Does early childhood BCG vaccination improve survival into adulthood in a population with a low tuberculosis prevalence? Quasi-experimental evidence on non-specific effects from 39 Swedish birth cohorts

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 or the bones (Raviglione Mario C., 2018). TB is highly fatal if untreated but can be 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 that are immunocompromised (e.g. living with HIV), living with diabetes, undernourished, and those that use alcohol or tobacco (Lönnroth, Jaramillo, Williams, Dye, & Raviglione, 2009).

The mycobacterium tuberculosis has been around for thousands of years but tuberculosis prevalence increased to an epidemic level only in the 18^th century, reaching an estimated death rate of 1,000 per 100,000 inhabitants in 1800 (Barberis, Bragazzi, Galluzzo, & Martini, 2017; Chisholm, Trauer, Curnoe, & Tanaka, 2016; Hershkovitz et al., 2015; Murray, 2001). This dramatic increase in TB mortality is thought to be related to increasing urbanization and the close agglomeration of individuals in cities as well as poor living standards (Lönnroth et al., 2009). In the 19^th 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 only accounting for 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 & Smith, 2017). While evidence is mixed, it is generally believed that the BCG vaccine only protects against the development of active TB but not infection itself, and that the vaccine has the greatest benefits when administered after birth rather than to adolescents or adults (World Health Organization; 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 which we will return to later on when discussing potential issues of confounding in prior research.

In Europe, the BCG vaccine was widely adopted in the 1940s (Dara et al., 2014). 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 high-risk groups only, such as healthcare 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 lead most of the countries where vaccination was compulsory to discontinue their policies (Böttiger, Romanus, Verdier, & Boman, 1982). In Sweden, for example, TB mortality decreased from around 22 per 100,000 in 1950 to less than 4 per 100,000 in 1975 (Committee on Government Operations, United States Congress Senate; Sakari Härö, 1994). In Norway, mortality attributable to TB was 29 per 100,000 people in 1950 and decreased to 2 per 100,000 in 1975 (Central Bureau of Statistics Norway, 1961).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, the country we study here, compulsory vaccination of neonates was introduced in 1940 and then reversed in April 1975 because of the persistently low TB prevalence (The Public Health Agency of Sweden, 2020; Bergström, Mäkinen, & Romanus, 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 covered at-risk groups such as children born to parents from high TB incidence countries (Bergström et al., 2001). In Norway, BCG vaccinations became compulsory in 1947 for students leaving elementary school, usually at age 13 (Waaler, Galtung, & Mordal, 1971). Mass vaccination campaigns started in 1947 (Tverdak & Funnemark, 1988). However, unlike Sweden, mandatory vaccination in Norway remained until 2005, when the policy changed to a voluntary scheme, and again in 2009, when the vaccination recommendation was limited to at risk groups only (Norwegian Institute of Public Health, 2008; Norwegian Institute of Public Health, 2016). BCG vaccine coverage in Norway increased from 88.6% among adolescents aged 13 in 1959 to 92.2% in 1973 and was above 97% since 1975 (Norwegian Institute of Public Health; Tverdak & Funnemark, 1988).

Non-Specific Effects of the BCG Vaccine

Non-specific effects are health effects of a vaccine that are unrelated to the targeted disease. The first non-specific effects were reported on the vaccinia vaccine against smallpox, which 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 non-specific 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 adaption of the innate, or unspecific, 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 (Roberts, Alberts, Bruce, Johnson, Alexander, Lewis, Raff, & Walter, 2015). This trained immunity, i.e. the change in the innate immunity after BCG vaccination, is one of the potential channels for the BCG vaccine related non-specific effects (Moorlag, Arts, van Crevel, & Netea, 2019). This physiological mechanism is consistent with a broader body of evidence from demography and population health which finds a robust link between early childhood health and changes to the immune system and later life (Bonanni, 1999; Drummond, Chevat, & Lothgren, 2007; Karafillakis, Hassounah, & Atchison, 2015; Mercer, 1985; Michiels, Govaerts, Remmen, Vermeire, & Coenen, 2011). 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 & Morales, 2021).

