key: cord-1050863-e3fn0v7y authors: Brilli, Ylenia; Lucifora, Claudio; Russo, Antonio; Tonello, Marco title: Vaccination take-up and health: Evidence from a flu vaccination program for the elderly() date: 2020-09-29 journal: J Econ Behav Organ DOI: 10.1016/j.jebo.2020.09.010 sha: fbf4c31bb67922b94393ab28a9dd8a6a3397cb95 doc_id: 1050863 cord_uid: e3fn0v7y We analyze the effects of a vaccination program providing free flu vaccine to individuals aged 65 or more on take-up behavior and hospitalization. Using both administrative and survey data, we implement a regression discontinuity design around the threshold at age 65, and find that the effect of the program on take-up ranges between 70% and 90% of the average vaccination rate for individuals aged less than 65. We show that this effect is not entirely driven by an income channel, but also depends on the expected benefits of vaccination. The results on health outcomes provide suggestive evidence that the program reduces the likelihood of emergency hospitalization. Viruses represent a serious public health issue, not only because of their effects on the individuals' health, but also for their implications on the health sectors and, ultimately, on the overall society ( Adda, 2016 ) . Vaccination has been the most effective instrument used to eradicate or, at least, limit the diffusion of viruses, which otherwise would cause severe illness and death. 1 The influenza viruses cause annual epidemics which are estimated to result in about 3-5 million cases of severe illness, and about 290,0 0 0-650,0 0 0 deaths ( WHO, 2017 ) . Influenza epidemics also have substantial implications for the health sectors, as clinics and hospitals can be overwhelmed during illness periods ( WHO, 2017 ) . The elderly are the individuals most affected by the influenza virus, and by the development of severe complications in case they are infected. In the latest influenza seasons in the United States, the hospitalization rate of the elderly has been four times the overall hospitalization rate. In Europe, nearly half of the hospitalization and death cases refer to the oldest age group. 2 The main strategy to protect the most vulnerable individuals against seasonal influenza has been to implement vaccination programs targeted toward the elderly population. In 2003, the World Health Organization (WHO) urged to increase vaccination coverage to 75% among older persons, and, in 2009, the European Union (EU) Council issued a recommendation encouraging Member States to implement policies aimed at reaching this target. 3 Even though the influenza vaccination remains non-mandatory, many countries have thus attempted to increase the coverage by offering the flu vaccine free to individuals above a certain age, which, depending on the country, ranges between 59 and 65 years ( ECDC, 2018a ) . This paper assesses the effects of a flu vaccination program for the elderly, implemented in Italy, on vaccination takeup and health outcomes. 4 According to the Italian National Plan for Preventive Vaccination (NPPV hereafter), the influenza vaccine is freely provided, during a single visit with the general practitioner (GP), to individuals aged 65 or more. The program is likely to affect the individual's propensity to get the shot, because it lowers both the monetary and nonmonetary time costs associated with the vaccination decision. We identify the effects of the NPPV and, in particular, of universal eligibility to free vaccination at age 65, by adopting a regression discontinuity (RD) strategy around the threshold at 65 years of age. We use two different sources of data: (i) administrative individual-level data from the Health Authority of the Milan Metropolitan Area (in the North of Italy); and (ii) data from the Italian Survey on Health (2012) (2013) , provided by the Italian Institute of Statistics. The combination of these data sources provides a unique set of information to estimate the effect of the program on vaccination take-up and health, and allows us to validate the results against the different type of information provided. Our work relates to three strands of the literature. First, this paper is related to studies exploiting eligibility rules based on age thresholds to assess the effects of health insurance, or free health provision, on health services consumption and health outcomes. Card et al. (20 08, 20 09) document that individuals eligible for Medicare from age 65 have higher use of medical services, and face a significant reduction in mortality. Card et al. (2008) also find that the effect is highly heterogeneous in the population, and that individuals without health insurance before age 65 increase more than the others the use of low-cost services, such as doctor's visits. Second, this paper relates to the economic literature which assesses the effectiveness of vaccination programs, in terms of both vaccination take-up and health. 5 In the context of pediatric vaccinations, Chang (2016) estimates that state legislation mandating private insurers to cover pediatric immunizations increases the vaccination take-up rates substantially, suggesting that individuals are responsive to policies lowering the cost of immunization. Van Ourti and Bouckaert (2020) consider a vaccination program similar to ours, implemented in the Netherlands, and estimate that the program is associated with a substantial rise in vaccination take-up, as well as reduced mortality and health care consumption. Ward (2014) , instead, estimates that an influenza vaccination program expanding coverage outside the typical target groups (i.e. children and individuals aged 65 or more) leads to higher vaccination rates in the newly eligible age groups, which brings health improvements for older individuals as well. 6 Third, our study builds on the medical and economic studies that investigate the determinants of flu vaccination decisions. The medical literature suggests that demographic characteristics (such as gender and socio-economic status), as well as features of the health care system (such as vaccination cost) are strongly correlated with the flu vaccination decision ( Nagata et al., 2011; Daniels et al., 2004 ) . Within the economic literature, Mullahy (1999) and Schmitz and Wübker (2011) analyze the microlevel determinants of flu vaccine take-up, and find that the most important correlates of individuals' flu vaccination decision are age, health status, and physicians' quality. In an experimental setting, Bronchetti et al. (2015) show that small financial incentives proved to be effective in increasing flu vaccination intentions and actual take-up. Mullahy (1999) also suggests that, in addition to the out-of-pocket price of the vaccine, individuals may also respond to the nonmonetary time cost of getting the vaccination. Our paper contributes to the economic literature on the determinants of vaccination and health-related behaviors in two ways. First, our study provides new and clear evidence on the relationship between a specific health prevention policy, take-up and health outcomes, thereby improving our understanding of the implications of policies aimed at increasing vaccination rates. Unlike policies based on age-threshold studied by most prior work, that imply eligibility for insurance or free provision for a broad set of health services, in the policy considered in this paper eligibility for free vaccination is the only change occurring at the threshold, and thus represents a particularly powerful intervention for the evaluation of vaccination programs. By examining the health consequences of such program, the paper also contributes to a better understanding of the benefits associated with vaccination programs targeted toward the elderly population. Even though vaccination policies based on age-threshold eligibility are implemented in many developed countries, this study is among the very few providing causal estimates of their effects on the individual's vaccination decision and health. Second, our analysis sheds light on heterogeneous responses to universal eligibility to free vaccination. Importantly, by using both administrative and survey data, we observe a broad set of individual's and GP's characteristics, which make it possible not only to identify the subgroups of the population that comply more with the policy, but also to provide evidence on potential channels driving the effects on vaccination take-up. Our results show that the effect of universal eligibility for free vaccination on the individual's vaccination take-up ranges between 70% and 90% of the average vaccination rate for individuals below age 65. We interpret this estimate as a Local Average Treatment Effect (LATE) of the vaccination program on the take-up. Importantly, we show that the cutoff age does not refer to any other policy or behavioral changes (e.g. retirement), and we do not find any change in vaccination take-up at placebo ages before and after age 65. The analysis on heterogeneous effects shows that low-income individuals respond to the program only in case they are affected by poor health conditions, and that the program induces a