International Political Alignment During the Trump Presidency: Voting at the UN General Assembly

We examine voting behavior of Western allied countries in line with the United States over the period 1949 until 2019. Descriptive statistics show that voting in line with the United States on resolutions in the United Nations General Assembly (UNGA) was on average 7.2 percentage points lower under Donald Trump than under the preceding United States presidents. The policy shift is especially pronounced for resolutions dealing with the Middle East. The decline in common UNGA voting behavior is significant for the resolution agreement rate and the absolute difference of ideal points. The results suggest that the alienation of Western allies is not driven by ideological distance based on a classical leftwing-rightwing government ideology scale.

A substantial literature has examined the impact of stress during the early stages of life on later-life health. This paper contributes to that literature by examining the later-life health impact of stress during adolescence and early adulthood, using a novel proxy for stress: risk of military induction in the United States during the Vietnam War. The paper estimates that a 10 percentage point (2 standard deviation) increase in induction risk in young adulthood is associated with a 1.5 percentage point (8 percent) increase in the probability of being obese, and a 1 percentage point (10 percent) increase in the probability of being in fair or poor health later in life. These findings do not appear to be due to cohort effects; the associations exist only for men who did not serve in the war, and are not present for women or men who did serve. These findings add to the evidence on the lasting consequences of stress, and indicate that induction risk during the Vietnam War may, in certain contexts, be an invalid instrument for education or marriage, because it appears to have a direct impact on health.

Introduction
Many studies have estimated the long-term consequences of in utero exposure to stress and illness (see e.g. Almond and Currie 2011;Aizer et al., 2016). In contrast, the long-term consequences of stress experienced during late adolescence and early adulthood has received relatively little attention, despite that this is an important developmental time for the brain (NIMH, 2011;Kaestner andYarnoff 2011, Dahl 2004).
In this paper, we estimate the long-term health consequences of stress during adolescence and early adulthood. We use a novel measure of stress: the risk of being inducted into the military during the Vietnam War. 3 We exploit the substantial year-to-year variation in induction risk between 1955 (after the Korean War) and 1973 (when the US switched to an all-volunteer force) that resulted from both changes in demand for military manpower and changes in which ages were eligible for the draft. We focus on the entire period of the war , not just the period of the draft lottery , which is important because the pre-lottery years were a particularly morbid and lethal time for American soldiers. 4 We argue that the prolonged uncertainty about whether one would be conscripted into military service between the ages of 18½ and 26 caused large amounts of stress on young men. Highly publicized calls for increased troop levels in Vietnam and increasingly large numbers of U.S. servicemen being killed or wounded were well documented by media outlets, which made this risk particularly salient.
We focus on men who did not serve in the military during the Vietnam War; for these men, we can accurately estimate their lifetime exposure to induction risk. For men who did serve, we do not know their induction risk because we do not know at what age they were inducted, nor whether they volunteered or were drafted. Moreover, those who served faced much greater stress (e.g., due to combat) that is unrelated to their induction risk and is unquantifiable by us.
We find consistent evidence that early adult stress, as measured by induction risk, is associated with worse later-life health. A 10 percentage point increase in induction risk, which is a 2 standard deviation increase, is associated with a 0.2 unit (0.08%) increase in BMI, a 1.5 percentage point (1%) increase in the probability of being obese, a 1 percentage point (1%) increase in the probability of being in fair or poor health and a 2 percentage point (0.3%) decrease in the probability of being in very good or excellent health. This does not appear to be due to cohort effects or other trends; these associations exist only for men who did not serve in the military during the Vietnam War, not for men who did serve in the military or women.
In Section 2, we review the relevant literature. Section 3 describes the data and in Section 4 we describe our methods. We present our results in Section 5 before concluding in Section 6.

