Essays on Veterans Disability Compensation and the Effects of Military Service by Kyle Greenberg ARCHIVES MASSACHUSETTS INSTITUTE OF TECHNOLOGY B.S. Mathematics United States Military Academy, 2005 OCT 152015 Submitted to the Department of Economics in Partial Fulfillment of the Requirements for the Degree of LIBRARIES Doctor of Philosophy in Economics at the Massachusetts Institute of Technology September 2015 2015 Kyle Greenberg. All right reserved. The author hereby grants to MIT permission to reproduce and to distribute publicly paper and electronic copies of this thesis document in whole or in part in any medium now known or hereafter created. Signature redacted Signature of Ai uthor: 'f Certified by: Signature redacted ,Signature redacted Department of Economics August 15, 2015 Joshua Angrist Ford Professor of Economics Thesis Supervisor - Certified by: Signature redacted David Autor Professor of Economics Thesis Supervisor Accepted by: Ricardo Caballero Ford International Professor of Economics Chairman, Departmental Committee on Graduate Studies 2 Essays on Veterans Disability Compensation and the Effects of Military Service by Kyle Greenberg Submitted to the Department of Economics on 15 August, 2015 in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy in Economics ABSTRACT This dissertation consists of three empirical studies, each using administrative data from the U.S. Army, the Department of Veterans Affairs (VA), and the U.S. Social Security Administration. The first chapter investigates the correlation between local labor markets and VA Disability Compensation (DC) receipt among National Guard veterans who deployed to a combat zone between 2003 and 2006. I find that veterans from hometowns with low employment-to-population ratios are more likely to receive DC for both PTSD and physical conditions than veterans from hometowns with high employment-to-population ratios, but this association is stronger for PTSD than it is for physical conditions. PTSD awards that result in monthly benefit payments of at least $1,500 account for most of the correlation between employment-to-population ratios and PTSD, while only physical awards that generate relatively low payments are associated with employment-to-population ratios. The second chapter, a joint project with David Autor, Mark Duggan, and David Lyle, analyzes the effect of the DC program on Vietnam veterans' labor force participation and earnings. Exploiting the 2001 Agent Orange decision, which expanded DC eligibility for Vietnam-era veterans who served in-theater but not for other Vietnam-era veterans, we assess the causal effects of DC eligibility by contrasting the outcomes of these two Vietnam-era veteran groups. We estimate that benefits receipt reduced labor force participation by 18 percentage points among veterans enrolled due to the policy, though measured income net of transfer benefits rose on average. The third chapter exploits enlistment test score cutoffs in a fuzzy regression-discontinuity (RD) design to evaluate the causal effect of military service on mortality for individuals who applied to the active duty U.S. Army from 1996 through 2007. Fuzzy RD estimates suggest that military service does not increase mortality, but I cannot rule out positive effects. Ordinary Least Squares (OLS) estimates that compare applicants who did not enlist to soldiers with similar characteristics provide additional evidence that military service does not increase mortality. OLS estimates also indicate that a soldier's first year in the Army is associated with substantial reductions in mortality relative to nonveteran applicants. Thesis Supervisor: Joshua Angrist Title: Ford Professor of Economics Thesis Supervisor: David Autor Title: Professor of Economics 3 4 Acknowledgements I am grateful to Joshua Angrist and David Autor for their invaluable guidance and support throughout my time in graduate school. Josh and David are incredible economists, inspiring teachers, and probably the hardest workers I know. Having the opportunity to learn from them has not only influenced this dissertation, but also shaped my approach to research. I would also like to thank several other faculty members who always found time to offer me valuable advice, including Jim Poterba, Heidi Williams, Daron Acemoglu, Amy Finkelstein, Michael Greenstone, and Parag Pathak. I would not be writing this were it not for the help and friendship of my fellow graduate students. Alex Bartik, Brendan Price, and Richard McDowell continue to amaze me with their knowledge, generosity, and humor. I generated many ideas for research while talking to Sally Hudson, Ashish Shenoy, and Miikka Rokkanen, often on our way back from intramural hockey games. I am also privileged to be friends with people like Enrico Cantoni, Annalisa Scognamiglio, Ludovica Gazze, Tommaso Denti, Manasi Deshpandi, Melanie Wasserman, Adam Sacarny, Christopher Palmer, and many others. Luke Gallagher, Whitney Dudley, Johan Gorr, and Kevin Friedman of the U.S. Army's Office of Economic and Manpower Analysis provided outstanding research assistance throughout my time in graduate school. I would also like to thank Mike Yankovich and David Lyle for their valuable advice and mentorship over the past four years. I also thank Colonel (Retired) Jeffery Peterson for giving me the opportunity to study economics at MIT. Most importantly, I must thank Anne for her love, patience, and encouragement during the many ups and downs I experienced while writing this dissertation. I dedicate this to her. 5 6 Contents 1. Why Do So Many Post-9/11 Veterans Receive Disability Compensation? Evidence from Local Labor Markets 1.1 Introduction.................................................................................................. 10 1.2 Disability Compensation Program Details.............................................................. 15 1.3. Theoretical Framework.....................................................................................18 1.4. Data and Descriptive Statistics ............................................................................... 1.5. Results ....................................................................................................... 1.6. Can Health Factors Explain the Correlation Between PTSD and Local Economic Conditions......37 1.7. C onclusion ................................................................................................... 1.8. Chapter 1 Appendix.........................................................................................44 22 . 27 41 2. The Impact of Disability Benefits on Labor Supply: Evidence from the VA's Disability Compensation Program 65 2.1 Introduction.................................................................................................. 2.2 The Veterans Disability Compensation Program: Eligibility, Benefits, and Work Incentives......70 2.3. The 2001 Agent Orange Decision, Type 2 Diabetes, and 'Service-Connectedness'..................74 2.4. Data and Analytic Sample................................................................................ 2.5. The Impact of the Agent Orange Policy on Receipt of Disability Benefits ......................... 79 2.6. Consequences for Labor Force Participation.............................................................85 2.7. Comparing Labor Supply and Enrollment Impacts: Instrumental Variables Estimates ............. 88 2.8. Impacts on Total Earnings................................................................................ 91 2.9. C onclusion ...................................................................................................... 92 76 2.10. Chapter 2 Appendix.........................................................................................94 3. The Impact of Voluntary Military Service on Mortality: Evidence from Active Duty U.S. Army Applicants 3 .1 Introduction .................................................................................................... 3.2 Institutional Background.....................................................................................126 3.3. Data and Descriptive Statistics..............................................................................127 3.4. E mpirical Strategy............................................................................................130 3 .5. R esults..........................................................................................................134 3.6. Mortality Trends over Time by Application Cohorts....................................................142 3.7. C onclusion ..................................................................................................... 3.8. C hapter 3 Appendix..........................................................................................146 7 123 145 8 Chapter 1. Why Do So Many Post-9/11 Veterans Receive Disability Compensation? Evidence from Local Labor Markets* Kyle Greenberg July 2015 Abstract U.S. veterans who served after September 2001 are nearly twice as likely to receive Disability Compensation (DC) as veterans of earlier wartime eras. Post-Traumatic Stress Disorder (PTSD) award rates are almost double those of Vietnam-era veterans even though casualty rates in Iraq and Afghanistan were a fraction of those in Vietnam. Using administrative data from the U.S. Army and the Department of Veterans Affairs, I investigate the correlation between local labor markets and DC receipt among National Guard veterans who deployed to a combat zone between 2003 and 2006. I find that veterans from hometowns with low employment-to-population ratios are more likely to receive DC for both PTSD and physical conditions than veterans from hometowns with high employment-to-population ratios, but this association is stronger for PTSD than it is for physical conditions. PTSD awards that result in monthly benefit payments of at least $1,500 account for most of the correlation between employment-to-population ratios and PTSD, while only physical awards that generate relatively low payments are associated with employment-to-population ratios. Combat injuries, mortality, and Department of Defense PTSD diagnoses are not correlated with labor market conditions. Furthermore, the effects of combat exposure on DC receipt do not vary with local employment rates, suggesting that vulnerability to combat does not explain the relationship between local economic conditions and PTSD. An alternative explanation is that imprecise screening and the implicit employment criteria associated with mental disorders make veterans in weak labor markets more likely to receive PTSD benefits than veterans in strong labor markets. * I am grateful to Joshua Angrist and David Autor for invaluable guidance and support. I would like to thank Alex Bartik, Sally Hudson, Chris Palmer, Brendan Price, Miikka Rokkanen, Annalisa Scognamiglio, Ashish Shenoy, and seminar participants at MIT and the United States Military Academy. Luke Gallagher of the U.S. Army's Office of Economic and Manpower Analysis provided outstanding research assistance. The findings and conclusions expressed herein are those of the author and do not represent the position of the Department of Defense, the United States Army, or the United States Military Academy. 9 1.1 Introduction Nearly one-third of all veterans who served after September 2001 received Disability Compensation (DC) from the Department of Veterans Affairs (VA) in 2013, an unprecedented share (Veterans Benefits Administration, 2014). Table 1.1 shows that post-9/11 veterans are more likely to receive DC than are veterans of the Vietnam, Korea, or World War II eras despite suffering fewer fatalities and wounds than previous wartime cohorts. Concern for the physical and emotional well-being of soldiers and young veterans has elicited important policy changes. In March 2008, Secretary of Defense Robert Gates developed the Wounded Warrior Task Force to improve services for soldiers suffering from physical wounds, posttraumatic stress, and traumatic brain injury (Gates, 2014, Chapter 4). In February 2015, the U.S. Senate unanimously passed the Clay Hunt Suicide Prevention for American Veterans Act, which requires annual independent evaluations of VA mental health care and suicide prevention programs. Compensation for Post-Traumatic Stress Disorder (PTSD) accounts for much of the unusually high rate of disability receipt among recent veterans. Post-9/11 veterans are twice as likely to receive payments for PTSD as are Vietnam-era veterans (Table 1.1). This is particularly noteworthy because the number of Vietnam-era veterans receiving compensation for PTSD increased almost four fold between 1999 and 2013, from 90,695 to 348,164 (Veterans Benefits Administration, 2000 and 2014) A veteran's monthly disability payment increases with his Combined Disability Rating (CDR), which, in turn, is a function of conditionspecific ratings for each disability that is present. Table 1.2 reveals that service-connected mental health conditions, nearly 60 percent of which are PTSD, result in much higher ratings than common serviceconnected physical conditions. Unsurprisingly, veterans with high disability ratings work less. Table 1.3 reports the share of employed veterans by CDR ranges for male, post-9/11 veterans in the 2013 American Community Survey (ACS) who were born between 1974 and 1986. Forty-two percent of veterans who reported a combined rating of 70 or higher were employed in 2013, which was half of the employment rate of young veterans who did not receive VA disability payments. Both health and the complicated incentives built into the DC program may 10 contribute to this association. In 2013, the average annual monetary compensation for veterans with combined ratings between 70 and 100 was over $29,800 (Veterans Benefits Administration, 2014) Such large payments, which are not subject to federal or state taxes and include free access to VA health care, could reduce labor supply through a non-distortionary income effect. Additionally, two key aspects of the program generate an implicit tax on income. First, veterans with a single disability rated at 60 or higher, or with a combined rating of 70 or higher and at least one disability rated at 40 or higher, are eligible for an Individual Unemployability (IU) designation, which grants payments at the 100 CDR level but prohibits participation in substantially gainful employment. Second, ratings for mental health disorders are based not only on the severity of symptoms, but also on the degree of social and occupational impairment due to such symptoms. Thus, veterans who claim mental disorders have a clear incentive to reduce their labor supply to achieve larger benefits. If nonemployment increases the appearance of occupational impairment, thus leading to higher disability ratings and potentially IU status, then veterans from hometowns with weak labor markets should be more likely to receive benefits for PTSD than veterans from hometowns with strong labor markets. This same result could occur if DC application costs are large and veterans from regions with better economic conditions face a higher opportunity cost of applying. If screening for PTSD is just as precise as screening for physical conditions, and if the application costs for PTSD are equal to the application costs for physical conditions, then local economic conditions should exhibit the same correlation with PTSD as they do for physical conditions that have ratings similar to PTSD. However, under the assumption that application costs for PTSD and physical conditions are the same, but screening for PTSD is less precise, the magnitude of the correlation between local employment rates and physical conditions provides a lower bound on the correlation between local employment rates and PTSD that is due to application costs. In other words, if some veterans obtain PTSD benefits because nonemployment increases the appearance of occupational impairment and screening for PTSD is relatively imprecise, then the association between PTSD and local economic conditions should be stronger than the association between physical conditions and local economic conditions. 11 This chapter uses individual-level data from the Army National Guard and the Department of Veterans Affairs to investigate whether the employment criteria associated with mental disorders contributes to the high rate of PTSD benefits among male National Guard soldiers who deployed to a combat zone between 2003 and 2006. The results of my investigation suggest that the employment-to-population ratio in a soldier's hometown is more strongly correlated with benefits for PTSD than benefits for physical conditions. The impact of local employment rates on receipt of any PTSD benefits is 50 percent larger than the impact of local employment rates on receipt of any benefits for physical conditions. Adding the intensive margin greatly strengthens this result. The correlation between local economic conditions and PTSD ratings, where ratings combine the extensive (i.e. receipt of benefits) and intensive (i.e. severity of rating) margins, is statistically significant and substantial. The impact of a standard deviation decrease in local employment rates on PTSD ratings is 65 percent as large as the effect of a standard deviation increase in military unit casualty rates, where casualty rates reflect an arguably exogenous measure of exposure to situations that should elicit symptoms of post-traumatic stress.' In contrast to the sharp correlation between local employment rates and PTDS ratings, local employment rates exhibit no statistically significant correlation with physical disability ratings, where a soldier's physical rating is his CDR based on physical conditions alone. Part of the reason local employment rates are more strongly correlated with PTSD ratings than they are with physical ratings is mechanical. If few physical conditions result in ratings commensurate with ratings for PTSD, then PTSD ratings will naturally exhibit stronger intensive margin correlation with local employment rates than physical ratings will exhibit even if PTSD and physical disabilities are equally correlated with local employment rates on the extensive margin. Yet almost all of the correlation between local employment rates and PTSD is concentrated in PTSD awards rated 70 or higher, the ratings that require the greatest appearance of social and occupational impairment. Moreover, the association between local employment rates and PTSD ratings of 70 or higher is significantly stronger than the association between Prior to October 2013, the VAk's criteria for a PTSD diagnosis was based on the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV). DSM-IV defines PTSD as "intense fear, helplessness, or horror" resulting from a person experiencing, witnessing or being "confronted with an event or events that involved actual or threatened death or serious injury, or a threat to the physical integrity of self or others" (pp. 427--428). 12 local employment rates and physical ratings of 30 or higher, even though the percentage of veterans within my sample who have PTSD ratings of 70 or higher is less than the percentage of veterans with physical ratings of at least 30. I therefore interpret the responsiveness of relatively high PTSD ratings to local employment rates as primarily the result of the implicit employment criteria associated with benefits for PTSD rather than DC application costs. I also investigate leading alternatives to this interpretation. One alternative is that soldiers from regions with poor economic conditions are exposed to more combat than soldiers from regions with strong economic conditions. I find no correlation between local employment rates and casualties due to hostile action, which argues against this hypothesis. A second alternative is that the same traumatic experience might be more likely to trigger PTSD in veterans from weak labor markets than in veterans from strong labor markets. This is not consistent with my results because the effect of combat exposure on PTSD ratings does not vary with local employment rates. A third alternative is that local employment rates are correlated with factors that influence behavioral health. I find little evidence of this hypothesis as local employment rates are not associated with three measures of health that are independent of the DC program: 1) PTSD disability awards granted by the far more stringent DoD Disability Evaluation System (DES), 2 2) mortality as of May 2012, and 3) the percentage of veterans within the same geographic area used to construct employment rates who report symptoms consistent with post-traumatic stress. Recent research on the DC program has found that Vietnam-era veterans with lower Armed Forces Qualification Test (AFQT) scores and lower education levels are more likely to receive DC, but were no more likely to serve in combat, than veterans with higher AFQT scores and education levels (Angrist, Chen, and Frandsen, 2010, Autor Duggan, and Lyle, 2011). A negative correlation between disability receipt and measures of human capital is potentially consistent with the health effects of military service because low cognitive ability is a risk factor for post-traumatic stress (McNally and Shin, 1995, Macklin et al., 1998). I also find a negative relationship between casualty rates and AFQT scores among National Guard soldiers, which raises the possibility that lower-skilled Vietnam veterans were more likely to suffer injuries The DES usually only grants disability payments for conditions that make a soldier unfit for service in a combat zone. Additional details on this program are discussed below. 2 13 conditional on service in a combat zone. My findings therefore provide more credible evidence that factors other than health have contributed to the substantial growth in DC rolls over the past decade. Additionally, while several studies have documented the responsiveness of Social Security Disability Insurance (SSDI) to local economic conditions (Black, Daniel, Sanders, 2002, Autor and Duggan, 2003), few have investigated this relationship in the context of DC. 3 My results also complement recent work by Coile, Duggan, and Guo (2015), who find that since 2010, the labor force participation of veterans between the ages of 25 and 34 has been more sensitive to economic shocks than that of non-veterans of the same age. This investigation also provides estimates of the effects of combat exposure on disability receipt. Several studies have estimated the effects of combat exposure on diagnoses or symptoms of post-traumatic stress, but not on disability receipt for PTSD (Cesur, Sabia, and Tekin, 2013, Hoge et al., 2004, Hoge, Auchterlonie, and Milliken, 2006, Kolkow, Spira, Morse, and Grieger, 2007, Tanielian and Jaycox, 2008). Most of these studies rely on self-reported measures of exposure, which could be endogenous to DC if surveys about combat experience are conducted after a veteran has had an opportunity to apply for disability benefits, a possibility discussed in Benitez-Silva et al. (2004). But even surveys that have no bearing on DC can be very inaccurate. For example, the highly publicized RAND report, The Invisible Wounds of War, derived conclusions from a survey where 23 percent of service members reported being injured in a manner that did not require hospitalization, 10.7 percent reported being injured severely enough to require hospitalization, and 9.5 percent reported "engaging in hand to hand combat" (Tanielian and Jaycox, 2008, Table 7.3). These rates seem extraordinarily high considering that fewer than 2.3 percent of the 1.9 million service members deployed to Iraq or Afghanistan prior to April 2009 suffered wounds as a result of hostile action (Institute of Medicine, 2010).4 Understanding how combat exposure and factors other than health contribute to disability receipt is 3A 2006 working paper by Duggan, Rosenheck, and Singleton (2006) suggests that total DC program expenditures became more sensitive to local economic conditions after a 2001 policy change that expanded DC eligibility to Vietnam veterans who served in theater. The final version of this paper, Duggan, Rosenheck, and Singleton (2010), does not report this finding. 4 The Defense Manpower Data Center (DMDC) maintains a monthly tally of military casualties by conflict at https://www.dmdc.osd.mil/dcas/pages/casualties.xhtml (accessed 5 March 2015). This tally indicates that 42,000 service members suffered wounds prior to January 2011. 14 critical for determining whether disability payments are the most effective way to rehabilitate veterans with service-connected injuries. The large financial awards associated with high disability ratings could act as a disincentive for treatment of post-traumatic stress, a possibility discussed in Gros, Price, Yuen, and Acierno (2013), Bradley, Greene, Russ, Dutra, and Westen (2005), and McNally and Frueh (2013) (among others). The work disincentives built into the DC program could also counteract policies designed to increase veteran employment, such as the VOW to Hire Heroes Act and the Post-9/11 G.I. Bill. Using the 2001 Agent Orange decision that expanded DC eligibility for Vietnam-era veterans who served in theater, Autor, Duggan, Greenberg, and Lyle (2015a) estimate that DC receipt reduced labor force participation of older veterans by 18 percentage points. If similar effects persist among more recent veterans, then there could be negative repercussions on the U.S. economy's aggregate labor supply given that 8 percent of American men between the ages of 25 and 34 have served or are currently serving in the military (2013 ACS). Furthermore, the longterm costs of veteran disability payments are substantial and growing. Between 2000 and 2013, real annual DC payments to all veterans increased by 146 percent, from $20.0 billion to $49.2 billion, despite a 20 percent decline in the overall veteran population during the same time period (Veterans Benefits Administration, 2013 and 2014). Section 1.2 of this chapter describes the DC program and section 1.3 outlines a theoretical framework for characterizing the relationships between local economic conditions, combat exposure, application costs, and decisions to apply for benefits. Section 1.4 describes my data and the construction of unit-level casualty rates as a measure of combat exposure. Section 1.5 discusses estimates linking local employment rates and combat exposure to DC take-up. Section 1.6 investigates whether health factors contribute to the correlation between labor market conditions and PTSD, and section 1.7 concludes. 1.2 Disability Compensation Program Details DC provides monetary compensation and health care to veterans suffering from diseases or injuries that are sustained or aggravated during military service. Veterans apply for all conditions that are potentially 15 service-related to one of 56 VA Regional Offices (VARO), which generally service veterans who live in a particular state.5 For each condition claimed, a Rating Veterans Service Representative (RVSR) determines if there is sufficient evidence to support the claim and whether the disability is service-connected. The RVSR then assigns a rating on a scale of 0 to 100, in increments of 10, for each verified condition. A rating of 0 affords an applicant free VA health care to treat that particular condition, but does not result in monthly benefit payments. Disability payments for DC are primarily a function of a veteran's Combined Disability Rating (CDR), which is an increasing, concave function of condition-specific disability ratings. As an example, a veteran awarded a 70 percent rating for PTSD and a 20 percent rating for back pain will receive a CDR of 80: 70 for PTSD and 6 percent more for the back pain, which represents 20 percent of the "remaining" 30 percent of ability. Since all CDRs are in increments of 10, 76 is then rounded up to 80. In 2014, a veteran with a spouse and one child received $131 per month for a CDR of 10, $1506 per month for a CDR of 70, and $3134 per month for a CDR of 100. Veterans with a CDR of 30 or higher receive modest additional payments if they have dependents.6 Unlike SSDI, DC is not explicitly work-contingent: a veteran may participate in the labor force while still receiving benefits. A key exception to this rule is the Individual Unemployability (IU) benefit. Veterans who receive a disability rating of 60 or more for one condition, or who have a CDR of 70 or more and at least one condition with a disability rating of at least 40, may qualify for the IU benefit. IU status results in payments at the 100 CDR level, but prohibits participation in "substantially gainful employment." DC payments continue throughout a veteran's lifetime, unlike SSDI benefits, which terminate when a recipient turns 65. DC payments are also exempt from federal and state taxation. Thus, each dollar of DC payments is equivalent to $1.30 to $1.50 in pre-tax income, depending on a recipient's marginal tax bracket. Access to VA healthcare increases with a veteran's CDR. The priority for enrollment in VA health care is based on a veteran's Priority Group, which ranges from 1 through 8, with 1 being the highest priority for enrollment. After leaving the military, all veterans receive 5 years of free VA health care under VA Priority 5 California has three VAROs while Texas, New York, and Pennsylvania all have two. Washington, D.C., has one. 6 The VA also adjusts payments annually according to the Consumer Price Index. An exact table of payments can be found at http://www.benefits.va.gov/COMPENSATION/resources 16 comDO I11 .asD. Group 6. The VA places veterans with a CDR of 10 or 20 in Priority Group 3, veterans with a CDR of 30 or 40 in Priority Group 2, and veterans with a CDR of 50 or higher in Priority Group 1. Service members on active duty are not allowed to receive DC until they have departed the service, but National Guard soldiers may receive benefits as long as they temporarily terminate their benefits while performing active duty service through either a deployment or a full-time National Guard position. There are no time limits associated with disability applications: a veteran may submit an application just prior to leaving active duty service, several years after leaving the service, or both if multiple conditions develop at different times. Once awarded, DC benefits are rarely retracted. In fact, CDRs usually increase over time as veterans request a reevaluation of existing conditions or apply for new conditions (Autor et al., 2015a, Congressional Budget Office, 2014). Mental disorders, 58 percent of which are PTSD, receive much higher ratings than other conditions. Table 1.2, which reports the distribution of ratings for the five most common types of disabilities among all DC recipients as of September 2013, indicates that 36 percent of mental disorders are rated 70 or higher, relative to less than 2 percent of conditions for the other common disability types. There is also evidence that concurrent SSDI receipt is more common among DC beneficiaries with mental disorders than it is among DC beneficiaries with physical disorders. A 2009 General Accounting Office report found that the majority of wounded veterans who received SSDI benefits were eligible for SSDI because of mental disorders, predominately PTSD (General Accounting Office, 2009). Unlike ratings for physical disabilities, which, except for IU status, are based exclusively on the severity of symptoms, ratings for mental disorders are contingent on both the severity of symptoms and the degree of social and occupational impairment resulting from those symptoms. The general rating formula for mental disorders indicates that a rating of 100 is warranted for mental disorders that result in "total occupational and social impairment, due to such symptoms as: gross impairment in thought processes or communication; persistent delusions or hallucinations; grossly inappropriate behavior; persistent danger of hurting self or others; intermittent inability to perform activities of daily living (including maintenance of minimal personal hygiene); disorientation to time or place; memory loss for names of close relatives, own occupation, or own 17 name." Likewise, a rating of 70 is warranted for mental disorders resulting in "occupational and social impairment, with deficiencies in most areas, such as work, school, family relations, judgment, thinking, or mood" due to symptoms such as suicidal ideation, obsessive rituals, and near-continuous panic or depression affecting the ability to function independently, among others. The three other possible ratings for mental disorders, 50, 30, and 10, have gradually lower thresholds, both in terms of occupational impairment and symptom severity.7 Thus, mental disorders could potentially distort labor supply decisions more than physical disorders do, partly because nonemployment increases the appearance of occupational impairment and partly because higher ratings for mental disorders make these conditions more likely to qualify for the IU benefit. 