from 1950 to 1981. For Sweden, this comprises 3,456,602 people in total (1,779,913 males and 1,676,689 females), ranging from 93,248 in the cohort born in 1978 to 123,354 in the 1966 cohort. In Norway, 1,959,730 people were born in this period (1,008,178 males and 951,552 females), ranging from 50,708 in 1981 to 67,746 in 1969. Figure 2 shows the probability of survival to age 40 in Sweden. Survival to age 40 increased from 95.0% in the 1950 birth cohort to 97.8% in the 1981 birth cohort. Across all years, the probability of survival was higher among females than males. For example, in 1975 the probability of survival to age 40 was 98.2% among females and 97.1% among males. Despite this difference in the absolute level, survival in the male and female subpopulations followed a similar trend. Visually, we observe no evidence of an abrupt change in the probability of survival in the whole population when comparing the cohort born in the year of the BCG vaccine discontinuation (97.6% in 1975) and the first cohort for whom the vaccine was optional (97.7% in 1976). This smooth trend around 1975 is also seen when separately looking at the male and female subpopulations.
Regression Discontinuity Design Results
The RDD analysis confirms the results of the visual inspection: we do not find evidence of an effect of the BCG vaccination discontinuation on survival to age 40 in the birth cohort born in 1976. We estimate a precise null effect (−0.03 percentage points (hereafter “pp”); 95% CI, −0.44 to 0.16) on the probability of survival to age 40 at the point of BCG vaccination discontinuation (Table 1 and Figure 3). As with the visual results, this finding was not driven by just males or females: we also find no evidence that discontinuation affected the probability of survival to age 40 among the male (−0.03 pp; CI, −0.52 to 0.38) or female (−0.03 pp; CI, −0.59 to 0.18) subpopulations.
Multiperiod Difference-in-Differences Results
While the RDD estimates potential nonspecific effects for the cohort born in the year of the vaccine discontinuation, the DiD allows us to generate estimates for each birth cohort and, thus, yields the opportunity to detect lagged effects. The event-study plot in Figure 4 illustrates the difference in survival between birth cohorts born in Norway and Sweden for each cohort born before and after the BCG discontinuation. The coefficients are shown net of an overall level difference in survival between the two countries and a time trend. In support of the parallel trends assumption, we find that most point estimates in the prediscontinuation period are not statistically different from 0 (exact values are shown in online appendix Table S1). This suggests that prior to the discontinuation of BCG vaccination in Sweden, both countries were following a parallel trend in cohort survival to age 40. The coefficients after the dashed vertical line in 1975 capture any effect of BCG discontinuation on survival in Sweden. Similar to our RDD results, we find no evidence of differences in cohort survival to age 40 in the post-BCG-discontinuation period between individuals born in Sweden and those born in Norway across most birth cohorts.
By splitting the estimations by birth cohort, the event-study may result in underpowered estimates of the effect of BCG discontinuation. When pooling all data from the pre- and postdiscontinuation periods, respectively, we also find no evidence of an effect of the BCG vaccine on the probability of survival to age 40 (−0.01 pp; CI, −0.09 to 0.07; see Table 2).
We draw a similar set of conclusions when examining males and females separately. Among both subpopulations, there was no difference in the probability of survival for most of the cohorts born before the BCG vaccine discontinuation, supporting the parallel trends assumption (see Figure 4). In the female subpopulation, there was also no difference in the probability of survival between Sweden and Norway after the vaccine discontinuation (net of the time and level trends). This finding is confirmed in the pooled analysis (0.02 pp; CI, −0.09 to 0.12; see Table 2). In the male subpopulation, Figure 4 shows negative, statistically significant point estimates for three of the six postdiscontinuation periods (1976, 1978, 1980), indicating that there might have been a nonspecific effect of the BCG vaccine on survival to age 40. However, the point estimates are marginal and, thus, of limited practical relevance. Pooling all pre- and postdiscontinuation birth cohorts reveals that the probability of survival did not statistically differ from 0 (−0.03 pp; CI, −0.16 to 0.10; see Table 2) among males.
To be consistent with the RDD and include the same birth cohorts in both approaches, we restricted our sample to the cohorts born within the MSE optional bandwidth calculated in the RDD analysis, which was ±4 years in the total population and the male and female subpopulations (Table 2). The results show that for the birth cohorts born within the MSE optimal bandwidth, there was no difference in the probability of survival between the total, male, and female populations in Sweden and Norway.