Early-Life Stress and Later-Life Health
A large and growing literature in economics estimates the effect of early life stressors on a wide variety of later-life outcomes, including educational attainment, earnings, and health (see Almond and Currie 2011;Aizer et al., 2016;Camacho, 2008). This literature generally focuses on stressors in utero or during infancy, but other research, examining the long-term effects of leaving school during a recession on later life health, has found significant negative effects on health for adult men (Maclean 2013, Maclean et al. 2016. Another line of research seeks to 5 including psychological and biochemical responses, as well as behavioral changes that may be initiated through these channels (Schneiderman et al. 2005).
Exposure to stress leads to biological responses from the nervous, cardiovascular, endocrine, and immune systems (Schneiderman et al. 2005). Stressors lead to the release of hormones such as epinephrine and cortisol, which increase sources of energy through higher blood sugar and the breaking down of fats into useable energy (Schneiderman et al. 2005, Sapolsky et al. 2000. The body diverts this energy to tissues that become more active during stress (skeletal muscles and brain) and away from less critical activities like eating, growth, and sexual activity (Schneiderman et al. 2005). This temporarily increases blood pressure through increased heart rate and stroke volume (the amount of blood pumped with each beat). Chronic mobilization of these processes is associated with high blood pressure, cardiac hypertrophy (the thickening of the heart muscle, which reduces the heart chamber size), damaged arteries and plaque formation, suppressed immunity including atrophied wound healing, slower surgery recovery, higher susceptibility to viruses including upper respiratory infections, worse antibody response to vaccines, and increased inflammation which can exacerbate many of the aforementioned conditions (Schneiderman et al. 2005, Sapolsky et al. 2000. Importantly, stress can lead to psychological problems like anxiety and depression, which are associated with sleep problems, substance abuse, heavier cigarette consumption, and higher alcohol consumption (Schneiderman et al. 2005).
Additionally, stress affects individuals' eating patterns.  find chronic stress, as measured by hair cortisol levels, to be associated with numerous measures of increased body weight, including persistence of obesity. A potential pathway through which stress affects obesity is by diminishing an individual's self-control leading to non-utility maximizing, 6 potentially impulsive food decisions like eating more energy-dense, high-calorie foods (Ruhm 2012, Torres andNowson 2007).

The Effects of Military Service on Later-Life Outcomes
Researchers have studied the effect of military service on labor market outcomes (Angrist 1990), education outcomes (Card and Lemieux 2001), and violent crimes (Lindo and Stoecker 2014). See Dobkin and Shabani (2009) for a discussion of the health effects of military service, and Gimbel and Booth (1996) for a discussion of stress among veterans. However, we focus on men who did not serve in the military, so the mechanism of actual military service is not relevant.

The Effects of Draft Avoidance Behaviors on Later-life Outcomes
For certain periods of the Vietnam War, otherwise draft-eligible men could receive deferments for being enrolled in college; for this reason, college attendance became a draftavoidance behavior. This led to an increase in college attendance, and several studies have used induction risk as an instrumental variable for educational attainment (see e.g. Card and Lemieux 2001), 5 in some cases to estimate the effect of college education on later-life smoking (de Walque 2007, Grimard andParent 2007) or mortality (Buckles et al. 2016). This literature suggests that draft risk is associated with increased education and improved health.
In certain years, deferments were also available for men who were married or had children. As a result, draft avoidance behaviors also included getting married and having children (Kutinova 2009, Hanson 2011. Kutinova (2009) finds evidence of increased first births occurring between 8 and 10 months following an executive order ending marital deferments for childless men because it created an incentive to have children to qualify for the deferment that 7 remained available for fathers. Hanson (2011) finds that young men married at younger ages during periods in which marital deferments existed. While a large literature supports the notion that married individuals are generally healthier than single individuals, this is not necessarily a causal relationship (Wood, Goesling, and Avellar 2007).
To address this mechanism, our models control for education, marital status, and number of children. We are not able to control for age at first marriage or age at first birth, so our controls for these variables may be imperfect. However, to the extent that education and marriage promote health, failure to perfectly control for them should bias our estimates toward finding that exposure to induction risk is associated with better health later in life.

Data
The primary data used in this project are: 1) the National Health Interview Survey (NHIS) and 2) induction data taken from Reports of the Director of the Selective Service. In this section, we describe these data and document how we constructed the variables used in our analysis.