1.3 Theoretical Framework This section offers a theoretical framework for understanding how a veteran's employment status and exposure to trauma influence his decision to apply for benefits. The model explains how a combination of imprecise screening and a clear financial gain from disability benefits could cause veterans with poor employment prospects to receive benefits at higher rates. However, the same outcome is consistent with precise screening, but large application costs. I therefore conclude this section by offering a set of conditions that enables an empirical test to distinguish between these two explanations. A veteran decides to apply for a disability condition, denoted by d, then an RVSR determines whether an award should be granted. For each award granted, the RVSR also assigns a disability rating, r. RVSRs base their decisions on a veteran's condition-specific health, hd. If hd falls below a specific threshold, Td r), then the veteran is granted a monetary payment of Br, which increases with r. If hd < Td(10), then the veteran receives no monetary payment since 10 is the lowest possible rating. A veteran's health is weakly decreasing in his exposure to trauma, m. When the condition of interest is PTSD, exposure to trauma can be thought of as proximity to fatal or dangerous events because this is the primary channel through which a soldier can 7 The general rating formula for mental disorders can be found in Subpart B-Disability Ratings, 38 U.S.C. 4.130 (2015). 18 acquire a clinical diagnosis of PTSD. When the condition pertains to a physical impairment, however, trauma reflects physical injuries from dangerous situations and other aspects of military service, such as wearing heavy body armor, participating in strenuous physical training, or exposure to potentially dangerous chemicals. I initially assume that application costs are zero and that VA evaluators are perfectly informed of h'. In this scenario, a veteran will apply for conditions for which hd < Td(r). The fraction of veterans who receive disability for condition d at rating r or higher is: (1.1) f d(r)(m) = Pr(hd(m) < Td(r)) Under the assumptions of zero application costs and perfect screening, fd(r) (m) is increasing in exposure to trauma. I now relax the assumption of no application costs. A veteran who applies for disability condition d pays application cost, cd (e), which represents the time costs associated with applying for condition d. Application costs are decreasing in a veteran's employment status e, where a larger value of e represents more hours worked, because veterans who are not in the labor force, unemployed, or working part-time face a lower opportunity cost of applying. A veteran will apply for condition d if the expected utility of applying, minus application costs, exceeds the utility of not applying, which I normalize to zero. Thus, a veteran will apply for condition d if the expected utility gain exceeds the cost of applying: (1.2) E[UdIApplyd] > cd(e) The fraction of veterans who receive disability for condition d at rating r or higher is now: 19 (1.3) fdM)(e,m) = Pr(Receive Award Apply) x Pr(Apply) = Pr(hd(m) < Td(r)) x Pr(cd(e) < [Br x Pr(hd(m) < Tdar) With application costs, f d(r) (e, m) is increasing in m and decreasing in e. I now relax the assumption that RVSRs are perfectly informed of an applicant's health by allowing applicants to pay a cost, gd(e), to seek higher disability payments. A positive value of gd(e) could result from seeking additional medical examinations if the first one is not favorable to the applicant, appealing ratings decisions, or exaggerating symptoms. It could also result from veterans who stop working to increase the appearance of occupational impairment, which could generate higher ratings for mental disorders or possibly IU status. gd (e) is decreasing in e for two reasons. First, the implicit tax on labor force participation generated by an IU benefit and high mental disorder ratings is lower for veterans with poor employment prospects. Second, nonemployed veterans face a lower opportunity cost of seeking more favorable disability ratings. When an applicant pays gd(e), the RVSR's assessment of the applicant's health decreases by ydgd(e), where yd represents the effectiveness, or rate of return, of seeking benefits. Conditions that are more difficult to diagnose (i.e., conditions where screening is less precise) have large values of yd while conditions with perfect screening have yd = 0. It is important to emphasize that a positive value of gd (e) does not necessarily imply malingering on the part of the veteran. RVSRs who adjudicate mental health decisions could also influence gd(e) if they incorrectly assume a veteran's financial situation is directly attributable to post-traumatic stress, a possibility discussed in Worthen and Moering (2011). Likewise, since higher combined ratings provide DC recipients with better access to VA health care, some medical examiners might produce more favorable reports for jobless disability applicants. In the context of this model, exaggeration on the part of the RVSR or the medical examiner can be thought of as a very high rate of return (i.e. high yd) at minimal additional cost (i.e. gd(e)) to the applicant. 20 Now the fraction of veterans who receive disability for condition d at rating r or higher is: (1.4) fd( (e,m) = Pr(Receive Award IApply) x Pr(Apply) = Pr(hd(m) - ydgd(e) < Td(r)) x Pr(cd(e) + gd(e) < [Br x Pr(hd(m) - ydgd(e) < Td(r)]) In section 5, I use local employment rates as a source of variation in a veteran's potential employment status that is not correlated with the health effects of military service. Equations (1.3) and (1.4) illustrate how correlation between disability receipt and local labor markets is consistent with the presence of application costs, but could also result from a combination of imprecise screening and the possibility of financial gain from disability benefits. However, under the assumption that application costs for PTSD and physical conditions are the same, but screening for PTSD is less precise than screening for physical conditions, the correlation between local employment rates and physical conditions provides an upper bound on the impact of application costs. To see this last point more clearly, consider a veteran whose opportunity cost of applying for DC is equal to the present-discounted value of a lifetime stream of $200 per month when he lives in a hometown with high employment rates, but is only equal to the present-discounted-value of a lifetime stream of $100 per month when he lives in a hometown with low employment rates. Suppose that this veteran has a physical ailment that will result in a rating of 10 if he applies for DC. Since a single disability rated at 10 results in a monthly benefit of $131, this veteran will not apply for DC when he lives in a hometown with high employment rates but will apply when he lives in a hometown with low employment rates. In other words, application costs make DC benefits for physical conditions less likely when economic conditions are good. Now suppose this veteran has a set of physical ailments that would result in a CDR of 30. In this case, the veteran will apply for DC regardless of the economic conditions in his hometown because a CDR of 30 results in a monthly benefit of more than $400. However, if this same veteran does not suffer from any 21 physical ailments, but instead receives a rating of 30 for PTSD only when he lives in a hometown with low employment rates, then application costs could not explain the correlation between local economic conditions and the probability of receiving a PTSD rating of 30. Assuming equal application costs for PTSD and physical conditions is probably reasonable because the VA's application process is similar for all types of conditions. A 2005 Office of the Inspector General (OIG) survey of RVSRs suggests that RVSRs have more difficulty rating claims for mental disorders than claims for cardiovascular, respiratory, auditory, and eye conditions, supporting my assumption that screening for PTSD claims is less precise (VA Office of the Inspector General, 2005). Nevertheless, physical conditions like back pain, joint pain, migraine headaches, or Traumatic Brain Injury (TBI) can be difficult to diagnose, which is why attributing all correlation between $e$ and physical disabilities to application costs is likely to be an overestimate. 1.4 Data and Descriptive Statistics A. Data and Sample Selection My data combine National Guard administrative records with DoD casualty information and VA disability rolls. The National Guard data are from a monthly panel of enlisted military records that contain demographic information, military service information, state and zip code of residence, and mobilization information. 8 The United States Military Academy's Office of Economic and Manpower Analysis used social security numbers to merge military data to VA disability records, DoD casualty records, and the SSA Death Master File (DMF). VA disability records consist of a single cross-section of individuals who were enrolled in the DC program as of January 2014. Data on each beneficiary include a list of all disabilities for which the VA approved compensation and the rating for each disability. Unfortunately, VA records do not indicate whether a veteran applied for an award but was ultimately denied. They also do not indicate when a veteran first started to receive compensation. Casualty data come from the Defense Casualty Information and Processing System, the official casualty database of the DoD. The appendix to this chapter describes in detail 8Unlike the Active Duty Army, where assignment of new soldiers is based on the needs of the Army, National Guard soldiers sign enlistment contracts to join a specific military unit that is usually near their residence. 22 how casualties are recorded in DCIPS and explains why casualty records are not easy to manipulate. The DMF indicates whether an individual was no longer alive as of May 2012. I use employment information of non-veteran men in the American Community Survey (ACS) between the years 2005 and 2009 to construct employment-to-population ratios (local employment rates) at the race (white or non-white) by Super Public Use Microdata Area (PUMA) level.9 A veteran's Super PUMA is based on his zip code of residence at the start of his first deployment. I mapped each five-digit zip code to a specific Super PUMA using the zip-code-to-PUMA crosswalk file (based on year 2000 geographic boundaries) provided by the Geographic Correspondence Engine of the Missouri Census Data Center.' 0 For each combination of race and Super PUMA, I constructed the mean employment rate using non-veterans born between 1974 and 1986, which captures 92 percent of the birth years for veterans used in the sample for the analysis below. Part B of the appendix describes the construction of local employment rates in more detail. Part C of the appendix describes the construction of the analysis sample, which I henceforth refer to as the National Guard sample. The final sample consists of 55,939 male, enlisted National Guard soldiers who deployed to a combat zone between January 2003 and December 2006. It only includes veterans who joined the National Guard with no prior military service and soldiers who deployed within ten years of joining the Guard." Since I exploit within-state variation in employment rates at the Super PUMA level, the sample does not include veterans from Alaska, North Dakota, South Dakota, Wyoming, Vermont, or Washington, D.C., all of which have only one Super PUMA. 9 Super PUMAs are geographic areas with 400,000 or more residents. Each Super PUMA is comprised of two to six PUMAs, which are nested within super PUMAs. 10 This file is available at http://mcdc2.missouri.edu/websas/geocorr2k.html. If a zip code spanned more than one Super PUMA, I assigned it to the Super PUMA with the largest allocation factor (i.e., the Super PUMA where most individuals within the zip code belonged). " Soldiers who serve beyond ten years have reenlisted at least twice and are likely to serve until they qualify for retirement benefits. This restriction is also a data necessity because I do not have service information on soldiers who joined the National Guard prior to 1993. 23 B. Descriptive Statistics Table 1.4 reports summary statistics for soldiers in the National Guard sample. Column 1 presents statistics for all in the sample and columns 2 through 5 present statistics for the following subsets: non-DC recipients, DC beneficiaries for conditions other than PTSD, DC beneficiaries for PTSD and possibly other conditions, and DC beneficiaries with a CDR of 100 or IU status. Nearly 33 percent of soldiers received DC payments from the VA in January 2014, with 19 percent receiving compensation for PTSD and 1.5 percent receiving an IU benefit. DC recipients were exposed to more combat, are slightly older, and have lower human capital attributes than non-DC beneficiaries. Whereas 40 percent of veterans who did not receive DC served in a combat occupation during their first deployment, this fraction is 51 percent among veterans receiving PTSD benefits and 50 percent among veterans receiving benefits at the 100 CDR level. Similarly, even though only 3.3 percent of veterans in my sample were ever wounded in action (WIA), WIA rates were 5.1 percent for veterans with DC for conditions other than PTSD, 8.8 percent for veterans with PTSD benefits, and 12.8 percent for veterans with an IU benefit or a CDR of 100.12 DC receipt, especially when it involves PTSD, IU benefits, and Combined Disability Ratings of 100, is also negatively correlated with test scores and education levels. The average AFQT score for veterans who did not receive DC in January 2014 is 60.4, but the average score is 59.0 among DC recipients without PTSD benefits, 54.2 among DC recipients with PTSD benefits, and 51.4 among DC recipients with an IU benefit or a CDR of 100. PTSD beneficiaries are disproportionately more likely to receive high combined ratings. On average, veterans with PTSD have combined ratings that are twice as large, and have monthly payments that are 2.5 times as large, as DC recipients who are not compensated for PTSD. Veterans in the former group are also six times as likely to have an IU designation. Interestingly, the two groups have similar physical ratings, where a veteran's physical rating is the CDR he would receive after excluding ratings for mental health conditions and tinnitus (ringing in the ears).1 3 Thus, differences in combined ratings are due almost entirely No DC beneficiaries were Killed in Action (KIA) or deceased because benefits terminate with death. 13 For example, a veteran with a 30 percent rating for PTSD, a 20 percent rating for knee arthritis, and a 10 percent rating for tinnitus would have a CDR of 50 and a physical rating of 20. The physical rating is meant to have a lower probability of type 2 screening errors than are ratings for PTSD. I exclude tinnitus from the construction of physical 12 24 . JL.T to benefits for PTSD. Similarly, PTSD is the primary condition for 81 percent of DC beneficiaries with IU status or a CDR of 100 (column 5) and for 78 percent of beneficiaries with a CDR of 60 or higher (not included in Table 1.4). Although not indicated in Table 1.4, it is worth noting that 22 percent of veterans with a physical rating of 100 were wounded in combat, compared to only 9 percent of veterans with a PTSD rating of 100. Figure 1.1 plots PTSD and physical disability rates for veterans in the National Guard sample against their local employment rates. The plots indicate a stronger correlation between local employment rates and receipt of any PTSD award than between local employment rates and receipt of any physical award. Figure 1.2 plots average PTSD ratings and average physical ratings against local employment rates. Veterans who do not receive benefits for PTSD have a PTSD rating of 0 and veterans who do not receive benefits for physical conditions have a physical rating of 0. Thus, Figure 1.2 combines both the extensive (receipt of benefits) and intensive (degree of rating) margins. Although veterans from Super PUMAs within the highest quintile of employment rates have average PTSD ratings that are comparable to their average physical ratings (7.6 for PTSD ratings, 6.6 for physical ratings), veterans from Super PUMAs within the lowest quintile of employment rates have average PTSD ratings that are 65 percent larger than their average physical ratings (13.2 for PTSD ratings, 8.0 for physical ratings). The plots in Figure 1.3 further investigate the possibility of an intensive margin relationship between PTSD or physical ratings and local employment rates by displaying the correlation between local employment rates and PTSD or physical ratings of 70 or more, both of which make a veteran eligible to receive an IU award. The incidence of PTSD ratings of at least 70 is larger than the incidence of physical ratings of at least 70 for veterans of all employment rates, but this is especially true for veterans from hometowns with low employment rates. The plots in Figures 1.1, 1.2, and 1.3 provide some initial descriptive evidence that PTSD is more strongly correlated with local economic conditions than are physical conditions, and that much of the correlation between PTSD and local employment rates is concentrated in PTSD awards rated 70 or higher. I explore these possibilities in greater detail below. ratings because Title 38, Chapter 1, Section 4.87 of the electronic Code of Federal Regulations (accessed 10 December 2014) instructs medical examiners to evaluate tinnitus based on whether an applicant perceives ringing in one ear, both ears, or the head. 25 C. Company Casualtiesas a Proxyfor Combat Exposure Estimating the effects of combat exposure on disability receipt not only has important policy implications, but also provides a useful benchmark for interpreting the association between local employment rates and disability receipt. Since self-reported measures of combat exposure are susceptible to bias and misinterpretation, as discussed in the introduction, I construct a proxy measure of combat exposure using administrative casualty records that are less prone to error. The measure reflects the percentage of casualties within each soldier's company during his first deployment. Specifically, I sum the number of wounded in action and killed in action (KIA) in soldier i's company, c, over each month, t, that i is deployed, leaving out the individual observation for i. Then I divide by the average number of deployed soldiers in i's company over each month that i is deployed: E 1 = 100 T=1T t Z e cit)(KIAjt + WIAjt) ifj T 1 T 1 Ni ifj E c t) where T is the number of months of the first deployment for i and N is the number of soldiers in i's company in month t.14 From an individual level, there is little potential for selection into combat exposure. E only includes the casualty rates of other soldiers within each individual's company. Even if less risk-averse soldiers exhibit more dangerous behavior while deployed, this will not directly influence their level of E. Similarly, since some soldiers might volunteer for additional deployments, only considering company casualties from a soldier's first deployment avoids this sort of selection-bias. Consistent with this, Appendix Table 1.1 shows that E is not strongly correlated with baseline characteristics after controlling for a soldier's occupation and rank at the start of his first deployment between 2003 and 2006.15 Notably, Appendix Table 1.1 also reveals Yankovich (2011) also uses administrative casualty data to investigate of the effects of combat exposure on military experience. My measure, E , differs from the measure of combat exposure constructed by Yankovich in three ways. First, my measure is based on the percentage of service members in each soldier's company who suffer a casualty while Yankovich's measure is the number of monthly casualties sustained in a soldier's company, averaged over each month the soldier is deployed. Second, my measure only includes company casualty rates during each soldier's initial deployment; Yankovich incorporates casualties from all deployments. Lastly, my measure only includes casualties of other personnel in a soldier's unit, whereas Yankovich includes casualties sustained by the individual soldier. '5 I define a soldier's occupation as his Career Military Field (CMF), which is the first two digits of his Military Occupation Specialty (MOS). 14 26 that local employment rates exhibit no correlation with Ei. Soldiers with higher values of E are, on average, more likely to have some college credits, but this correlation is economically small. A standard deviation increase in local employment rates is only associated with a half a percentage point increase in some college experience. Nevertheless, Eg is still an imperfect measure of combat exposure. Companies with high casualty rates might have poor leadership, which could be correlated with disability receipt or other outcomes. Ei is also better suited to measure the effects of emotional trauma than physical trauma. A soldier who was not injured himself but who witnessed several of his friends die or sustain wounds is more likely to develop symptoms of post-traumatic stress and might be more likely to develop hearing problems or back pain later in life (e.g., if his military unit participated in more offensive actions). However, witnessing such trauma would not make the soldier more likely to have lost a limb in battle. To address this last concern, I also control for wounds sustained as a result of hostile action in many of the regressions that follow. When both measures of combat exposure are included in the same regression, I interpret estimates of Ei as the effect of exposure to trauma that is not due to physical combat injuries. 1.5 Results A. Casualties, Local Employment Rates, and AFQT Scores A negative relationship between casualties and local employment rates would suggest that veterans from areas with poor economic conditions experienced more dangerous situations in combat. I investigate this possibility by estimating the following equation: (1.5) CAS = a + flEMPgr + Osjr + XL' + EL CASg indicates the number of casualties (WIA plus KIA) for individual i of race r (white or non-white) who lived in Super PUMA g and state s at the start of his first deployment year j.16 EMPgr is the Casualties are summed over all deployments between 2003 and 2011 for each soldier, not just his first deployment. Although estimated by ordinary least squares, this is technically not a linear probability model because approximately 16 27 employment rate for veterans from Super PUMA g and race r. No Super PUMA spans multiple states, so g uniquely identifies s. A fixed effect for each combination of state, deployment year, and race ensures that variation in employment rates comes from veterans who deployed from different Super PUMAs, but who came from the same state, deployed in the same year, and have the same race. Xi is a vector of individual controls that includes a linear term for AFQT and indicators for year of birth, military occupation, military rank, and education level.1 7 I cluster standard errors on each combination of Super PUMA and race, the level of variation in local employment rates. Table 1.5 reports estimates of (1.5). To make the table easier to read, I have multiplied all left-hand side variables by 100. Estimates excluding Xi, reported in column 1, panel A, indicate that a one percentage point increase in employment rates predicts a 0.007 percentage point increase in battlefield casualties, but this estimate is not statistically different from zero. The estimated effects of local employment rates on casualties hardly change with the inclusion of AFQT scores (column 2) and additional controls for year of birth, occupation, rank, and education (column 3). Estimates of equation (1.5) with Ei included as a control, reported in column 4, show no change in the correlation between local employment rates and casualties, but reveal that veterans who served in companies that experienced high casualty rates were more likely to be wounded than veterans who served in companies that experienced few casualties, as expected. Panels B and C break down casualties into lethal (panel B) and nonlethal (panel C) components. Panel B actually indicates a positive association between employment rates and combat fatalities, although this estimate is only marginally significant.' 8 Overall, the results presented in Table 1.5 do not suggest that veterans from hometowns with weak labor markets experienced greater battlefield exposure, injury, or death than veterans from hometowns with strong labor markets. 0.15 percent of soldiers in my sample suffered more than one wound from hostile action. I consider a soldier's first deployment between 2003 and 2006 as his first deployment during his military service. National Guard deployments to combat zones were very rare prior to 2003 (Bonds, Baiocchi, and McDonald, 2010). '7 Within my sample there are four education levels (GED or high school dropout, high school graduate, some college, college), 31 possible years of birth (1958-1988), 38 possible military occupations, and 3 possible ranks (Private First Class (PFC) or below, Specialist, and Sergeant or above). Education, occupation, and rank information are determined at the time a soldier starts his first deployment. 18 It is worth noting that panel B of Table 1.5 reveals a negative relationship between Ei and hostile deaths. Veterans who were killed in action had their deployments cut short. Thus, their values for Ei are mechanically smaller. 28 Interestingly, Table 1.5 reveals that soldiers with low AFQT scores were more likely to be injured or killed in battle than soldiers with high AFQT scores, even after controlling for age, education, occupation, and rank. The estimates reported in column 4, which show that the negative relationship between AFQT scores and casualties persists with the inclusion of company casualty rates, does not indicate that soldiers with low AFQT scores sorted into companies that were exposed to more combat.1 9 Low cognitive ability could be a risk factor for being killed or wounded in combat. Soldiers with high AFQT scores might have greater ability to avoid or react to dangerous situations than soldiers with low AFQT scores, thus reducing their likelihood of suffering an injury. This is consistent with studies that have found a positive correlation between AFQT scores and soldier performance on tanker gunnery ranges and skill qualification tests (Scribner, Smith, Baldwin, and Phillips, 1986, Home, 1987, Winkler, 1999). B. The Effects ofLocal Employment Rates and Combat Exposure on DisabilityReceipt: Extensive Margin As discussed in the theoretical section, the presence of either application costs or imprecise screening for some disability conditions could cause veterans from weak labor markets to receive disability at higher rates than veterans from strong labor markets. To investigate this possibility, Table 1.6 reports ordinary least squares (OLS) estimates of the following equation: (1.6) Yi = a + fEMPgr + yEi + SWIA 1 + ipsjr + Xi6 + eC In panel A, Y is an indicator for receiving any DC benefits. As above, g and s index veteran i's Super PUMA and state of residence, respectively, at the start of his first deployment between 2003 and 2006, j indexes year of deployment, and r indexes race (white or non-white). Ei is the combat exposure measure described above and WIAi is the number of wounds that i sustained over all deployments. Including a fixed effect for each combination of state, deployment year, and race ensures that any correlation between employment rates and disability receipt is not confounded by variation in VARO approval procedures or 19 To further investigate this possibility, I also estimated equation (1.5) with company fixed effects, which hardly change the negative relationship between AFQT scores and casualties. These results are available from the author upon request. 29 other factors that potentially vary at the state and year level, such as state-level National Guard policies or Unemployment Insurance replacement rates.20 Controlling for deployment year is also critical because veterans who deployed in 2003, for example, had three more years to apply for disability, and potentially upgrade their ratings, than veterans who deployed in 2006. As with Table 1.5, all outcomes have been multiplied by 100. Estimates excluding E1 , WIA , and X1 , reported in column 1, indicate that a one percentage point increase in employment rates is associated with a 0.294 percentage point decrease in the probability of receiving any DC benefits. The estimated effect of employment rates on DC receipt decreases by 23 percent (three-fourths of a standard error) when AFQT is included as a control (column 2). This is not surprising: veterans from regions with poor economic conditions are more likely to have low AFQT scores and the estimate on AFQT is consistent with previous research which has documented that low cognitive ability is a strong predictor of DC receipt (Autor et al., 2011). The estimated effect of employment rates on DC receipt hardly changes with the inclusion of year of birth, military occupation, military rank, and education controls (column 3), or with the inclusion of E and WIAi (column 4). With the rich set of controls included in column 4, a standard deviation decrease in employment rates is associated with a 1.23 percentage point decrease in the probability of receiving any DC benefits, which is roughly 40 percent as large as the effect of a standard deviation increase in Ei and 4 percent as large as the effect of being wounded on any disability receipt. Panels B and C of Table 1.6 break down any DC receipt into receipt of any compensation for PTSD and physical conditions, respectively. The estimates reported in column 2 of both panels indicate that controlling for AFQT scores attenuates the estimated effect of local employment rates on disability receipt for both PTSD and physical conditions. Likewise, the estimated effects of employment rates on both outcomes hardly change with the inclusion of Ej, WIA 1 , and Xj, consistent with estimates reported in panel A. The estimate reported in column 4 of panel B indicates that a one percentage point increase in local employment rates is As discussed above, VAROs typically service all veterans in a particular state. All results are similar when I remove veterans from California, Texas, New York, and Pennsylvania, the four states that have more than one VARO. 21 The mean (standard deviation) of local employment rates and E for soldiers in the National Guard sample are 80.73 (5.91) and 2.95 (5.20), respectively. 