Existing evidence of non-specific BCG effects

There is a small number of studies, most of them observational, estimating non-specific effects of the BCG vaccine on child survival (Biering-Sørensen et al., 2018; Breiman et al., 2004; Castro, Pardo-Seco, & Martinón-Torres, 2015; Elguero, Simondon, Vaugelade, Marra, & Simondon, 2005; Garly et al., 2003; Lehmann, Vail, Firth, Klerk, & Alpers, 2005; Nankabirwa, Tumwine, Mugaba, Tylleskär, & Sommerfelt, 2015; Roth et al., 2005; Schaltz-Buchholzer, Aaby, et al., 2021; Schaltz-Buchholzer, Kjær Sørensen, Benn, & Aaby, 2021; Steenhuis et al., 2008; Thysen et al., 2020; Vaugelade, Pinchinat, Guiella, Elguero, & Simondon, 2004). The majority of these studies was conducted in low- and middle-income countries and suggest a positive association between BCG vaccine and child survival. Evidence on the vaccine’s non-specific effect on adult morbidity and mortality is also limited (Giamarellos-Bourboulis et al., 2020; Kölmel et al., 2005; Lankes et al., 2009; Pfahlberg et al., 2002; Rieckmann; Rieckmann, Villumsen, Jensen, et al., 2017; Rieckmann, Villumsen, Sørup, et al., 2017; Villumsen et al., 2009; Villumsen et al., 2013; Wardhana et al., 2011). All studies but one are based on small non-representative samples or are limited to specific health outcomes (Rieckmann, Villumsen, Sørup, et al., 2017). Most importantly, all studies except for one clinical trial do not sufficiently account for potential confounding and can thus only estimate associations but not establish causal relationships (Giamarellos-Bourboulis et al., 2020).

For example, Pfahlberg et al. (2002) study the relationship between BCG vaccination and malignant melanoma among 1,230 individuals from eleven health centers in seven European countries between 1994 and 1997. Using a case-control design, the authors compare those that did and did not receive the BCG vaccine and conclude that the BCG vaccine might reduce the risk of malignant melanoma. In contrast, Riekmann et al. (2019) link data on 5,090 adults from the Copenhagen School Health Records to the Danish Cancer Registry and find no association between BCG vaccination and malignant melanoma. Focusing on lymphoma and leukaemia, Villumsen et al. (2009) also use data from the Copenhagen School Health Records and find that individuals who had received the BCG vaccine had half the risk of developing lymphoma compared to those who had not. However, there was no difference in the risk of developing leukaemia.

Two observational studies analyse the association between BCG vaccination and adult survival. Kölmel et al. (2005) analysed 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 analysing the relationship between the BCG vaccination and long-term survival used the Copenhagen School Health Records Register used as part of the prior mentioned papers. Rieckmann et al. (2017) linked these data to the Civil Registration System and the Danish Register of Causes of Death databases, focusing on death by natural causes (excluding tuberculosis or smallpox). Comparing those that did receive the BCG vaccine to those that did not, they found that vaccinated individuals had approximately half the risk of dying from natural causes than unvaccinated individuals.

While some of these studies suggest that the BCG vaccine has non-specific 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, Kjær Sørensen, et al., 2021). 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 around side effects or because these parents have a lower health service uptake in general. 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 non-specific 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 into adulthood.

Data

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 which 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 (University of California, Berkeley, and Max Planck Institute for Demographic Research). We identified six potential countries with a change in vaccination policy that also had data available in the HMD (Austria, Denmark, Finland, France, Great Britain, Norway, 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 into adulthood.

One of our study designs, the multi-period 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 vaccines at the same or similar time as Sweden. Based on 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 1988 from the Human Mortality Database. To minimize confounding from the aftermath of World War II, we only included birth cohorts born in 1950 or later. Thus, our observation period covers cohorts born from 1950 to 1988, which allows us to study the impact of BCG vaccination on the cohort probability of survival to age 30 (for those born in the final 1988 birth cohort, we have data on their survival up to the year 2018, when cohort members turned 30).

Unfortunately, we were not able to use Swedish register data, which is commonly used in demographic analyses of the Swedish population, (Barclay & Kolk, 2018; Bucher-Koenen, Farbmacher, Guber, & Vikström, 2020; Cook, Fletcher, & Forgues, 2019; Helgertz & Bengtsson, 2019; Lazuka, 2019; Oudin Åström, Edvinsson, Hondula, Rocklöv, & Schumann, 2016) for the following two reasons: (1) our analysis requires survival data for cohorts born several years before the BCG discontinuation in 1975, but the Demographic Data Base only includes data from 1973 and later. (2) It is not currently possible to link the Demographic Data Base to BCG vaccination data.