We conducted several sensitivity analyses to test the robustness of our results. First, it could be the case that the rate of change of survival across cohorts depends on the absolute level of survival and that this functional form affects our estimates. To investigate this, we plotted the log probability of survival and found that flattening of the cohort survival trends was not driven by the fact that the probability of survival approached 100% (see online appendix Figure S2). Second, the RDD estimates whether there was a sudden change in the level of the probability of survival after the BCG vaccine discontinuation. It could, however, also be the case that the discontinuation slowed down the rate of change in the increase of the probability of survival to age 40. To investigate this, we applied a regression kink design (RKD). The RKD relies on the same assumptions as the RDD but is applied when there is a discontinuity not in the probability of being treated but in its first derivative, that is, in the slope of the relationship between the treatment and the running variable. In the context of our analysis, this would be the case if the probability of receiving the vaccine was still high in 1975, the year of the discontinuation, and only gradually decreased in the following years. Another potential scenario could be that the probability of being vaccinated already declined before the official discontinuation. Rather than causing a shift in the probability of survival, the gradual discontinuation would cause a change in its slope. In the RKD analysis, we also found no evidence of an effect of the BCG vaccine discontinuation on the rate of change in the probability of survival in the total, male, and female populations (Table 3). Third, to address this same issue and the fact that more individuals born in 1975 were born after the discontinuation (Romanus 1983), we set the cutoff to be one year earlier (1975 in the RDD and 1974 in the multiperiod DiD), which produced results consistent with those of the main analysis (see Tables S2–S4 and Figure S3 in the online appendix). Fourth, in the RDD analysis, the point estimate and inference might depend on the bandwidth within which the LATE is estimated. Increasing the bandwidth will reduce variance as more observations are included in the analysis. However, it increases the bias because these observations are less similar to each other than those closer to the cutoff. Decreasing the bandwidth would have the opposite effect and increase variance while reducing the bias. We increased and decreased the bandwidth by one to three years in the total population and in the female and male subpopulations. Across all alternative bandwidths, the point estimates remained relatively stable and statistically insignificant (see online appendix Table S5).
There is continued scientific interest in the nonspecific effect of the BCG vaccine (Goodridge et al. 2016). If the BCG vaccine had nonspecific effects and increased adult survival among the general population, it could be a simple and cost-effective intervention to improve population health. Mechanistically, such an effect seems plausible since the BCG vaccine is thought to improve general childhood immunity (Cirovic et al. 2020). Other studies focused on other early-life conditions also find evidence of life course consequences to changes in early-life immunity and health (Almond, Currie, and Duque 2018; Almond and Mazumder 2005). Using population data and two complementary quasi-experimental methods that take advantage of the sudden discontinuation of the mandatory BCG vaccination in Sweden, we find no evidence for nonspecific effects of the BCG vaccine on the probability of survival to age 40. In other words, up to this age and at the population level, the BCG vaccine does not appear to have substantial protective effects on mortality from diseases other than TB. Importantly, our null findings are not a by-product of low statistical power. We had data with one observation for each individual born between 1950 and 1981, resulting in a sample of 3,456,602 individuals. Thus, we were able to estimate highly precise nulls with confidence interval widths of just 0.6 to 0.9 pp. This result is robust to multiple sensitivity and robustness checks and remains valid when estimating the effect separately for the male and female subpopulations. One exception is that in the DiD analysis, some male birth cohorts have statistically significant point estimates in the postdiscontinuation period. However, this pattern is inconsistent and the point estimates are less than 0.6 pp.
Our results are in stark contrast to a previous study from Copenhagen, Denmark, which found that BCG vaccination was associated with an approximately 40% lower probability of dying (Rieckmann, Villumsen, Sørup et al. 2017). The differences between our studies are likely due to residual confounding. In particular, those authors compared children who did and did not receive the vaccination during a period in Denmark when BCG vaccination became voluntary. Thus, parents self-selected their children into vaccination, and it is likely that the decision to vaccinate was related to characteristics that also influenced the probability of survival. Such factors could include parental knowledge about the health of their child (as mentioned previously, BCG vaccination can have strong side effects for less healthy children) and the background and socioeconomic status of parents who decide to vaccinate their children compared with those who do not. The authors undertook several steps to adjust for such potential sources of confounding; however, there is still a likelihood that the results were driven by unobserved systematic differences between the vaccinated and unvaccinated that affected both the vaccination status and survival, since vaccination status was related to a direct parental choice and not a mandatory law, as in our study. Another reason for the dramatic difference between our results and this earlier study could be that the authors included only natural causes of death, which may result in a larger point estimate than when including all causes of death. In Sweden, natural causes of death among individuals up to age 39 composed approximately 30–40% of all deaths between 1997 and 2021 (National Board of Health and Welfare 2019). Thus, even when using data on all-cause mortality for the outcome construction, we would expect to see nonspecific effects on natural causes of death reflected in the point estimate. Given that our point estimates are almost 0 (−0.03 pp) and our confidence intervals are precise, we conclude that the difference between our results and those of the study by Rieckmann, Villumsen, Sørup et al. (2017) are likely not caused by differences in the outcome definition.