3.1.National Health Interview Survey (NHIS)
The NHIS is a cross-sectional, nationally representative data set of the US noninstitutional population. Each year, the NHIS interviews approximately 100,000 individuals, asking questions about basic demographic and socioeconomic characteristics, military service, height and weight (from which we calculate body mass index or BMI), 6 and self-reported health.
We use data from the 1982-1996 NHIS for our main analysis sample. Based on our sample of interest -men born from 1937 to 1956 -our main analysis sample is aged 25 to 59 during 1982-8 1996. The sample size varies between 119,000 and 140,000 depending on the regression model specification, due to missing values for certain variables.
Because height and weight measures come from self-reports, they likely suffer from reporting error (Bound et al. 2001;Cawley et al., 2015). We adjust these measures for reporting error using data from the National Health and Nutrition Examination Survey (NHANES) III (1988)(1989)(1990)(1991)(1992)(1993)(1994), which include both self-reports and measurements of weight and height and are from a similar time period as our NHIS data (Burkhauser and Cawley 2008). 7 From 1982-96, NHIS respondents were asked whether they had ever served in the US military and, if so, during which period (e.g. World War II, Korean War, Vietnam War). We do not use NHIS data from before 1982 because in those earlier years the NHIS data do not include month of birth, which we use to construct a more accurate measure of induction risk. We do not use NHIS data from after 1996 because in those later years the NHIS only asked whether respondents had been "honorably discharged" from the military, not whether they had served during the Vietnam War era (or served at all but without an honorable discharge). We focus only on the years in which we have the most complete military data , however we extend our sample through 2012 as a sensitivity analysis.
Our sample is limited to individuals born in the calendar years 1937 to 1956. This is to avoid any contamination of our main analysis sample with individuals serving in the Korean War, which ended in July 1953. Thus, any individual born in 1937 would not have been eligible 7 Using the NHANES data for the same age range as our NHIS data (age 25-59), we estimate the following regressions separately by gender: ℎ = + + + + (1) Where SelfReport refers to self-reported weight, and X includes age and race (non-Hispanic Black, non-Hispanic other race, and Hispanic). We then save the constant and coefficient estimates from these regressions and create an adjusted measure for weight in the NHIS sample by multiplying the NHIS self-reported weight values and demographic characteristics by their coefficients from NHANES regressions, and adding them together, along with the constant from the regression model. We perform a similar adjustment for self-reported height. Using these constructed values, we create an adjusted measure of BMI and obesity (BMI≥30) which we use throughout the rest of this manuscript. We also estimate all models using the non-adjusted BMI and obesity measures and find slightly larger and more statistically significant results using unadjusted variables. to serve in Korea. de Walque (2007) also imposes this sample restriction. 8 Because we are interested in educational attainment, we limit our sample to those over age 25 at interview as most individuals have completed their schooling by this age. We further restrict our sample to those with valid measurements of height and weight because we use these to create BMI. 9 In certain analyses, we use a restricted-use version of the NHIS which provides us with data on state of birth (blinded), 10 whether and in which specific year an individual died if he or she died by 2011 (NHIS-Mortality Linked File), and his or her age of death. These mortality data were collected by matching NHIS participants annually to the National Death Index system death certificates by a combination of social security number, gender, first and last name, date of birth, or birth month and year.

3.2.Selective Service Reports
We use annual (1955)(1956)(1957)(1958)(1959)(1960)(1961)(1962)(1963)(1964)(1965)(1966) and semi-annual (1967-1975) reports of the Director of the Selective Service for induction eligibility, the number of monthly inductions, as well as additional institutional details and troop level data. These induction data are available by month, which we use to construct our measure of induction risk. We calculate risk based on the relevant laws governing both the pre-lottery years and the lottery years of the draft (see Appendix for more details on the history of the draft). We calculate induction risk as the number of inductions in a given 12 month period, using month and year of birth and monthly induction numbers from annual reports to construct a measure of induction risk by month of birth, divided by the number of individuals at risk or: Where RAm,y is the risk at age 18 for those of age A born in month m in year y. We calculate the denominator, Size of Cohort at Risk, using the induction reports for the official number of individuals who have registered for selective service and are between the ages of eligibility (e.g. 18½ to 26) for the given period. 11 We calculate risk separately for ages 18½, 19, …, 26, which we then sum to calculate induction risk from 18½ to 26. We focus on 18½ to 26 rather than college years (e.g. 19-22), which was the focus of previous research (Card and Lemieux 2001, de Walque 2007, Grimard and Parent 2007, Buckles et al. 2016, because ages 18½ -26 were the actual, legislated ages at which an individual was at risk of induction. 12 We calculate the total risk as follows, taking into account that one remains at risk of induction only if one was not previously inducted: 11 We also perform this calculation using measures of the cohort of age 17 year olds as reported by the Department of Education (and creating measures of those between the ages of 18½ -26, by assuming the cohort of 17 year olds in a given year will be the correct number of 18 year olds in the following year). Our results are robust to these multiple methods, so we use the official number of individuals registered via the selective service throughout. 12 For years in which the draft lottery system was in place, we only consider eligibility for those in the birth cohorts at risk.
Where R18 denotes risk at age 18½, while all other measures denote risk for the full year a person is a given age, for month and year of birth m,y. For years in which the US used a draft lottery system, we calculate ex ante risk (i.e., the risk prior to the resolution of the lottery): for those subject to the lottery, we set their risk of induction as equal to the number of inductions divided by the number of men of draft-eligible age in that year. 13 We multiply induction risk by 100 in all specifications so that all coefficients can be interpreted as the association of the outcome with a one-percentage-point change in risk.