2 30 associated with a 0.220 percentage point decrease in the probability of receiving any PTSD benefits. This is approximately 50 percent larger in magnitude than the estimated effect of employment rates on receipt of any benefits for physical disabilities, as reported in column 4, panel C. Notably, unit casualty rates, as measured by Ei, are more strongly correlated with receipt of any PTSD benefits than with receipt of any physical benefits. A one percentage point increase in Ei is associated with a 0.622 percentage point increase in the probability of receiving any benefits for PTSD and a 0.290 percentage point increase in the probability of receiving any benefits for physical conditions. As discussed above, it is not surprising that E is more strongly correlated with PTSD than with physical conditions because higher unit casualty rates should directly increase a soldier's propensity to develop post-traumatic stress, but only indirectly increase his propensity to develop physical disabilities. In contrast, the effect of wounds sustained through hostile action, WIAi, on PTSD receipt is 30 percent smaller than the effect of wounds on receipt of any physical award (22.52 relative to 32.53), although it is very clear that wounded veterans are substantially more likely to receive benefits for PTSD and physical conditions than are uninjured veterans. C. The Effects of Local Employment Rates and Combat Exposure on Disability Receipt: Intensive and Extensive Margins The results reported in panels B and C of Table 1.6 suggest that application costs contribute to, but do not fully explain, the correlation between local economic conditions and receipt of PTSD benefits on a purely extensive margin. However, the evidence for this interpretation is less than conclusive. On one hand, the estimated effect of local employment rates on receipt of any PTSD award (column 4, panel B) is within the margin of error of the estimated effect of local employment rates on receipt of any physical award (column 4, panel C). On the other hand, PTSD awards receive much higher ratings than awards for physical ailments: more than half of veterans who receive compensation for physical conditions have a physical disability rating of 10 or 20, whereas 95 percent of DC beneficiaries with a PTSD award have a PTSD rating of 30 or higher. Thus, application costs might deter some veterans from applying for physical disabilities that would result in ratings of 10 or 20, but would not necessarily deter the same veterans from applying for PTSD awards that 31 would typically result in a rating of 30. Additionally, a purely extensive margin comparison ignores the possibility that local labor markets might influence disability receipt on the intensive (severity of rating) margin, and that intensive margin effects could differentially influence PTSD and physical ratings. For example if a disability applicant's nonemployment status increases his appearance of occupational impairment or encourages him to increase his initial disability rating to offset financial hardship, then local labor markets would exhibit correlation with PTSD ratings on both the extensive and intensive margins. To investigate the possibility that DC responds to local labor markets on an intensive as well as extensive margin, Table 1.7 reports OLS estimates of (1.6) where the outcomes are a veteran's CDR (panel A), a veteran's PTSD rating (panel B), and a veteran's physical rating (panel C). 22 The estimates reported in column 1, panel A of Table 1.7 indicate that a one percentage point increase in local employment rates is associated with a 0.22 average decrease in PTSD ratings. Similar to the pattern exhibited in Table 1.6, estimates decrease by 23 percent with the inclusion of a control for AFQT (column 2), but remain relatively consistent with the inclusion of additional individual-level covariates (column 3) and with the inclusion of of E and WIAi. Using the estimates reported in column 4, the effect of a standard deviation decrease in employment rates on a veteran's CDR is 5.7 percent of the mean CDR in the sample (= 1 Z 59 ) 5.7). By contrast, the estimates reported in panel A, Table 1.6 indicate that the effect of the same decrease in employment rates on DC receipt is 3.8 percent of the percentage of DC beneficiaries within the sample (-.2 0(-.9) = 3.8). The larger effect of employment rates on CDRs, relative to the effect of employment rates on receipt of any DC benefits, suggests that local labor markets influence disability receipt on both the extensive and intensive margins. The estimates reported in panels B and C of Table 1.7 reveal that PTSD ratings explain most of the correlation between local labor markets and combined ratings. The estimated effect of local employment 22 CDRs are 0 for veterans who received no DC and are 100 for veterans who receive an IU benefit. Likewise, PTSD ratings nd a0 for veterans who received no comupensation for PTSD and are 100 for veterans designated as IU for PTSD, and physical ratings are 0 for veterans with no physical disabilities and 100 for veterans designated as LU for a physical condition. Since all estimates include both DC beneficiaries and non-beneficiaries, the results reported in Table 1.7 combine both the extensive (receipt of disability) and intensive (disability rating) margins. 32 rates on PTSD ratings (column 4, panel B) is almost identical to the estimated effect of local employment rates on CDRs (column 4, panel A). Comparing the estimated effect of employment rates on PTSD ratings to the effect of E on PTSD ratings also suggests that the intensive margin influence of employment rates is larger than the intensive margin influence of unit casualty rates. Whereas the effect of a standard deviation decrease in local employment rates is 40 percent as large as the effect of a standard deviation increase in Ej ((=220)( L \\k(.622) (5.2)/ 0.40 (column 4, panel B, Table 1.6), the effect of the / on receipt of any PTSD benefits same decrease in local employment rates on PTSD ratings is 65 percent as large as the effect of a standard deviation increase in E (=(-5) 0.65 (column 4, panel B, Table 1.7). Unlike PTSD ratings, physical ratings exhibit little correlation with local labor markets. The estimated effect of employment rates on physical ratings is not statistically significant when all controls are included (column 4) and is one-fifth of the magnitude of the estimated effect of employment rates on PTSD ratings (-0.169 for PTSD ratings, -0.034 for physical ratings). Furthermore, I can rule out an effect of employment on physical ratings that is of equal value to the effect of employment on PTSD ratings. To further investigate the intensive margin responsiveness of PTSD and physical ratings to local labor markets, panel A of Table 1.8 reports estimates of equation (1.6) where the outcome indicates receipt of a PTSD rating above 0 (column 1), or PTSD ratings of at least 30, 50, 70, or 100 (columns 2, 3, 4, and 5, respectively).23 This follows the analysis of distribution effects described in Angrist (2001). Panel B is analogous to panel A, except the outcomes are indicators for having physical ratings at or above the rating cutoff for each column. To make the table easier to read, all outcomes have been multiplied by 100. The estimates reported in column 2 of panel A indicate that a one percentage point increase in local employment rates is associated with a 0.21 percentage point decrease in the probability of receiving a PTSD rating of 30 I have chosen these ratings cutoffs because nearly all mental disabilities result in condition-specific ratings of 10, 30, 50, 70, and 100. 23 33 or higher, which is nearly identical to the estimate reported in column 1 of panel A.24 On the other hand, a one percentage point increase in local employment rates is associated with a 0.075 percentage point decrease in the probability of receiving a physical disability rating of at least 30 (column 2, panel B), which is substantially smaller in magnitude than the corresponding estimate for PTSD ratings of at least 30. Importantly, the estimates reported in columns 3 through 5 of Table 1.8 (panel A) suggest that many veterans who apply for PTSD benefits because of weak local labor markets receive high ratings. A one percentage point increase in local employment rates is associated with a 0.184 percentage point increase in the probability of receiving a PTSD rating of 70 or higher (column 4), which is close to the estimated effect of local employment rates on receipt of any PTSD benefits (column 1, panel A). Benchmarking the effect of local employment rates against the effect of E and WIAi for different PTSD ratings tells a similar story. The association between a standard deviation decrease in employment rates and receipt of any PTSD award is 40 percent as large as the effect of a standard deviation increase in Ej ((-.220)(-5.) 0.40). However, the association between a standard deviation decrease in employment rates and PTSD awards rated 100 or IU is nearly 160 percent as large as the effect of a standard deviation increase in E ( = 1.57). Despite strong correlation between local employment rates and PTSD ratings of 70 or 100, the estimates reported panel B indicate no correlation between local employment rates and physical ratings of 50 or more. Unlike PTSD, ratings for physical conditions depend exclusively on the severity of observable symptoms. The precisely estimated zero effect of employment rates on physical ratings of at least 50 (column 3, panel B) therefore suggests that veterans from weak labor markets did not see a substantial deterioration in physical health following their deployments. It also suggests that high application costs do not deter veterans from applying for disabilities that are likely to result in CDRs of 50 or higher, although this interpretation should be made with caution because few veterans have physical ratings of 50 or higher. The similar estimates reported in columns 1 and 2 of panel A, Table 1.8, are not surprising because nearly 95 percent of veterans who receive disability for PTSD have ratings of 30 or higher. 24 34 Appendix Table 1.2 distinguishes between IU status and ratings of 100. Column 1, which reports estimates of equation (1.6) where the outcome is an indicator for receiving a rating of 100 or IU status for either PTSD (panel A) or physical conditions (panel B), is identical to column 5 of Table 1.8. Columns 2 and 3 subsequently break down this outcome into IU status and a rating of 100, respectively. The estimates reported in panel A indicate that local employment rates are correlated with both IU awards for PTSD and PTSD ratings of 100. PTSD ratings of 100 are more strongly correlated with local employment rates than are IU awards for PTSD, but this comparison is not very meaningful because, unlike physical disorder ratings, mental disorder ratings of 100 require "total occupational and social impairment." In other words, both PTSD IU awards and PTSD ratings of 100 have a subjective, occupational-impairment component. The estimates reported in panel B reveal that local employment rates are neither correlated with IU status for physical conditions nor physical ratings of 100. Interestingly, ratings constructed from all conditions other than PTSD (not just physical conditions) exhibit little correlation with local employment rates, particularly on the intensive margin. This can be seen in panel C of Appendix Table 1.3, which reports estimates of equation (1.6) where the outcome is an indicator for receiving any non-PTSD disability (column 1), or non-PTSD ratings of at least 30, 50, 70, or 100 (columns 2, 3, 4, and 5, respectively). A veteran's non-PTSD rating represents his CDR after excluding any rating for PTSD. To facilitate comparison, panels A and B of Appendix Table 1.3 reproduce estimates of the effect of local employment rates on PTSD and physical outcomes, respectively. All estimates reported in panel C are very close to the estimated effects of employment rates on physical disabilities, as reported in panel B. Panel D reports estimates where the outcomes are based on ratings from mental disorders other than PTSD while panel E reports estimates where the outcome is tinnitus, which is also not included in physical disability ratings, as discussed above. Since all diagnoses for tinnitus result in a rating of 10, no estimates are reported in columns 2 through 5 of panel E. Notably, the results reported in panel D of Appendix Table 1.3 do not indicate substantial correlation between employment rates and mental disorders for conditions other than PTSD. These results appear to contradict the assumption that screening for mental conditions is less precise than screening for physical 35 conditions. However, the diagnostic criteria for PTSD is more broad, and therefore less precise, than the diagnostic criteria for other anxiety disorders, a point discussed in Brewin, Lanius, Novac, Schnyder, and Galea (2009). Moreover, the key diagnostic criterion associated with PTSD, exposure to a traumatic event, could lead to over-diagnosis in situations where a traumatic event occurred, as discussed in Maercker et al. (2013). Taken together, I interpret the results from this section as evidence that imprecise screening and the employment criteria associated with PTSD benefits, rather than application costs, drive most of the correlation between local labor markets and PTSD receipt. This is not to say that application costs are not present. A plausible interpretation for the extensive margin correlation between employment rates and physical disabilities, and the lack of correlation when both the extensive and the intensive margins are combined, is that application costs deter some veterans in strong labor markets from applying for physical disabilities that are likely to result in low ratings, but not physical conditions that are likely to result in high ratings. Another explanation is that screening procedures for low-rated physical disabilities are imprecise. The latter explanation is a distinct possibility given that nearly 50 percent of veterans in my sample with any physical disability receive benefits for back pain, knee pain, or migraines, three conditions that predominately rely on self-reported symptoms of pain. In contrast to physical conditions, PTSD awards not only exhibit strong correlation with local employment rates on the extensive margin, but much of this correlation is concentrated in PTSD awards rated 70 or 100, the ratings that require the most evidence of social and occupational impairment. This result is consistent with the hypothesis that the employment criteria associated with the mental disorder rating schedule encourages many veterans from weak labor markets to seek high PTSD ratings. Additionally, the correlation between local employment rates and PTSD awards rated 70 or higher (column 4, panel A, Table 1.8) is significantly larger than the association between local employment rates and physical ratings of 30 or higher (column 2, panel B, Table 1.8), even though the incidence of physical ratings of 30 or higher is greater than the incidence of PTSD ratings of 70 or higher. This last point suggests that application costs do not fully explain the relationship between local economic conditions and PTSD. It is also worth noting that veterans 36 with PTSD ratings of 70 or higher receive a lifetime stream of between $1,300 and $3,200 in monthly benefits, a very substantial sum. Overall, it seems unlikely that the time costs associated with applying for DC drive the correlation between local employment rates and high PTSD ratings, but could potentially explain the relationship between local employment rates and physical conditions. 1.6 Can Health Factors Explain the Correlation Between PTSD and Local Economic Conditions? It is natural to ask whether exposure to the same traumatic experience is more likely to cause post- traumatic stress in veterans from weak labor markets relative to veterans from strong labor markets, which could potentially explain the results in the preceding section. Financial hardship or the lack of a social support structure due to joblessness could exacerbate stress associated with a traumatic event, which is consistent with evidence that PTSD symptoms worsen when individuals stop working (Schnurr, Lunney, Sengupta, and Spiro III, 2005, Frueh, Grubaugh, Elhai, and Buckley, 2007). If poor economic conditions aggravate symptoms of PTSD, then the effect of combat exposure should be larger for veterans from hometowns with lower employment rates than for veterans from hometowns with higher employment rates. I investigate this possibility by estimating the following equation: (1.7) Yi = a + 3EMPgr + yEi + 6(EMPgr x Ei) + Osjr + XL'9 + Ei Column 2 of Table 1.9 reports estimates of (1.7) where the outcome is a veteran's PTSD rating. The small, imprecise estimate of employment rates. 8 does not indicate an interaction between combat exposure and local Since estimates of the interaction term in (1.7) attribute any differences in the effects of combat exposure by race, deployment year, or geographic region to differences in predicted employment rates, column 3 reports estimates of the following equation: (1.8) Yi = a+ yEi + S(EMPgr x EI)+ pgjr + X:6 + Ei Column 1 of Table 1.9 reports estimates of (1.7), but excluding the interaction term. EMPr has been demeaned for all estimates reported in Table 1.9. 25 37 The main effect of EMP, is excluded from (1.8) because it is absorbed in the (Super PUMA) by (deployment year) by (race) fixed effects, which replace the (state) by (deployment year) by (race) fixed effects in (1.7). Estimates of (1.8) still do not indicate a significant interaction between employment rates and combat exposure. Columns 4 through 6 repeat columns 1 through 3 but replace E with WIA to investigate whether wounded veterans from areas with relatively poor economic conditions obtain higher PTSD ratings than do wounded veterans from strong labor markets. The point estimates reported in column 6 indicate that the effect of being wounded on PTSD ratings is 9 percent smaller for veterans from regions with an employment rate that is a standard deviation above the mean employment rate (13.89 - 0.22 x 5.9 = 12.59, which is 9 percent smaller than 13.89), although this estimate is not statistically significant. Overall, the estimates in Table 1.9 argue against the interpretation that veterans from regions with poor economic conditions are more psychologically vulnerable to combat exposure than veterans from regions with strong economic conditions. Another potential explanation for the results in the previous section is that low employment rates are correlated with factors that negatively influence mental health. Ideally, my data would contain a precise measure of veteran mental health that is independent of the VA disability program. Absent this, I investigate whether there is a relationship between local economic conditions and diagnoses from DoD disability exams. The DoD Disability Evaluation System follows the same rating schedule as the VA's DC program, but screening standards for DoD disability are substantially more stringent. A soldier is not eligible to apply for DoD disability unless his military commander or a medical provider refers him to a Medical Evaluation Board (MEB). Furthermore, DoD disability is usually only granted for conditions that make a soldier unfit for service in a combat zone. Whereas 32.5 percent of veterans in the National Guard sample received DC for any condition in January 2014, only 2.5 percent of the sample received DoD disability. On the other hand, over 90 percent of veterans who received DoD disability also received DC. Column 1 of panel A, Table 1.10, reports estimates of (1.6) where the outcome is 100 for veterans who receive any DoD disability pay and 0 otherwise. In column 2 the outcome indicates receipt of DoD disability for PTSD and in column 3 the outcome indicates receipt of DoD disability for any physical condition. The 38 results reported in all three columns indicate no relationship between local employment rates and any of the three DoD disability outcomes. Thus, while employment rates appear to predict a VA diagnosis of PTSD, they do not appear to predict a DoD diagnosis of PTSD that is based on the more strict screening standard. On the other hand, wounded veterans are substantially more likely to receive DoD disability for PTSD and physical conditions, a result that is consistent with the results reported in Table 1.6. There is also a positive, albeit statistically insignificant, relationship between Ei and receipt of DoD disability for PTSD. Still, the effects of wounds and combat exposure on DoD disability receipt are much smaller in magnitude than the effects of wounds and combat exposure on DC receipt, reflecting the more stringent screening standards for DoD disability relative to those for DC. To the extent that mortality is a downstream consequence of poor mental health, the absence of an association between local employment rates and mortality further argues against the hypothesis that local employment rates are negatively correlated with factors that aggravate mental health. This result is from column 4 of panel A, Table 1.10, which reports estimates of (1.6) where the left-hand side variable is 100 if a veteran was no longer alive as of May 2012 and 0 otherwise. The positive point estimate actually indicates a positive association between employment rates and mortality, although this estimate is only marginally significant. The low incidence of DoD disability and mortality makes these measures imperfect reflections of health. I therefore constructed geographic measures of self-reported health using the sample of post-9/l1 veterans born in 1974 or later within the 2009-2011 ACS. As an example, in column 1, panel B of Table 1.10, the outcome for veteran i who lives in Super PUMA g and is of race r is the percentage of post-9/l1 veterans in the ACS who live in Super PUMA g, are of race r, and who self-report having any disability for a condition other than hearing or eyesight. If veterans from weak labor markets have less access to quality health care than do veterans from strong labor markets, then we would expect to see a negative correlation between local employment rates and these measures of self-reported health. However, it is important to emphasize that the group of post-9/11 veterans in the ACS does not perfectly represent National Guard veterans in my sample because the ACS also includes veterans who served in the Active Duty Army, the Army Reserve, or other 39 military branches. Thus, any correlation between local economic conditions and the self-reported health of veterans in the ACS could also reflect differences in demographics or military experiences between National Guard soldiers and members of other services.26 The point estimate of -0.10 in column 1, panel B is marginally significant but still less than half the magnitude of the point estimate reported in column 4 of Table 1.10 (panel A). In other words, veterans from weak labor markets might be more likely to self-report disabilities than veterans from strong labor markets, but the magnitude of this relationship is not likely to explain the correlation between local labor markets and DC receipt within my National Guard sample. Column 2 of panel B is analogous to column 1, except the outcome in column 2 is the percentage of post-9/11 veterans in the ACS, for each combination of race and Super PUMA, who report having cognitive difficulty due to physical, mental, or emotional conditions. The point estimate reported in column 2, -0.050, is insignificant and less than one-fourth the magnitude of the correlation between employment rates and PTSD awards, as reported in panel B of Table 1.6. This result does not suggest that veterans from weak labor markets are substantially more likely to have cognitive difficulty than veterans from strong labor markets. Furthermore, the point estimate reported in column 3, where the outcome is the percentage of post-9/11 veterans in the ACS who report having a disability that substantially limits a basic physical activity, is similar in magnitude to the point estimate reported in column 2 (-0.050 in column 2 and -0.055 in column 3). Thus, it does not appear that labor market conditions are correlated with factors that influence mental health but not physical health. As a final check, I investigate whether the correlation between local employment rates and PTSD ratings is due to post-deployment military outcomes that influence disability receipt. Although National Guard veterans may receive DC while not performing an active duty mission, Guard veterans are still more likely to I used the 2009-2011 ACS to construct these indexes because the ACS did not track self-reported disability prior to 2009. 2011 was also the last year the ACS used the Super PUMA as a geographic unit. I also do not report estimates of Ei and WIA1 in panel B because geographic variation in combat exposure among National Guard veterans is not likely to represent the same variation among all post-9/11 veterans. 27 Cognitive difficulty is a common symptom of post-traumatic stress (American Psychiatric Association, 1994). 26 40 receive disability upon exiting the military. 8 Estimates of (1.6), where the left-hand side variable equals 100 for veterans who left the military prior to June 2013 and 0 otherwise, are reported in column 1 of Appendix Table 1.4. These estimates do not indicate that local employment rates are correlated with attrition from the military. Column 2 of the same table also indicates no significant correlation between local employment rates and the propensity for soldiers to deploy more than once. The results from this section do not suggest that health factors drive the correlation between local economic conditions and PTSD benefits. Similarly, the lack of correlation between employment rates and battlefield casualties (Table 1.5) argues against the hypothesis that the direct effects of combat exposure explain the relationship between local employment rates and PTSD benefits. An alternative explanation is that the large monetary value of PTSD benefits induces many veterans with poor employment prospects to file claims for PTSD. As discussed above, the correlation between local employment rates and PTSD benefits could also be the result of medical examiners or RVSRs who incorrectly interpret a veteran's employment status to be the direct consequence of post-traumatic stress. Similarly, medical examiners and RVSRs might produce reports that are more favorable to jobless disability applicants because higher CDRs result in higher priority access to VA healthcare. Unfortunately, since my data does not contain information on claims that do not result in disability ratings, I cannot distinguish between behavior on the part of veterans and behavior on the part of examiners. 1.7 Conclusion Disability benefits for PTSD are highly correlated with labor market conditions in a veteran's hometown while equally large benefits for physical disabilities are unrelated to the same conditions. This pattern seems unlikely to reflect elevated health risks for National Guard soldiers from regions with weak labor markets for several reasons. First, injuries sustained in combat are not correlated with labor market conditions. Second, veterans from regions with low employment-to-population ratios were no more likely to receive DoD PTSD 11 percent of veterans who were still in the military as of June 2013 received payments for PTSD while 23 percent of veterans who had left the military prior to June 2013 received benefits for PTSD. June 2013 is the last month for which I have a military record on file. 28 41 diagnoses, nor were they more likely to be deceased by 2012, than veterans from regions with high employment-to-population ratios. Third, the effects of combat exposure appear unrelated to local economic conditions, a result that weighs against the hypothesis that weak labor markets exacerbate symptoms of posttraumatic stress. These findings suggest that the relative attractiveness of DC when job opportunities are scarce contributes to the unusually high incidence of PTSD awards among recent veterans. At the same time, these results are also consistent with the possibility that medical examiners or VA ratings adjudicators incorrectly interpret joblessness to be a consequence of post-traumatic stress. The thrust of contemporary policy has been to expand benefits for PTSD. In July 2010, President Obama announced that the VA would ease the evidentiary standards required for veterans to receive compensation for PTSD (White House, 2010). Since veterans who were discharged from the military under other than honorable circumstances are not eligible for VA benefits, the Secretary of Defense recently simplified the application process for veterans seeking to upgrade their discharge status "based on claims of previously unrecognized post-traumatic stress" (Hagel, 2014). In response to statistics showing low earnings among those qualified for DC by virtue of PTSD, the Congressional Budget Office recommended supplementing disability payments to veterans with mental disorders (Congressional Budget Office, 2014). The results presented here suggest that low earnings among PTSD beneficiaries are not entirely due to health, but rather reflect a combination of generous benefits and an implicit tax on earnings resulting from the determination of mental disorder ratings and IU status based on the appearance of occupational impairment. A key result from my investigation is that much of the correlation between PTSD and weak labor markets is concentrated in PTSD awards rated 70 or higher, and thus eligible for IU status. The potential for larger payments could encourage many veterans to apply for increases in their PTSD ratings, a possibility that is consistent with the substantial CDR growth seen in Vietnam-era veterans immediately after they enroll in DC (Autor et al., 2015a). Likewise, the financial incentives associated with DC could act as disincentives for rehabilitation. There is evidence that disability status predicts treatment withdrawal among OIF and OEF veterans diagnosed with PTSD (Gros et al., 2013). Earlier studies indicate that treatment for PTSD is less effective for veterans than for other trauma victims (Bradley et al., 2005). Similarly, a 2005 investigation by 42 the VA Office of the Inspector General reveals that attendance at VA behavioral health treatment sessions decreased by 80 percent when veterans received the maximum level of DC payments. For many veterans who receive PTSD benefits, labor market reintegration could prove even more difficult than psychological rehabilitation. Recent research on SSDI applicants indicates that the probability of future employment declines by 0.4 to 0.5 percentage points each month that an applicant waits for a decision, during which time he is not allowed to work (Autor, Maestas, Mullen, and Strand, 2015b). There is also substantial evidence that employers discriminate against job applicants with long unemployment spells (Eriksson and Rooth, 2014, Kroft, Lange, and Notowidigdo, 2013, Oberholzer-Gee, 2008). This presents a difficult challenge for the VA. On one hand, DC benefits contribute to an improved quality of life for those who indeed struggle in the labor market. On the other hand, the strong anti-work incentives inherent in the DC program, especially for PTSD, could have long-term repercussions on the health and well-being of post9/11 veterans. 43 1.8 Chapter 1 Appendix A. CasualtyInformation When a company suspects that a soldier has been either killed or wounded, it is required to fill out a Casualty Feeder Card (DA Form 1156) in accordance with AR 600-8-1. The Casualty Feeder Card requires units to indicate when and where a casualty occurred, whether a casualty was the result of hostile or nonhostile action, and whether the casualty is deceased or injured. They then forward this information to their battalion's personnel shop, which subsequently puts this information into DCIPS. 29 Casualties who are declared deceased as a result of hostile action are reported as Killed in Action (KIA) and those who are declared wounded as a result of hostile action are reported as Wounded in Action (WIA). Battalion personnel shops will initiate a casualty report in DCIPS for any soldiers who are involved in a hostile action and subsequently determined by medical personnel to have been injured or killed. Likewise, battalion personnel shops will update DCIPS entries for any soldiers initially entered into the database as WIA or KIA, but who are subsequently determined either not to have sustained injuries or not to have been injured or killed as the result of hostile action. I implicitly assume soldiers cannot manipulate the casualty process in any systematic way. This assumption might be problematic because the Casualty Feeder Card, the form used to initially label a soldier as wounded before he or she is entered into DCIPS, is one of several pieces of information required to receive a Purple Heart Award. The Purple Heart does not come with any monetary awards but is often viewed with great pride in military and civilian circles. Some might argue that soldiers with lower potential earnings might have more of an incentive to be labeled as wounded, which might explain why soldiers with lower AFQT scores suffer more wounds. I do not believe this potential bias presents a major problem, for three reasons. First, there is a negative relationship between both KIA and AFQT scores and WIA and AFQT scores. Second, the military has an incentive to tally accurate casualty records because the DCIPS database is "used by DoD organizations, external government agencies, both houses of Congress, the President, the news media, and the general public.. .to understand trends in casualties as they relate to terrain, advances in 29 A battalion is the higher headquarters for a company. Most battalions consist of 3 to 5 companies. 