higher increase in vaccination for individuals living in large families: we interpret this as evidence that individuals value the expected benefits associated with the immunization, both from an individual and a collective perspective. We also find that individuals who do not suffer from poor health conditions or low income increase their vaccination take-up once they become eligible, which may suggest that the reduction in the nonmonetary cost of vaccination implied by the program matters for their vaccination decisions. Finally, we evaluate the effect of the vaccination program on the probability of hospitalization during the same flu epidemic season. The results, though not precisely estimated, provide suggestive evidence that the program is associated with a reduction in the probability of emergency hospitalization, which supports the idea that the influenza infection leads to complications, for which the elderly need immediate care. We interpret these changes in hospitalizations at age 65 as the Intention-To-Treat (ITT) effect of the vaccination program: in fact, the consequences that we observe on the health measures may be due to the change in vaccination behavior, as well as to any potential indirect effects on health-related behaviors (e.g. better information through GPs or spillover effects from vaccinated peers). The rest of the paper is structured as follows. Section 2 describes seasonal influenza, the vaccination program under study, and the Italian national healthcare system. Section 3 presents the empirical strategy, Section 4 describes the data, and Section 5 discusses the main identification assumptions. Section 6 presents the baseline results on vaccination take-up, while Section 7 presents the analysis on heterogeneous effects. Section 8 presents the results on health outcomes. Section 9 concludes and derives policy implications. Seasonal influenza is an acute and highly contagious infectious disease with mostly respiratory symptoms. It is caused by the influenza virus and is easily transmitted, predominantly via the droplet and contact routes and by indirect spread from respiratory secretions ( WHO, 2017 ) . Each year, influenza causes substantial morbidity and mortality, particularly in elderly individuals and those with poor or chronic health conditions, who face the highest risk of developing subsequent serious complications. Vaccination is the safest and most recommended strategy to reduce the epidemics ( WHO, 2017 ). Thus, despite being typically non-mandatory, it is strongly recommended for elderly and high-risk individuals. According to recommendations of the CDC and the ECDC, the flu vaccination should be repeated every year because the influenza virus constantly evolves. Every spring, WHO establishes the types of vaccines to be used in the next season, according to the predictions on the type of virus most likely to be circulating. 7 In order to increase the vaccination coverage, countries have adopted different policies aimed at lowering the cost of immunization. In the United States, flu vaccination is part of the preventive services provided within Medicare Part B, to which individuals become eligible when they turn 65. In Europe, the majority of countries provide free flu vaccination (either within the public national healthcare system or in a national health insurance scheme) to individuals who are above a certain age threshold, which, depending on the country, varies between 59 and 65 years. 8 However, while being usually higher than that of other age groups, the elderly vaccination rate is still far from reaching the WHO target of 75% in many developed countries. In Italy, vaccination against seasonal influenza is regulated by the National Plan for Preventive Vaccination (NPPV), which is established by the Italian Ministry of Health. The flu vaccination campaign starts in late October and finishes by the end of December, while the circulation of the influenza virus occurs between October/November and April of the following year ( ECDC, 2017 ) . 9 The NPPV establishes that flu vaccination shall be completely free for selected categories of individuals, who may be at risk of complications in case of flu infection. The first category refers to the elderly: individuals are entitled to free vaccination from the campaign which starts in the calendar year in which they turn 65. The age-based eligibility to free vaccination only depends on the year of birth, and not on the month of birth; moreover, it is determined at the beginning of the campaign (in October), and does not change during the same campaign. For example, individuals born in January 1948 are not eligible for free vaccination the month they turn 65, i.e., on January 2013, at the end of the 2012 campaign, but become so in October 2013, when the 2013 campaign starts. 10 Other categories of individuals are offered a free flu vaccine, regardless of their age, because they may be at-risk of complications in case of contagion: (i) individuals affected by a certified chronic disease, especially those of the respiratory and cardiovascular systems, diabetes, or other diseases determining a weakening of the immune system; (ii) women in their second or third trimester of gestation; (iii) individuals institutionalized into nursing homes. As a preventive measure, the flu vaccine is also offered free to individuals working in the health, education or military sectors, care-takers of individuals at risk of complications, and individuals working in contact with animals ( Ministero della Salute, 2013 ). According to the NPPV regulations, individuals eligible for free flu vaccination get the immunization from the GP during one visit. More precisely, they pay neither the vaccine shot, nor the injection nor the visit(s) from the GP, and thus face a zero out-of-pocket price. In order to increase the vaccination take-up in the at-risk categories above, the NPPV regulations also call for an active role of the GP, who should actively offer them the vaccine ( Ministero della Salute, 2013 ). All the individuals aged less than 65, not included in any of the categories above and wishing to vaccinate themselves, have to get a prescription from their GP in order to buy the vaccine at the pharmacy and refer back to the doctor or to professional nurses to get the shot. 11 This implies that individuals not eligible for free vaccination under the NPPV bear not only the monetary cost of immunization, 12 but also a nonmonetary cost, mainly due to the time spent in the immunization process. Fig. 1 displays the age profile for the flu vaccination take-up, using the data sources employed in the analysis. Panel A reports the proportion of individuals who are vaccinated against the seasonal flu computed from administrative data from the Milan MA,-which cover vaccinations provided within the NPPV program -, while Panel B shows the same statistics drawn from the Italian Survey on Health, which covers all vaccinations -either obtained through the NPPV or purchased privately -for the whole of Italy. 13 Both data sources, despite covering different types of vaccinations, indicate that the most sizable jump occurs at age 65. In addition to the vaccination against influenza, the NPPV also stresses the importance of non-pharmaceutical measures to decrease the likelihood of virus transmission, such as washing hands, or covering the mouths when sneezing. During the influenza campaign, all health facilities show posters advertising these measures, and GPs as well are expected to inform their patients about their use and efficacy. The Italian national healthcare system (NHS hereafter) is mainly public and managed by the regional governments, while minimum quality standards are defined at the state level for all the regions. Under the NHS, all residents can freely consult a general practitioner (GP), who is responsible for prescriptions of drugs and requests for specialist visits. Hospitalizations are also freely provided by the NHS to all residents. Cost-sharing is instead required for specialist visits, diagnostic checks, and drugs. Individuals are exempted from the cost-sharing in case of (i) poor health conditions (i.e. a certified chronic disease or a disability), (ii) a combination of both poor health and low income, or (iii) low income only. 14 Notice that the NHS regulations for the exemption from cost-sharing do not apply to flu vaccination, as the free provision of the vaccine is strictly regulated by the NPPV. In other words, individuals exempted from cost-sharing are not eligible for 9 In the paper, we will refer to the vaccination campaign as campaign and to the period of circulation of the influenza virus as season . For instance, the 2013 vaccination campaign starts at the end of October and finishes by the end December 2013, while the 2013 season implies that the influenza virus circulates between October 2013 and April 2014. 10 This implies that individuals born in January 1948 cannot postpone strategically the decision to get vaccinated toward the end of the 2012 campaign in order to benefit from free provision. 