Methods
We estimate a reduced-form model of later-life health as a function of cumulative induction risk during the Vietnam War. We estimate an ordinary least squares model of the form: Where is the dependent variable denoting various later-life health outcomes, including a continuous variable for BMI and an ordinal variable for self-reported health (1 -5, with 1 denoting poor and 5 denoting excellent health). 14 We estimate probit models for the binary dependent variables obesity (BMI≥30), fair or poor health (self-reported health ≤ 2), and very good or excellent health (self-reported health ≥ 4) of the following form: For each of these models, are demographic and human capital characteristics described below, δ are fixed effects for age at interview, γ are fixed effects for year of interview, and ε is an error term. We report marginal effects from the probit models. We limit our main analysis sample to men who did not serve in the military during the Vietnam War Era. 15 In all specifications, we cluster standard errors at the birth year cohort level.
We estimate three models, adding progressively more covariates in each specification.
We start with the most exogenous set of regressors and then add regressors that control for ways in which induction risk could affect later-life health (e.g. education, marital status). We interpret any remaining association of induction risk with later-life health, after controlling for education and marital status, as the result of stress, although we acknowledge that we cannot measure stress directly.
The basic specification is limited to the most exogenous regressors: race (non-Hispanic Black, non-Hispanic other race, and Hispanic, with non-Hispanic Whites the omitted reference group), a birth cohort specific trend, defined as birth year minus 1937, and indicator variables for age and year. Specification 2 adds covariates for family size and marital status at the time of the NHIS interview. These additional covariates reflect the fact that the Vietnam War has been 14 We also estimate regressions with self-reported health as the dependent variable using an ordered probit and find qualitatively similar results; the results are available upon request. 15 We can calculate the aggregate probability of induction for men who did not serve in the military. We cannot do this for men who did serve -we do not know when they enlisted or whether they were drafted or in what year they were drafted. Moreover, the relevant measure of stress for them is likely not their risk of induction but their experiences during the war, about which we have no information.
linked to increased marriage and fertility rates (Kutinova 2009, Hanson 2011. While stress may affect later life health directly, in the form of induction risk, stress will also affect health through an increased likelihood of marriage and having children. Specification 3 additionally includes years of education past high school, and log income, which are endogenous variables but possible mechanisms through which induction risk may affect health (see e.g. Card and Lemieux 2001).
We also investigate the effect of stress in early adulthood on the risk of mortality, using a restricted-use version of the NHIS that contains data on mortality. Using a Cox proportional hazards model, we estimate time to death for men who never served during the Vietnam War.
This equation takes the form: where ( ) is the hazard of dying by 2011, the year in which the NHIS-Mortality linked file follows individuals. Induction risk and other covariates are exactly the same as described above and we perform the same three specfications.
As a falsification test, we also estimate our models for male veterans. Men who served in the military during the Vietnam War at some point joined the military (we do not know whether they voluntarily enlisted or were drafted, or at what age they joined the military) and were then no longer at risk of being inducted. Moreover, those who served in the military during the Vietnam War faced many other stressors during these ages (e.g. combat, separation from loved ones, injury, deaths of fellow soldiers); we have no way of quantifying this stress but there is no reason to believe that it is correlated with induction risk for others of the same age who did not serve.
14 As an additional falsification test, we estimate models for women. Women were at zero risk of induction throughout this era, so as a falsification test we assign them the induction risk experienced by men of the same month and year of birth. This is a strong falsification test in that women of this age cohort may have experienced stress due to the risk that their brothers, boyfriends, or husbands might be inducted. However, the induction risks of those people would be based on their own birth month and year, not the birth month and year of the woman. Still, women may have experienced stress about their classmates, who are of similar age.
If we find a similar correlation between induction risk and later-life health for these placebo samples (veteran men, women) as we find for our main sample of men who did not serve in the military during the Vietnam War, that would suggest that the correlation is due to cohort effects or other trends at the time. In contrast, if we find a considerably different pattern between induction risk and later-life health for non-veteran men than for the placebo samples, that would be consistent with stress affecting later-life health. Falsification tests are never conclusive; one cannot prove the null hypothesis of no bias, but failure to reject the null is informative.