44 medicine, the advent of better technology that has enhanced the safety of the war fighter, or the challenges brought about by new threats" (DMDC Defense Casualty Analysis System website). Third, medical personnel and military command teams are wearisome of soldiers who might attempt to fake an injury for personal gain. It seems reasonable to assume any attempts to falsify casualty reports are not large enough to significantly bias my results. B. ConstructingLocal Employment Rates with the ACS I construct local employment rates using the 2005-2009 five-year American Community Survey (ACS). I first limited the ACS to men born between 1974 and 1986. I excluded individuals who were not in the labor force, who belonged to institutional group quarters, who served in the active duty military at the time of the survey, and who had served in the military at any point after September 2001. I then grouped individuals into cells of race (non-Hispanic whites and non-whites) and Super Public Use Microdata Area (PUMA). To construct local employment rates, I divided the number of employed persons by the total number of persons in each cell, weighting observations according to the ACS "perwt" variable. After constructing local employment rates for each combination of (race)x(Super PUMA) within the 2005-2009 ACS, I assigned an employment rate to everyone in the National Guard sample. I defined a soldier's Super PUMA based on the zip code of his mailing address at the start of his first mobilization. I mapped each five-digit zip code to a specific Super PUMA using the zip-code-to-PUMA crosswalk file (based on year 2000 geographic boundaries) provided by the Geographic Correspondence Engine of the Missouri Census Data Center (http://mcdc2.missouri.edu/websas/geocorr2k.html). Each PUMA belongs to only one Super PUMA. For zip codes that span two or more PUMAs, I assigned each zip code to the PUMA with the largest allocation factor, where an allocation factor is the portion of each zip code that corresponds to each PUMA. Half of zip codes that span multiple PUMAs have maximum allocation factors above 90 percent, meaning 90 percent or more of such zip codes belong to one PUMA, and 98 percent of zip codes that span multiple PUMAs have maximum allocation factors above 50 percent. 45 About 2 percent of soldiers in the National Guard sample lived in zip codes that did not map to a Super PUMA. I assigned these soldiers to the Super PUMA corresponding to the following zip code, but conditional on the following zip code having the same first three digits as the original zip code. For example, if veterans from zip code 11101 did not match to a Super PUMA, then I mapped them to the Super PUMA corresponding to zip code 11102. However, if veterans from zip code 11199 did not match to a Super PUMA, then they were excluded from my final sample. Overall, I was able to assign all but seven soldiers to a Super PUMA. C. Sample Selection Details My original sample consists of 99,010 National Guard soldiers who enlisted in the National Guard between fiscal years 1993 and 2006 and who deployed to a combat zone (mostly Iraq and Afghanistan) between January 2003 and December 2006. I first remove 9,869 female soldiers, then 25,599 soldiers with prior military experience. Soldiers with prior military service are older than non-prior-service soldiers, have different work experiences, and have different incentives for serving in the National Guard.' I then remove 1,952 soldiers who deployed to a combat zone more than ten years after their initial entry into the National Guard. Soldiers who serve beyond ten years have reenlisted at least twice and are likely to serve until they qualify for retirement benefits. I further exclude an additional 1,128 soldiers who have missing AFQT information, missing education information, or who enlisted before 17 years of age or after 35 years of age.3 Since I exploit within-state variation in employment rates at the Super PUMA level, I remove 1,667 veterans from U.S. territories, 7 veterans who had zip codes that I could not match to a particular Super PUMA, and This number is comparable to a 2010 RAND study that used data managed under the Office of the Assistant Secretary of Defense for Reserve Affairs to find that 168,213 Army National Guard soldiers were mobilized under Federal Title 10 authority between September 2001 and May 2007 (Bonds et al., 2010, p.8). The RAND number is larger than the number of soldiers in my original sample because it includes officers, soldiers who joined the National Guard before October 1992, soldiers who mobilized in support of operations conducted within the Continental U.S. (these mobilizations do not constitute a combat deployment within my data), and a few soldiers who deployed to combat zones before 2003 or between January and May 2007. 30 Prior service soldiers can join the National Guard as a way to finish out their eight-year military commitment (at least four years of which must be on Active Duty) while still receiving heavily discounted health care through TRICARE. 32 National Guard soldiers are not allowed to enlist in the military before turning 17 or after turning 35. The few in my sample who did were either granted special waivers or had incorrectly coded years of birth. 31 46 2,679 veterans from Alaska, North Dakota, South Dakota, Wyoming, Vermont, or Washington, D.C., all of which have only one Super PUMA. I finally remove 176 veterans who were the only soldiers in their company during this first deployment. 3 My final sample consists of 55,936 enlisted National Guard soldiers. Above I describe how I use casualty rates for other soldiers in the same company to construct a measure of combat exposure. Soldiers who are the only member of a company will mechanically have no value for this measure. It makes sense to exclude them from my analysis because these rare instances usually only occur for soldiers assigned to higher headquarter staffs. 3 47 Figure 1.1. Any PTSD and Any Physical Disability by Local Employment Rates B. Percentage with any DC for physical conditions - A. Percentage with any DC for PTSD . 0 - 0 0 ,0 0 ,e L() . 0 4 e 0 0 0 C S *0 4 0e@ a 0 0 40 - LO) Nl 0- 68 72.4 76.8 81.2 Local Employment Rate 85.6 90 68 72.4 76.8 81.2 Local Employment Rate Notes: This figure plots the percentage of veterans in the National Guard sample who receive DC for PTSD (panel A) and the percentage who receive DC for physical conditions (panel B) by local employment rates. 85.6 90 Figure 1.2. Mean PTSD Rating and Mean Physical Rating by Local Employment Rates B. Physical Rating A. PTSD Rating (0~ CO. 0@e 0 000 CqJ 0 0 .0 0 .s CO- Do 4 S. 0 0- 0- 68 72.4 76.8 81.2 Local Employment Rate 85.6 90 68 72.4 76.8 81.2 85.6 Local Employment Rate Notes: This figure plots mean PTSD ratings (panel A) and mean physical ratings (panel B) by local employment rates for veterans in the National Guard Sample. 90 Figure 1.3. High PTSD Ratings and High Physical Ratings by Local Employment Rates B. Percentage with Physical Rating>60 A. Percentage with PTSD Rating>60 0 0 0)- to (0- 0 0 0@e q M 0 w 000 0- 0- 68 72.4 76.8 81.2 Local Employment Rate 85.6 90 68 72.4 76.8 81.2 85.6 Local Employment Rate Notes: This figure plots the percentage of veterans in the National Guard sample with a PTSD rating above 60 (panel A) and the percentage with a physical rating above 60(panel B) by local employment rates. 90 ............ Any DC Receipt PTSD Benefits -'A- Table 1.1. DC Receipt and Casualty Rates By Wartime Era World War II Korea Vietnam (3) (2) (1) A. DC Receipt as of 2013 0.120 0.070 0.171 0.013 0.006 0.047 Wounded in Action (WIA) Killed in Action (KIA) B. Casualty Rates 0.042 0.018 0.0059 0.0181 0.035 0.0054 Post-9/1 1 (4) 0.310 0.084 0.014 0.0015 Notes: This table reports DC receipt and casualty rates by wartime era. In panel A, DC receipt rates are the number of veterans from a wartime era who receive DC divided by the total number of veterans from that wartime era. In panel B, casualty rates are the number of casualties divided by the total number of persons who served during each wartime era. The number of veterans from each wartime era who receive DC or specifically PTSD benefits is from the Veterans Benefits Administration 2013 Annual Report. The number of casualties during WWII, Korea, and Vietnam come from Leland and Oboroceanu (2010). The number of casualties for Post-9/1 I veterans is available at http://www.defense.gove/news/casualty.pdf (accessed 12 February 2015). The number of veterans from each wartime era is from the 2014 Veteran Population Projection Model, available at http://www.va.gov/vetdata/Veteran-population.asp. The denominator for Post-9/1 1 DC rates excludes 1.4 million members who were serving in the military in September 2012, which comes from Defense Manpower Data Center Historical Reports. 51 -AAwwk Table 1.2. Disability Rating Distributions for Most Common 0 Rating 10 - 20 Rating 30 - 40 Rating 50 Body System (1) (2) (3) Musculoskeletal 0.276 0.644 0.062 Auditory 0.340 0.605 0.029 Skin 0.732 0.228 0.031 Neurological 0.169 0.651 0.141 Mental 0.025 0.113 0.269 Body Systems - 60 Rating 70 - 100 Rating (4) (5) 0.011 0.006 0.012 0.013 0.008 0.001 0.012 0.013 0.214 0.359 Notes: This table reports the share of disabilities that fall within the specified rating range for the five most common body systems with disabilities. The data is from the VBA 2013 Annual Report. DC recipients from all wartime eras are included in the reported statistics because VA Annual Benefits Reports do not distinguish these statistics by wartime era. Table 1.3. Employment Rate by Combined Disability Rating - Male, Post-9/1 1 Veterans No DC 10-20 Rating 30-40 Rating 50-60 Rating 70 - 100 Rating (5) (4) (3) (2) (1) Share Employed 0.828 0.834 0.761 0.701 0.424 Notes: This table reports the share of employed veterans by Combined Disability Rating (CDR) for malb post-9/1 I DC recipients who were born between 1974 and 1986, based on data from the 2013 American Community Survey. It excludes 130 veterans with no reported CDR. 52 Table 1.4. Summary Statistics White (Non-Hispanic) GED or High School Dropout High School Diploma Some College Baccalaureate Degree Combat Occupation Military Rank PFC or Below Noncommissioned Officer DC Receipt All No DC Receipt (No PTSD) (1) (2) (3) A. Demogranhic Variables 33.3 33.1 33.6 [4.1] [4.3] [3.9] 59.0 60.4 59.0 [19.5] [19.6] (19.5) 0.805 0.805 0.828 0.172 0.161 0.154 0.713 0.721 0.711 0.095 0.097 0.110 0.020 0.021 0.025 0.431 0.402 0.466 0.295 0.283 0.292 0.166 0.174 0.173 Wounded in Action (WIA) Killed in Action (KIA) Deceased by May 2012 B. Casualty and Mortality Rates 0.033 0.014 0.050 0.003 0.004 0 0.008 0.012 0 Age in 2014 AFQT Any DC Receipt Any PTSD Benefits PTSD Is Primary Condition IU Benefit Combined Disability Rating PTSD Rating Physical Rating Mean Monthly DC Payment Observations DC Receipt (PTSD) (4) CDR=100 or IU Benefit (5) 33.8 [4.4] 54.2 [18.2] 0.787 0.221 0.686 0.079 0.014 0.511 0.339 0.132 34.7 [5.0] 51.4 [17.3] 0.776 0.302 0.625 0.061 0.012 0.493 0.379 0.110 0.088 0 0 0.128 0 0 C. Disability Compensation Variables 0.325 1 0.186 0 0.172 0 0.015 0.012 16.2 31.0 [28.1] [23.5] 10.0 [23.6] 7.0 20.0 [16.8] [21.4] 321 529 [655] [590] 0.928 0.071 63.9 [21.6] 53.6 [25.5] 22.6 [25.0] 1334 [788] 55,936 10,380 37,779 7,777 1 1 1 0.866 0.808 0.407 91.7 [11.4] 82.1 [35.5] 39.5 [34.9] 2858 [0] 2,049 Notes: This table reports descriptive statistics for veterans in the National Guard sample. In panel A, a veteran's education level, occupation, and military rank is determined at the start of his first deployment to a combat zone between 2003 and 2006. In panel C, DC receipt information is as of January 2014. A veteran's primary condition is the condition for which he receives the highest disability rating. A veteran's PTSD rating is 0 if he receives no benefits for PTSD and 100 if he receives an IU award for PTSD. A veteran's physical rating is the CDR he would receive excluding any ratings for mental health conditions and tinnitus. Veterans who receive an IU award for a physical condition are given a physical rating of 100 and veterans with no physical conditions are given a physical rating of 0. Mean monthly DC payments represent the payment a veteran would receive in 2014 based on his combined disability rating and assuming no dependents. Veterans with no disability awards are given a mean monthly DC payment of 0. Standard deviations are reported in brackets. 53 Table 1.5. Relationship Between Local Employment Rates and Casualties Local Employment Rate AFQT (3) (2) (1) A. Dependent Variable: 100 x All Casualties (KIA + WIA) 0.007 0.013 0.024 (0.027) (0.027) (0.026) -0.025*** -0.021*** (0.004) (0.004) Combat Exposure (E,) Outcome Mean Local Employment Rate AFQT 3.59 3.59 3.59 B. Dependent Variable: 100 x 1[Killed in Action] 0.0089 0.0097 0.0104* (0.0061) (0.0061) (0.0061) -0.0028*** -0.0025** (0.0011) (0.0012) Combat Exposure (E,) Outcome Mean Local Employment Rate AFQT 0.28 0.28 0.28 C. Dependent Variable: 100 x Wounded in Action -0.002 0.003 0.014 (0.025) (0.025) (0.024) -0.022*** -0.018*** (0.004) (0.004) Combat Exposure (E;) Outcome Mean 3.31 Individual Controls Clusters Observations 1,048 55,936 3.31 3.31 1,048 55,936 X 1,048 55,936 (4) 0.019 (0.023) -0.021*** (0.004) 0.605*** (0.031) 3.59 0.0105* (0.0062) -0.0025** (0.0012) -0.0103** (0.0052) 0.28 0.009 (0.022) -0.019*** (0.004) 0.616*** (0.030) 3.31 X 1,048 55,936 Notes: This table reports estimates of the relationship between local employment rates and all casualties sustained by hostile action, fatal casualties, and non-fatal casualties for veterans in the National Guard sample. The dependent variable in each regression is 100 times the number of casualties sustained over all deployments after 2002. The local employment rate is the employment-to-population ratio for nbn-veteran men born between 1974 and 1986 by race (white or non-white), within each Super PUMA of the 2005-2009 ACS. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race). Individual controls (columns 3 and 4) include year-of-birth indicators, military occupation indicators, military rank indicators, and education indicators. A veteran's state and all individual controls are determined at the start of the veteran's first deployment. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 54 Table 1.6. Impact of Local Employment Rates on any DC, any PTSD, and any Physical Disability (4) (3) (2) (1) A. Dependent Variable: 100 x I [Any Disability Compensation] -0.227*** -0.199*** -0.208*** Local Employment Rate -0.294*** (0.086) (0.084) (0.077) (0.078) -0.219*** -0.214*** -0.253*** AFQT (0.011) (0.011) (0.011) Combat Exposure (E,) 0.601*** (0.053) 28.98*** Wounded in Action (1.10) Mean Outcome 32.46 32.46 32.46 32.46 B. Dependent Variable: 100 x I [Any Compensation -0.234*** -0.294*** Local Employment Rate (0.071) (0.073) -0.227*** AFQT (0.009) Combat Exposure (E,) for PTSD] -0.212*** (0.065) -0.192*** (0.009) Wounded in Action Mean Outcome 18.56 18.56 18.56 C. Dependent Variable: 100 x I [Any Compensation for a Physical Disability] -0.154** -0.205*** -0.141** Local Employment Rate (0.067) (0.068) (0.061) -0.191*** -0.169*** AFQT (0.009) (0.009) Combat Exposure (E ,) -0.220*** (0.066) -0.188*** (0.009) 0.622*** (0.051) 22.52*** (1.16) 18.56 Mean Outcome 22.58 22.58 22.58 -0.148** (0.062) -0.163*** (0.009) 0.290*** (0.049) 32.53*** (1.17) 22.58 Individual Controls Clusters Observations 1,048 55,936 1,048 55,936 X 1,048 55,936 X 1,048 55,936 Wounded in Action Notes: This table reports estimates of the effects of local employment rates on any DC receipt, any PTSD disability receipt, and any physical disability receipt for veterans in the National Guard sample. All dependent variables reflect disability ratings as of January 2014. Physical disabilities are all disabilities other than mental disorders and tinnitus (ringing of the ears). The local employment rate is the employment-to-population ratio for non-veteran men born between 1974 and 1986, by race (white or non-white), within each Super PUMA of the 2005-2009 ACS. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race). Individual controls (columns 3 and 4) include year-of-birth dummies, military occupation dummies, military rank dummies, and education dummies. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 55 Table 1.7. Impact of Local Employment Rates on CDRs, PTSD Ratings, and Physical Disability Ratings (1) (2) (3) (4) A. Dependent Variable: Combined Disability Rating (CDR) Local Employment Rate -0.218*** -0.168*** -0.154*** -0.160*** (0.061) (0.059) (0.053) (0.054) AFQT -0.188*** -0.157*** -0.153*** (0.007) (0.007) (0.007) Combat Exposure (E) 0.342*** Wounded in Action Mean Outcome Local Employment Rate AFQT 16.46 16.46 B. Dependent Variable: PTSD Rating -0.2l1*** -0.174*** (0.049) (0.048) -0.140*** (0.006) 16.46 -0.164*** (0.044) -0.115*** (0.006) Combat Exposure (E) Wounded in Action Mean Outcome Local Employment Rate AFQT 9.95 9.95 9.95 C. Dependent Variable: Physical Disability Rating -0.053* -0.033 -0.031 (0.029) (0.029) (0.027) -0.074*** -0.064*** (0.004) (0.004) Combat Exposure (E) Wounded in Action Mean Outcome Individual Controls Clusters Observations 6.97 6.97 6.97 1,048 55,936 1,048 55,936 X 1,048 55,936 (0.035) 23.76*** (0.88) 16.46 -0.169*** (0.045) -0.113*** (0.006) 0.296*** (0.029) 12.96*** (0.77) 9.95 -0.034 (0.027) -0.061*** (0.004) 0.076*** (0.020) 17.58*** (0.74) 6.97 X 1,048 55,936 Notes: This table reports estimates of the effects of local employment rates on CDRs, PTSD ratings, and physical ratings for veterans in the National Guard sample. All dependent variables reflect disability ratings as of January 2014. CDRs, PTSD ratings, and physical ratings equal zero for veterans who did not receive DC, benefits for PTSD, or benefits for physical conditions, respectively. Veterans with an IU benefit are given a CDR of 100 and a PTSD or physical rating of 100 if the IU benefit is because of PTSD or a physical condition. A veteran's physical rating is his CDR after excluding all mental disabilities and tinnitus. The local employment rate is the employmentto-population ratio for non-veteran men born between 1974 and 1986, by race (white or non-white), within each Super PUMA of the 2005-2009 ACS. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race). Individual controls include year-of-birth dummies, military occupation dummies, military rank dummies, and education dummies. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 56 Table 1.8. Impact of Local Employment Rates on Disability Compensation - Extensive and Intensive Margins 100x 1 [PTSD Rating>0] Local Employment Rate Combat Exposure (E1 ) Wounded in Action Mean Outcome Clusters (Company) Observations A. PTSD Disability Rating Outcomes 100x 100x 1 [PTSD Rating>50] I [PTSD Rating?30] 100x 1 [PTSD Rating>70] 100x I [PTSD Rating= 00/IU] (5) -0.098*** (0.024) 0.071*** (0.021) 4.52*** (0.68) (1) -0.220*** (0.066) 0.622*** (0.051) 22.52*** (1.16) (2) -0.208*** (0.067) 0.564*** (0.046) 21.74*** (1.15) (3) -0.193*** (0.063) 0.336*** (0.040) 16.24*** (1.13) (4) -0.184*** (0.046) 0.162*** (0.031) 8.80*** (0.93) 18.56 1,048 55,936 17.57 1,048 55,936 12.15 1,048 55,936 6.53 1,048 55,936 1,048 55,936 2.82 B. Physical Disability Rating Outcomes 100 x I [Phys. Rating>0] Local Employment Rate Combat Exposure (E,) Wounded in Action Mean Outcome Clusters (Company) Observations (1) -0.148** (0.062) 0.290*** (0.049) 32.53*** (1.17) 22.58 1,048 55,936 100x 100x 1 [Phys. Rating;>30] (2) -0.075 (0.050) 0.136*** (0.037) 27.49*** (1.24) 1 [Phys. Rating>50] (3) 0.012 (0.029) 0.038 (0.026) 19.30*** (1.04) 100 x I [Phys. Rating>70] (4) 0.003 (0.017) 0.011 (0.018) 11.01*** (0.83) 100 x 1 [Phys. Rating=100/IU] (5) -0.003 (0.008) 0.002 (0.007) 2.39*** (0.36) 10.96 1,048 55,936 5.03 1,048 55,936 2.11 1,048 55,936 0.46 1,048 55,936 Notes: This table reports estimates of the effects of local employment rates on PTSD disability awards and physical disability awards. The dependent variable in each regression equals 100 for veterans with an award rating greater than or equal to the cutoff specified in each column, and 0 otherwise. The notes in Table 7 describe the construction of PTSD ratings, physical disability ratings, and local employment rates. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race), a linear term for AFQT, year-of-birth indicators, military occupation indicators, military rank indicators, and education indicators. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 57 Table 1.9. Interaction of Local Employment Rates and Combat Exposure Measures Dependent Variable: PTSD Disability Rating (2) -0.165*** (0.047) 0.376*** (0.030) 0.376*** (0.030) -0.001 (0.005) (3) (4) -0.166*** (0.044) (5) -0.15 (0.044 14.03*** (0.76) 14.06*** (0.75) -0.25* (0.13) 13.89*** (0.80) -0.22 (0.14) 9.95 1,048 55,936 9.95 1,048 55,936 X 9.95 1,048 55,936 (6) ) Local Employment Rat e (1) -0.167*** (0.046) Combat Exposure (E,) (E,) x (Local Emp. Rat e) 0.361*** (0.032) 0.002 (0.006) Wounded in Action (WIA) (WIA) x (Local Emp. Rate) (Super PUMA) x (Race) x (Deployment Year) FE Mean Outcome Clusters Observations 9.95 1,048 55,936 X 9.95 1,048 55,936 9.95 1,048 55,936 Notes: This table reports estimates of equations (7) and (8) for veterans in the National Guard sample. The dependent variable is a veteran's PTSD rating in January 2014. Veterans who do not receive compensation for PTSD have a rating of zero and veterans who receive an IU award for PTSD are given a PTSD rating of 100. The local employment rate is the employment-topopulation ratio for non-veteran men born between 1974 and 1986, by race (white or non-white), within each Super PUMA of the 2005-2009 ACS. The results reported in columns 1, 2, 4, and 5 include a fixed effect for each combination of (state) x (deployment year) x (race) while the results reported in columns 3 and 6 include a fixed effect for each combination of (Super PUMA) x (deployment year) x (race). All regressions include a linear term for AFQT, year-of-birth indicators, military occupation indicators, military rank indicators, and education indicators. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 58 1- 1.... . .. . -- .......... - . Table 1.10. Impact of Local Employment Rates on DoD Disability, Mortality, and Veteran Health A. DoD Disability and Mortality Outcomes 100 x 1 [Any DoD Disability] Local Employment Rate Combat Exposure (E,) Wounded in Action Mean Outcome Clusters Observations (1) -0.010 (0.018) 0.013 (0.018) 11.43*** (0.87) 2.426 1,048 55,936 100 x 1 [DoD Disability for PTSD] (2) -0.001 (0.011) 0.018 (0.013) 4.43*** (0.55) 100 x 1 [DoD Physical Disability] (3) 0.004 (0.016) 0.011 (0.016) 10.93*** (0.86) (4) 0.019* (0.011) -0.008 (0.009) -0.05 (0.26) 0.833 1,048 55,936 2.081 1,048 55,936 0.828 1,048 55,936 100 x 1 [Deceased by May 2012] B. Veteran Health Index Outcomes Veteran SelfReported Disability Index Veteran Physical Health Index (0.060) Veteran Cognitive Health Index (2) -0.050 (0.045) 7.210 1,047 55,928 4.979 1,047 55,928 2.939 1,047 55,928 (1) Local Employment Rate Mean Outcome Clusters Observations -0. 100* (3) -0.055 (0.039) Notes: This table reports estimates of equation (6) for veterans in the National Guard sample where the dependent variable is described by the header for each column. A veteran is considered deceased if he was identified as such in the SSA Death Master File prior to June 2012. The Veteran Self-Reported Disability Index is the percentage of male, post-9/11 veterans in each Super PUMA (by race) of the 2009-2011 ACS who were born in 1974 or later and who reported having any disability other than one involving hearing or eyesight. The Veteran Cognitive Health Index and the Veteran Physical Health Index are the percentage of male, post-9/11 veterans in each Super PUMA (by race) of the 2009-2011 ACS who reported having cognitive difficulties and physical disabilities, respectively. The local employment rate is the employment-to-population ratio for nonveteran men born between 1974 and 1986, by race (white or non-white), within each Super PUMA of the 20052009 ACS. All regressions include year-of-birth indicators, a linear term for AFQT, military occupation indicators, military rank indicators, education indicators, and a fixed effect for each combination of (state) x (deployment year) x (race). Standard errors are clustered on each combination of Super PUMA and race (white or non-white). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 59 Appendix Table 1.1. Covariate Balance on E Baseline Characteristic Mean (1) AFQT Age in 2012 Local employment rate GED or less than high school 59.02 [19.50] 33.33 [4.07] 80.73 [5.91] 0.17 Some college or more 0.12 No dependents 0.70 (State) x (Race) x (Dep Year) FE Occupation and Rank FE P Val (Joint x 2 Test) Clusters Observations 55,936 Regression of LHS Variable on E; (3) (2) -0.0488** (0.0215) -0.0189*** (0.0035) -0.0006 (0.0093) 0.0000 (0.0004) 0.0003 (0.0003) 0.0014*** (0.0004) 0.0218 (0.0207) -0.0028 (0.0036) 0.0052 (0.0095) -0.0005 (0.0004) 0.0008** (0.0003) 0.0002 (0.0004) X X X 0.000 0.172 1048 55,936 1048 55,936 Notes: This table reports coefficients from regressions of the baseline characteristic listed on E j, the combat exposure measure described in section 3.3. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race). All baseline characteristics, including a veteran's state of residence, are determined at the start of a veteran's first deployment between 2003 and 2006. The construction of local employment rates is described in section 3.2. Joint X 2 tests are for the null hypothesis that the coefficients on combat exposure in all regressions are equal to zero. Standard deviations are reported in brackets. Standard errors, clustered on each combination of (Super PUMA) x (race), are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 60 Appendix Table 1.2. Impact of Local Employment Rates on IU and 100 Ratings A. PTSD Disability Rating Outcomes Local Employment Rate Combat Exposure (E,) Wounded in Action Mean Outcome Clusters (Company) Observations Local Employment Rate Combat Exposure (E,) Wounded in Action Mean Outcome Clusters (Company) Observations 100x 100x I [PTSD Rating= 00/IU] 1 [PTSD Rating=IU] (2) -0.024** (0.012) 0.021* (0.012) 2.451*** (0.498) (1) -0.098*** (0.024) 0.071*** (0.021) 4.519*** (0.677) 2.82 1,048 55,936 1.24 1,048 55,936 B. Physical Disability Rating Outcomes 100x 100 x 1[Phys. Rating= I00/IU] I [Phys. Rating=IU] (2) (1) -0.003 -0.001 (0.008) (0.005) 0.002 0.004 (0.007) (0.005) 2.389*** 0.743*** (0.356) (0.217) 0.46 1,048 55,936 0.18 1,048 55,936 100 x I [PTSD Rating=100] (3) -0.074*** (0.019) 0.050*** (0.017) 2.068*** (0.464) 1.58 1,048 55,936 100 x I [Phys. Rating=100] (3) -0.002 (0.006) -0.002 (0.005) 1.646*** (0.291) 0.28 1,048 55,936 Notes: This table reports estimates of the effects of local employment rates on PTSD disability awards and physical disability awards. The dependent variable in each regression equals 100 for veterans with an award specified in each column, and 0 otherwise. The notes in Table 7 describe the construction of PTSD ratings, physical disability ratings, and local employment rates. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race), a linear term for AFQT, year-of-birth indicators, military occupation indicators, military rank indicators, and education indicators. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 61 Appendix Table 1.3. Impact of Local Employment Rates on Disability Compensation Outcomes - Extensive and Intensive Margins 100x 100x 100x 100x 100x 00/lU] 1 [Rating=1 1 [Rating>70] 50] [Rating 1 I [Rating>30] 1 [Rating>0] (5) (4) (3) (2) (1) A. PTSD Disability Rating Outcomes -0.098*** -0.184*** -0.193*** -0.208*** -0.220*** Local Employment Rate (0.024) (0.046) (0.063) (0.067) (0.066) 2.82 6.53 12.15 17.57 18.56 Mean Outcome Local Employment Rate Mean Outcome Local Employment Rate Mean Outcome Local Employment Rate Mean Outcome Local Employment Rate Mean Outcome -0.148** (0.062) 22.58 B. Physical Disability Rating Outcomes 0.012 -0.075 (0.029) (0.050) 5.03 10.96 0.003 (0.017) 2.11 -0.003 (0.008) 0.46 -0.141** (0.071) 28.43 C. Non-PTSD Rating Outcomes 0.009 -0.075 (0.036) (0.053) 7.25 14.79 0.009 (0.021) 3.22 0.003 (0.010) 0.74 0.041* (0.022) 3.66 D. Mental Disorder Rating (excluding PTSD) Outcomes 0.007 0.023 (0.012) (0.019) 1.24 2.84 0.001 (0.008) 0.58 0.006 (0.006) 0.27 E. Tinnitus Outcome (all tinnitus diagnoses results in a rating of 10) -0.106** (0.053) 16.69 Notes: This table reports estimates of the effects of the local employment rates on PTSD disability awards, physical awards, non-PTSD awards, mental awards for conditions other than PTSD, and tinnitus. The dependent variable in each regression equals 100 for veterans with an award rating greater than or equal to the cutoff specified in each column, and 0 otherwise. The notes in Table 7 describe the construction of PTSD ratings, physical disability ratings, and local employment rates. Non-PTSD ratings represent a veteran's CDR from all conditions other than PTSD while mental disability ratings represent a veteran's CDR based on all mental conditions other than PTSD. All regressions include a fixed effect for each combination of (state) x (deployment year) x (race), a linear term for AFQT, year-of-birth indicators, military occupation indicators, military rank indicators, and education indicators. The number of observations in each regression is 55,936 and the number of clusters is 1,048. Standard errors are clustered on each combination of (Super PUMA) x (race). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 62 Appendix Table 1.4. Impact of Local Employment Rates on Military Outcomes 100 x 1 [Exit Army by June 2013] Local Employment Rate Combat Exposure (E,) Wounded in Action Mean Outcome Clusters Observations 100 x 1 [More than One Deployment] (1) -0.074 (0.071) 0.187*** (0.046) 1.83 (1.12) (2) 0.053 (0.065) 65.130 1,048 55,936 31.742 1,048 55,936 -0.306*** (0.059) 3.55*** (1.22) Notes: This table reports estimates of equation (6) for veterans in the National Guard sample where the dependent variable is described by the header for each column. The local employment rate is the employment-to-population ratio for nonveteran men born between 1974 and 1986, by race (white or non-white), within each Super PUMA of the 2005-2009 ACS. All regressions included, year-of-birth indicators, a linear term for AFQT, military occupation indicators, military rank indicators, education indicators, and include a fixed effect for each combination of (state) x (deployment year) x (race). Standard errors are clustered on each combination of Super PUMA and race (white or nonwhite). ***,**, and * denote significance at the 1%, 5%, and 10% levels, respectively. 