11 The injection must be performed by a doctor or professional nurse, in order to check for potential side effects of the vaccine. 12 The price of the vaccine shots sold in pharmacies might vary from 12 to 30 euros, depending on the year, type, and producer. The price for an injection can vary between 10 and 30 euros, depending on the doctor and on whether it is done at home or in the clinic. 13 Both data sources are described in detail in Section 4 . In Section 5.1 we explicitly tackle the issue that private vaccinations are observed in the survey data only. 14 Cost-sharing exemptions for low income apply to individuals with income below a certain threshold, unemployed, or retired with a pension below a minimum level. free vaccination, unless they belong to one of the categories listed in Section 2.2 . However, there may be some overlaps between the two regulations, especially for individuals with poor health conditions. Exemption from cost-sharing because of health conditions applies to all individuals affected by any chronic disease, while the free provision of the flu vaccine prioritizes individuals with diseases of the respiratory or cardiovascular system or diseases determining a weakening of the immune system ( Ministero della Salute, 2013 ). Hence, we may expect individuals exempted from cost-sharing because of health conditions (categories (i) and (ii) above) to also have the highest probability of being eligible for free vaccination regardless of their age. 15 The aim of the paper is to estimate the effect of universal eligibility for free vaccination on influenza vaccination takeup. To this purpose, we exploit the fact that, in Italy, under the NPPV, flu vaccination is free for all individuals aged more than 65 and we use information on age to determine the individual's assignment to the treatment in a RD setting. This framework makes it possible to estimate the effects of the age-related eligibility change, by comparing individuals who are essentially identical under all characteristics, but differ for being born on opposite sides of the cutoff. The baseline RD equation for the estimation of the effect on vaccination take-up takes the following form: in which d i is the running variable, indicating the distance between the individual's date of birth or age in years -depending on the data source used -from the cutoff; ET i is a dummy defining the eligibility for the treatment status; f R and f L are unknown smoothing functions of the running variable d i , and Pr ( V i ) is a dummy indicating whether individual i is vaccinated against seasonal influenza. Eq. (1) is estimated using both administrative data from the Milan MA, for the 2013 flu vaccination campaign, and survey data from the Italian Survey on Health. Given the different information available in the two data sources, the variables in Eq. (1) are defined as follows. In the administrative data, the running variable d i is defined as the distance in days between the individual's birth date and the threshold point represented by January 1st, 1949, and ET i takes value 1 for those born before January 1, 1949, -i.e. aged 65 and thus eligible for free vaccination in the 2013 campaign. In the survey data, the running variable d i is defined as the distance in years from the cutoff age of 65, and the treatment variable ET i takes value 1 for those aged 65 or more. Given that the assignment to the treatment is deterministic and based on the date of birth/age of the individual, which, in a sufficiently small neighborhood of the cutoff, can be considered as good as random, the parameter β RD is an estimate of the causal effect of universal free vaccination on the outcomes of interests: We interpret the effect on vaccination take-up as a Local Average Treatment Effect (LATE), because we observe both the eligibility status (determined by the age threshold) and the individual's direct compliance with the treatment. For the analysis of the effect of the NPPV on health outcomes, we only use the administrative data from the Milan MA and define the dependent variable in Eq. (1) as the hospitalization probability during the 2013 flu season, Pr ( H i ). In this case, the estimated effect should be interpreted as an Intention-To-Treat (ITT), because it may also capture changes in the individual's health behavior or spillover effects associated with the change in vaccination take-up. For example, eligible individuals, along with the free flu vaccination, could receive additional information from their GP about non-pharmaceutical preventive measures, which may induce a change in health behaviors and affect their short-term health status. In this section we describe in details the two data sources used in the empirical analysis: (i) administrative individuallevel records of residents in the Milan Metropolitan Area (MA), in the North of Italy; and (ii) data from the 2012-2013 wave of the Italian Survey on Health, a nationally representative survey run by the Italian Institute of Statistics. The administrative data are provided by the Agency for Health Protection of the Milan MA. 16 The Agency for Health Protection is the lowest health administrative level. The Milan MA includes the municipality of Milan and 133 surrounding municipalities and is located in the Lombardy region, in the North of Italy. 17 The data refer to the 2013 influenza season, and is drawn from three administrative sources: (i) the General Health Register (GHR), (ii) the Register of General Practitioners (GP), and (iii) the Hospitalization Records (HR). The GHR includes basic demographic characteristics, such as gender and municipality of residence, and the date of birth, which represents the running variable used in the analysis. Importantly, the data provide information on the exact date of vaccination against seasonal influenza, if that occurred within the NPPV program. More precisely, we define the vaccination probability Pr ( V i ) as equal to 1 if the individual received the flu vaccination between October and December 2013, and zero otherwise. 18 A drawback of the information on vaccinations provided by the GHR data is that we can only observe vaccinations provided within the NPPV program, while vaccinations privately purchased are excluded. Since this is a relevant feature for our analysis, we carefully discuss its implications in Section 5.1 . 19 The GHR data also provide information on the type of exemption from cost-sharing for each individual (i.e. due to chronic health conditions, low income or combination of both conditions). We define a dummy variable equal to 1 in case of exemption because of health issues only, which is intended to proxy for the individual's health status. The GHR data was then merged to the GP register, which provides information on the GP (such as experience, age and number of patients), with whom each individual is assigned. In addition to using these variables as controls in the baseline analysis, we also exploit them to study heterogeneous effects. The HR reports all hospitalizations that occurred in the territory of the Agency. 20 Importantly, the data provides information on the date of hospitalization(s), which we use to link the event to the 2013 flu season. More precisely, we define a dummy variable indicating whether or not at least one hospitalization event occurred over the period when the influenza virus was circulating during the 2013 season, i.e. from week 43 in 2013 (mid-October) until week 17 in 2014 (end of April). 21 The analysis on the administrative data is performed on individuals aged 64 or 65 in 2013 (i.e. born between January 1, 1948 and December 31, 1949) . In order to keep the individual's vaccination decision as much homogeneous as possible, we 16 Access to the data was granted within a memorandum of agreement between the Catholic University -Milan and the Milan MA Agency for Health Protection (in Italian, Agenzia di Tutela della Salute ). 17 The Milan MA includes approximately 4.2 millions inhabitants, and represents the seventh largest metropolitan area in the European Union (Eurostat data on 2012). 18 In the analysis, we only consider vaccinations obtained by December 31st, 2013. The vaccinations performed after December 31 in the 2013 campaign represent the 0.11% of the sample. Results do not change if these vaccinations are included in the analysis, and are available upon request from the authors. 19 It should be noticed that this feature is not distinctive of our datasource, but rather common in administrative data provided by health authorities. See, for instance, Anderberg et al. (2011) . 20 In the data, we also observe whether the hospitalization was planned (e.g. requested by the GP), or access was made via the Emergency Rooms. 