4.1.Sample Selection
A potential concern with this study is that there will be sample selection on health status in who serves in the military based on the level of induction risk. This selection arises because of potential composition effects caused by variation in military force needs. As the need for more soldiers increases, the induction risk also increases, leading to more inductions and thus potentially culling more of the healthier individuals into service. In this case, during periods in time in which induction risk was particularly great, the health composition of men not inducted into the armed forces might be slightly lower, which would bias us towards finding a correlation of induction risk on health. We look for evidence of this in the results of the regression models estimated for veterans; if induction takes place in order of healthiness, then when induction risk is higher, it should be associated not only with lower average health among the non-Vietnamveterans, but also lower average health among the veterans, who had less healthy individuals join their ranks.
Premature mortality due to stress is another potential source of selection bias. If those who were severely stressed by the possibility of being inducted into the military during the Vietnam War were more likely to die before our sample period, then our analyses would be biased away from finding that stress worsens later-life health. We explore this possibility by estimating models of mortality for the time period we can observe.
Finally, those with higher socioeconomic status may have been more adept at avoiding military service either through family connections or greater resources. If higher socioeconomic status individuals tend to have better later-life health (Grossman, 2015), and were more likely to avoid military service during periods of increased induction risk, then our estimates would be biased against finding a negative health effect of induction risk later in life. However, those of higher SES may have consistently used any advantages to avoid military service; i.e. that such avoidance did not vary with the overall induction risk.

Empirical Results
We report summary statistics in Table 16 For women, true induction risk is zero throughout the war, but in Table 2 the induction risk listed for women is that for men of the same birth month and year; this is used as a falsification test later in the paper.  year as calculated using equation (3). The extreme variation and non-monotonicity in induction risk is very useful, as it implies that our regressor of interest is unlikely to be correlated with omitted variables such as general trends in health. Table 1 shows the years and ages at which a birth cohort was at risk of induction and provides a clearer picture of the source of variation in risk due to the number of years a given birth cohort was at risk of military induction. It also separately identifies the type of induction 16 We remove those individuals who served in the military at any time in a subsequent sensitivity analysis, but these men are included in our main sample because they may have been stressed by the risk of being inducted during the Vietnam War.
system being used during the cohorts' age-eligible years, with ages in italics denoting that the draft lottery system was in place from 1970-1972. Table 3 presents results for men who did not serve during the Vietnam War, controlling for time trends, race and ethnicity, and age and year of survey fixed effects, while Table 4 also includes controls for family size and marital status at the time of interview. Results from these models are similar, so we discuss only the Table 4 results. Among male non-Vietnam-veterans, a 10 percentage point (2 standard deviation) increase in induction risk is associated with a 0.5 percentage point (3%) increase in obesity, a 0.7 percentage point (8%) increase in fair or poor health and a 0.5 percentage point (1%) decrease in the probability of being in very good or excellent health. All these results are statistically significant at least at the 10 percent level.

5.1.Results: Men Who Did Not Serve in Vietnam
In Table 5 we also include controls for years of education completed beyond high school and log income. This is an important change to the model, because college education was a draft-avoidance behavior, and educational attainment is consistently associated with better laterlife health (see, e.g., Cutler and Lleras-Muney, 2010). The coefficient on induction risk becomes larger and more statistically significant in this model. A 10 percentage point (two standard deviation) increase in induction risk is associated with a 0.2 unit (1%) increase in BMI, a 1.5 percentage point (8%) increase in the probability of obesity, a 0.05 unit (1%) decrease in selfreported health, a 1 percentage point (10%) increase in the probability of being in fair or poor health and a 2 percentage point (3%) decrease in the probability of being in very good or excellent health. Income and education are both associated with better self-assessed health, and education is also associated with a lower probability of obesity (income is negatively correlated with obesity, but it is not statistically significant). Table 6 presents results from a Cox proportional hazard model of mortality following equation (6). Panel A reports results for men who did not serve in Vietnam, our main analysis sample. The results of specification 1 and specification 2 (which adds controls for marital status and family size) suggest a slight, statistically significant protective role of induction risk on the hazard of death. However, in specification 3, when we include educational attainment and income variables, this association becomes smaller and statistically insignificant.

5.2.Mortality
The negative correlation between exposure to stress and later-life mortality is the opposite of what one would expect. In the next section, devoted to falsification tests, we will provide evidence that suggests that these mortality results, but not the results for weight and selfassessed health, are due to cohort effects.