63 Chapter 2. The Impact of Disability Benefits on Labor Supply: Evidence from the VA's Disability Compensation Program* Co-Authored with David H. Autor (MIT and NBER), Mark Duggan (Stanford and NBER), and David S. Lyle (United States Military Academy) April 2015 Abstract Combining administrative data from the U.S. Army, Department of Veterans Affairs (VA) and the U.S. Social Security Administration, we analyze the effect of the VA's Disability Compensation (DC) program on veterans' labor force participation and earnings. The largely unstudied Disability Compensation program currently provides income and health insurance to almost four million veterans of military service who suffer service-connected disabilities. We study a unique policy change, the 2001 Agent Orange decision, which expanded DC eligibility for Vietnam veterans who had served in-theatre to a broader set of conditions such as type 2 diabetes. Exploiting the fact that the Agent Orange policy excluded Vietnam era veterans who did not serve in-theatre, we assess the causal effects of DC eligibility by contrasting the outcomes of these two Vietnam-era veteran groups. The Agent Orange policy catalyzed a sharp increase in DC enrollment among veterans who served in-theatre, raising the share receiving benefits by five percentage points over five years. Disability ratings and payments rose rapidly among those newly enrolled, with average annual non-taxed federal transfer payments increasing to $17K within five years. We estimate that benefits receipt reduced labor force participation by 18 percentage points among veterans enrolled due to the policy, though measured income net of transfer benefits rose on average. Consistent with the relatively advanced age and diminished health of Vietnam era veterans in this period, we estimate labor force participation elasticities that are somewhat higher than among the general population. * This research was supported by the U.S. Social Security Administration through grant #10-P-98363-1-05 to the National Bureau of Economic Research as part of the SSA Retirement Research Consortium. The views and findings expressed herein are those of the authors and do not purport to reflect the position of the U.S. Military Academy, the Department of the Army, the Department of Defense, the SSA, or the NBER. We are grateful to Josh Angrist, Orley Ashenfelter, Mary Daly, and seminar participants at Arizona State University, the Federal Reserve Board, Princeton University, and Stanford University for helpful comments. We are indebted to Luke Gallagher of the Army Office of Economic Manpower Analysis for outstanding research assistance and to Mike Risha of the Social Security Administration for assistance with all aspects of data development and interpretation. 64 2.1 Introduction This chapter investigates the effect of the U.S. Department of Veterans Affairs' (VA's) Disability Compensation (DC) program on the labor supply of military veterans. Since the ratification of the U.S. Constitution in 1789, the federal government has provided cash benefits to disabled veterans. During the Civil War, the benefit was revised from a flat payment scheme to a graduated schedule based on disability severity. Multiple governmental agencies administered Veterans' benefits until the summer of 1930, when they were consolidated under a new federal agency called the Veterans' Administration.34 We focus on a major legislative change that took effect in 2001, which generated a plausibly exogenous increase in the generosity of disability benefits for one group of Vietnam Era veterans but not another.35 Motivated by an Institute of Medicine study that linked exposure to Agent Orange and other herbicides used by the U.S. military during the Vietnam War to the onset of type 2 diabetes, the VA in July of 2001 expanded the medical eligibility criteria for Vietnam veterans to include diabetes as a covered condition. This coverage expansion applied to veterans who served "in-theatre" in Vietnam, Cambodia, or Laos during the 1964 to 1975 period. It did not, however, apply to the approximately 55 percent of Vietnam era veterans who did not serve in theatre during the war. The 2001 policy change coincided with a sharp acceleration in the number of veterans receiving DC benefits, documented in Figure 2.1. Some of this overall increase was attributable to a much higher rate of DC enrollment among veterans serving in the 1990s and 2000s than among their counterparts from earlier service eras.36 But much of it was driven by the rise in DC enrollment among Vietnam era veterans. As shown in Figure 2.2, the fraction of Vietnam era veterans receiving DC benefits had been trending up gradually prior to the 2001 policy change so that 9 percent of Vietnam-era veterans received DC benefits in that year. But there was a significant break in that trend after 2001 so that by 2013 more than 17 percent of 3 This was changed to the Department of Veterans Affairs in 1989. See Gruber (2000) for an analysis of a reform to the federal government's disability program in all parts of Canada except for Quebec that is estimated to have reduced labor supply. 36 Veterans from the Gulf War and Global War on Terror are 2-3 times more likely than veterans from WWII or the 3 Korean War era to receive DC benefits. Shifts in the composition of veterans (as those from older eras die and the recent era join the ranks) have contributed to a substantial increase in total DC enrollment. 65 Vietnam era veterans were receiving DC benefits. No similar changes in rates of DC enrollment occurred for veterans from earlier service eras. 37 The policy-induced increase in DC enrollment provides an unusual opportunity to study how disability benefits affect the labor supply of near elderly male veterans. Adopting the terminology used by the military, we distinguish among 'boots on the ground' (BOG) Vietnam era veterans-the veterans directly affected by the Agent Orange policy-with 'not on ground' (NOG) veterans, whose DC benefit eligibility was not expanded. We analyze unique administrative data for a sample of more than 4 million U.S. Army veterans to compare the evolution of labor market outcomes among BOG veterans to other Vietnam era veterans who did not serve in the Vietnam theatre during the conflict there. By using other Vietnam era veterans as our comparison group, we account for the possibility that veterans would have retired sooner (or later) than nonveterans for reasons unrelated to the DC program. And given the large number of years of pre-2001 data included in our research data, we can control for possible differential trends between BOG and NOG veterans. A large body of research investigates the effects of U.S. federal disability programs on health, employment, poverty, consumption and welfare. 38 This research has focused almost exclusively on the Social Security Disability Insurance (SSDI) program, with early studies considering the effects on labor force participation (Parsons, 1980; Bound, 1989; Bound and Waidmann, 1992) and subsequent studies exploring the sensitivity of the program to economic conditions (Black et al, 2002) and the labor market effects of changes in the program's medical eligibility criteria and in effective replacement rates (Autor and Duggan, 2003).39 One challenge for this research is that, because SSDI is a federal program, there is no natural comparison group against which to estimate the effects of the program. To address this issue, more recent research has estimated the effect of SSDI on the labor supply of applicants and beneficiaries by using variation in the propensity to award disability benefits across disability examiners (Maestas et al, 2013; 37 38 See Duggan, Rosenheck, and Singleton (2010) for a comparison to veterans from earlier service eras. See for example the Handbook chapter by Bound and Burkhauser (1999). The labor supply consequences of the federal Supplemental Security Income (SSI) program are rarely studied because SSI largely serves individuals with extremely limited prior work histories (those with significant work histories normally qualify for SSDI). 39 66 Autor et al., 2015b) and administrative law judges (French and Song, 2014). The findings from these studies, which utilize large-scale administrative data sets on both SSDI enrollment and earnings, suggest that labor force participation among marginal applicants-that is, those who would receive an SSDI award from a lenient judge or examiner but not from a stricter one-declines by about thirty percentage points as a result of receiving an SSDI award.4 There has been much less research on the VA's Disability Compensation program. Autor and Duggan (2007), Autor et al (2011), and Coile et al (2015) use data from the Current Population Survey (CPS) to explore how labor force participation changed for male Vietnam-era veterans relative to similarly aged nonveteran males after the 2001 policy change. All three studies demonstrate a significantly larger decline in labor force participation among Vietnam-era veterans in the post-2001 period, though their confidence intervals are compatible with a wide range of effect sizes. A concern with this body of work is that veterans might retire sooner than non-veterans for reasons unrelated to the DC program. 41Using the Vietnam-era draft lottery as an instrumental variable for Vietnam-era military service, Angrist et al (2010) estimate that employment was lower and transfer income receipt was higher among low skilled veterans than among lowskilled non-veterans. They hypothesize that the lower employment rate of Vietnam-era veterans is due to the availability of Veterans Disability Compensation benefits. In related work on the labor supply of veterans, Boyle and Lahey (2010) study the expansion of the Veterans Health Insurance program to non-disabled veterans in the mid-1990s to analyze changes in labor force participation stemming from increased real incomes and reductions in "job lock." Using individual-level data from the U.S. Department of Veterans Affairs, we document that DC receipt and enrollment growth were higher among BOG than NOG veterans prior to 2001, but these gaps were Using an approach similar to Bound (1989), Chen and van der Klauuw (2008) estimate even smaller effects of SSDI enrollment, with an upper bound of a 20 percent reduction in labor force participation. Results from Von Wachter et al (2011) suggest labor supply effects may be larger for younger SSDI recipients. While SSDI reduces the incentive to work among recipients, the results of a recent policy change in Norway suggest that disability insurance recipients are responsive to changes in the magnitude of this incentive (Kostol and Mogstad, 2014). 40 4' Duggan et al (2010) use five years of data (odd-numbered years between 1997 and 2005) from the Veterans' Supplement to the CPS to compare changes in labor force participation among Vietnam veterans who served in theatre and other Vietnam era veterans. Given the small sample size, their estimates are very imprecise, as their confidence interval includes the full range of possible effect sizes (zero effect or a one-for-one reduction in labor force participation). 67 relatively stable. After 2001, however, the rate of DC enrollment grew much more rapidly among BOG veterans, as shown in Figure 2.3. Between 2000 and 2006, the ratio of DC receipt among BOG relative to NOG veterans rose from approximately two-to-one to three-to-one: almost one-in-four BOG veterans in our analysis sample received DC benefits in 2006 versus one-in-twelve among veterans in our NOG sample. This trend break was driven primarily by a sharp increase in the number of diabetes awards to BOG veterans as shown in Figure 2.4. DC enrollment growth among BOG veterans shows essentially no break in trend when one excludes DC recipients with a diabetes diagnosis (Figure 2.5). Using matched data from the Social Security Administration, we also document a differential increase in SSDI enrollment after 2001 among BOG veterans. A plausible channel for this effect, which is about one-tenth as large as the corresponding change in DC enrollment, is that receipt of veterans disability benefits eases financial constraints associated with exiting the labor force and applying for SSDI benefits. Moreover, since SSA regulations require disability adjudicators to consider disability decisions of other federal agencies (U.S. General Accounting Office, 2009), a veteran's enrollment in DC may also increase the likelihood that he or she applies for and ultimately receives Social Security disability benefits. By generating a sharp differential increase in DC enrollment among BOG veterans for reasons unrelated to changing health, the Agent Orange policy change permits causal estimation of the effect of DC program participation on the labor supply of near-elderly Vietnam veterans. Analyzing this policy-induced variation, we find that labor force participation, defined as having strictly positive earnings for the year in our administrative data, declined sharply among BOG relative to NOG veterans soon after the 2001 policy change. The results are similar for younger (born 1949-51) and older (born 1946-48) Vietnam veterans in our analysis sample, suggesting that this pattern reflects an effect of the policy rather than a tendency of BOG veterans to retire sooner than their NOG counterparts for reasons unrelated to the DC program. For every 100 individuals who entered the DC program as a result of the policy change, we estimate that 18 drop out of the labor force. The magnitude of this decline reflects in part the size of the cash transfers that DC beneficiaries receive. Among veterans who entered the DC program after 2001, annual benefits averaged $1 OK in the first year of enrollment and $17K in the fifth year of enrollment. Since DC benefits are not subject to state or 68 federal taxation, their after tax value is 30 to 40 percent greater than nominally equivalent labor income. Indeed, the increase in disability benefit income among BOG veterans more than offset (on average) their reduction in earnings among BOG veterans, so that total incomes of BOG relative to NOG veterans rose steadily after 2001. Combining our labor force participation and benefit receipt estimates, we obtain a non-participation elasticity of 0.50. This estimate is larger than the figure of 0.16 reported by Coile and Gruber (2007) for the elasticity of retirement of near-elderly adults with respect to Social Security and other retirement wealth. 2 Our estimates are highly consistent with the participation elasticities calculated by Boyle and Lahey (2010), however, who studied labor supply of older non-disabled veterans ages 55 through 64 who were granted access to VA health insurance in the mid- 1990s. Distinct from the SSDI program, which provides income replacement for beneficiaries who are unable to work due to disability, DC benefits are awarded as compensation for service-related reductions in health and thus, for the most part, are not contingent on veterans' past or present employment. This observation would suggest that any DC-induced reduction in labor supply that we detect would be attributable to a nonincentive income effect. However, nearly 14 percent of BOG veterans who were receiving DC by the end of our sample window were receiving maximum DC benefits of approximately $2,900 monthly because they were deemed unable to work ('Individually Unemployable' or IU) due to their disability. For veterans receiving the IU benefit and those seeking it, the labor supply effects that we estimate are likely to encompass both income and incentive effects. The labor supply effects that we document acquire added significance in light of the rapid growth in Disability Compensation enrollment, with annual benefit payments of $49.2 billion in 2013 (U.S. Veterans Benefits Administration, 2014). Today's veterans are substantially more likely than U.S. veterans of earlier cohorts to obtain DC benefits; indeed, veterans who have exited military service since 2001 are more than three times as likely as were Vietnam era veterans to receive disability compensation benefits soon after the A venerable literature estimates labor supply elasticites, including Frisch (1959), Ashenfelter and Heckman (1974), Abbott and Ashenfelter (1976), and Chetty (2012) among many key contributions. Very few studies estimate income elasticites of participation, however. McClelland and Mok (2012) provide a recent review. 42 69 completion of service. This cross-cohort contrast suggests that Veterans Disability Compensation program costs may rise substantially beyond what would be predicted based on earlier generations of veterans, a concern highlighted by Bilmes and Stiglitz (2008). The analysis proceeds as follows. Section 2.2 discusses the financial and labor force participation incentives created by the Veterans Disability Compensation program. Section 2.3 describes the 2001 Agent Orange decision that made type 2 diabetes a service-connected disability for veterans who served in the Vietnam theatre. Section 2.4 details the construction of our data, which we use in Section 2.5 to analyze the impact of the Agent Orange decision on veterans' enrollment in DC and their receipt of transfer income from DC and two other federal disability programs, SSDI and Supplemental Security Income (SSI). Section 2.6 presents reduced form estimates of the impact of the Agent Orange policy on labor force participation and total labor earnings. Section 2.7 combines these margins to provide instrumental variables estimates of the impact of DC enrollment and total disability benefits payments on labor force participation. Section 2.8 documents the effect of the Agent Orange policy on total measured income inclusive of disability benefits and net of any induced change in labor supply. Section 2.9 concludes. 2.2 The Veterans Disability Compensation Program: Eligibility, Benefits and Work Incentives The DC program pays cash benefits and provides prioritized access to VA health facilities to military veterans with service-connected medical conditions, meaning that they are caused or aggravated by their military service. Unlike SSDI and SSI, federal programs that classify disability using a categorical (all-ornothing) determination, the DC program rates disability on a discrete scale with eleven gradations ranging from zero to 100 percent in ten percent increments. Ratings depend on the type and severity of the disability, with more severe conditions receiving a higher rating. If the recipient receives ratings for multiple disabilities, the recipient's Combined Disability Rating (CDR) is an increasing, concave function of the 4 The range of possible ratings differs among disabilities. For example, type 2 diabetes can have a rating of 10, 20, 40, 60, or 100 percent. Arthritis can be assigned a rating of 10 or 20 percent. For a list of conditions and ratings see http://www.warms.vba.va.gov/bookc.html. A disability with a 0 percent rating would not increase the monthly cash benefit but would entitle the veteran to priority for health care through the Veterans Health Administration. 70 individual ratings, where concavity prevents the combined rating from exceeding 100 percent.44 A. Eligibilityand benefits Veterans seeking DC benefits apply to one of 56 regional offices of the Veterans Benefit Administration (VBA), which collects necessary information and forwards the information to a Rating Board. For each disability claimed, the Rating Board determines whether the disability is verified, whether it is service connected, and if so, what rating it merits. During the 2000 fiscal year, more than 70 percent of those applying for DC sought benefits for more than one medical condition (U.S. Veterans Benefits Administration, 2001). Applicants face one of three possible outcomes: outright rejection, an award for some but not all conditions, or an award for all conditions. In 2006, current DC beneficiaries averaged 2.97 disabilities per recipient, with the highest number of disabilities per capita among Gulf War and Vietnam Era veterans, and the lowest number among WWII veterans (U.S. Veterans Benefits Administration, 2007). Monthly benefits awarded by DC are an increasing and convex function of the veteran's CDR. In 2014, a 10 percent award provided a monthly payment of $131 whereas a 100 percent award provided a monthly payment of $2,858.4 Veterans receiving a CDR of 30 percent or higher and who have spouses, dependent children, or surviving parents also receive modest additional benefits. The VBA also considers employment capability for veterans with severe disabilities. Veterans who have single disabilities rated at 60 percent or above or a combined disability rating of at least 70 and one disability rated at least 40 can qualify for the Individual Unemployability (IU) designation if VBA determines that they are unable to "to secure and follow a substantially gainful occupation by reason of service-connected disability." Veterans receiving the IU designation are provided cash payments at the 100 percent CDR level even if their CDR is less than 100 If a claimant has multiple disabilities, only the claimant's 'residual ability' is considered when determining the effect of each additional disability on the CDR. For example, if a veteran has two disabilities rated at 50%, his CDR would be he equal to the sum of 50% for the first disability and 50% of his residual capacity of 50% for the second disability, all rounded to the nearest increment of 10%. Thus, two disabilities rated at 50% results in a CDR of [0.5 + (1 - 0.5) * 0.5] = 0.75, which is then rounded up to 0.80. 4 The stated policy of the VBA that the DC benefits schedule reflects the average reduction in earnings capacity for each value of the CDR. Since benefits determination depends only on CDR and family status, it is clear that the benefit payment will exceed the earnings loss for some veterans and fail to meet the earnings loss of others. In 2014, the monthly benefit schedule (by CDR) was: $131 (10%), $259 (20%), $401 (30%), $578 (40%), $822 (50%), $1,041 (60%), $1,312 (70%), $1,526 (80%), $1,714 (90%), and $2,858 (100%). 4 71 percent. Veterans Disability Compensation benefits typically have longer award durations and fewer work restrictions than other federal disability benefits. DC benefits are also not subject to federal income or payroll tax; hence, a dollar in DC income is roughly equivalent to $1.30 to $1.50 in pre-tax earned income, depending upon the recipient's marginal tax rate. DC benefits generally do not offset and are not offset by other federal transfer benefits, and, once awarded, are rarely retracted. 46 Unlike federal SSDI benefits, DC benefits do not terminate when a recipient reaches retirement age, even for recipients receiving the IU benefit. Moreover, a veteran's ongoing receipt of DC benefits is neither work-contingent nor incomecontingent, except for veterans who have received the IU rating. Thus, DC benefits are roughly akin to an inflation-indexed annuity that provides monthly payments for as long as a veteran remains alive. Appendix Table 2.1 summarizes DC cash benefits paid in fiscal year 2006, the final year for which we have individual-level DC data in the analyses below. The first three columns enumerate the count of recipients, the total dollars paid, and the average annual benefit in each CDR category at the end of fiscal year 2006. The average annual payment to the 2.73 million DC recipients in this year was $10,862 per capita, totaling approximately $29.6 billion for the year. Veterans with ratings between 0 and 20 percent accounted for 44 percent of recipients but just 8 percent of dollars paid. Those with ratings at 70 percent or above comprised 21 percent of the population and received 62 percent of the benefits payments. 48 Total DC benefits payments rose from $20.8 billion to $49.1 billion between 2001 and 2013 (in constant 2013 dollars). Simultaneously, the estimated veteran population declined from 26.1 million to 22.1 million (VBA, 2002 and 2014). As a result of these changes, real annual DC expenditures per living veteran A Veteran may receive both DC and SSDI payments without any reduction in benefits from either program, though SSI payments will generally be reduced or eliminated by DC payments. 46 A veteran can lose the IU rating if his annual labor market earnings (measured by SSA earnings data) exceeds a threshold amount, which was equal to $6,000 in 2004 and 2005. The General Accounting Office notes, however, that "this process relies on old data, outdated and time-consuming procedures, insufficient guidance, and weak eligibility criteria" (GAO 2006, p. 23). 48 The average monthly benefit amounts for those with ratings between 0 and 20 percent are very close to the baseline amounts because veterans with these ratings are not eligible for dependent benefits. The average amounts paid for those rated 60 percent and higher are substantially greater than the baseline amounts because many of these recipients are eligible for the 100 percent payment amount because they are receiving the IU benefit. 47 72 increased by 180 percent (from $798 to $2,234). B. Work Incentives under DC The graduated scale of DC disability ratings creates a complex set of incentives. Though disability ratings for DC recipients notionally depend exclusively on medical criteria rather than employment status, veterans may nevertheless perceive that their disabilities will receive higher ratings if they are not employed when applying to obtain or increase benefits. Veterans also face an incentive to repeatedly reapply to increase their Combined Disability Ratings-and therefore their benefits-as their health conditions evolve. 49 One consequence is that veterans' CDRs and benefit levels tend to rise steeply in the years following enrollment, as shown in panels A and B of Table 2.1.50 This pattern of rapidly escalating benefits following enrollment suggests that policies that induce veterans to obtain an initial DC award, even at a low CDR, may lead to substantially larger claims over the longer term and discourage labor force participation. The availability of the Individual IU designation is likely to amplify these incentives. The IU benefit has significant monetary value: a 2006 General Accounting Office report found that the average present discounted incremental value of receiving an IU award in 2005 was approximately $300 to $460 thousand for veterans age 20 (net of existing benefits), and was $89 to $142 thousand for veterans age 75 (U.S. GAO, 2006).5 The availability of this benefit appears likely to induce at least some subset of work-capable veterans to curtail labor force participation to qualify. Once the IU designation is awarded, veterans face an incentive to maintain low earnings since the benefit is technically only available for those with labor market earnings at or below the poverty level for a single individual (U.S. GAO, 2006).52 In fact, we observe very few reductions in CDRs in our data, and it is possible that those few that exist reflect coding errors. Veterans face little risk of having their CDRs reduced after the initial award. 4 We describe the sample used to construct Table 2.1 in Section III below. Although our data codes DC receipt in each year from 1998 through 2006, we can only determine what year DC was awarded if a veteran is observed not receiving DC in a prior year. We can thus identify DC enrollment cohorts from 1999 forward, but not for 1998. 50 51 Among veterans in our sample who received benefits at the 100 percent disability level in 2006, about half were designated as LU. We henceforth do not distinguish between the IU benefit and 100 percent disability since many DC recipients with 100 percent disability may have previously qualified for the IU benefit with a lower CDR. 52 The fact that only veterans with severe disabilities (a CDR of 60 or higher) are eligible for the IU benefit might be expected to deter all but the most disabled veterans. But the data in Table 2.1 indicate that very high CDRs are not 73 The DC program may also alter work incentives through its interactions with other federal benefits programs, SSDI in particular. Though the DC and SSDI programs have distinct disability screening criteria, the medical information generated by the DC award may alert some veterans that they suffer from impairments that could merit an SSDI award (and vice versa). Receipt of DC benefits may also render the SSDI application process less financially onerous, since SSDI applicants must remain out of the labor force for at least five months before receiving SSDI benefits. Because cash benefits from the two programs are additive rather than offsetting, it is plausible that a veteran's receipt of either DC or SSDI benefits increases his odds of applying for the other." 2.3 The 2001 Agent Orange Decision, Type 2 Diabetes and 'Service-Connectedness' For a disability to be classified as service-connected, it must be "a result of disease or injury incurred or aggravated during active military service." This criterion makes it straightforward for a veteran to obtain compensation for a tangible injury that occurs during military service but significantly more difficult to obtain compensation for a disease that develops later in life, such as cancer or heart disease. Thus in 2006, the five most prevalent service-connected disabilities were primarily battle traumas: hearing defects, tinnitus, general musculoskeletal disorders, arthritis due to trauma, and scars (U.S. Veterans Benefits Administration, 2006). Nevertheless, disabilities that typically develop post-service are also prevalent: post-traumatic stress disorder (PTSD) and hypertensive vascular disease (high blood pressure) were the sixth and ninth most prevalent service-connected disabilities in 2006. In November of 2000, type 2 diabetes was added to the list of compensable and presumptively serviceconnected impairments for Vietnam veterans who had served in theatre due to their potential exposure to the herbicide Agent Orange. This policy change substantially weakened the link between service-connectedness uncommon, even for veterans that initially enter with low or moderate CDRs. Among veterans awarded DC benefits in 1999, only 15 percent qualified for either the IUbenefit or 100 percent disability (panel C). Seven years later, in 2006, three times that number (45 percent) of the 1999 DC enrollment cohort was either receiving the IU benefit or was 100 percent disabled. 1 The combination of VA health benefits and Medicare benefits from SSDI may also be more attractive than either individually since VA and Medicare differ in ailments covered, rapidity of access to treatment, size of co-pays, and coverage of prescription drugs. 74 and DC benefits for eligible veterans. The Agent Orange policy was years in the making. The U.S. military applied more than 19 million gallons of herbicides to defoliate Vietnamese jungle areas between 1962 and 1971, a quantity of defoliant sufficient to cover 8.5 percent of the country's land area (U.S. Department of Veterans Affairs, 2003). 5 After the war ended, concern grew among veterans that their exposure to the dioxins in Agent Orange would have long-term adverse consequences. The VA responded by establishing the Agent Orange Registry in 1978. In 1991, Congress enacted the Agent Orange Act, which charged the National Academy of Sciences' Institute of Medicine (IOM) with reviewing the evidence for a link between Agent Orange exposure and the prevalence of certain medical conditions. In a series of reports, the IOM found insufficient evidence to establish an association between dioxin exposure and diabetes. But the publication of two new studies in 1999 and 2000 that found an association between dioxin exposure and diabetes turned the tide against this longstanding consensus (Calvert et. al., 1999; Air Force Health Study, 2000). In October of 2000, the IOM concluded that there was "limited/suggestive evidence of an association between exposure to the herbicides used in Vietnam or the contaminant dioxin and type 2 diabetes" (IOM, 2000). " This in turn prompted the Secretary of Veterans Affairs' decision to classify type 2 diabetes as presumptively service connected. The VA's adoption of the Agent Orange policy offers an unusual opportunity: almost three decades after the end of the Vietnam War, veterans who served in theatre were unexpectedly granted presumptive eligibility for financially significant Disability Compensation benefits without a precipitating change in health. This chapter exploits the contrast in expanded benefits eligibility between in-theatre and not-intheatre veterans-whom we refer to as 'boots on ground' (BOG) and 'not on ground' (NOG) veterans-to study the impact of benefits receipt on veterans' labor supply. 5 Agent Orange accounted for more than 80 percent of the total amount of herbicide dispensed. The same 2000 report by the National Academy of Sciences' Institute of Medicine explained that any increased risk of Type 2 diabetes due to Agent Orange appeared to be small and that family history, physical inactivity, and obesity were far greater predictors of diabetes (IOM, 2000). 