Data preparation

Our outcome of interest is the probability of survival to age 30 in each birth cohort by country. Constructing this outcome requires information on the number of individuals that survived to age 30 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. We, therefore, 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, Heuveline, & Guillot, 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 30” and used the cohort life table probability of survival from birth to age 30 to determine what share of this initial birth cohort survived to age 30, and correspondingly assigned values to the “alive at age 30” variable. For example, if among the 115,414 births for the 1950 birth cohort, the cohort-life table probability of survival to age 30 was 90%, we would assign values of 1 for “alive at age 30” to (0.9 x 111,414 =) 100,273 observations. This procedure was repeated for each birth cohort, resulting in a final dataset of 4,152,211 observations, which equals the total number of births in Sweden from 1950 to 1988. We repeated this approach to generate individual-level datasets for Norway as well as separately for males and females in each country.

Identification strategies

We used two complementary quasi-experimental designs to assess the effect of the sharp BCG discontinuation in Sweden on cohort survival to age 30: the regression discontinuity design (RDD) and the multi-period difference in differences (DiD) approach. The RDD estimates the causal effect of an intervention or policy in situations where exposure to the intervention/policy is based on whether individuals fall on one side or another of some type of cut-off. This cut-off point is usually referred to as the “threshold” and the continuous variable on which the cut off is based is known as the “running variable.” For example, individuals may be eligible for free health care (the policy) based 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 an outcome would have evolved continuously. Heuristically, this can be expressed that 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 not mandatory anymore. There are no yearly data on BCG coverage for the years around the threshold. However, Romanus et al. (2006) report 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 30 induced by an event other than the BCG vaccination discontinuation. To our knowledge, there have been no major shifts in policies nor national-level events, such as natural disasters, which could have affected the probability of survival of the cohorts born just before the discontinuation differently to 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 neither above nor below the threshold (Figure S 1).

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), including triangular Kernel weights (to give more weight to observations closer to the threshold), and a data-driven approach for identifying the bandwidth around the threshold (to remove arbitrary bandwidth selection and potential manipulation of the bandwidth size by researchers) (Imbens & Kalyanaraman, 2012). For inference, we used robust bias-corrected standard errors and confidence intervals (Cattaneo et al., 2019). Our estimation model is:

Here Y_i is the probability of survival to age 30 for new born i, X_i is a continuous measure of year of birth (the running variable), and D_i is a binary indicator that takes the value 1 if a birth cohort was born after the discontinuation of mandatory BCG vaccination (Cunningham, 2021). β₃ is the causal effect of the discontinuation of the BCG vaccine on survival to age 30 for the cohort born in the threshold year, i.e. 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 only emerges after several birth cohorts (for example if it took a few birth cohorts to fully discontinue the vaccine). Although it is 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 multi-period DiD model (sometimes also referred to as leads and lags model) with Norway as the comparison country.

The multi-period DiD estimates the effect of the vaccine discontinuation by comparing the probability of survival in Sweden and Norway in each birth cohort accounting for temporal trends, i.e. the general positive trend in the probability of survival, as well as the baseline difference between the two countries. The central identifying assumption of the multi-period DiD is that in the absence of BCG discontinuation, the cohort probability of survival to age 30 in Sweden would follow the same trend as the cohort probability of survival to age 30 in Norway (the so-called parallel trends assumption). This assumption is commonly assessed by comparing the trends in the outcome variable in the pre-treatment 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.

• Download figure
• Open in new tab

Figure 1: Probability of survival to age 29 in Sweden and potential control countries

Note: We plotted the probability of survival to age 29, rather than 30, because data needed to construct survival to age 30 were not available for France and Great Britain. The vertical line divides the birth cohorts into those born before the vaccine discontinuation (1975 and earlier) and those born after.

The multi-period DiD regression model is set up as follows: with c indexing countries and t indexing years (Cunningham, 2021). μ_c is a country fixed effect which accounts for the baseline difference in the probability of survival between the treatment (Sweden) and control (Norway) countries. λ_t is a vector of year fixed effects which account for the time trend common to both countries (1975, the year of the BCG vaccine discontinuation, is the reference year). μ_t λ_t is a vector of the interaction between the country and year dummies and 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 provides 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 with a post dummy (=1 for all cohorts born in or after the year of discontinuation) and including year fixed effects as individual dummy variables (Angrist & Pischke, 2008):

Regression Discontinuity 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 long-term survival in the birth cohort born in 1976. We estimate a precise null effect (0.00 percentage points (pp) (95% CI -0.20 – 0.17)) on the probability of survival to age 30 at the point of BCG vaccination discontinuation (Table 1, Figure 3). As with the visual results, this finding was not driven by just males or females. We also find no evidence that BCG vaccine discontinuation affected the probability of survival to age 30 for among the male (0.05pp (−0.17 - 0.25)) or female (0.11pp (−0.46 – 0.15)) subpopulations (Table 1, Figures S 2 and S 3).