5.3.Falsification Tests: Male Vietnam War Veterans and Women
In Table 7 we present results for the same models that were estimated using samples of:  for the placebo samples of veteran men and all women as we found for our sample of interest who are non-veteran men. In fact, the correlation for the placebo samples was generally of opposite sign; whereas we found that non-veteran men exposed to greater induction risk are in 20 worse later-life health, the opposite was true for the placebo samples. The results of these falsification tests yield no evidence that the results for the sample of interest are due to cohort effects or unobserved trends.
The results of the falsification test do, however, raise the question of why there is a positive correlation for the placebo samples. It is possible that there are birth cohort effects that are visible in the placebo samples and that cause attenuation bias in our sample of interest.
For example, previous research (e.g. Robinson et al. 2012) has documented evidence of birth cohort effects on abdominal obesity, especially among women, which may partially explain the positive association between health and induction risk for women, although we find very little consistent evidence of an association between BMI or obesity and induction risk for women or male veterans in Table 7. Overall, the results of these falsification tests suggest that the negative association observed for male non-Vietnam-veterans between induction risk and worse later-life health is not due to a cohort effect or trend in unobserved variables. Although falsification tests cannot be definitive, these are consistent with stress having an adverse impact on later-life health.
The falsification tests for mortality yield a different result. Results of hazard models of mortality are listed in Table 6, Panel B for veterans, and Table 6, Panel C for women. For both placebo samples, induction risk has a similar correlation with later-life mortality as it has for our main analysis sample of male non-veterans. For all three groups, higher induction risk is associated with a lower probability of mortality. In fact, the point estimates are larger for the two placebo samples than for the main analysis sample of male non-veterans. Based on the results of these falsification tests, we conclude that the correlation is not an effect of exposure to stress but is instead a cohort effect.

State Fixed Effects
In Appendix Table A1, we present results from a specification similar to Tables 3 -5 would expect from adding men to the sample whose stress is not well captured by the average induction risk). Overall, we continue to find evidence that stress, in the form of induction risk, worsens later-life health.

Variations in Sample Inclusion Criteria
We also examine the sensitivity of results to variation in the sample inclusion criteria.
These results are presented in Appendix sample. In Panel E, we limit the sample to those born between 1930 and 1936 who did not serve in the military, but were age-eligible to serve during the Korean War. These results are generally similar to those of the main specification; exposure to a higher induction risk tends to be associated with a higher probability of obesity and worse self-assessed health.

Model Estimates for Veterans of Various Wars
We estimate additional falsification tests, using data for male veterans of wars other than Vietnam in Appendix Table A4. The results generally confirm that veterans, no matter which war they served in, do not exhibit the negative correlation between induction risk (that was found for men who did not serve) and later-life health; the exception is that Korean veterans exhibit a positive correlation between induction risk and later-life BMI and obesity.

Alternate Measure of Stress: Risk of Being Killed in Action
We also estimate models using an alternate measure of stress. Specifically, the new proxy for stress is the interaction of induction risk with the probability of being killed in action, calculated as the number of deaths in a calendar year divided by the level of American troops in Vietnam in that year (de Walque 2007). The logic is that stress is raised not simply by the probability of serving in the military but also by the chances that one could be killed. We create this interaction in a similar fashion to equation 3. We present results of these estimations in Appendix Table A5, Panel A. These results are qualitatively similar to those in the main specification: greater stress (proxied by higher risk) is associated with a higher probability of obesity and worse self-assessed health.

Alternate Measures of Expected Induction Risk
We also explore other plausible proxies for stress. It may be that individuals are not able to perfectly forecast their risk of induction; instead, they may observe the rates of induction for the cohorts ahead of them and assume that they will face similar risks. To explore this possibility, we estimate our base model, but assigning individuals the induction risk of cohorts that came of military age 1, 2, or 3 years before them. In Appendix Table A6, the results suggest that the risk of military induction of one's own cohort has the largest effect on later life health.
The size of the effect on health using risk of induction for older cohorts tends to decrease with the distance of the cohort, suggesting previous induction risk affects later cohorts but that they update their own risk calculations as they get closer to induction age.