One important consideration is that our results are for Sweden, a country with a near negligible burden of TB and where infectious disease mortality, excluding deaths from COVID-19, is extremely low (National Board of Health and Welfare 2019). The lack of a strong relationship between early childhood BCG vaccination and adult mortality in this context is not necessarily evidence that changes to early-life immunity do not affect adult mortality risk, but rather that such life course effects may become less important as the general level of infectious mortality in a population decreases. BCG vaccination may have beneficial nonspecific effects for individuals at elevated risk of mortality from infectious diseases, such as those in many low- and middle-income countries or for specific at-risk groups within low-TB countries. Evaluating the nonspecific effects of BCG vaccination among such populations is an important area of future work.
A second important consideration with RDD analyses is that the effect estimate of a treatment or policy applies only to those within the bandwidth around the threshold point. In our case, the RDD estimates the effect of BCG discontinuation on the first birth cohort to not receive the vaccine. If the BCG vaccine protected against infectious diseases, in particular during childhood as suggested by previous studies (Biering-Sørensen et al. 2018; Giamarellos-Bourboulis et al. 2020), there might be a time lag between the discontinuation of the vaccine and the realization of the nonspecific effects as younger cohorts become exposed to larger shares of unvaccinated individuals. In this case, cohorts born after infectious diseases have crossed a critical prevalence threshold would be affected by the BCG discontinuation, whereas those born during or just after the discontinuation would not. Another reason for lagged effects would be a gradual decrease in the BCG coverage over several birth cohorts rather than an abrupt change. However, both of these scenarios would be captured in our multiperiod DiD analysis. Therefore, the null results in both the RDD and the multiperiod DiD suggest that there were no immediate or lagged nonspecific effects of the BCG vaccine.
Our study has several limitations and considerations. If an exogenous factor, which we did not account for in the analysis, caused a decline (increase) in the probability of survival at the same time as the discontinuation of BCG vaccination, it could have offset potential beneficial (harmful) nonspecific effects of the BCG vaccine. To rule this out, we undertook several steps. First, in our sensitivity analysis, we varied the cutoff year. The conclusions of both quasi-experimental method analyses remain unaffected by this change. Second, we conducted an extensive search to identify potential large-scale changes in the health, economic, and political sectors in Sweden and Norway. We found no evidence of any changes in the living situation in the countries, which could have differentially affected survival at the population level in either country. Another aspect to consider is that the vaccinia vaccine was discontinued in Sweden in 1976, that is, shortly after the BCG vaccine (Public Health Agency of Sweden 2020). However, given the fact that the vaccinia vaccine was discontinued around the same time in Norway and that we also find no statistically significant effect in the multiperiod DiD analysis, we do not expect the vaccinia discontinuation to have biased our results (Norwegian Institute of Public Health 2010). Because of data constraints, we only estimated the effect of the BCG discontinuation on survival to age 40. Because mortality before these ages in Sweden is largely due to external and accidental causes of death rather than causes of death related to the immune system, we cannot exclude the possibility that BCG vaccination will have nonspecific effects at older ages, when infectious diseases are more detrimental. Rather, our results are best interpreted in the context of prior studies that do find potential evidence of nonspecific effects among younger, middle-aged adults.
Our study finds that the BCG vaccination did not have nonspecific effects on survival to age 40 in Sweden. Thus, our research suggests that reintroducing or continuing national mandatory BCG vaccination in countries with similar population mortality and morbidity profiles as Sweden is unlikely to lead to substantial or cost-effective improvements in population health up to midlife.
To our knowledge, no data on the TB-related mortality rates were available. Thus, we used the absolute numbers of TB cases and divided them by the population size in the respective year (Central Bureau of Statistics Norway 1961; Statistics Norway 2022; World Bank 2022).
Giamarellos-Bourboulis et al. (2020) conducted a clinical trial with 202 participants aged 65 years or older who received either the BCG vaccine or a placebo at hospital discharge. The trial showed that the short-term risk of infection was reduced among the treatment group. However, these results cannot be generalized to a broader population nor to survival to midlife.
If there were multiple potential counterfactual countries, we could have combined them using a synthetic control approach. This was our initial intention after identifying seven potential control countries: Austria, Finland, France, Great Britain, Japan, Norway, and Portugal. Following initial synthetic control analyses, we proceeded with the multiperiod DiD rather than the synthetic control approach as no country other than Norway had a pretrend that resembled Sweden’s, and hence the synthetic control approach gave nearly 100% of the weight to Norway.