21 The definition of this period comes from the analysis of the Italian National Health Institute (NHI), which is in charge of documenting the epidemiological characteristics of each flu seasonal epidemics ( ISS, 2014 ) . In addition to measuring the hospitalization outcome at the extensive margin, we also define a variable measuring the intensive margin of the phenomenon, by using the number of days at the hospital in the same period. This analysis gives estimates which are never statistically significant at conventional levels; the results are available upon requests from the authors. exclude individuals with a certified disability or institutionalized in nursing homes, who are likely to receive the vaccination at their home or at the nursing institution. We also exclude individuals for whom we do not have reliable information on their GP. 22 An important concern when performing a RD analysis relates to the density of the observations around the threshold, which may indicate manipulation in the running variable ( McCrary, 2008 ) . Figure A . 1, Panel A , in Online Appendix, reports the number of individuals born in each calendar day. The figure shows an unexpectedly high number of births on January 1, 1949, and, to a lower extent, on the days before and after. The McCrary test, reported in Figure A .2, Panel A, in Online Appendix, confirms that there is a statistically significant jump in the number of births at the cutoff date. 23 Even though it is implausible that this pattern is systematically linked to the program under study, the peak in the number of observations at the cutoff may still bias our estimates ( Barreca et al., 2016 ) . Thus, in our analysis on the administrative data we adopt a donut specification, by excluding the individuals who are born at the cutoff date (January 1, 1949) and in the day before and after. 24 The final sample in the administrative data from the Milan MA consists of 68,962 observations. Panel A in Table 1 reports the descriptive statistics of the main variables used in the analysis. Individuals aged 65, our Treated indicator, account for half of the sample. The flu vaccination rate is around 12%, while the probability of hospitalization within the flu season is around 4%. Half of the sample is composed by females and by individuals living in a urban area. GP's characteristics show an average age of 58, almost 25 years of experience in the practice and a high number of patients (over, 1480). 22 Individuals with disability, institutionalized in a nursing home and without reliable information on their GP represent, respectively, 9.6%, 0.4% and 2.3% of the entire population aged 64 or 65 years. Table A .4 in Online Appendix shows that these characteristics are smooth around the cutoff date, while Table A .8 in Online Appendix shows that the baseline results are confirmed when these categories are included in the estimation sample. 23 The estimated discontinuity is −0.183 (0.029), with a t -statistics of 6.284. This pattern seems consistent with a framework where birth registers were manually filled: in the years after World War II it was standard to give birth at home and declare the birth to the Register Office of the municipality of residence in the days following the event. Further data inspection shows that this pattern is present for cohorts born in the years during or close to World War II, while it tends to reduce for younger cohorts. 24 In Section 6.2 we show that the results are robust to using different donut specifications or the full sample. The Italian Survey on Health (ISH henceforth) provides information on the health status, preventive behaviors and use of health services for a nationally representative sample of the Italian population. For our analysis we use the 2012-2013 wave, composed of four interview stages held in September 2012, December 2012, March 2013 and June 2013, which sampled about 120,0 0 0 individuals. The running variable in the analysis using ISH data is the age of the individual, measured in years. Since the age of the respondent is self-reported and the survey does not provide information on exact year of birth, we may face measurement error in the assignment to the treatment. For instance, individuals in their 65th year of age who may be eligible for free vaccination according to the NPPV program may declare themselves as 64-year-old in case their birthday has not occurred yet when interviewed. In order to minimize the likelihood of age misreporting, we use only individuals interviewed in December 2012 or March 2013, i.e. at the end or at the beginning of the year. 25 The relevant feature of the ISH data, for our analysis, is that every individual is asked whether he/she obtained the vaccination against seasonal influenza in the previous 12 months; from this survey question, we define the vaccination probability Pr ( V i ). Importantly, this information covers all vaccinations, i.e. those provided within the NPPV program, as well as the vaccinations privately purchased. Given the timing of the interviews and the wording of the question, the information on vaccination against seasonal influenza refers to the 2012 campaign (held between October and December 2012), rather than the 2013 campaign considered in the administrative data. 26 The ISH data also provides information on demographic and socio-economic characteristics, which we use as control variables and also exploit for the analysis on heterogeneous effects. More precisely, in addition to the individual's health status and gender, we observe the level of education, employment status and family size of the individual, from which we derive dummy variables indicating, respectively, whether the individual is a High School graduate, retired, or living in a single household. Finally, we observe the sector of occupation, which we use to control for whether the individual works or has worked (in case of retirement) in the health and education sectors. 27 For the analysis on vaccination take-up using ISH data, we use individuals aged between 40 and 90, and, as with the administrative data, we exclude individuals with a disability. 28 The final sample used in the analysis with ISH data consists of 31,033 observations. Table 1 , Panel B, reports the descriptive statistics of variables used in the analysis. Treated individuals, aged 65 or more, accounts for 33% of the sample, while the flu vaccination rate in the sample is above 23%. Half of the sample is composed by females (53%), while 32% of individuals has a certified chronic disease, 31% of the sample is retired, and 35% has a high level of education. Finally, 17% of individuals in the sample live alone and about 20% are currently working or have worked in the education or health sectors. Before presenting the baseline results on vaccination take-up, in this section, we discuss the assumptions required for the identification of the effects of universal eligibility to free vaccination. The identification of the effects of interest relies primarily on the fact that the cutoff at age 65 generates a sizable variation in the vaccination take-up. This is of paramount importance for the analysis on vaccination take-up, in which our estimates represent Local Average Treatment Effects. Incidentally, this is also relevant for the analysis on health outcomes, for which we expect the change in vaccination behavior to be one of the channels driving the results. Because the administrative data only include vaccinations provided within the NPPV, a concern may be that this data under-report the number of vaccinations for individuals aged 64, and thus overestimate the jump at the cutoff. This is not an issue for the ISH data, in which all flu vaccinations in Italy are reported -i.e. both under the NPPV program and privately purchased. In what follows, we do take a number of steps to argue that this limitation of the administrative data does not affect our results. First, we compare the probability of vaccination around the cutoff in the administrative and survey data. Fig. 2 shows that in both data sources there is a sizable discontinuity at age 65. In the administrative data (Panel A), the proportion of 25 As a further check, in Section 6.2 , we perform the analysis excluding individuals who report 65 years of age, and we label it as donut specification. 26 The 2012 and 2013 vaccination campaigns are similar on many dimensions. First, the vaccine composition was the same in both campaigns ( Ministero della Salute, 2013 ) . Second, the vaccination rates both below and above age 65 are also very close: the vaccination rate for the age group 45-64 is 9% in the 2012 campaign and 9.5% in the 2013 campaign; the vaccination rate above age 65 is 54.2% in the 2012 campaign and 55.4% in the 2013 campaign ( Ministero della Salute, 2019 ). 27 As reported in Section 2.2 , these individuals are eligible to free flu vaccination even before age 65. 