24
We find evidence consistent with the hypothesis that stress in adolescence and early adulthood worsens later-life health. Specifically, a greater risk of military induction during the Vietnam War is associated with worse later-life health for men who did not serve in the military.
These adverse health effects are modest, but are relatively robust, and include a higher risk of obesity and higher probability of reporting being in worse health. Falsification tests conducted with samples of women and male veterans do not show the same pattern for obesity and selfassessed health, which suggests that those results are not due to cohort effects or trends in unobservables. Moreover, it does not appear to be due to compositional effects (i.e., more healthy men being removed from the ranks of non-veterans when induction risk is high), because the negative correlation is still detectable (though, naturally, weaker) in a sample in which the non-veterans are pooled with veterans. The finding that stress is associated with a greater risk of obesity and worse self-assessed health in later life is robust to alternate proxies for stress (such as the probability of being killed in action) and various sample inclusion criteria.
For the outcome of mortality, however, the falsification samples exhibit the same patterns as the main analysis sample, which leads us to conclude that the correlation between induction risk and later-life mortality is due to cohort effects or other omitted variables.
This study has a number of limitations. First, we cannot directly measure stress. No largescale survey that we know of has measured the stress of individuals using cortisol laboratory measurements at the time of the Vietnam War. However, the measure we use, which we limit to those who did not serve in Vietnam, captures an important source of uncertainty faced by individuals at that time. For those who do not serve, we do not have baseline health measures to investigate whether they did not serve due to preexisting health conditions that would bias these individuals towards worse health later in life. However, this should not bias our results unless 25 health varies by month and year of birth in a way that is correlated with induction risk, which seems unlikely.
Among veterans, we cannot distinguish those who volunteered immediately for service from those who were inducted after experiencing the stress of the draft. However, this should bias the coefficient on risk in the veterans model towards showing ill effects of draft risk on later-life health, which we do not find for the veteran sample. Finally, we are limited in the laterlife health outcomes that we can examine by the questions that were asked in NHIS.
Overall, we find a consistent association between a proxy for stress and worse later-life health. One possible mechanism by which stress affects later life health is diet. Reviews have concluded that chronic stress is associated with eating energy-dense foods (i.e., foods higher in sugar and fat) and weight gain, with greater effects for men (Torres and Nowson, 2007).
Another possible mechanism is the stress hormone cortisol; it is believed that long-term hyperactivation of the system that regulates cortisol can contribute to the development of obesity and metabolic syndrome, which includes hypertension, diabetes, and high cholesterol . These findings are consistent with our estimates linking stress in young adulthood to obesity and worse self-reported health later in life.
This study contributes to the literature on the adverse later-life consequences of stress, most of which has focused on the effect of stress or insults at the earliest ages (Almond and Currie, 2011;Aizer et al., 2016;Camacho, 2008). Specifically, this study establishes that even stress in adolescence and early adulthood may be associated with worse later-life health.
This paper also contributes to the previous literature on the effects of education on laterlife health. Specifically, while some previous studies used induction risk as an instrument for education, the results of this paper suggest that in some contexts it may be an invalid instrument, 26 as induction risk may affect outcomes such as health directly, not only through the mechanism of education.
The policy implications of this study are that even systems that seek to allocate burdens in a fair and transparent way (e.g., draft lottery) may impose unanticipated costs on participants through the mechanism of stress. In our study, we show that this is true even of those who never serve in the military.