5 75 2.4 Data and Analytic Sample Our research draws on a unique set of linked administrative data sources. The first is a near census of approximately four million veterans who left the Army between 1968 and 1985. The U.S. Army's Office of Economic and Manpower Analysis (OEMA) constructed this database by combining two files from the Defense Manpower Data Center (DMDC): the first DMDC file enumerates essentially every person who left the Army between 1968 and 1985 (designated as the service member's "loss year"); OEMA then merged this "loss year" file with DMDC's Vietnam file, which identifies the vast majority of veterans who served in the Vietnam theatre.56 Approximately 36 percent of this sample served with boots on the ground in Vietnam, Cambodia, or Laos during the Vietnam War era, and is therefore potentially directly affected by the 2001 expansion of the DC program's medical eligibility criteria. More than 35 percent of the NOG sample had a start year of 1976 or later and thus did not serve during the Vietnam War era.5 Appendix Table 2.2 and Table 2.2 show summary statistics on the distribution of loss years, years-of-birth, and start years for military service. We use three additional data sources to measure mortality, employment, and disability outcomes for Vietnam era veterans. To measure DC participation, OEMA obtained from VA detailed information about veterans' enrollment and DC benefits received from VA programs in September of each year from 1998 through 2006. To account for mortality, OEMA merged their data to the Social Security Administration Death Master File (DMF), which includes the year of death for any individual in the sample who died prior to 2008. According to DMF data, approximately 13 percent of the 4.1 million individuals in the sample were deceased by late 2007. To collect employment, earnings, SSDI, and SSI information, OEMA contracted with the U.S. Social The file does not include the comparatively small number of Army service members who died during service. U.S. government archives record 38,224 Army service members who were killed in action in Vietnam. 56 7Veterans who were in the "loss year" file and the Vietnam file are in the BOG sample while veterans who were in the "loss year" file but not the Vietnam file are part of the NOG sample. The Vietnam file contained loss year information for approximately 25 percent of veterans, which explains why a small fraction of the BOG sample has a loss year before 1968 or after 1985. Our final analysis sample, described below, excludes all veterans with loss years before 1968 or after 1985, but does contain a few veterans who were in the Vietnam file but not in the "loss year" file. This group comprises only 1.25 percent of our final sample and our results are nearly identical when we exclude them from our analysis. 76 Security Administration (SSA) to match veterans in the OEMA data set to enumerate wage earnings and Social Security benefits in each year from 1976 through 2007.58 This resulted in a successful match for 3.8 of the 4.1 million veterans in the full data set, with overall match rates exceeding 90 percent for both BOG and NOG samples, as detailed in Appendix Table 2.3. Confidentiality rules prevent SSA from disclosing individual earnings or benefits data. SSA instead provided statistics on earnings and benefits for cells containing five to nine veterans. These statistics include the number of cell members with zero earnings, mean labor earnings, the number receiving SSDI and SSI, and mean SSDI and SSI benefit amounts. In constructing cells, we grouped individuals with similar background characteristics, including gender, race, BOG and NOG status, and year of birth. Our final analysis sample consists of veterans who joined the military between 1966 and 1971, and were born between 1946 and 1951 (see Table 2.2). The appendix to this chapter contains more details on the cell construction and sample selection. Panel A of Table 2.3 compares the BOG and NOG samples. The fraction nonwhite is approximately equal in the two samples (11.3 and 11.8 percent, respectively) as is the fraction with positive earnings in 1998 (84.4 and 85.1 percent). Among those with a non-missing Armed Forces Qualification Test score (AFQT), the average scores are also relatively close (52.1 and 53.4). And by construction, the average yearof-birth and the average start year are comparable in the two groups. There are clear differences between the BOG and NOG samples as well. BOG veterans are more than twice as likely as NOG veterans to be receiving DC benefits in 2000 (14.3 versus 6.5 percent), just prior to the 2001 policy change described above. This in part reflects the greater toll that military service took on those who served in the Vietnam theatre. An examination of trends in key outcome variables in the BOG and NOG samples prior to the 2001 policy change reveals many similarities: the fraction with zero earnings increased by similar, though not identical, amounts for both samples between 1998 to 2000 (by 1.5 and 1.0 points for BOG and NOG 58 SSA required a match on social security number, last name, and date of birth of each individual to make a match. Additionally, members of the NOG sample are more likely to have missing data on education and less likely to be missing AFQT score data. 5 77 samples, respectively), as did the fraction receiving SSDI benefits (1.2 and 0.9 points for BOG and NOG samples). Our analysis will control for any differential trends in outcome variables that precede the policy change. A remaining concern with the primary analytic sample is that individuals who are matched in the SSA data may systematically differ from those who are not. This is especially an issue for veterans who died prior to 1997. As shown in columns 3 and 4 of panel B in Table 2.3, the fraction of veterans in the full sample who were deceased as of 1997 was 6.2 and 6.1 percent in the full BOG and NOG samples but only 4.8 and 2.9 percent in the SSA verified samples-though, notably, the fraction of veterans who died between 1997 and 2007 in the BOG and NOG samples is closely comparable in both the full sample and the SSA verified sample. 60 The lower SSA match rate for deceased veterans is a consequence of SSA's record matching criteria, which require a match on subjects' full names as well as SSN and date of birth. Due to poor optical character recognition, the NOG data contained a relatively high frequency of garbled names. We worked with the credit information provider TransUnion to obtain names for those with incomplete information.61 Due to limited availability of archival credit information prior to 1997, TransUnion could not provide names for most veterans who passed away before that same year, leading to a low overall match rate for NOG soldiers who died prior to 1997. Differential mortality match for the BOG and NOG samples is not a major threat to the validity of our research design, however, since our primary focus is on outcomes from 1998 forward. Thus, soldiers who were deceased as of 1997 are excluded from the analysis. To benchmark the representativeness of the sample, we compare the OEMA data with a similarly drawn group of males from the 2000 IPUMS Census file. Using the 5 percent Census IPUMS extract, we draw a group of all males born between 1946 and 1951, and further limit the sample to (self-reported) Vietnam-era We constructed the sample summarized in panel B of Table 2.3 in the same manner as the sample summarized in panel A, except that we did not exclude veterans who did not match with SSA earnings information. 61 TransUnion performed this work pro bono. Our original sample had 1.7 million observations with a missing name. 60 TransUnion was able to provide names for 1.5 million of these observations upon confirming a match with date of birth and social security number. 78 veterans.62 Appendix Table 2.4 provides a side-by-side comparison of age, race, schooling, annual earnings, and share with non-zero earnings in the OEMA and Census samples. Race and labor force participation rates are closely comparable: the fractions of the OEMA and Census samples with non-zero earnings are 84.1 percent and 82.2 percent, respectively, while the percent nonwhite are 11.3 and 13.3. Reflecting the fact that OEMA data code education at the time of military enlistment (average age of 20) while the Census data code educational attainment in late adulthood, the OEMA sample reports considerably lower educational attainment than the Census sample. Average earnings in the OEMA sample are also about 10 percent lower than in the Census sample. This gap may reflect earnings differences between Army veterans and those of other branches of the military. SSA data may also fail to capture some earnings sources, including selfemployment and non-covered work. Overall, our comparison of OEMA and Census data provides some assurance that the OEMA sample is representative of the target population of Vietnam era Army veterans, measured in terms of age, race, labor force participation and earnings. 2.5 The Impact of the Agent Orange Policy on Receipt of Disability Benefits The Agent Orange policy spurred a steep rise in Disability Compensation enrollment, and may potentially have had spillover effects on enrollment in other federal disability programs as well. We begin by estimating impacts on DC enrollment, followed by SSDI and SSI enrollment, and finally, total federal disability benefits. A. Enrollment in Veterans Disability Compensation Figure 2.3 plots the fraction of BOG and NOG veterans receiving DC benefits in September of each year from 1998 through 2006. Prior to the Agent Orange change in 2001, DC enrollment was rising somewhat more rapidly among BOG than NOG veterans. But DC enrollment among BOG veterans accelerated substantially after 2001. Column 2 of panel A in Table 2.4 shows that BOG DC enrollment increased by 0.4 The Census data do not allow us to distinguish among veterans according to their branch of military service. To the extent that Army veterans are different from their counterparts serving in the Navy, Air Force, Marines, or Coast Guard, we would expect some differences between the Census and OEMA samples. 62 79 percentage points per year between 1998 and 2000 (13.5 to 14.3) and by 1.6 percentage points annually between 2001 and 2006 (15.0 to 23.0). In contrast, DC enrollment growth rates among NOG veterans remained small and relatively steady, increasing from 0.1 percentage points per year between 1998 and 2000 (6.3 to 6.5) to 0.2 percentage points per year between 2001 and 2006 (6.6 to 7.6).63 Data from the National Health Interview Survey (Schiller et al, 2010) indicate that the fraction of individuals with diabetes varies substantially by race, with rates among blacks substantially higher than among whites. Consistent with this fact, an examination of the trends in DC enrollment in Table 2.4 reveals substantial differences by race in the BOG sample, with DC enrollment increasing about 40 percent more among BOG nonwhites (19.3 to 30.2, 10.9 percentage points) than among BOG whites from 2001 to 2006 (14.4 to 22.1, 7.7 percentage points). These raw differences in DC enrollment trends between BOG and NOG veterans may reflect differences in veteran characteristics in addition to any impact of the Agent Orange policy. To account for these factors, we estimate a set of OLS models that regress DC enrollment on a full set of controls for veterans' year-ofbirth, race, and AFQT score quintile.64 For consistency with the subsequent analysis of labor market outcomes, our DC variables are calculated as means over five to nine veterans grouped at the level of SSA reporting cells. 65 We estimate the following equation for 1998 through 2006, weighting each cell-year by the number of individuals in the cell: (2.1) 2006 yt x BOGj + XJflt + Ejt yit = at + yo x BOGj + 1999 The outcome variable Yjt is the percentage of cell j enrolled in the DC program in September of year t, and The policy change took effect in July of 2001 and we measure DC enrollment in September. Thus, our 2001 data is arguably more "pre" than "post," though it's likely that the policy change did contribute to DC enrollment growth between September 2000 and September 2001 in our data. 64 Veterans with low AFQT scores are more likely to enroll in DC (Autor et al 2011), and average scores differ slightly between BOG and NOG veterans. 65 SSA outcomes are only available at the cell level due to confidentiality restrictions, as discussed above. Because we apply OLS models to cell means and weight by the number of individual observations in each cell, the cell level estimates will be in most cases algebraically identical to those that we would obtain if cells were instead disaggregated to individual level rows. The one exception to this dictum arises from the fact that ten percent of cells in have more than one AFQT quintile represented within the cell. We assign the cell to the AFQT quintile nearest to the cell's mean AFQT quintile in these cases, meaning that the cell-level and corresponding individual level regressions will differ slightly. 63 80 BOG is an indicator variable that is set equal to one if veterans in cell j are in the BOG sample and is otherwise equal to zero (cells include either all BOG or all NOG veterans). The term at is a vector of nine indicator variables for each year considered, and Xjt is a vector set of 14 variables corresponding to the possible values of year-of-birth, AFQT quintile, and race.66 We interact each of these 14 indicator variables with nine year-specific indicator variables to account for differential levels of growth rates in DC enrollment by age, race, or AFQT level during the 1998 through 2006 period. The coefficient yo corresponds to the (conditional) baseline DC enrollment gap between BOG and NOG veterans in the base year of 1998, while the coefficient vector Yt estimates the difference in this gap in each subsequent year 1999 through 2006 relative to the enrollment gap in 1998. The statistically significant estimate of 6.97 for yo in the first column of panel A, Table 2.5 implies a 7 percentage point gap in DC enrollment between the BOG and NOG samples in the baseline year (1998) after controlling for race, year-of-birth, and AFQT quintile, quite similar to the unconditional estimated difference of 7.2 percentage points in columns 1 and 2 of Table 2.4 (13.5 for BOG versus 6.3 for NOG). The next eight rows of the first column in Table 2.5 display the estimates for Yt in each year from 1999 through 2006. The statistically significant estimates of 0.33 and 0.59 for y1999 and y2000 imply that DC enrollment was increasing more rapidly (by 0.3 percentage points annually) for the BOG than NOG sample prior to the 2001 policy change. Beginning in 2001, these coefficients increase much more rapidly, by about 1.4 percentage points per year, and reach a cumulative differential of 7.98 percentage points by September of 2006. As shown in panels B and C of Table 2.5, the point estimates differ only modestly by birth cohort (1946-48 and 1949-51), with slightly larger effects for the older than the younger group (8.28 and 7.65 percentage points, respectively). The sharp break in trend for DC enrollment among BOG relative to NOG veterans motivates a In our sample there are six possible values of YOB (1946 through 1951), two possible values of race (white and nonwhite), and six possible values of AFQT quintiles (we group those with a missing AFQT into a sixth category). Given there are nine years of data used in these estimates, we are including 126 interactions. Veterans within each cell have the same year-of-birth and race (by construction) and the vast majority also have the same AFQT quintile. Data for each veteran is included in each of these nine years unless the veteran dies at some point between 1998 and 2006, in which case the veteran is dropped from the sample. Year-of-death is one of the variables used to construct the cells, so typically the entire cell is dropped. 66 81 parameterized version of equation (2.1) found in even-numbered columns of Table 2.5. For this specification, we replace the full set of year-by-BOG interactions with two linear time trends: a pre-2001 trend and a post2001 trend change, estimated relative to the pre-2001 trend: (2.2) Yt = at + yo X BOG + So X BOGj X (t - 1998) + 61 X BOG X (t - 2001) X 1[t > 2002] + XjJlt + Ejt Here, 65 captures the pre-existing trend in BOG relative to NOG DC participation just prior to the policy change while 61 estimates any additional change in the BOG relative to the NOG trend following the policy. We define 2002 to be the first post-policy year in this specification given that most of the time from September 2000 to September 2001 occurred before the policy took effect in July 2001. To interpret 61 as the causal effect of the Agent Orange policy change on DC enrollment, we must assume that the pre-existing trend in the difference in DC enrollment between the BOG and NOG samples would have continued after 2001. The estimates in the even-numbered columns of Table 2.5 highlight the significant trend increase in DC enrollment among BOG relative to NOG veterans following the 2001 policy change. Because we treat 2001 as a pure pre-policy year, the estimate of S in each case is slightly larger and the estimate of 61 slightly smaller than is suggested by the year-by-year estimates. The magnitude of the trend break is similar between the two birth cohort groups (1946-48 and 1949-51). In panel A of Table 2.6, we split the sample by race (white/nonwhite) and by AFQT score. The acceleration in DC enrollment is largest among BOG veterans who are nonwhite and have relatively low (below the 4 0 th percentile in the sample) AFQT scores.67 Figures 2.4 and 2.5 relate these enrollment trends to the Agent Orange policy. Figure 2.4 documents a steep increase in the likelihood that BOG (but not NOG) veterans receive compensation for diabetes, both in the year of current DC receipt and in the year of initial DC enrollment. As documented in Panel C of Table 2.4, the fraction of DC beneficiaries receiving compensation for diabetes among BOG veterans rose from 2.4 to 28.1 percent (25.7 points) between 2001 and 2006 versus 1.6 to 6.4 percent among NOG veterans (4.8 67 We drop cells with missing AFQT scores for these specifications. 82 points). 68 Yet, as shown in Figure 2.5, there was only a modest increase in the fraction of veterans receiving DC benefits absent a diabetes award. Thus, the rapid growth in DC enrollment appears substantially accounted for by an influx of new diabetes awards. Figure 2.6 underscores that the increase in DC enrollment was similarly rapid among older (YOB 1946-48) and younger (YOB 1949-51) Vietnam veterans in our sample, as indicated by the regression estimates in Table 2.5. The estimates so far capture the extensive margin impact of the Agent Orange policy on DC participation. Panels B and C of Table 2.6 consider outcome measures that combine responses along the extensive (enrollment) and intensive (rated severity) margins: Combined Disability Ratings (CDRs) and annual benefits payments, both of which are continuous measures of DC participation. 69 Both CDRs and annual benefit payments rose significantly more rapidly for BOG than NOG veterans after 2001. Panel B of Table 2.6 shows that prior to the Agent Orange policy, average CDRs of BOG relative to NOG veterans were rising at 0.34 percentage points annually; after 2001, this differential trend tripled to 1.00 percentage points annually (0.34 + 0.66). Similarly, panel C indicates that the year-over-year increase in annual mean DC benefits paid to BOG relative to NOG veterans rose from $113 prior to the policy change to $305 per year after 2001 ($113 + $192). Consistent with the estimates for DC enrollment by race and AFQT score, the remaining columns of panels B and C demonstrate that CDRs and benefits payments increased more rapidly among nonwhite than white BOG veterans, and more rapidly among low-AFQT than high-AFQT BOG veterans. B. Enrollment in other FederalDisability Programs A significant constraint that some workers face when applying for SSDI benefits is that they are unlikely There are three likely reasons that DC receipt for diabetes increased in the NOG. One is that our data only imperfectly classify BOG and NOG veterans; some veterans categorized as NOG may in fact have served in theatre and hence presumptively qualified for service-connectedness for their type 2 diabetes. A second cause is that administration of DC benefits appears to be somewhat discretionary. Even prior to the Agent Orange decision in 2000, 1.4 percent of BOG and 0.7 percent of NOG DC recipients were receiving compensation for service-connected diabetes (panel C of Table 2.4). Finally, some NOG veterans may have served in Korea in 1968 or 1969. This source of slippage does not invalidate our identification strategy provided that the post-2001 trend break in DC receipt in BOG versus NOG veterans is induced by the Agent Orange Policy change, which seems quite plausible. 68 Veterans receiving the level. 69 IU rating are coded with a CDR of 100 since they receive benefits at the 100 percent disability 83 to qualify for an award if they participate gainfully in the labor force while their case is being decided (Autor et al., 2015b). Knowledge of this fact may deter workers with residual work capacity from applying for benefits since they would have to forfeit potential earnings in pursuit of an uncertain benefit. Since, as we document below, many veterans did leave the labor force after becoming eligible for DC benefits, it is plausible that a subset that would otherwise have been deterred from seeking SSDI benefits would choose to pursue these benefits as an indirect consequence of the Agent Orange policy.70 To examine these potential spillovers, we estimate models analogous to equation (2.2) where the outcome variable is the probability of SSDI receipt among BOG and NOG veterans (again measured at the level of SSA reporting cells). Reflecting the longer sample window available for the SSA outcomes, these models include two additional "pre" years, 1996 and 1997, and one additional "post" year, 2007. Panel A of Table 2.7 shows that, like DC enrollment, SSDI receipt was trending upward faster among BOG than NOG veterans by 0.13 percentage points per year prior to the 2001 policy change. But this trend accelerated after 2001, increasing to 0.25 percentage points (0.13 + 0.12) annually. Notably, the magnitude of the estimated effect on SSDI is only about 10 percent as large as the corresponding estimate for DC enrollment (compare column 1, panel A of Tables 2.5 and 2.6).7' These results suggest that the Agent Orange policy likely spurred additional SSDI enrollment, which may in turn have magnified any reduction in labor force participation among eligible veterans. In panel B of Table 2.7, we estimate models for receipt of SSI. Here, the estimates point in the opposite direction: SSI enrollment was growing more rapidly among NOG than among BOG veterans prior to 2001, and this trend accelerated after 2001, in all likelihood because the additional DC transfer income made some veterans ineligible for means-tested SSI benefits. The magnitude of the estimated break in trend for SSI is only one-tenth as large as the corresponding estimate for SSDI, however. The final panel of Table 2.7 estimates equation (2.2) for disability income combining benefits payments Borghans et al (2014) find related evidence of program spillovers in disability receipt. On average, Dutch DI recipients offset a El.00 loss in DI benefits with a 0.30 increase in other social support programs. Duggan et al (2007) find evidence that reductions in Social Security retirement benefits stemming from the rising U.S. Full Retirement Age have lead to an increase in SSDI enrollment among cohorts facing a higher full retirement age. 71 In estimates not displayed, we find impacts on SSDI receipt that are similar for the 1946-48 and 1949-51 cohorts. 70 84 from all three federal disability programs: DC, SSDI, and SSI. These estimates only include data from 1998 through 2006 since we do not have DC data for 1996, 1997, and 2007. Inclusion of SSDI and SSI benefits with DC benefits raises the differential trend increase in benefits payments for BOG relative to NOG veterans by approximately an additional 15 to 20 percent (compare to panel C of Table 2.6). Across all veterans in our sample, combined disability benefits were rising by $128 annually among BOG relative to NOG veterans prior to the 2002 Agent Orange policy change. Commencing in 2002, this differential trend rose to $354 ($128 + $226). This average masks considerable heterogeneity. For veterans with high CDRs, DC benefits payments substantially exceed SSDI benefits for all but the highest (pre-disability) earners; moreover, DC benefits are not subject to federal taxation. Thus, the bulk of the impact of the Agent Orange policy on veterans' transfer income accrues through the DC program. 2.6 Consequences for Labor Force Participation We now consider impacts of the Agent Orange policy on veterans' employment and earnings. Using models analogous (2.1) and (2.2), we test whether the fraction of BOG veterans in the labor force declined differentially relative to NOG veterans following the 2001 policy change. We again include data from 1996, 1997, and 2007 because of the longer sample window available for SSA outcomes. A. Impacts on laborforce participation Table 2.8 considers the impact of the Agent Orange policy on labor force participation, measured as the percentage of living veterans in an SSA reporting cell who have positive labor earnings. The first column finds that labor force participation among BOG veterans in the baseline year of 1996 was slightly lower (by 0.32 points) among BOG than NOG veterans. This participation gap was modestly expanding (i.e., becoming more negative) in the early part of our sample period, falling by an additional 0.81 points between 1996 and 2001. This expansion accelerated after 2000, with the BOG relative to NOG labor force participation rate falling by an additional 2.02 percentage points between 2001 and 2007. Column 2, which fits the parameterized model in equation (2.2), shows the rate of divergence between BOG and NOG veterans 85 roughly doubled after the Agent Orange decision, from 0.15 percentage points per year to 0.33 percentage points (0.15 + 0.18) annually after 2001. Coupled with the evidence in Table 2.5 on DC enrollments, the results in Table 2.8 suggest that the differential increase in DC enrollment among BOG veterans spurred a reduction in their labor supply. A natural alternative interpretation, however, is that BOG veterans were in worse health than NOG veteransperhaps due to the rigors of in-theatre military service-leading to comparatively earlier retirements in the 2000s for reasons unrelated to the 2001 policy change. If so, the accelerating decline in LFP among BOG veterans after 2001 should primarily impact the oldest Vietnam veterans, with younger veteran cohorts exhibiting a similar LFP falloff as they approached these early retirement years in subsequent years. Figure 2.7 and the remaining two panels of Table 2.8 explore these competing hypotheses by again subdividing the sample into two birth cohorts of average ages 54 and 51 years at the time of the policy change. Both the figure and the table reveal that labor force participation of BOG relative to NOG veterans for both sets of birth cohorts dropped sharply soon after the policy change. The year-over-year decline in BOG LFP relative to NOG LFP doubled after 2001 for both older (born 1946-48) and younger (born 1949-1951) cohorts, rising from 0.17 to 0.36 percentage points (0.17 + 0.19) among the former group, and from 0.14 to 0.32 percentage points (0.14 + 0.18) among the latter group. This trend break in BOG relative to NOG labor force participation occurring simultaneously across birth cohorts after 2001 suggests that benefits expansion rather than early retirements explains the sharp drop in BOG labor force participation following the Agent Orange decision. The upper panel of Table 2.9 explores these employment patterns by race and AFQT subgroup. Declines in LFP generally mirror uptake of DC across demographic groups. Consistent with the Table 2.6 findings for DC enrollment, the fall in labor force participation among nonwhite and among low-AFQT BOG relative to NOG veterans following the 2001 policy change is more pronounced than among white and higher-AFQT veterans. These differences across groups are not statistically significant, however. 86 B. Impacts on labormarket earnings Alongside its extensive margin impact, receipt of DC benefits may spur veterans to reduce their work hours or switch from full to part-time employment, thus reducing earnings conditional on ongoing employment. We explore the effect of the Agent Orange policy on total labor force earnings by estimating equation (2.2) for the log of mean cell earnings, excluding (of necessity) cells where all veterans have exited the labor force.72 Panel B of Table 2.9 shows that earnings averaged about 2.6 log points lower among BOG than NOG veterans in our baseline year, and this gap was expanding by about 0.2 log points per year through 2001. The earnings gap grew significantly more rapidly-by about 0.7 log points annually (0.23 + 0.45)following the 2001 policy change. Once again, the effects are larger for nonwhite and low-AFQT veterans. Since these total earnings impacts combine extensive margin (participation) and intensive margin (hours and wage) responses, we cannot directly infer whether they reflect changes in earnings among incumbents or merely a reduction in the number of veterans working. The evidence in panel A of Table 2.9 suggests, however, that minority and low-AFQT veterans were particularly likely to exit the labor force upon receiving DC benefits. If we hypothesize more generally that veterans with lower potential earnings are most likely to exit the labor force in response to increased transfer income, we can compare extensive margin and total earnings responses to assess evidence for intensive margin responses. Concretely, suppose that earnings among those dropping out of the labor force are below the average of those still working within the cellthat is, income effects are larger for low earnings workers. In this case, the decline in log cell earnings will be smaller than the decline in log cell labor force participation. Conversely, if the decline in log cell earnings exceeds the decline in log cell participation, this indicates that either earnings is also falling on the intensive margin or, contrary to our assumption, higher earning workers are exiting employment disproportionately. To facilitate earnings and employment comparisons, panel C of Table 2.9 reports estimates where the dependent variable is the log of the fraction of veterans in a cell working (rather than simply the percentage working, as in panel A). Comparing across panels B and C reveals that the decline in relative earnings of We use the log of the mean rather than the mean of the log because of the cell-level aggregation of our earnings data. We do not, however, adjust cell means for non-participation; if a cell contains five veterans each earning $25K annually and one leaves the labor force, mean cell earnings falls to $20K. Thus, this measure incorporates intensive and extensive margin adjustments. 