28 The larger age-interval in the ISH data compared to the administrative data is mandated by the fewer observations available in the survey data around the cutoff age. However, as we show in Section 6.1 , the actual bandwidth used in the estimation is at most within five years from age 65. Concerning the exclusion of individuals with a disability, Table A .5, Column 1, in Online Appendix, shows that the proportion does not vary discontinuously at age 65, while Table A .9, Columns 1-2, in Online Appendix, shows that the results do not change if we include disabled individuals in the sample. Finally, we check whether the ISH data shows any evidence of discontinuity in the number of observations at the cutoff age, which may indicate manipulation in the running variable, but we find no evidence of this (see Panel B in Figures A.1 and A.2 in Online Appendix). vaccinated individuals ranges between about 7% at 64 years of age and 17% at 65 years of age, which implies a raw jump by 10 percentage points. In the survey data (Panel B), the vaccination rate increases from 22% for 64-year-olds to 31% for 65year-olds, which implies a change by 9 percentage points. The figure indicates that the administrative data are characterized by a lower vaccination rate than the survey data, both below and above age 65, despite the fact that above the threshold all the flu vaccines are provided within the NPPV. This may indicate that the Milan MA is characterized by a lower take-up rate than the Italian average. This pattern is also confirmed by the figures reported in Table A .1 in the Online Appendix, in which we compare the administrative and survey data with official statistics from the Ministry of Health ( Ministero della Salute, 2019 ). In particular, the data from the Ministry of Health, reported in Panel C of Table A .1, show that the Lombardy region (Panel C.1), where the Milan MA is located, is characterized by a lower vaccination rate than the rest of Italy (Panel C.2), both below and above age 65. Within the Lombardy region, data on pediatric vaccinations indicate that the Milan MA has the lowest take-up rate (see Table A .2 in Online Appendix). 29 In other words, this evidence suggests that the lower vaccination rate that characterizes the Milan MA, compared with the rest of Italy, is associated with a different propensity to get vaccinated across areas. Still, we find reassuring that the discontinuity in vaccination take-up reported in Fig. 2 shows up with a similar magnitude in both administrative and survey data (10 percentage points in the Milan MA data vs 9 in the ISH data). While a different propensity to get vaccinated in the Milan MA can explain the lower vaccination rate that we observe in the administrative data, this does not rule out that there may be a proportion of individuals not eligible for free vaccination who get it privately. Thus, our second step to address the above concern consists in performing the baseline analysis on both administrative and survey data. Since the ISH data provide information on all types of vaccinations, the estimated effect delivers the impact of the program on the overall vaccination rate in the population. Moreover, the direct comparison of the results obtained from the different data sources provides indirect evidence in support of the hypothesis that the privately-purchased component of vaccines is negligible. The analysis on the ISH data also answers to potential concerns about the external validity of the results from the administrative data, because it allows us to check whether the results vary across different areas of the country. Fig. 3 . Tests of continuity of the probability of retirement: age profile and RD estimates by age. Notes : Panel A reports the share of individuals who are retired from the labor force by age; the horizontal axis indicates the age in years. Panel B reports the RD robust estimates with Triangular Kernel and local polynomials of order 1 ( Calonico et al., 2014; of the probability to be retired at different age thresholds, reported in the X-axis, and the corresponding 95% confidence intervals. Source : own elaborations on 2012-2013 ISH data. The regression discontinuity design gives an estimate of the effect of universal eligibility for free vaccination under the assumption that 64-and 65-year-olds do not differ in any other observable or unobservable characteristics. This assumption further implies that no other relevant policy change occurs at the same cutoff age. An important concern is whether age 65 also coincides with changes in the probability to retire: if that were the case, the estimated effect on vaccination take-up would also incorporate the potential effect of a change in labor force participation. In Italy, an individual is entitled to retirement from work in two cases: (i) at a given age, provided that the individual has a minimum of 20 years of contributions ( statutory retirement ); (ii) at any age, if the individual reaches a minimum number of years of contributions ( early retirement ). In recent years the age threshold for statutory retirement has changed several times. In our analysis we consider the vaccination decisions of individuals during the 2013 campaign (when using the administrative data) and the 2012 campaign (when using ISH data). Table A .3 in Online Appendix reports the age thresholds for eligibility to statutory retirement in the years used in the analysis, and makes clear that in no case 65 is an age threshold considered for eligibility to statutory retirement in the calendar years used. Given that the administrative data does not provide information on the labor market participation of individuals, we use the ISH data to test whether age 65 coincides with a sizable discontinuity in actual retirement decisions. Fig. 3 , Panel A reports the retirement age profile, and suggests that the largest increase in retirement occurs at age 60, while the probability to retire seems quite smooth around age 65. We formally test for the existence of discontinuities in retirement at different ages between 60 and 70, by estimating an equation similar to Eq. (1) with the probability to retire at each age as dependent variable: the results are reported in Fig. 3 , Panel B, which confirms that there is no change in retirement between age 64 and 65. Thus, we can conclude that our estimates are not affected by changes in labor supply or retirement status. 30 We also test whether other observable characteristics of the individuals change discontinuously at the cutoff, by estimating a version of Eq. (1) with the individual characteristics observed in the administrative and survey data as dependent variables. Table A .4 in the Online Appendix reports the results for the variables that we observe in the administrative data, while Table A .5 in the Online Appendix reports the results for the variables in the ISH data. The tables show that there are no significant changes at the age threshold of 65 for any of the variables considered. 31 Finally, the presence of spillover effects could represent an additional threat to our identification strategy if individuals internalize the perceived level of herd immunity of one's peers (e.g. colleagues at work, neighbors or family members). 32 However, in our setup this can occur only if the level of herd immunity changes discontinuously at the cutoff date, which may happen if 65-year-olds expect to interact only with other eligible individuals -thus facing higher herd immunity among their peers -and not with 64-year-olds non-eligible for free vaccination. While we believe that such separation between 65-and 64-year-olds is rather unrealistic, such free-riding behavior is also strategically unsound since it would eventually drive vaccination take-up above 65 years of age to zero. This section presents our baseline results on the effect of universal eligibility for free flu vaccination on vaccination takeup. We perform a RD robust estimation following the non-parametric optimal bandwidth selection procedures suggested by Calonico et al. (2014) and Calonico et al. (2017) , and using a triangular kernel and a coverage error rate (CER) bandwidth selector. For each estimation, we report the results from the administrative data from the Milan MA and from the Italian Survey on Health. 33 The results reported in Table 2 show that the estimated parameter is about 6 percentage points, when using the Milan MA data (Panel A), and 7 percentage points when using ISH data (Panel B), regardless of whether we include additional controls, or whether we use a polynomial of order 0 or 1. The effect is sizable: considering the average vaccination rate for the control group in the two datasets, these estimates correspond to an increase in vaccination take-up of about 90% in administrative data and 70% in ISH data. The bandwidth used for the estimation is very small: it ranges between 28 and 60 days from the cutoff date in the administrative data estimation and between 1 and 3 years from age 65 in the ISH data estimation, which strengthen the plausibility of the assumption that individuals on both sides of the cutoff are comparable on all dimensions but eligibility to free vaccination. 34 Given that the ISH data provides information on all types of vaccinations, either obtained through the NPPV or purchased privately in the market, the estimated coefficients reported in Panel B of Table 2 can be interpreted as the effect of the NPPV on the overall vaccination rate in the population. However, the fact that the analysis with administrative data, which only include vaccinations obtained through the NPPV, delivers estimates which are quantitatively similar provides indirect 32 For herd immunity it is intended in the medical literature the protection against a certain disease that any individual gets as a spillover effect that comes from the fact that a substantial share of the population is immune to that disease because of the vaccination. 