Appendix A. The Vietnam Conflict and the Risk of Induction into the U.S. Military
Following the French military departure from Vietnam in 1954, the US began to directly aid the South Vietnamese government, and US military advisors began training the South Vietnamese military in 1955 ("Chronology of Events Relative to Vietnam, 1954Vietnam, -1965Vietnam, ," 1965.
This growing influence in Vietnam did not affect US military inductions until the Gulf of Tonkin Incident in 1964 in which a US ship engaged the North Vietnamese navy in the Gulf of Tonkin ("Chronology of Events Relative to Vietnam, 1954Vietnam, -1965Vietnam, ,"1965. Following this incident, induction rose rapidly, peaking in October 1966, and again in the first half of 1968 before decreasing rapidly; see Figure 1. Prior to U.S. involvement in Vietnam, induction levels also were high in the second half of 1961 due to the building of the Berlin Wall, and in the early 1950s due to the Korean conflict.
The risk of fighting in the Vietnam War took on a particular salience unseen before due to its novelty as America's first "televised war" (Hallin 1986, pp. 105). Before the first official exchange of fire in 1964, US publications printed disturbing images of a Buddhist monk's selfimmolation in protest of the US backed Diem government (Hallin 1986 (Engelhardt 2007, Hallin 1984. As the U.S. death toll rose in Vietnam, political unrest in the US grew with student protests and draft card burnings (Appy 2003, Engelhardt 2007 Angrist, 1991). Voluntary inductees, on the other hand, served a shorter tour of duty with no choice of military occupation and while they were certain to serve, they controlled the timing of their service (Annual Report, 1955, Semi-Annual Report, 1968, Angrist, 1991. 18  Local boards conducted pre-induction examinations in order to have induction-eligible individuals ready when they received calls from the State Director. Approximately half of all examinees were cleared for service (Annual Report, 1966). An additional 20 percent were rejected for service after being delivered for induction (Annual Report, 1966;Angrist, 1991).
Those passing pre-induction medical exams were sent to Armed Forces induction stations.
A priority system dictated the order of induction; generally, delinquents (a classification of involuntary inductee) received the highest draft priority, followed by volunteers for induction. 19 The priority of involuntary induction after the two classes above varied during the Vietnam War period, with various marital, paternal, and student deferments created, modified, and repealed. The lexis chart in Table 1 summarizes which ages of young adults were eligible for induction by year, and Appendix Figures 1 and 2 give a brief summary of changes in the 19 Delinquency is defined as failure to comply with the Universal Military Training and Service Act. Examples are refusal to register, failure to supply board with information, failure to report for pre-induction examination, or failure to report for induction (Annual Report 1952). Any person of ages 18-26, under provision part 1630 of SS regulations, can offer themselves for induction at any point in time. Persons between age 17 and 18, with the approval of a guardian, also can volunteer for induction (Annual Report 1955).
induction system from 1948 through the end of the induction system and the beginning of an allvolunteer force in 1973. Our data do not allow us to determine an individual's eligibility for specific deferments, so they are not used in our calculation of induction risk.
Executive order 10659 provided a new order for inductions, one created to prevent older registrants (those 26 or older) and fathers from being high priority inductees (Annual Report 1955, pg. 27 The variation in probability of induction for our analysis sample is due to: 1) changes in the number of inductions by year that were driven by political decisions about the Vietnam War; and 2) changes over time in the ages that were eligible to be inducted or were subject to the draft lottery. 23  25-40 Source: Ages at which the birth cohorts were interviewed in the National Health Interview Survey are displayed in the right-most column. Ages in italics denote the draft lottery system was in place at this point in time (1970)(1971)(1972).   (4) and (5) in the text using 25-59 year old men who did not serve in the military during the Vietnam War Era from the 1982-1996 NHIS. Each column in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include age and year of interview fixed effects. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (4) and (5) in the text using 25-59 year old men who did not serve in the military during the Vietnam War Era from the 1982-1996 NHIS. Each column in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include age and year of interview fixed effects. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.   (4) and (5) in the text using 25-59 year old men who did not serve in the military during the Vietnam War Era from the 1982-1996 NHIS. Each column in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include age and year of interview fixed effects. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (6) in the text using 25-59 year old individuals from the 1982-1996 NHIS. Each cell in the table comes from a separate regression in which the main independent variable is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation (3) in the text. All models are estimated using Cox proportional hazards models. All estimates include age and year of interview fixed effects. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (4) and (5) in the text using 25-59 year old men who served in the military during the Vietnam War Era in Panel A, and women in Panel B from the 1982-1996 NHIS. Each cell in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include controls for race and ethnicity, time trend, and age and year of interview fixed effects. Specification 2 also includes controls for family size and marital status, while Specification 3 includes log family income and educational attainment controls as well. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (4) and (5) in the text using 25-59 year old men who served in the military during the Vietnam War Era in Panel A, and all men in Panel B from the 1982-1996 NHIS. Each cell in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include controls for race and ethnicity, time trend, and age and year of interview fixed effects. Specification 2 also includes controls for family size and marital status, while Specification 3 includes log family income and educational attainment controls as well. Naive regressions include a binary indicator for whether the individual served in the military during the Vietnam Era. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (4) and (5) in the text using 25-59 year old men who did not serve in the military during the Vietnam War Era from the 1982-1996 NHIS. Each cell in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. Birth cohorts included in the analysis are listed in Panel titles. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include controls for race and ethnicity, time trend, family size, and marital status and age and year of interview fixed effects. Specification 3 includes log family income and educational attainment controls as well. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (4) and (5) in the text using 25-59 year old men who served in the military during the Vietnam War Era from the 1982-1996 NHIS. Each cell in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. Birth cohorts included in the analysis are listed in each row. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation 3 in the text. All estimates include controls for race and ethnicity, time trend, family size, marital status, log family income, and educational attainment, and age and year of interview fixed effects. Standard errors clustered at the birth cohort level are in parentheses: *** indicates significance at the 1% level, ** indicates significance at the 5% level, and * indicates significance at the 10% level.  (4) and (5) in the text using 25-59 year old men who did not serve in the military during the Vietnam War Era from the 1982-1996 NHIS. Each cell in the table comes from a separate regression. Estimates for BMI and health are from OLS models while estimates for obese, fair or poor, and very good or excellent health are from probit models and coefficients represent marginal effects. Birth cohorts included in the analysis are listed in Panel titles. The main independent variable in all equations is the risk of being inducted into the army between the ages 18 1/2 and 26 based on equation