• Download figure
• Open in new tab

Figure 3: Non-specific effects of the BCG vaccine discontinuation on the cohort probability of survival from birth to age 30 in Sweden

Note: Pink lines represent linear trends fitted separately for cohorts born before and after vaccine discontinuation; blue lines are local linear trends fitted within the bandwidth and separately for cohorts born before and after vaccine discontinuation; dashed lines represent the bandwidth determined by the data; purple vertical line represents the year 1976, year of birth of the cohort from whom the local average treatment effect is estimated.

Multi-period difference in differences results

While the RDD estimates potential non-specific 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 shows 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 pre-discontinuation period are not statistically different from zero (exact values are shown in Table S 1). This suggests that prior to the discontinuation of BCG vaccination in Sweden, both countries were following a parallel trend in cohort survival to age 30. The coefficients after the 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 30 in the post-BCG-discontinuation period between individuals born in Sweden and Norway across most birth cohorts. For Swedish cohorts born after 1983, we see some evidence of lower survival to age 30; however, these differences are for cohorts born nearly 10 years after the discontinuation of BCG. For this reason, we consider it more likely that the negative point estimates for younger birth cohorts were caused by other changes at the population level that differed between Norway and Sweden at that time rather than being caused by the discontinuation of the BCG vaccine.

• Download figure
• Open in new tab

Figure 4: Event study plot displaying multi-period difference in differences results

Note: Coefficients (with 95% confidence intervals) display the difference in the probability of survival between cohorts born in the respective year in Sweden and Norway net the secular time trend and fixed countries differences. The dashed vertical line marks the reference group, which is the cohort born in 1975.

By splitting the results by birth cohort, the event-study results may result in underpowered estimates of the effect of BCG discontinuation. When pooling all data from the pre- and post-discontinuation periods, respectively, we find evidence of a very small negative effect of the BCG vaccine on the probability of survival to age 30 by -0.06pp (−0.11 – 0.00) (Table 2). While these results are statistically precise due to our large sample size (over six million observations), the effects we estimate are extremely small in magnitude at less than a 10th of a percentage point and thus of minimal policy significance. Further, based on the event-study plot, this minimal effect is likely driven by the lower survival among the 1983+ birth cohorts.

We find a similar set of conclusions when examining males and females separately. Among both the male and female population, there was no difference in the probability of survival for the cohorts born before the BCG vaccine discontinuation, indicating support for the parallel trend assumption. 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 visual finding is confirmed in the pooled analysis (0.00pp (−0.07 - 0.07)). In the male subpopulation, the results also suggest that there were no non-specific effects of the BCG vaccine on survival. However, as observed in the total population, cohorts born after 1983 in Sweden had a lower probability of survival in Sweden than in Norway. Furthermore, pooling all pre- and post-discontinuation birth cohorts reveals that the probability of survival decreased by 0.11pp (−0.19 – -0.02) among men.

It is likely that this result is driven by the youngest birth cohorts, which represent approximately one third of the post-discontinuation birth cohorts, but were born nearly 10 years after the vaccine discontinuation and thus the validity of the DiD assumption that any difference was due to the discontinuation alone for this group is questionable. To test this, we restricted our sample to the cohorts born within the MSE optional bandwidth calculated in the RDD analysis, which was +/-six years in the male sample (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 (Table 2).

Sensitivity analyses

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% (Figure S 6). 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 30. 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 not a discontinuity in the probability of being treated but in its first derivative, i.e. in the slope of the relationship between the treatment and the assignment 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 and female populations (Table 3). Among men, we found evidence of a slight decrease in the rate of change (−0.10 (−0.24 – 0.00)) of mortality. However, although statistically precise, the magnitude of this decrease is extremely small and thus of limited policy significance. 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 cut-off to be one year earlier (1975 in the RDD and 1974 in the multi-period DiD), which produced consistent results with those of the main analysis (Tables S 2 to S 4, Figures S 7 to S 9). 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 cut-off. Decreasing the bandwidth would have the opposite effect and increase variance while reducing the bias. It was not possible to decrease the bandwidth due to the limited number of mass points in our analysis but increasing the bandwidth by one, two, and three years, did not change the results and all point estimates remained statistically insignificant (Table S 5).