33 All regressions with ISH data use sampling weights. 34 Table A .7 in Online Appendix reports the results from a parametric analysis in which we vary the bandwidth. We use a linear specification, with or without interaction terms between the treatment and the running variables, with heteroskedasticity-robust standard errors and a triangular weight, which is decreasing in the distance from the cutoff. The results confirm the estimates from the non-parametric estimations: in particular, the estimates using a bandwidth of 1-3 months in administrative data and 1-3 years in ISH data are very close to the ones reported in Table 2 . Analysis on vaccination take-up for placebo age groups before and after age 65. Notes : the figure reports RD robust estimates with Triangular Kernel and Coverage Error Rate (CER) optimal bandwidth (BW) selector ( Calonico et al., 2014; and the corresponding 95% confidence intervals, from regressions on placebo ages between 63 and 67 using a polynomial of order 1 and controlling for the covariates. For each estimation, the older cohorts plays the role of the placebo treated group, the younger cohorts that of the placebo control group. The circles indicate RD robust estimates from placebo ages, while the black diamonds refer to the age cutoff used in our analysis. For the set of covariates included in the estimations and their definitions, see the footnote to evidence that the share of privately-purchased vaccinations below age 65 is negligible and does not hamper the validity of the estimates that we obtain for the Milan MA (reported in Panel A). Nonetheless, the discussion in Section 5.1 also points to a lower propensity to get vaccinated in the Milan MA compared to the rest of Italy. The similarity between the estimates reported in Panels A and B suggests that our findings are externally valid. We further check for this, by exploiting the fact that the ISH data refers to all Italy and by analyzing the effect of the program on different areas of the country. The results reported in Table A .6 in Online Appendix show that the estimated effects of the NPPV in the North and Centre-South of the country are very close, hence reassuring us that the effect does not differ across areas that may be characterized by a different propensity to vaccinate. We test the robustness of the results on vaccination take-up on four main dimensions. First, we replicate the analysis by using placebo ages, in order to make sure that the estimated change in vaccination probability is due to the flu vaccination program at age 65, and not to underlying age trends. We consider two placebo ages before age 65 (64 and 63) and two placebo ages after age 65 (66 and 67), and assign the placebo treatment status to individuals to the right of the placebo cutoff. The results reported in Fig. 4 show that in no case the placebo treatment matters for the individual's vaccination decision, either with the administrative data from the Milan MA or with the ISH survey data from Italy. Second, we check that the results are not sensitive to our sample selection criteria, for which we exclude individuals with a disability or (in the administrative data) institutionalized in nursing homes or without information on their GP. Third, we test the sensitivity of the estimated effect to the specifications adopted in the baseline RD, by changing the clustering of the standard errors or the donut criterion. More precisely, we repeat the analysis (i) by clustering the standard errors at the running variable level (date of birth or age, as suggested by Lee and Card (2008) ), at the available geographical level of aggregation (municipality or region), or (in the administrative data) at the GP's level; and (ii) by changing the donut criterion. The results reported in Tables A.8 and A.9 in the Online Appendix are very similar to the baseline reported in Table 2 . 35 In the previous section, we have estimated that universal eligibility to free vaccination at age 65 induces an increase in take-up ranging between 70% and 90% of the vaccination rate of individuals not eligible to the program. In what follows, we shed light on the channels driving this result, by analyzing heterogeneous effects in the population. We consider several subgroups differing according to individual characteristics, e.g. health status, income, gender, education, family size and sector of occupation, and GP's characteristics. For the analysis we use both the administrative data from the Milan MA and the survey data from Italy, depending on which type of information is available in each dataset. The literature has shown that individuals value the decision to get vaccinated against seasonal influenza, depending on their health status, which determines the benefits associated with the immunization ( Nagata et al., 2011 ) , and to the cost of vaccination, which can be either monetary or nonmonetary ( Mullahy, 1999 ) . In order to study the role played by these characteristics for the estimated effect of the NPPV on vaccination take-up, we use information on cost-sharing exemptions available in the administrative data from the Milan MA. More precisely, we use information on whether the individual is exempted from cost-sharing and, if so, which is the type of exemption. We identify four mutually-exclusive categories of individuals: (i) those exempted because of a health condition only, (ii) those exempted because of health issues and low income, (iii) those exempted because of low income only, and (iv) a residual category for individuals with no certified exemptions. We interpret the four categories above as providing a monotonic ordering in the health-income space which decreases in severity going from (i) to (iv), because, typically, individuals with exemptions because of health issues onlycategory (i) -suffer from more severe chronic conditions (see Section 2.3 ). It is important to remind that exemption from cost-sharing does not entitle to free flu vaccination. The vaccination against influenza is regulated by the NPPV according to the guidelines outlined in Section 2.2 : in particular, the flu vaccine is provided free to individuals aged less than 65 in case of health conditions determining a weakening of the immune system (especially chronic diseases of the respiratory or the cardiovascular systems). Thus, it is plausible to assume that the probability of free flu vaccination for individuals under 65 is higher for categories (i) and (ii), and lower for (iii) and (iv). This is supported by the average vaccination rates among 64-year-olds that we observe in the data, which is higher for categories (i) and (ii). Results are presented in Table 3 . Panel A shows no effect of universal eligibility for free vaccination on take-up for individuals who are exempted from cost-sharing because of a certified chronic condition. This result might seem counterintuitive, given that individuals in this group could benefit the most from flu vaccination. However, it may indicate that the category of individuals exempted because of health issues overlaps to a large extent with the category of individuals eligible to free flu vaccine because of chronic conditions (regardless of age), so that individuals in this group are not really affected by the change in eligibility status at age 65. Panels B, C, and D report the effect of universal eligibility for free vaccination on individuals exempted from cost-sharing because of poor health status and low income (Panel B), low income only (Panel C), or without any exemptions (Panel D). We find that individuals characterized by both poor health and low income (Panel B), as well as individuals without any exemptions from cost-sharing (Panel D), are more responsive to free vaccination eligibility at age 65, with an increase in the vaccination probability ranging between 7 and 9 percentage points. In the context of the flu vaccination program under study, it should be noticed that the age threshold induces a reduction in the monetary and nonmonetary time cost associated with the immunization. Individuals who are exempted from cost-sharing because of low income may value to a greater extent the reduction in monetary cost. The fact that we find an effect of the program on individuals exempted because of income and health issues (Panel B), and not on those exempted for low income only (Panel C), seems to suggest that the reduction in monetary cost matters only if individuals also value the benefit of not getting the flu, as in case of poor health conditions. On the contrary, the statistically significant effect that we find for individuals without any exemptions (Panel D) may suggest that these individuals are more responsive to the reduction in the nonmonetary cost associated with the flu immunization program. However, the large effect we find for this category may also be due to the extremely small vaccination rate at age 64 (around 2%). The medical literature on the determinants of the flu vaccination decision has also stressed the relevance of sociodemographic characteristics ( Nagata et al., 2011 ) . In what follows, we use the ISH data to shed light on whether sociodemographic conditions affect the individual's response to the program. More precisely, we check whether the effect of free eligibility to flu vaccination changes depending on the individual's gender, level of education, family size and sector of occupation. The results are reported in Table 4 . The coefficients reported in Panels A and B show that the effect does not differ between females and males, and by level of education. 