72 87 BOG versus NOG veterans after 2001 is, in most cases, slightly larger than the relative decline in labor force participation. In column 1, for example, the relative decline in log earnings of BOG relative to NOG veterans expands by 0.45 log points per year after 2001 whereas the relative decline in log LFP increases by only 0.36 log points per year. This pattern also holds for black and white veterans considered separately, though it does not hold for lower and higher-AFQT veterans. We infer that the Agent Orange policy likely induced a mixture of labor force exit and earnings reductions among beneficiaries, though the evidence on the latter margin is far from clear cut. 2.7 Comparing Labor Supply and Enrollment Impacts: Instrumental Variables Estimates Motivated by the evidence that the Agent Orange policy both raised DC enrollment and lowered earnings among eligible veterans, we now use the Agent Orange policy as an instrumental variable for receipt of DC benefits. This instrumental variables approach provides a valid estimate of the causal effect of DC benefits on labor supply under the assumption that the Agent Orange policy only affects employment and earnings through its impact on benefits enrollment and transfer payments. This exclusion restriction is untestable of course, but the panoply of findings above confer credibility: we find a sharp rise in DC receipt, DC payments, and total disability income among BOG relative to NOG veterans following adoption of the Agent Orange policy; the rise in take-up was equally steep and pronounced among younger and older veterans, and was greatest among nonwhite and low-AFQT veterans; and in all cases, take-up responses are paralleled by differential employment and earnings drops among BOG relative to NOG veterans that occur among both older and younger veterans, and which are slightly more pronounced among nonwhite and low-AFQT vets. Table 2.10 presents instrumental variables of equation (2.2) for the impact of DC participation on labor force participation, where the endogenous variable in panel A is the percentage of a cell enrolled in DC. The 2SLS estimate in column 1 finds that each percentage point increase in DC enrollment reduces the fraction of veterans working by 0.18 points-that is, approximately one veteran exits the labor force for We limit the sample to the years 1998 through 2006 for which DC measures are available, rather than the longer sample window of 1996 through 2007 for which SSA earnings data are available. The instrumental variable for DC enrollment is BOG; x (t - 2001) x 1[t > 2002]. 73 88 every five veterans newly awarded DC benefits. Comparing across rows of this table, we find slightly larger employment effects for nonwhite and low-AFQT veterans, consistent with our maintained hypothesis that participation among these groups is more elastic. Notably, the (scaled) labor force participation impact is estimated to be slightly (though not significantly) larger for younger than older veterans, again underscoring that the results are unlikely to be simply driven by differential trends in early retirement among BOG relative to NOG veterans. By specifying DC participation as the endogenous variable, the panel A estimate does not account for policy-induced increases in DC benefits along the intensive margin-higher benefits among incumbents, and greater potential benefits among new enrollees. We relax this restriction in panel B by making our endogenous variable the sum of disability income (DC, SSDI and SSI benefits). This variable captures changes in transfer payments among both new enrollees and incumbents.74 The panel B estimates find that each thousand dollars in disability benefits payment reduces the probability of a veteran working by approximately 0.80 percentage points, with slightly larger impacts among nonwhite veterans (0.97 points) and slightly smaller impacts among older veterans (0.70 points). Scaling these estimates by the mean rate of labor force participation (84.7 points) and mean year 2000 earnings (in constant $2013 dollars) of $52.5K (Appendix Table 2.4) implies an extensive margin labor force participation elasticity of -0.50. This is similar in magnitude to the estimates reported by Boyle and Lahey (2010), who study the labor supply of a relatively comparable set of veterans who were nearing retirement in the mid- 1990s. We also cannot rule out that the participation elasticity we obtain incorporates both pure (non-incentive) income effects from increased transfer income and additional incentive effects stemming from the potential availability of IU and SSDI benefits. Fully distinguishing these incentive and non-incentive effects remains a ripe topic for future work. To what degree is this causal effect driven by changes in labor force participation among incumbent DC recipients-many of whom received a higher combined disability ratings and corresponding benefit increases as a consequence of the Agent Orange policy-rather than by changes in participation among new enrollees? Appendix Table 2.5 probes this question by comparing the evolution of CDRs among veterans according to Because cells receive no disability payments in a given year, we specify the disability payment measure in real dollars rather than taking the logarithm. 74 89 DC enrollment status as of 1998 (the earliest year for which we have enrollment data). Column 1 reports estimates of equation (2.2) for all veterans in our sample where the left hand side variable is a veteran's combined disability rating. Column 2 includes only veterans enrolled in the DC program as of 1998, while column 3 excludes veterans not enrolled in DC as of 1998. Comparing the column 2 and 3 estimates reveals that CDRs were increasing more rapidly among 1998 BOG incumbent DC recipients than among their NOG counterparts prior to the policy change, but that there is only a modest acceleration in this trend after 2001. Specifically, CDRs were increasing by about 0.8 percentage points more annually among the BOG incumbents than the NOG incumbents between 1998 and 2001, and this differential increased only modestly to 1.1 percentage points annually from 2002 forward (0.84 + 0.26). In contrast, the acceleration among new BOG entrants after 2001 (column 2) is much larger, increasing from 0.20 to about 1.0 percentage points annually after 2001(0.22 + 0.73). Thus the Agent Orange policy change generated a much larger impact on the DC status of BOG veterans who had not already enrolled in DC prior to the policy change than it did on incumbent recipients, suggesting that labor supply impacts are likely driven primarily by new DC entrants. Panel B of Appendix Table 2.5 complements this evidence by comparing labor force participation trends between cells with a high density of DC incumbents and those with a low density of DC incumbents.75 Column- 1 reports estimates of equation (2.2) for all veterans in our sample, while column 2 restricts our sample to cells where at least one third of veterans ("high density") were enrolled in DC as of 1998 and column 3 restricts our sample to cells where less than one third of veterans ("low density") were enrolled in DC as of 1998.76 Following the 2001 policy change, the rate of divergence among BOG relative to NOG veterans actually attenuated slightly for cells with a high density of incumbents (from 0.35 to 0.29 points per year), but nearly tripled for cells with a low density of incumbents (from 0.11 to 0.32 points per year). This pattern again suggests that the differential decline in labor force participation among BOG relative to NOG While the SSA outcome cells never combine BOG and NOG veterans, they often combine a mixture of DC and nonDC enrollees, and this fraction changes over time as enrollment evolves. It would have been invalid to construct cells based on eventual DC enrollment, of course, since DC enrollment is an outcome of the analysis. 76 We cannot fully distinguish labor force participation effects for incumbents and potential entrants because labor force participation rates are determined at the cell level and cells are not stratified on DC enrollment. For example, consider a cell with two veterans who were enrolled in DC as of 1998 and three veterans who were not enrolled as of 1998. If all five veterans had positive earnings in 1998, but only four had positive earnings as of 2006, we are unable to determine whether the veteran who dropped out of the labor force was enrolled in DC as of 1998 or not. 75 90 veterans after 2001 is driven primarily by veterans induced to enroll in DC by the Agent Orange policy rather than by policy-induced changes in behavior among incumbent recipients. 2.8 Impacts on Total Earnings Since the Agent Orange policy precipitated a rise in transfer payments and fall in employment and earnings among eligible veterans, its net impact on incomes of BOG relative to NOG veterans is ambiguous. We compare trends in total measured income (earnings plus the sum of DC, SSDI and SSI benefits) between BOG and NOG veterans in Table 2.11, again limiting to the years 1998 through 2006, where we have all earnings and benefits variables available. Since we have already established that employment and earnings respond endogenously to policy-induced shifts in disability transfer income (as theory would predict), the current exercise should be viewed as descriptive rather than inferential; the net impact of the Agent Orange policy on veterans' income combines exogenous policy impacts and endogenous behavioral responses. The first column of Table 2.11 (panel A) shows that, at the start of sample window in 1998, measured incomes of BOG veterans were slightly lower than those of NOG veterans, with gaps of 0.75 log points overall, 0.80 log points among nonwhites, and 0.70 log points among whites.77 Notably, incomes were significantly higher (by 3.7 log points) among low-AFQT BOG veterans and significantly lower (by 2.5 log points) among high-AFQT BOG veterans (relative to their NOG counterparts). Recall, we saw earlier in panel B of Table 2.9 that labor earnings of BOG veterans averaged up to 5 log points below that of NOG veterans. Thus, comparing Tables 2.9 and 2.11 reveals that disability transfer income largely made up the difference between BOG and NOG veterans, and resulted in substantially higher income for low-AFQT BOG veterans relative to low-AFQT NOG veterans, consistent with the findings of Angrist et al (2010) for loweducation veterans. The second and third rows of Table 2.11 summarize the evolution of the BOG/NOG income differential before and after the Agent Orange policy. Prior to the policy change, BOG veterans were on a slightly more positive total income trend than NOG veterans, with average annual incomes in the full sample rising by 0.20 These reported earnings are covariate adjusted since they derive from the Table 2.11 regression estimates. Raw differences are similar, however, as may be inferred from Table 2.3. 77 91 log points annually in relative terms. This differential trend steepened significantly following the policy change, jumping from 0.20 to 0.71 log points (0.20 + 0.51) annually. For nonwhites and low-AFQT veterans, the increase in the trajectory was steeper, rising from 0.27 to 1.31 log points annually for the former group, and from 0.33 to 1.19 log points for the latter. While the Agent Orange policy sharply reduced employment and earnings among eligible veterans, the Table 2.11 estimates make clear that the policy-induced rise in transfer income more than offset these earnings reductions. As a final characterization of these net policy and behavioral effects, the lower panel of Table 2.11 summarizes the net impact of DC enrollment on total earnings, where DC enrollment is instrumented with the Agent Orange policy. The outcome variable for these estimates is cell total income-earnings plus disability benefits-and the endogenous variable is again DC enrollment. The point estimate of 0.52 for the full sample implies that - if a cell increased from 0 to 20 percent enrolled in DC - then average total veteran income in the cell would increase by about 10.4 percent. In interpreting this sizable impact, note that we found above that only one in five (18 percent of) newly enrolled veterans exited the labor force as a consequence of the Agent Orange policy. For the remaining four-fifths, enrollment in DC raised transfer income while generating only minimal offsetting reductions in labor income, thus unambiguously raising total income. Of course, average incomes of Vietnam era veterans were modest in the early 2000s and remained modest net of transfer benefits following the Agent Orange policy change. Nevertheless, this policy had the unusual effect of increasing average veteran incomes, inclusive of the foregone earnings of the approximately one in five newly enrolled DC recipients induced to exit the labor force. 2.9 Conclusion This chapter provides evidence that the policy-induced increase in enrollment in the VA's Disability Compensation program had a significant effect on the labor supply of Vietnam-era veterans. We estimate that 18 percent of individuals who became eligible for the DC program as a result of the policy change dropped out of the labor force. Analyses that explore the effect on average earnings support the conclusion that DC enrollment substantially lowered labor supply among Vietnam-era veterans who served in the 92 Vietnam theatre during the conflict there. These effects appear to be driven by those made newly eligible for DC benefits rather than by increases in benefits among those already receiving DC prior to the 2001 policy change. The increase in transfer income more than offsets the reduction in earnings, so that average income among BOG veterans increased relative to their counterparts who did not serve in theatre. Combining policyinduced increases in transfer income and reductions in labor force participation, we estimate an extensive margin labor supply elasticities of -. 50, which is somewhat larger than consensus estimates, though not implausible for this near-elderly population of Vietnam era veterans in diminished health. It is also likely that the income effect of DC income on employment and earnings is augmented by the interaction between DC enrollment and veterans' potential eligibility for Individual Unemployability and SSDI benefits, both of which provide strong incentives against gainful labor force participation. Consistent with this hypothesis, we find that the Agent Orange decision affected enrollment in two other federal disability programs, raising enrollment in SSDI by approximately one recipient for every ten veterans newly enrolled in DC, and slightly reducing enrollment in the means-tested SSI program (though the estimated effects are an order of magnitude smaller than for SSDI). In addition, 14 percent of Vietnam era veterans within our sample received IU status by 2006. Among veterans who enrolled in DC the year after the Agent Orange policy change, fully one-third received IU status or a CDR of 100 by 2006 . Since IU benefits are only available to veterans with earnings at or below the poverty threshold for a single individual, the rapid growth of IU designations among newly enrolled DC beneficiaries underscores the potential for the graduated DC benefit program to generate unintended labor supply impacts. Analyzing the dynamic interaction between these incentive and non-incentive effects of disability benefit programs is a worthy goal for future research. 93 2.10 Chapter 2 Appendix A. Construction of Cells in the BOG and NOG Samples To construct cells for matching individuals in the BOG and NOG samples with SSA data, we group individuals based on their values of certain background characteristics, including gender, race, year of birth, and so forth. The list of variables used to form cells, ranked in the order in which the grouping occurred, is: 1. Gender (male or female) 2. 3. 4. 5. 6. 7. 8. 9. Race (white or nonwhite) Death year (e.g. 1985) Year-of-birth (e.g. 1946) Start year (e.g. 1966) Education at entry (e.g. 0, 1, 2, 3, 4 for hsd, hsg, smc, clg, clg-plus) AFQT score quintile (e.g. 2) Loss year (e.g. 1972) Region of residence (e.g. Northeast) Before forming cells, we determined which individuals could be verified upon matching to the SSA data. A match would be verified if the social security number, date of birth, and at least six letters of the last name could be matched in the two data sets. We construct cells from the verified BOG and NOG samples so that there are between five and nine individuals in each cell. Each cell consists only of individuals in the BOG or in the NOG. The number of variables used in the grouping varies across cells. This occurs because in some cases, a cell reaches a size of five to nine after grouping on a relatively small number of variables. If a cell is between 10 and 29 individuals, we do not split further, but instead split the cell into the maximum number of cells with size five or more after sorting on the next matching variable to maximize similarity within the cells. If a cell has fewer than five individuals then we re-merge it with an adjacent cell. If a cell with fewer than five individuals is merged with an adjacent cell to form a new cell with more than nine individuals, the merged cell is split in two. The distribution of the cell size for the verified NOG sample is: In Cell of Size 5 6 7 8 9 Total Frequency 1,388,440 657,696 171,269 70,048 71,271 2,358,724 Percentage 58.86 27.88 7.26 2.97 3.02 100.00 94 The corresponding distribution for the verified BOG sample is: In Cell of Size 5 6 7 8 9 Total Frequency 942,415 328,338 92,785 41,120 40,878 1,445,536 Percentage 65.19 22.71 6.42 2.84 2.83 100.00 The frequency distribution for the number of variables used in the BOG and NOG cell formation is as follows (thus 206,780 in the BOG matched on variables Number of Vars 1 2 3 4 5 6 7 8 9 Total 1 through and including 8): # in BOG 0 61 4,149 30,716 62,807 65,140 162,909 206,780 912,974 1,445,536 # in NOG 0 89 3,391 38,087 86,671 80,909 174,281 397,667 1,577,629 2,358,724 B. Constructionof the Analysis Sample Appendix Figure 2. 1A diagrams the construction of the sample used in our analysis. The "loss year" file and the Vietnam file combine to form our baseline OEMA sample of 4.1 million veterans. After merging the OEMA sample to VA disability data, the SSA Death Master File, and SSA earnings information, we restrict our sample to 3.8 million veterans who had valid SSA earnings data. As seen in Table 2.2, the number of BOG veterans in our database is largest in the 1966 through 1971 start years; veterans who began their service after 1971 tended not to serve in the Vietnam theatre while those entering before 1966 are much less likely to be included in our sample since many in the latter group left the Army prior to 1968, the first year of our "loss year" file. We therefore restrict both the BOG and NOG samples to veterans who have a start year 95 between 1966 and 1971 inclusive, which reduces our sample to 1.9 million veterans. 78 We further restrict attention to individuals born between 1946 and 1951 inclusive, with 1.5 million veterans remaining. We next exclude approximately 150,000 veterans who have missing loss years, loss years before 1968, or loss years after 1985. We finally drop an additional 8,000 veterans who are in SSA cells where not all veterans have the same birth year, at least one veteran in the cell has a start year before 1966 or after 1971, or at least one veteran in the cell has a loss year before 1968 or after 1985. Our final sample includes 1.351 million veterans of the U.S. Army who began their service between 1966 and 1971 and were born between 1946 and 1951. For the analysis, we converted this final sample into a 1996-2007 panel of 15.2 million (year) x (individual) observations, or 2.9 million (year) x (cell) observations, as depicted in Appendix Figure 2.1 B. 79 For the nearly 200,000 veterans with missing start years, we imputed their start year based on the median start year for other veterans in the sample with the same date of birth. This has little bearing on our final sample as nearly all veterans with missing start years have missing loss years and are therefore excluded from our analysis. 78 A veteran who dies in year t is dropped from the analysis sample in year t and in all subsequent years. Because yearof-death is one of the variables used when constructing cells, typically all members of a cell die in the same year (if any die). The number of surviving cells falls by approximately 4.8 percent from 1996 to 2007. We restrict our panel to 1998 through 2006 for all specifications where the outcome is based on VA disability data. 79 96 ....I.....I.I.,,..0 Figure 2.1. Number and Percentage of Veterans Enrolled in DC 4.0 - 18 3.5 - 15 3.0 12 0 3-r 2.5 9 CAC 2.0 6 1.5 3 1 .0 , 1988 . 1990 . 1992 . I 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 Year -4-Number Enrolled -a- Percent Enrolled (All Vets) Notes: This figure reports the number of veterans enrolled in the Veterans' Disability Compensation (DC) program (left axis) and the percentage of living veterans enrolled in DC (right axis). Prior to 1999, all data comes from Annual Statistical Abstracts of the United States. From 1999 through 2014, annual DC enrollment numbers are from Veterans' Benefits Administration Annual Reports and the number of veterans in the U.S. comes from VetPop 2007, VetPop20 10, and VetPop2014. 97 Figure 2.2. DC Enrollment as a Percentage of Vietnam Era Veterans - 20 - -o 16 0 Ci 12 - C W 8 0 W to 4 0 *1 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 Year Notes: This figure reports the annual percentage of Vietnam era Veterans who received any DC benefits. See sources from Figure 1. 98 Figure 2.3. Percentage of BOG and NOG Veterans Enrolled in DC 25 -1 20 m ' da 015 C E - 0 10 0 5- 0* 1998 1999 2000 2001 2002 2003 2004 2005 Year -* BOG Veterans -0- NOG Veterans Notes: This table reports the annual percentage of BOG and NOG veterans who received any DC benefits between 1998 and 2006. 99 2006 Figure 2.4. Prevalance of DC Benefits for Diabetes in BOG and NOG Samples - 7 E 5 - - 6 . 0 4 0~ o 0 - 3 --- 1999 2000 - 2001 . 0 2002 2003 2004 2005 Year -*-BOG, Diabetes Benefits S-NOG, Diabetes Benefits -- BOG, Diabetes Benefits During Year of DC Enrollment -+- NOG, Diabetes Benefits During Year of DC Enrollment Notes: This table reports the annual percentage of BOG and NOG veterans who were enrolled in DC and who received DC benefits for diabetes. The series "BOG, Diabetes Benefits During Year of DC Enrollment" refers to the percentage of veterans in the BOG sample who were enrolled in DC, received benefits for diabetes, and began receiving diabetes benefits the same year they enrolled in DC. 100 20106 Figure 2.5. Percentage of BOG and NOG Veterans Enrolled in DC without Diabetes 18 16 E 12 - 14 i 0 0 10 - 0 8 ---- -e--U s-I 0~ 6 ---- ----- - - 0 0 4 4 2 0 i I I I I I I I 1999 2000 2001 2002 2003 2004 2005 2006 Year -+BOG Veterans -0- NOG Veterans Notes: This table reports the annual percentage of BOG and NOG veterans who were enrolled in DC but did not receive compensation for diabetes. 101 Figure 2.6. BOG DC Receipt Relative to NOG by Year of Birth Cohort - 9 - 8 - 7 C: - 6 5 - 4 24 0 1998 1999 2000 2001 2002 2003 2004 2005 Year -8-2YOB 1946-1948 -+-YOB 1949-1951 Notes: This table reports the absolute percentage point difference between BOG DC receipt and NOG DC receipt for two veteran birth cohorts: 1946-1948 and 1949-1951. The vertical line indicates the year of the Agent Orange policy change. Each series is normalized to the BOG-NOG difference in DC receipt as of year 1998. 102 2006 Figure 2.7. BOG LFP Relative to NOG by Year of Birth Cohort - - 0 -0.5 a.) -1 - a.) a.) 0 -1.5 e' a.) a.) -2 ~.) I- a.) -2.5 9%e -3 -3.5 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Year -*- -+- YOB 1946-1948 YOB 1949-1951 Notes: This table reports the absolute percentage point difference between BOG labor force participation (LFP) and NOG LFP for two veteran birth cohorts: 1946-1948 and 1949-195 1. The vertical line indicates the year of the Agent Orange policy change. Each series is normalized to the BOG-NOG difference in LFP as of year 1996. Labor force participation is defined as positive earnings. 103 2007 Table 2.1. Evolution of Disability Compensation (DC) Benefits by Year of DC Enrollment DC Enrollment Cohorts 1999-2006 Outcome Year Year of Enrollment 1999 2000 2001 2002 2003 2004 2005 2006 1999 2000 2001 2002 2003 2004 2005 2006 1999 2000 2001 2002 2003 2004 2005 2006 1999 2000 2001 2002 2003 2004 2005 2006 1999 2000 2001 (1) (2) (3) 41.8 45.5 44.1 10,773 15.3 0.8 12,577 11,714 2002 (4) 2003 (5) A. Mean Combined Disability Rating 53.5 57.7 49.7 55.0 59.4 50.0 46.2 52.3 41.2 46.0 38.7 41.8 B. Mean Annual DC Payment ($2013) 14,389 15,734 17,325 14,295 16,090 17,756 10,709 12,603 14,991 9,611 12,432 10,650 C. Percentage Receiving IU Award 31.3 26.4 20.9 31.3 17.6 25.2 15.3 20.8 12.3 2004 (6) 2005 (7) 2006 (8) 60.3 62.3 56.6 50.9 47.8 43.0 62.3 64.6 60.6 54.8 52.2 48.5 41.3 63.6 66.3 62.6 57.6 55.2 51.5 45.1 39.1 18,347 18,923 16,715 14,368 13,007 11,114 19,083 19,799 18,329 15,878 14,735 13,222 10,335 19,822 20,701 19,317 17,153 16,099 14,521 11,922 9,590 43.2 44.6 38.6 30.4 27.0 22.5 14.4 44.7 46.5 40.7 33.2 30.0 25.6 17.7 11.5 13.8 14.6 39.9 59.7 49.6 43.7 39.9 14.8 15.8 40.4 59.9 50.1 44.5 40.8 40.9 or 100 CDF 40.5 36.7 41.5 37.1 33.5 28.0 25.9 20.1 21.9 15.1 16.8 D. Percentage Receiving Compensation for Diabetes 12.4 7.9 11.0 1.1 2.5 9.4 12.1 13.4 0.5 2.3 39.4 38.6 15.3 37.6 59.0 59.5 58.5 48.6 49.1 42.7 Notes: This table reports the progression of Disability Compensation (DC) benefits for veterans in the sample described in the data appendix. Each row corresponds to a DC enrollment cohort. Each column corresponds to a particular year. In panel A, veterans with an IU award are given a CDR of 100. In panel B, mean annual DC payments are in $2013. 104 Birth Year (1) Table 2.2. Distribution of Year of Birth and Start Year in OEMA Sample B. Start Year Distribution A. Year of Birth Distribution BOG Veterans NOG Veterans BOG Veterans NOG Veterans Percent of of Percent of Percent Percent of Observations BOG Sample Observations NOG Sample Start Year Observations BOG Sample Observations NOG Sample (5) (4) (3) (2) (1) (5) (4) (3) (2) 1.42 0.07 0.07 0.10 0.09 0.09 0.15 0.17 0.18 0.25 0.37 106,389 12,132 12,107 15,181 14,769 14,192 16,002 15,689 7.14 0.81 0.81 1.02 0.99 0.95 1.07 1.05 15,895 18,357 26,859 1.07 1.23 1.80 123,079 13,933 13,250 14,212 15,024 17,947 23,911 34,477 47,114 60,134 80,247 8.26 0.93 0.89 0.95 1.01 1.20 1.60 2.31 3.16 4.03 5.38 <1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 36,885 1,776 1,779 2,485 2,254 2,339 4,018 4,359 29,298 38,786 48,100 1.75 0.10 0.11 0.14 0.19 0.29 0.42 0.73 1.13 1.49 1.85 72,463 2.79 131,973 8.86 1966 38,014 1.46 88,681 5.95 3.74 7.56 7.71 6.30 174,641 263,378 263,700 149,738 11.72 17.67 17.69 10.05 <1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 45,287 2,665 2,840 3,548 4,813 7,522 10,865 18,988 1946 4,778 6,362 9,675 1947 1948 1949 1950 108,739 165,733 155,513 143,759 4.19 6.39 5.99 5.54 227,042 227,941 191,467 134,445 15.23 15.29 12.85 9.02 1967 1968 1969 1970 96,939 196,281 199,973 163,525 1951 1952 141,803 165,508 5.46 6.38 72,982 37,175 4.90 2.49 1971 1972 191,736 215,043 7.39 8.29 60,564 15,087 4.06 1.01 1953 141,310 5.45 15,720 1954 1955 >1956 157,134 162,502 968,018 6.05 6.26 37.30 5,849 919 1,518 1.05 0.39 0.06 0.10 1973 1974 1975 >1976 Missing 144,792 187,413 162,643 920,849 1,276 5.58 7.22 6.27 35.48 0.05 3,053 1,753 1,802 3,624 196,766 0.20 0.12 0.12 0.24 13.20 Total 2,595,194 100.00 1,490,359 100.00 Total 2,595,194 100.00 1,490,359 100.00 Notes: Panel A reports the distribution of birth years in the original OEMA sample for the BOG and NOG subsamples. Panel B reports the distribution of start years. A start year is the year a veteran entered the Army. 105 Table 2.3. Characteristics of the BOG and NOG Samples A. SSA Verified Sample B. Full Sample BOG NOG BOG NOG (2) (1) (2) (1) % Verified in SSA Data % with Start Yrs 1966-71 % with YOB 1946-1951 % Deceased by 1997 % Deceased by 2006 % Nonwhite % Missing Education % HS Dropout % HS Grad % Some College % College Grad % More than College %Missing AFQT Score Average AFQT Score Average Year of Birth Average Start Year Average Loss Year % on DC in 1998 % on DC in 2000 % on DC in 2006 Mean Annual DC Payment in 1998 Mean Annual DC Payment in 2000 Mean Annual DC Payment in 2006 Mean Annual DC Pay in 1998 if>0 Mean Annual DC Pay in 2000 if>0 Mean Annual DC Pay in 2006 if>0 % Positive Earnings in 1998 % Positive Earnings in 2000 % Positive Earnings in 2006 Ln(Cell Mean Earnings) in 1998 Ln(Cell Mean Earnings) in 2000 Ln(Cell Mean Earnings) in 2006 % on SSDI in 1998 % on SSDI in 2000 % on SSDI in 2006 % on SSI in 1998 % on SSI in 2000 % on SSI in 2006 100.0 100.0 100.0 4.8 9.9 11.3 1.1 31.4 48.1 14.5 4.8 0.2 26.4 52.1 1948.4 1968.4 1971.1 13.5 14.3 23.0 10,529 12,078 15,935 10,530 12,079 15,935 84.3 82.8 71.3 10.67 10.66 10.37 5.6 6.8 12.6 0.7 0.7 0.7 100.0 100.0 100.0 2.9 7.9 11.8 15.3 23.3 37.1 14.4 9.2 0.8 19.1 53.4 1948.6 1969.0 1971.4 6.3 6.5 7.6 9,993 10,750 12,543 9,993 10,751 12,543 85.1 84.1 74.4 10.70 10.69 10.44 4.6 5.5 9.7 1.0 1.0 1.3 95.7 100.0 100.0 6.2 11.5 11.8 1.2 31.4 48.3 14.3 4.7 0.2 25.8 51.7 1948.4 1968.4 1971.1 13.5 14.2 22.9 10,600 12,140 15,982 10,600 12,140 15,982 89.0 100.0 100.0 6.1 11.0 12.7 14.3 24.3 38.0 14.2 8.6 0.7 17.7 52.5 1948.7 1969.0 1971.4 6.2 6.4 7.5 10,137 10,888 12,658 10,137 10,889 12,658 Observations 765,852 585,017 806,698 659,283 Notes: Panel A reports summary statistics for veterans in the BOG and NOG samples described in Section III and the data appendix. Panel B includes veterans who did not verify with the Social Security Administration, but makes all other sample restrictions described in the data appendix. All earnings and DC payment values are in $2013. Disability, employment, and earnings information only reflect living veterans. 106 Table 2.4. DC Receipt and Diabetes Compensation among BOG and NOG Veterans, 1998-2006 Year 1998 1999 2000 2001 2002 2003 2004 2005 2006 1998 1999 2000 2001 2002 2003 2004 2005 2006 Full Sample BOG NOG (1) (2) 6.3 6.4 6.5 6.6 6.8 7.0 7.2 7.4 7.6 13.5 13.9 14.3 15.0 16.4 18.3 19.9 21.4 23.0 Whites NOG BOG (3) (4) A. Percentage Enrolled in DC 13.0 6.0 13.4 6.1 6.2 13.8 14.4 6.3 6.4 15.8 6.6 17.6 6.8 19.1 7.0 20.6 7.2 22.1 Nonwhites NOG BOG (6) (5) 9.0 9.2 9.3 9.4 9.7 10.0 10.3 10.6 10.9 B. Mean Annual DC Payment, Conditional on DC Enrollment 9,993 10,529 9,989 10,476 10,011 10,356 11,284 10,324 11,206 10,515 10,750 12,078 10,690 11,987 11,055 11,089 12,763 11,022 12,654 11,427 11,225 13,109 11,135 12,976 11,675 11,665 13,992 11,537 13,835 12,310 11,991 14,773 11,850 14,577 12,702 12,237 15,420 12,082 15,207 13,018 12,543 15,935 12,381 15,696 13,363 17.1 17.7 18.2 19.3 21.4 24.0 26.0 28.2 30.2 10,859 11,765 12,635 13,422 13,908 14,923 15,933 16,688 17,358 C. Percentage Receiving Compensation for Diabetes, Conditional on DC Enrollment 1998 1999 2000 2001 2002 2003 2004 2005 2006 1.4 1.4 1.6 3.3 4.4 5.2 5.8 6.4 0.7 0.7 2.4 13.3 19.6 22.9 25.6 28.1 1.2 1.2 1.4 3.0 4.0 4.8 5.4 6.0 0.6 0.6 2.2 12.4 18.4 21.8 24.4 26.9 2.2 2.2 2.5 4.8 6.3 7.2 7.8 8.4 1.3 1.3 3.9 18.3 26.2 30.0 32.8 35.2 Notes: Panel A of this table reports the percentage of veterans in the BOG and NOG samples who were enrolled in the DC program between 1998 and 2006. Panel B reports the mean monthly disability payments (in $2013) for veterans who were enrolled in the DC program between 1998 and 2006. Panel C reports the percenage of DC enrolled veterans who received compensation for diabetes. 1998 is not included in panel C because our data include specific conditions for 1999 through 2006 only. 107 Table 2.5. DC Receipt in the BOG versus NOG samples from 1998-2006 Dependent Variable: Percentage of Cell Enrolled in DC A. All B. YOB: 1946-1948 C. YOB: 1949-1951 (2) (1) (2) (1) (2) (1) BOG 6.97*** (0.06) BOG * (YR-1998) BOG * (YR-2001)* (YR>2002) BOG * (YR99) BOG * (YROO) BOG *(YRO1) BOG * (YR02) BOG * (YR03) BOG * (YR04) BOG *(YR05) BOG * (YR06) Outcome Mean (1998) # OBS (Cell)x(Year) 6.91*** (0.05) 0.40*** (0.02) 0.99*** (0.03) 7.12*** (0.09) 0.33*** (0.09) 0.59*** (0.09) 1.18*** (0.09) 2.46*** (0.09) 4.08*** (0.09) 5.39*** (0.09) 6.71*** (0.09) 7.98*** (0.09) 0.34*** (0.13) 0.58*** (0.13) 1.18*** (0.13) 2.55*** (0.13) 4.22*** (0.13) 5.58*** (0.13) 6.93*** (0.13) 8.28*** (0.13) 10.36 2,142,029 10.75 1,104,897 7.06*** (0.07) 0.40*** (0.03) 1.04*** (0.05) 6.83*** (0.09) 6.77*** (0.07) 0.40*** (0.03) 0.93*** (0.05) 0.32** (0.12) 0.61*** (0.12) 1.18*** (0.12) 2.37*** (0.12) 3.92*** (0.12) 5.18*** (0.12) 6.45*** (0.12) 7.65*** (0.13) 9.94 1,037,132 Notes: This table reports estimates of equations (1) and (2) where the dependent variable is the percentage of each cell enrolled in the Disability Compensation program each year. Panel A includes all veterans in our sample and other panels restrict the sample to veterans in the specified group. Each cell has one observation for each year between 1998 and 2006. All regressions include year fixed effects and fixed effects for each combationation of (AFQTquintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite and veterans with a missing AFQT score are grouped into a sixth AFQT category. All regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. * **, and * denote significance at the 1%, 5%, and 10% level, respectively. 