36 Interestingly, we find that individuals respond to the eligibility to free vaccination at age 65 if they live in households with more than one member (Panel C), which may suggest that they Table 3 Heterogeneous effects of eligibility for free vaccination at age 65 on take-up by type of exemption from cost-sharing. (1) (2) ( Calonico et al., 2014; ; for the list of covariates included and their definitions see the footnote to Table 1 . Panel A reports the estimates for patients who are exempted from cost-sharing because of a chronic disease; Panel B reports the estimates for patients with exemption from cost-sharing because of a serious health condition and low income; Panel C reports the estimates for patients who are exempted from cost-sharing rule because of low income only; Panel D reports the estimates for patients who have no exemptions. The vaccination rate for the non-treated individuals (i.e. those aged 64) in the four categories is reported in square brackets. Significance levels: * * * p < 0.01, * * p < 0.05, * p < 0.1. Source : own elaborations on administrative data from Milan MA. value the expected benefits associated with the immunization, especially in terms of positive spillovers to the other family members. Finally, the analysis by sector of occupation shows that individuals respond to the eligibility to free vaccination if they do not work or did not work (if retired) in the education or health sectors (Panel D). This could be because individuals working in the education or health sectors can obtain the flu vaccine free within the NPPV even before age 65 (see Section 2.2 ); moreover, individuals working in these sectors may be more aware of the benefits associated with the immunization, especially in contexts such as schools and hospitals where there may be relevant positive spillovers. For these reasons, individuals working in these sectors are likely to receive the flu vaccine even at younger ages, and may thus be less affected by the change in eligibility status at age 65. On the contrary, individuals working in other sectors, who are not eligible to free vaccination before age 65, increase their take-up by about 8 percentage points once they become eligible. This result seems also to suggest that the Health Authority is effective in reaching individuals without previous attachments to the program. The health economics literature suggests that the quality of physicians can be a strong determinant of the vaccination decisions ( Mullahy, 1999 ) . In particular, the role of physicians could be relevant in the context of the program under study because the NPPV calls for an active role of the GP, who should contact eligible individuals aged more than 65 and offer them the vaccine (see Section 2.2 ). We thus check, using the administrative data, whether the effect of the program varies by observable characteristics of the GP, which may proxy for the GP's quality, such as years of experience and number of patients. In order to do so we split the sample according to the median value of GP's characteristics. The results reported in Table 5 show no significant difference in take-up between the different groups. This may suggest that either the Health Authority uses other channels (rather than GPs) to reach eligible individuals, or that our proxies of GP's quality do not properly capture the ability of physicians to induce eligible individuals to get vaccinated. ( Calonico et al., 2014; ; for the list of covariates included and the definitions of the variables see the footnote to Table 1 . The vaccination rate for the non-treated individuals (i.e. those aged less than 65) in each category is reported in square brackets. Significance levels: * * * p < 0.01, * * p < 0.05, * p < 0.1. Source : own elaborations on 2012-2013 ISH data. Heterogeneous effects of eligibility for free vaccination at age 65 on take-up by GP's characteristics. (1) ( Calonico et al., 2014; ; for the list of covariates included and their definitions see the footnote to Table 1 . Panel A reports the estimates for GPs above and below the median of the years of experience (25), Panel B for GPs below and under the median number of GP's patients (1548). The vaccination rate for the non-treated individuals (i.e. those aged 64) in each category is reported in square brackets. Significance levels: * * * p < 0.01, * * p < 0.05, * p < 0.1. Source : own elaborations on administrative data from Milan MA. In this section, we document the effects of universal access to free vaccination for individuals aged more than 65, on health outcomes, measured by the probability of hospitalization. By using this measure, we focus on severe health shocks, so that many minor illnesses, associated with the influenza virus, end up undetected. However, these measures are likely ( Calonico et al., 2014; ; for the list of covariates included and their definitions see the footnote to Table 1 . The hospitalization rate for the nontreated individuals (i.e. those aged 64) is reported in square brackets. Significance levels: * * * p < 0.01, * * p < 0.05, * p < 0.1. Source : own elaborations on administrative data from Milan MA. to better capture the occurrence of complications in case of flu infection, compared to other outcomes, such as drugs consumption or sick leave absence ( ECDC, 2018b ). The medical literature has stressed the difficulty of identifying hospitalizations occurred as a consequence of flu complications ( Rothberg et al., 2008 ) , which is mainly given by the fact that influenza is rarely tested for in patients admitted to the hospital. One way to deal with this issue is to consider only hospitalizations occurred during the weeks of the virus diffusion, which are more likely to be related to the occurrence of a flu infection. For this reason, we use the administrative data from the Milan MA, which have the notable advantage of providing information on the date of hospitalization, thus allowing us to consider only hospitalizations occurred during the months of the virus diffusion in the 2013 season. Moreover, in the analysis, we distinguish between planned hospitalizations (e.g. if requested by the GP) or emergency hospitalizations, occurred through access to the Emergency Rooms. The analysis on health outcomes relies on the estimation of Eq. (1) with the probability of being hospitalized during the 2013 flu season as dependent variable. As pointed out in Section 3 , the effect on health outcomes cannot be unambiguously attributed to the increase in vaccination take-up documented in the previous sections, but should be considered as the overall effect of the vaccination program, including indirect and spillover effects. For instance, we cannot rule out that individuals eligible for free vaccination can be more likely to receive, either from the GP, or when attending the health facility to get the shot or from other peers, additional information on non-pharmaceutical measures, which can be used to decrease the likelihood of flu infection. Thus, any potential health effect documented in this section should be interpreted as an Intention-To-Treat (ITT) effect of the flu vaccination program. 37 Table 6 presents the results of the effects of eligibility for free vaccination on hospitalization outcomes using a nonparametric analysis. The results in Panel A show a negative relationship between eligibility for free vaccination and the overall probability of hospitalization, which is not statistically significant at conventional levels. The results in Panel B report a null effect for planned hospitalizations. The estimated coefficients are again negative for the probability of an emergency hospitalization, and become statistically significant when a polynomial of order one is used (Panel C). The fact that the program seems to induce a reduction in emergency hospitalizations only, supports the idea that the influenza virus in the elderly population can lead to complications for which patients require immediate access to the hospital. The estimate reported in Panel C, Column 4 of Table 6 indicates a reduction in the probability of emergency hospitalization by 1.4 percentage points, which correspond to 90% of the average hospitalization rate in the control group. The effect is large, especially considering that not all hospitalizations are due to flu infection or complications. However, it is worth emphasizing that these estimates have relatively wide confidence intervals due to sample size, which is too small to allow us Fig. 5 . Analysis on the probability of hospitalization for placebo age groups before and after age 65. Notes : the figure reports RD robust estimates with Triangular Kernel and Coverage Error Rate optimal bandwidth selector ( Calonico et al., 2014; and the corresponding 90% confidence intervals, from regressions on placebo ages between 63 and 67 using a polynomial of order 1, without covariates. For each estimation, the older cohorts plays the role of the placebo treated group, the younger cohorts that of the placebo control group. The circles indicate RD robust estimates from placebo ages, while the black diamonds refer to the age cutoff used in our analysis. For the set of covariates included in the estimations and their definitions, see the footnote to Table 1 . Source : own elaborations on administrative data from Milan MA. to estimate the effects precisely for rare outcomes such as hospitalization. 38 We shall also acknowledge that the magnitude and the significance of the effect appear to depend on the model specification. 39 However, in what follows we present the results from a battery of sensitivity checks, which still supports the suggestive evidence of a reduction in the probability of emergency hospitalization. First, we perform a placebo analysis, with the same placebo ages used for the placebo analysis on vaccination takeup: Fig. 5 confirms that the only statistically-significant decline in the probability of emergency hospitalization occurs for individuals aged 65, and not for the placebo ages. Second, we present the results on hospitalization measures net of cases of fractures, traumas, and sprains, which are less likely to be related to the flu. Table 7 reports the results and confirms that the most significant decline in hospitalization probability occurs for emergency accesses to the hospital. Third, we check the robustness of the results on emergency hospitalizations, by performing the same checks that we have performed for the vaccination measure. Table A .11 in Online Appendix shows that the results are qualitatively similar if we cluster the standard errors at the birth date, municipality or GP level (Panel A) or if we change the donut specification (Panel B). In this paper, we analyze the effects of a vaccination program implemented in Italy, as in several developed countries, which actively provides free flu vaccination to individuals aged 65 or more. We estimate that the program induces an immediate increase in vaccination take-up that ranges between 70% and 90% of the average vaccination rate of individuals aged less than 65. Our research design makes it possible to identify only the local effect of the program, while it does not allow us to quantify any persistent effect, which would make the overall effect of the policy plausibly larger. 40 We also study the mechanisms driving the increase in vaccination take-up. We find little evidence of an income effect, as low-income individuals only respond to the program in case they are affected by poor health conditions, and thus value 38 Unfortunately, we do not have information on GP visits or other more common health-related outcomes. 39 The results from a parametric analysis, reported in Table A .10 in the Online Appendix, show that the negative and statistically significant relationship between eligibility for free vaccination and the likelihood of emergency hospitalization remains only when a bandwidth of one month is considered. In this case, the estimated effect on emergency hospitalizations is in line with the estimates reported in Panel C, Column 4 of Table 6 . 40 As suggestive evidence, in the administrative data from Milan MA we observe that 80% of the 66-year-olds, in their second year of eligibility for the program, had the flu vaccination for the first time the year before. ( Calonico et al., 2014; ; for the list of covariates included and their definitions see the footnote to Table 1 . The vaccination rate for the non-treated individuals (i.e. those aged 64) is reported in square brackets. Significance levels: * * * p < 0.01, * * p < 0.05, * p < 0.1. Source : own elaborations on administrative data from Milan MA. the expected benefits associated with the immunization. Individuals without any cost-sharing exemption also increase their vaccination take-up significantly once they reach age 65: as these individuals are less likely to respond to a change in the out-of-pocket price of the vaccine, we interpret this as evidence that the nonmonetary cost reduction implied by the program also matters for the vaccination decision. Awareness of the positive externalities associated with the flu immunization seems also to trigger vaccination decisions, as individuals in large families are the ones increasing more their vaccination rate once they reach age 65. We evaluate the effects of the free vaccination program on individuals' short-term health status, measured by the likelihood of hospitalization during the same flu epidemic season. We find that eligibility to free vaccination at age 65 induces a reduction in the likelihood of emergency hospitalizations only, which supports the link between the influenza contagion and the occurrence of complications, for which the elderly need immediate care. Given that in most developed countries, including Italy, the vaccination rate of the elderly population is still far from reaching the 75% target recommended by WHO, our work bears important policy implications for the effectiveness of flu vaccination programs. Our results show that the price of the vaccine does not seem to be the main barrier to reaching the target. Individuals seem to respond also to the expected benefits associated with the vaccination, both on an individual and on a collective dimension. Thus, in addition to free provision, policymakers should consider other leverages, such as more effective information campaigns, that highlight the benefits of a higher vaccination rate in the population, or the reduction in nonmonetary time cost granted by the easier access to vaccination and GP's assistance. Our findings have become even more relevant with the Covid-19 pandemic, as policy makers and health authorities are now preparing for subsequent waves of the epidemics, which may coincide with the 2020 flu season ( CDC, 2020 ). In such a scenario, it is of paramount importance to increase the influenza vaccination rate, especially for the at-risk categories, in order to reduce the impact of respiratory illnesses in the population and the resulting burden on the healthcare system ( Belongia and Osterholm, 2020 ) . Effectiveness of state-level vaccination mandates: evidence from the varicella vaccine Economic activity and the spread of viral diseases: evidence from high frequency data Anatomy of a health scare: Education, income and the MMR controversy in the UK Heaping-induced bias in regression-discontinuity designs Covid-19 and flu, a perfect storm Attention, intentions, and follow-through in preventive health behavior: field experimental evidence on flu vaccination On the effects of bias estimation on coverage accuracy in nonparametric inference Robust nonparametric confidence intervals for regression-discontinuity designs The impact of nearly universal insurance coverage on health care utilization: evidence from medicare Does medicare save lives? Direct and spillover effects of Middle School vaccination requirements Vaccination guidance during a pandemic The effect of state insurance mandates on infant immunization rates Focus group interviews on racial and ethnic attitudes regarding adult vaccinations Seasonal influenza vaccination in Europe. Vaccination recommendations and coverage rates in the EU Member States for eight influenza sea Seasonal influenza vaccination and antiviral use in EU/EEA Member States Annual Epidemiological Report for 2017. European Centre for Disease Prevention and Control Effectiveness of pandemic and seasonal influenza vaccine in preventing pandemic influenza A(H1N1)2009 infection in England and Scotland Stagione influenzale 2013/2014. Sorveglianza epidemiologica Effectiveness of vaccination recommendations versus mandates: evidence from the hepatitis A vaccine Regression discontinuity inference with specification error Manipulation of the running variable in the regression discontinuity design: a density test It'll only hurt a second? Microeconomic determinants of who gets flu shots Social determinants of health and seasonal influenza vaccination in adults ≥ 65 years: a systematic review of qualitative and quantitative data Effectiveness of influenza vaccine in the community-dwelling elderly Complications of viral influenza Prevenzione e controllo dell'influenza: raccomandazioni per la stagione 2013-2014 Vaccinazione antinfluenzale in Italia: coperture vaccinali nella popolazione generale The Dutch influenza vaccination policy and medication use, outpatient visits, hospitalization and mortality at age 65 Influenza vaccination campaigns: is an ounce of prevention worth a pound of cure? The authors declare that have no relevant or material interests that relate to the research described in this paper. Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.jebo.2020.09.010 .