108 TTable 2.6. Disability Compensation Outcomes in the BOG versus NOG samples from 1998-2006 AFQT>45 AFQT<45 Whites All Nonwhites (5) (4) (3) (2) (1) BOG BOG * (YR-1998) * BOG * (YR-2001) (YR 2002) Outcome Mean (1998) # OBS (Cell)x(Year) A. Dependent Variable: Percentage of Cell enrolled in DC 7.72*** 6.80*** 8.33*** 6.91*** (0.10) (0.17) (0.05) (0.05) 0.61*** 0.37*** 0.52*** 0.40*** (0.08) (0.03) (0.05) (0.02) 0.99*** 1.27*** 1.19*** 0.95*** (0.04) (0.11) (0.06) (0.03) 10.36 2,142,029 13.44 232,540 9.96 1,909,489 11.55 660,866 B. Dependent Variable: Cell Mean Combined Disability Rating (CDR) 2.88*** 3.39* 2.81*** 3.83*** (0.10) (0.06) (0.03) (0.03) 0.50*** 0.34*** 0.47*** BOG * (YR-1998) 0.32*** (0.01) (0.05) (0.01) (0.03) 0.66*** 0.96*** 0.63*** 0.83*** BOG * (YR-2001) (0.04) (YR 2002) (0.02) (0.07) (0.02) * * BOG Outcome Mean (1998) # OBS (Cell)x(Year) BOG BOG * (YR-1998) * BOG * (YR-2001) (YR 2002) Outcome Mean (1998) # OBS (Cell)x(Year) 3.96 2,142,029 5.25 232,540 3.79 1,909,489 4.75 660,866 C. Dependent Variable: Cell Mean Annual DC Payment 1,072*** 930*** 766*** 744*** (11) (20) (10) (35) 107*** 158*** 159*** 113*** (10) (17) (5) (5) 239*** 192*** 290*** 179*** (14) (24) (7) (7) 1,076 2,142,029 1,425 232,540 1,031 1,909,489 1,323 660,866 5.93*** (0.07) 0.31*** (0.03) 0.81*** (0.05) 8.98 988,801 2.48*** (0.04) 0.26*** (0.02) 0.51*** (0.03) 3.19 988,801 673*** (13) 85*** (6) 145*** (9) 851 988,801 Notes: This table reports estimates of equation (2) where the dependent variable is the outcome indicated in the panel heading. Each cell has one observation for each year between 1998 and 2006. All regressions include year fixed effects and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite. Column I includes all veterans in our sample; all other columns restrict the sample to veterans in the specified group. Columns 1-3 group veterans with a missing AFQT score into a sixth AFQT category while columns 4-5 exclude veterans with missing AFQT scores. In panel B, veterans with an IU rating are coded as having a Combined Disability Rating (CDR) of 100. In panel C, mean annual DC payments are in $2013. All regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% level, respectively. 109 Table 2.7. SSDI, SSI, and Total Disability Income in the BOG versus NOG samples All Nonwhites Whites AFQT<45 (4) (3) (2) (1) BOG BOG * (YR-1 996) * BOG * (YR-2001) (YR:2002) Outcome Mean (1996) # OBS (Cell)x(Year) BOG BOG * (YR-1 996) * BOG * (YR-2001) (YR!2002) Outcome Mean (1996) # OBS (Cell)x(Year) A. Dependent Variable: Percentage of Cell Receiving SSDI 0.74*** 1.14*** 0.69*** 1.12*** (0.04) (0.12) (0.04) (0.07) 0.13*** 0.19*** 0.13*** 0.14*** (0.01) (0.04) (0.01) (0.02) 0.12*** 0.06 0.12*** 0.12*** (0.02) (0.06) (0.02) (0.04) 4.51 2,862,513 6.69 310,987 4.23 2,551,526 6.77 883,832 B. Dependent Variable: Percentage of Cell Receiving SSI -0.238*** -0.353*** -0.222*** -0.328*** (0.012) (0.055) (0.012) (0.028) -0.025*** -0.024 -0.025*** -0.038*** (0.004) (0.016) (0.003) (0.008) -0.013** -0.052** -0.008 -0.036*** (0.006) (0.026) (0.006) (0.013) 0.90 2,862,513 2.26 310,987 0.73 2,551,526 1.54 883,832 AFQT 45 (5) 0.81*** (0.04) 0.11*** (0.01) 0.10*** (0.02) 3.08 1,321,309 -0.063*** (0.013) -0.015*** (0.004) -0.004 (0.006) 0.50 1,321,309 * C. Dependent Variable: Annual Disability Income (DC Ben + SSDI Ben + SSI Ben), 1998-2006 only BOG 861*** 1,061*** 835*** 772*** 1,197*** (13) (44) (14) (26) (17) BOG * (YR-1998) 128*** 178*** 121*** 172*** 98*** (6) (21) (6) (12) (8) BOG * (YR-2001) 226*** 325*** 214*** 277*** 173*** (YR22002) (18) (11) (9) (30) (9) Outcome Mean (1998) Total Income Mean (1998) # OBS (Cell)x(Year) 1,608 53,639 2,142,029 2,181 38,276 232,540 1,535 55,601 1,909,489 2,093 39,549 660,866 1,225 59,663 988,801 Notes: This table reports estimates of equation (2) where the dependent variable is the outcome indicated in each panel heading. In panels A and B, each cell has one observation for each year between 1996 and 2007. Panel C only includes observations from 1998 through 2006 since DC outcomes are not available in other years. Annual disability benefits in panel C are in $2013. All regressions include year fixed effects and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite. Column I includes all veterans in our sample and columns 2-5 restrict the sample to veterans in the specified group. Columns 1-3 group veterans with a missing AFQT score into a sixth AFQT category while columns 4-5 exclude veterans with missing AFQT scores. All regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. * *, and * denote significance at the 1%, 5%, and 10% level, respectively. 110 Table 2.8. Labor Force Participation in the BOG versus NOG samples from 1996-2007 Dependent Variable: Percentage of Cell with Positive Annual Earnings B. YOB: 1946-1948 C. YOB: 1949-1951 A. All (2) (1) (2) (1) (2) (1) BOG -0.32*** -0.33*** 0.00 0.02 -0.82*** -0.87*** (0.11) (0.08) (0.05) (0.10) (0.07) (0.08) BOG * (YR-1996) -0.15*** -0.17*** -0.14*** (0.02) (0.02) (0.02) -0.18*** -0.18*** -0.19*** BOG * (YR-2001) * (YR>2002) (0.03) (0.04) (0.04) BOG * (YR97) -0.16 -0.15 -0.16 (0.11) (0.15) (0.16) BOG * (YR98) -0.31*** -0.29** -0.33** (0.11) (0.16) (0.15) BOG * (YR99) -0.45*** -0.38** -0.52*** (0.11) (0.15) (0.16) BOG * (YROO) -0.77*** -0.79*** -0.75*** (0.11) (0.15) (0.16) BOG * (YRO1) -0.81*** -0.87*** -0.76*** (0.11) (0.15) (0.16) BOG * (YR02) -1.01*** -1.06*** -0.96*** (0.11) (0.15) (0.16) BOG * (YR03) -1.34*** -1.45*** -1.23*** (0.11) (0.15) (0.16) BOG * (YRO4) -1.74*** -1.92*** -1.55*** (0.11) (0.15) (0.16) BOG * (YR05) -2.15*** -2.27*** -2.03*** (0.11) (0.15) (0.16) BOG * (YR06) -2.43*** -2.56*** -2.29*** (0.11) (0.15) (0.16) BOG * (YR07) -2.83*** -2.93*** -2.73*** (0.11) (0.15) (0.16) Outcome Mean (1996) # OBS (Cell)x(Year) 85.70 2,862,513 86.56 1,476,423 84.80 1,386,090 Notes: This table reports estimates of equations (1) and (2) where the dependent variable is the percentage of each cell with positive annual earnings. Panel A includes all veterans in our sample and panels B and C restrict the sample to veterans in the specified group. All regressions include year fixed effects and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite and veterans with a missing AFQT score are grouped into a sixth AFQT category. All regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. * **, and * denote significance at the 1%, 5%, and 10% level, respectively. III BOG * (YR-1996) * BOG * BOG * (YR-2001) (YRt2002) Outcome Mean (1996) # OBS (Cell)x(Year) 85.70 2,862,513 78.18 310,987 86.67 2,551,526 81.70 883,832 BOG * (YR-2001) (YR 2002) B. Dependent Variable: 100 x Ln(Cell Mean Earnings) -4.68*** -2.59*** -2.30*** 0.02 (0.18) (0.62) (0.18) (0.32) -0.23 * * -0.33* -0.21 *** -0.27*** (0.18) (0.05) (0.05) (0.10) -0.77*** -0.45*** -0.41 *** -0.40** (0.30) (0.09) (0.16) (0.08) Outcome Mean (1996) Earnings Mean (1996) # OBS (Cell)x(Year) 1,065.83 49,848 2,855,305 BOG * (YR-1996) BOG BOG * (YR-1996) * BOG * (YR-2001) (YR 2002) 1,032.25 35,449 308,923 1,070.12 51,691 2,546,382 1,038.31 36,656 879,962 C. Dependent Variable: 100 x Ln(Percentage of Cell with Earnings>0) -0.29*** -3.57*** 0.13 -1.20*** (0.31) (0.08) (0.09) (0.17) -0.3 1* ** -0.21 ** -0.20*** -0.24*** (0.09) (0.02) (0.05) (0.03) -0.36*** -0.52*** -0.35*** -0.44*** (0.04) (0.15) (0.04) (0.08) Outcome Mean (1996) # OBS (Cell)x(Year) * * * BOG -17.85 2,855,305 -28.26 308,923 -16.51 2,546,382 -23.28 879,962 -0.92*** (0.07) -0. 11 ** * A. Dependent Variable: Percentage of Cell with Earnings>0 -0.33*** -2.50*** -0.05 -0.88*** (0.05) (0.06) (0.18) (0.10) -0.19*** -0.15* -0.15*** -0.16*** (0.02) (0.05) (0.02) (0.03) -0.22** -0.18*** -0.18*** -0.20*** (0.03) (0.03) (0.09) (0.05) AFQT 45 (5) (0.02) -0.15*** (0.03) 89.08 1,321,309 -4.10*** (0.23) -0.13 (0.07) -0.27** (0.11) * Table 2.9. LFP and Earnings in the BOG versus NOG samples from 1996-2007 All Nonwhites Whites AFQT<45 (4) (3) (2) (1) 1,080.45 55,709 1,320,155 -1.14*** (0.10) -0.16*** (0.03) -0.29** (0.05) -13.25 1,320,155 Notes: This table reports estimates of equation (2) where the dependent variable is the outcome indicated in each panel heading. Each cell has one observation for each year between 1996 and 2007. Panels B and C exclude cells with 0 mean annual earnings. All earnings outcomes are in $2013. All regressions include year fixed effects and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite. Column 1 includes all veterans in our sample and columns 2-5 restrict the sample to veterans in the specified group. Columns 1-3 group veterans with a missing AFQT score into a sixth AFQT category while columns 4-5 exclude veterans with missing AFQT scores. All regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% level, respectively. 112 Table 2.10. 2SLS Estimates of DC Receipt and Annual Disability Benefits on Labor Force Participation Dependent Variable: Percentage of Cell with Positive Earnings. 1998-2006. DC Enrollment Outcome Mean (1998) # OBS (Cell)x(Year) (Disability Benefits)/1000 Outcome Mean (1998) # OBS (Cell)x(Year) YOB: 1946- YOB: 1949- All 1948 1951 Nonwhites Whites AFQT<45 (1) (2) (3) (4) (5) (6) A. Endogenous Variable: Percentage of Cell Enrolled in DC -0.17*** -0.25** -0.21*** -0.16*** -0.18*** (0.05) (0.11) (0.06) (0.07) (0.04) 84.7 2,142,029 85.5 1,104,897 83.9 1,037,132 77.1 232,540 85.7 1,909,489 B. Endogenous Variable: (DC Benefits + SSDI Benefits + SSI Benefits)/1000 -0.77*** -0.97** -0.94*** -0.70*** -0.80*** (0.19) (0.42) (0.28) (0.23) (0.18) 84.7 2,142,029 85.5 1,104,897 83.9 1,037,132 77.1 232,540 85.7 1,909,489 -0.20*** (0.07) 80.5 660,866 -0.84*** (0.28) 80.5 660,866 AFQT 45 (7) -0.17** (0.07) 88.3 988,801 -0.82*** (0.31) 88.3 988,801 Notes: This table reports 2SLS estimates of the effects of DC receipt on labor force participation (panel A) and annual disability benefits on labor force participation (panel B). The instrument for all estimates is (BOG)x(year - 2001)xl(year ' 2002). All regressions include a main effect for (BOG), a (BOG)x(year - 1998) time trend, year fixed effects, and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year). Annual disability benefits are in $2013. Each cell has one observation for each year between 1998 and 2006 and all regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% level, respectively. 113 Table 2.11. Earnings plus Disability Income in the BOG versus NOG samples Dependent Variable: 100 x Ln(Earnings + DC Ben + SSDI Ben + SSI Ben), 1998-2006 only All Nonwhites Whites AFQT<45 AFQT 45 (5) (4) (3) (2) (1) BOG BOG * (YR-1998) BOG * (YR-200l)* (YR 2002) Outcome Mean (1998) Total Income Mean (1998) # OBS (Cell)x(Year) A. Reduced Form Effect of Agent Orange Policy Change -0.75*** -0.80 -0.70*** (0.17) (0.53) (0.18) 0.20** 0.27 0.19** (0.08) (0.25) (0.08) 0.51*** 1.04*** 0.45*** (0.11) (0.36) (0.12) 1,074.05 53,639 2,137,665 1,043.42 38,276 231,235 1,077.96 55,601 1,906,430 3.66*** (0.28) 0.33** (0.13) 0.86*** (0.19) 1,048.71 39,549 658,480 253*** (0.23) 0.23** (0.11) 0.34** (0.16) 1,087.39 59,663 988,152 B. 2SLS Estimates of DC Receipt on 100 x Ln(Earnings + DC Ben + SSDI Ben + SSI Ben) Endogenous Variable: Percentage of Cell Enrolled in DC DC Enrollment Outcome Mean (1998) Total Income Mean (1998) # OBS (Cell)x(Year) 0.52*** (0.12) 1,074.05 53,639 2,137,665 0.83*** (0.29) 1,043.42 38,276 231,235 0.47*** (0.13) 1,077.96 55,601 1,906,430 0.73*** (0.16) 1,048.71 39,549 658,480 0.42** (0.20) 1,087.39 59,663 988,152 * Notes: Panel A of this table reports estimates of equation (2) where the dependent variable is 100 times the log of each cell's mean annual earnings and total disability income. Panel B reports 2SLS estimates of the effects of DC receipt on the same outcome. The instrument in panel B is (BOG) x (year - 2001) x 1(year > 2002) and the endogenous variable is the percentage of the cell enrolled in DC. Each cell has one observation for each year between 1998 and 2006, which are the years where DC information is available. Cells with 0 mean annual earnings are excluded. Earnings and disability income values are in $2013. All regressions include year fixed effects and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite. Column I includes all veterans in our sample and columns 2-5 restrict the sample to veterans in the specified group. Columns 1-3 group veterans with a missing AFQT score into a sixth AFQT category while columns 4-5 exclude veterans with missing AFQT scores. All regressions weight each observation according to the number of veterans in the cell. Standard errors are reported in parentheses. ***, and * denote significance at the 1%, 5%, and 10% level, respectively. 114 Appendix Figure 2. lA. Data Diagram Army Loss Files 1968-1985 Department of Defense Vietnam File Original OEMA Sample N = 4.086 million SSA Earnings Data VA DC Data N= SSA Mortality Data SSA Verified Sample 3.804 million, Cells= 694,589 YOB: 1946-1951, Start Year: 1966-1971 N = 1.508 million, Cells = 283,278 Exclude Veterans with Missing Loss Years, Loss Years before 1968, or Loss Years after 1985 N = 1.359 million, Cells = 255,222 N = Cell Restrictions* 1.351 million, Cells= 253,122 *Cell restrictions limit to the sample to cells where all veterans have the same birth year (466 observations dropped), all veterans have start years between 1966 and 1971 (2,677 observations dropped), all veterans have loss years between 1968 and 1985 (4,681 observations dropped), and no veterans have a missing loss year (80 observations dropped). Appendix Figure 2.1 B. Annual Observations Breakdown I I **A deceased veteran is removed from the panel in the year of his death and all years following. If only a fraction of the cell is deceased in a particular year, all cell level averages are recomputed to reflect averages among living veterans only. Appendix Table 2.1. Percentage of DC Recipients and Average Annual Benefit by CDR and Service Era in 2006 CDR Recipients Payments (2013 $ Millions) 0% 14,291 10% 20% 30% 40% 50% 60% 70% 80% 775,346 90% 60,546 100% 238,662 $15 $1,212 $1,275 $1,742 $1,956 $1,709 $3,236 $4,263 $3,252 $1,896 $9,051 Total 2,725,824 $29,608 417,721 334,931 259,834 161,568 184,264 165,257 113,404 Mean Annual Benefit Gulf Share with Each Rating by Service Era WW II Vietnam Korea $1,030 0.1% 0.3% 2.4% $1,563 21.8% 30.6% 13.3% 10.6% 9.0% 5.9% 7.0% 8.5% 5.7% 3.1% 14.7% 13.0% 12.4% 8.2% 5.4% 7.6% 5.5% 4.1% $37,926 27.3% 17.8% 15.3% 12.5% 7.2% 7.0% 4.8% 3.4% 1.7% 3.0% $10,862 694,813 947,598 $3,052 $5,202 $7,530 $10,577 $17,564 $25,799 $28,678 $31,310 1.4% 33.0% Peacetime 0.4% 37.3% 17.8% 2.3% 8.5% 12.5% 13.1% 8.1% 5.8% 7.1% 5.4% 4.1% 2.3% 7.2% 159,804 328,044 595,565 11.0% 8.0% 4.7% 5.7% 4.2% 2.7% 1.4% 7.0% Source: U.S. Department of Veterans Affairs, 2006 Annual Benefits Report. Average annual benefit payments are in $2013 (million 117 Appendix Table 2.2. Distribution of Loss Year in OEMA Sample A. NOG Veterans Loss Year 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 Missing Total Observations 17,787 83,222 265,968 241,772 122,400 164,026 175,414 162,983 160,387 146,516 129,296 141,220 137,754 119,916 128,458 141,806 131,090 125,179 2,595,194 B. BOG Veterans Percent of NOG Sample Observ ations Percent of BOG Sample 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.69 3.21 10.25 9.32 4.72 6.32 6.76 6.28 6.18 5.65 4.98 5.44 5.31 4.62 4.95 5.46 5.05 4.82 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 7,128 165,959 258,909 273,725 219,993 77,013 51,296 32,696 26,037 23,410 21,334 19,612 15,946 12,921 12,153 11,393 11,597 10,857 6,052 5,089 5,684 5,212 4,392 3,501 3,671 2,563 201,786 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.02 0.48 11.14 17.37 18.37 14.76 5.17 3.44 2.19 1.75 1.57 1.43 1.32 1.07 0.87 0.82 0.76 0.78 0.73 0.41 0.34 0.38 0.35 0.29 0.23 0.25 0.17 13.54 100.00 1,490,359 100.00 1 2 2 2 8 25 140 250 Notes: This table reports the distribution of loss years within the NOG and BOG groups of the original OEMA sample. A loss year is the year a veteran exited the Army. 118 4 Appendix Table 2.3. SSA Verification Rate by Start Year A. NOG Veterans Start Year S1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 >1976 Missing Obs B. BOG Veterans SSA Match Rate 86.5 85.3 88.2 87.3 87.8 88.6 82.2 82.4 82.8 81.5 82.4 84.3 84.8 88.1 88.4 91.0 93.5 93.6 91.4 91.4 91.5 91.6 94.4 93.3 92.3 87.0 87.1 86.6 86.2 85.6 85.8 88.2 96.3 81.7 6,970 1,321 2,057 2,027 3,752 3,846 5,400 4,275 1,925 1,983 1,522 1,807 1,776 1,779 2,485 2,254 2,339 4,018 4,359 4,778 6,362 9,675 38,014 96,939 196,281 199,973 163,525 191,736 215,043 144,792 187,413 162,643 920,849 1,276 Total Obs 8,209 2,808 3,943 3,754 5,988 4,794 8,286 13,569 13,125 15,593 13,370 12,950 12,132 12,107 15,181 14,769 14,192 16,002 15,689 15,895 18,357 26,859 88,681 174,641 263,378 263,700 149,738 60,564 15,087 3,053 1,753 1,802 3,624 196,766 SSA Match Rate 95.8 96.4 97.2 96.4 96.3 95.8 95.0 95.3 98.0 98.6 98.9 99.0 98.8 99.0 98.9 98.9 99.0 98.8 98.5 98.4 98.5 98.2 98.8 97.2 95.5 93.1 98.8 99.4 99.5 98.8 99.1 98.6 99.1 99.4 2,595,194 90.9 1,490,359 97.0 Notes: This table reports SSA verification rates for veterans by start year and BOG or NOG status. 119 Appendix Table 2.4. Comparison of Census and Army/SSA Demographics and Earnings Data for Year 1999 Age in 1999 A. Army/SSA Verified Sample (1999) Whites Nonwhites All (2) (1) (3) 50.5 50.5 50.3 (1.5) (1.5) (1.5) Race White Nonwhite 88.7 11.3 100.0 0.0 0.0 100.0 Education HS Dropout HS Grad Some College College Grad More than College 27.3 43.6 14.7 6.9 0.4 26.9 42.9 15.3 7.5 0.5 30.7 49.2 10.4 2.4 Positive Earnings 84.1 Annual Earnings ($2013) Mean Mean (if>0) Median Observations B. Vietnam Veterans born 1946-1951: 2000 Census All Whites Nonwhites (2) (3) (1) 50.7 50.7 50.5 (1.6) (1.6) (1.6) 86.7 13.3 100.0 0.0 0.0 100.0 0.1 5.2 39.3 29.7 15.9 9.9 4.7 39.1 29.1 16.6 10.5 8.3 41.0 33.4 11.3 6.0 85.0 76.5 82.2 83.1 76.3 52,512 (76,853) 61,053 (80,366) 45,530 54,570 (80,899) 62,898 (84,600) 47,123 36,376 (24,652) 46,566 (27,209) 33,879 58,488 (65,224) 71,158 (65,378) 48,941 60,909 (67,114) 73,303 (67,174) 49,220 42,703 (48,337) 55,935 (48,170) 36,356 1,286,921 1,141,340 145,581 203,781 178,601 25,180 Notes: This table reports summary statistics for veterans in the SSA verified sample and veterans in the 2000 census. Panel A maintains the same sample restrictions described in the data appendix and further restricts to veterans who were still alive as of 1999. All statistics in panel A are from 1999. Panel B reports summay statistics for veterans in the 5 percent 2000 Census IPUMS extract who were born between 1946 and 1951. All earnings are in $2013. 120 Appendix Table 2.5. DC Progression and LFP for Pre-Existing Beneficiaries and New Beneficiaries B. LFP (Cell Level) A. CDR (Individual Level) Not Enrolled Enrolled in in DC by All Veterans DC by 1998 1998 (1) (2) (3) High DC_98 Low DC_98 All Veterans Cell Cell (1) (2) (3) 2.87*** 2.55*** -0.01 -0.66*** -0.71*** 0.21*** (0.03) 0.34*** (0.01) 0.66*** (0.02) (0.18) 0.84*** (0.08) 0.26** (0.12) (0.02) 0.22*** (0.01) 0.73*** (0.01) (0.06) -0.15*** (0.03) -0.18*** (0.04) (0.27) -0.35*** (0.13) 0.06 (0.17) (0.07) -0.11*** (0.03) -0.21*** (0.04) 3.96 38.21 0.00 84.68 77.23 85.58 # OBS 11,376,787 1,165,699 10,211,088 2,372,076 244,709 2,127,367 Years 1998-2006 1998-2006 1998-2006 1998-2007 1998-2007 1998-2007 BOG BOG * (YR-1998) BOG * (YR-2001)* (YR>2002) Outcome Mean (1998) Notes: This table reports estimates of equation (2) where the dependent variable is a veteran's Combined Disability Rating (panel A), or the percentage of the veteran's cell with positive annual earnings (panel B). In panel A, column 1 includes all veterans in the sample, column 2 only includes veterans who initially enrolled in DC in 1998 or earlier, and column 3 only includes vetearns who were not enrolled in DC as of 1998. Veterans with an IU award are given a CDR of 100. In panel B, column 1 includes all cells, column 2 only includes cells where at least one third of veterans in the cell were enrolled in DC as of 1998, and column 3 includes cells where fewer than one third of veterans in the cell were enrolled in DC as of 1998. All regressions in panel A are at the individual level while regressions in panel B are at the cell level and are weighted by the number of veterans in a cell in a particular year. All regressions include year fixed effects and fixed effects for each combationation of (AFQT-quintile)x(year), (year of birth)x(year), and (race)x(year), where race is defined as white or nonwhite and veterans with a missing AFQT score * and * denote are grouped into a sixth AFQT category. Standard errors are reported in parentheses. * significance at the 1%, 5%, and 10% level, respectively. 121 Chapter 3. The Impact of Voluntary Military Service on Mortality: Evidence from Active Duty U.S. Army Applicants Kyle Greenberg July 2015 Abstract Since the advent of the all-volunteer force in 1973, military operations in Afghanistan and Iraq have produced more U.S. casualties than all other conflicts combined. This chapter exploits enlistment test score cutoffs in a fuzzy regression-discontinuity (RD) design to evaluate causal effects of military service for individuals who applied to the active duty U.S. Army from 1996 through 2007. Fuzzy RD results suggest that military service does not increase mortality, but I cannot rule out positive effects. Ordinary Least Squares (OLS) estimates that compare applicants who did not enlist to soldiers with similar characteristics provide additional evidence that military service does not increase mortality. OLS estimates also indicate that a soldier's first year in the Army is associated with substantial reductions in mortality relative to nonveteran applicants. * The views and findings expressed herein are those of the author and do not purport to reflect the position of the U.S. Military Academy, the Department of the Army, or the Department of Defense. I am grateful to Josh Angrist, David Autor, Alex Bartik, Brendan Price, Manasi Deshpandi, and seminar participants at M.I.T. and the United States Military Academy. I would also like to thank Luke Gallagher of the Army Office of Economic Manpower Analysis and Marisa Michaels of the Defense Manpower Data Center for outstanding research assistance and data collection. 122 3.1 Introduction With over 5,300 soldiers killed as a result of hostile action in Afghanistan and Iraq, military operations from September 2001 to date have marked the most lethal period of sustained conflict since the beginning of the all-volunteer force (U.S. Department of Defense, 2015). In addition to the immediate dangers of combat, political and military leaders widely interpret soldier suicides, which doubled between 2004 and 2009, as a symptom of the high operational tempo associated with recent conflicts (U.S. Army, 2010). Fear of respiratory complications from exposure to bum pits and concerns over the impact of improvised explosive devices on traumatic brain injury add to the public debate surrounding the long term health consequences of military service (Institute of Medicine, 2011, and CDC, NIH, DoD, and VA Leadership Panel, 2013) Attempts to quantify the effect of recent military service on mortality typically rely on comparisons of veterans to nonveterans. Using mortality data through 2005, Kang and Bullman (2009) find that veterans of Operation Enduring Freedom (OEF) and Operation Iraqi Freedom (OIF) are at significantly lower risk for all-cause mortality, but at slightly higher risk for suicide, than nonveterans after adjusting for age, race, and gender. A more recent study using mortality data through 2011 finds that OEF/OIF veterans are at greater risk of all-cause mortality than nonveterans, even after excluding combat deaths (Bollinger, Schmidt, Pugh, Parsons, Copeland, and Pugh, 2015). These studies suffer from two major drawbacks. First, they do not capture the entire association between military service and mortality because they ignore deaths that occur during combat. Second, their findings are difficult to interpret because comparisons of veterans to nonveterans suffer from selection bias, the direction of which is not immediately clear. On one hand, many potential recruits do not meet the military's strict medical standards, a fact that potentially explains why WWII veterans typically live longer than nonveterans of the same age (Seltzer and Jablon, 1974). On the other hand, veterans who volunteered for military service might be less risk-averse than nonveterans. This chapter uses plausibly exogenous variation in applicant test scores to avoid the selection bias inherent in comparisons of veterans to nonveterans. The U.S. Army requires most applicants to earn a 31 123 on their Armed Forces Qualification Test (AFQT) before enlisting. Moreover, many enlistment bonuses are only available to applicants with an AFQT of 50 or higher. Using these cutoffs as instruments for military service in a novel fuzzy regression discontinuity (RD) design, I find that military service potentially reduces all-cause mortality for U.S. Army soldiers who applied between 1996 and 2007.80 However, I cannot rule out positive effects. To increase precision, I construct Ordinary Least Squares (OLS) estimates of the effects of military service while controlling for applicant characteristics that predict enlistment (age, race, test scores, education, state of birth), similar to the identification strategy employed by Angrist (1998). OLS estimates indicate that veterans are 0.24 percentage points less likely to die (as of February 2010) than similar nonveteran applicants, a difference equal to 18 percent of the incidence of mortality among nonveteran applicants. Importantly, OLS estimates are similar to fuzzy RD estimates and hardly change with the inclusion of controls that indicate if an applicant tested positive for drugs or had at least one medically disqualifying condition. I also find that soldier's are substantially less likely to die, relative to similar nonveteran applicants, within a year of applying to join the military. Among potential recruits who applied prior to September 2001, the initial mortality gap between soldiers and nonveteran applicants widened (i.e. grew more negative) before leveling off around 4 years after applying. However, among potential recruits who applied between October 2001 and September 2004, the mortality gap narrowed before leveling off 4 to 5 years after applying. For soldiers in this latter group, the relative period of danger during their second, third, and fourth years of service coincided with the most lethal years of operations in Iraq. Still, the net effect of military service on mortality for the latter group of applicants is near zero, suggesting that military service does not substantially increase mortality even during times of relative conflict. I also estimate the effect of military service on receipt of Disability Compensation (DC) from the Department of Veterans Affairs (VA). Since the purpose of DC is to compensate veterans for injuries sustained or aggravated during military service, DC receipt should be strongly correlated with veteran As far as I know, this is the first attempt to use such an RD design to identify causal effects of military service. Unfortunately, the Social Security Administration's Public Death Master File (DMF), my source for mortality data, does not identify specific causes of death. 80 124 health. DC also helps verify the stability and precision of two stage least squares (2SLS) estimates for binary outcomes that have more variation than mortality. Fuzzy RD estimates imply that military service increases DC receipt by 24 to 28 percentage points. This is probably a close approximation of the effect of military service on receipt of any disability benefits, including DC, Social Security Disability Insurance (SSDI), and Supplemental Security Income (SSI), because only 3 percent of adults under the age of 40 receive SSDI or SSI.8 1 One potential explanation for the large effect of military service on DC receipt is that military service increases morbidity but not mortality. At the same time, factors other than health, such as the relative attractiveness of DC when job opportunities are scarce, could also contribute to high DC enrollment rates, a possibility discussed in Chapter 1 as well as in Angrist et. al (2010) and McNally and Freuh (2013). My results are broadly consistent with several studies that exploit Vietnam conscription lotteries to identify causal effects of military service prior to the advent of the all volunteer force. Although Hearst, Newman, and Hulley (1986) find an increase in deaths from suicides and motor vehicle accidents among draft-eligible men, more recent research suggests that any elevated death risks among draft eligible men, to include cause-specific death risks, had largely dissipated by 2002 (Conley and Heerwig, 2012). Angrist, Chen, and Frandsen (2010) find a causal effect of Vietnam-era military service on receipt of any Federal transfer income, particularly DC. Evidence from Australia's Vietnam draft lottery also indicates that military service substantially increases disability receipt, but not post-service mortality, among Australian conscripts (Siminski and Ville, 2011, and Siminski, 2013). The analysis proceeds as follows. Section 1 describes how potential recruits apply to enlist in the U.S. Army. Section 2 describes my data and discusses descriptive statistics. Section 3 explains my empirical strategy, a fuzzy RD design that exploits U.S. Army testing thresholds, and presents first stage results. Section 4 discusses 2SLS estimates and compares them to OLS estimates. Section 5 investigates how the effect of military service on mortality changes over time and Section 6 concludes. am currently in the process of acquiring SSDI, SSI, earnings, and employment outcomes from the SSA. In 2013, only 3.3 percent of nonveteran men between the ages of 25 and 39 received SSDI or SSI, compared to 3.1 percent for veteran men of the same age. (American Community Survey (ACS), 2013). 81 1 125 3.2 Institutional Background All persons interested in enlisting in the U.S. Army must first visit their local U.S. Army recruiting office. After determining that a potential recruit meets basic age, citizenship, and background requirements, the recruiter will schedule a two day appointment for the applicant at one of 65 Military Entrance and Processing Stations (MEPS). All applicants take the Armed Services Vocational Aptitude Battery (ASVAB) during their first day at the MEPS while the second day consists predominately of physical and medical examinations. Four of the 11 tests within the ASVAB contribute to an applicant's raw Armed Forces Qualification Test (AFQT) score, which is then converted to a scaled AFQT score that represents the percentile-rank (1-99) of an applicant's arithmetic and verbal reasoning skills relative to a nationally representative sample of 18-23 year olds. (U.S. Department of Defense, 2004).82 The AFQT has an important role in ensuring the quality of recruits for the All Volunteer Force. Law prohibits applicants with AFQT scores below 10 and non-high school graduates with AFQT scores below 31 from enlisting in any branch of the military. The Department of Defense further requires that 90 percent of non-prior service (NPS) recruits have high school diplomas, that at least 60 percent of recruits have AFQT scores of 50 or higher, and that no more than 4 percent of recruits have AFQT scores below 31 (U.S. Department of Defense 2004, p. 23). To meet DoD requirements, the U.S. Army rarely accepts applicants with AFQT scores below 31 and generally only offers enlistment bonuses to applicants with AFQT scores of 50 or higher (U.S. Army Recruiting Command, 2012). Applicants with low AFQT scores may retake the ASVAB one month after their initial examination and can retest again one month after the initial retest. Applicants who wish to retest a third time must wait an additional six months before doing so. Prior to July 2004, the DoD scaled AFQT scores against young adults selected to participate in the 1980 Profile of American Youth (PAY). After July 1"t, 2004, the DoD scaled AFQT scores according to 82 Unless otherwise indicated, all subsequent references to AFQT scores refer to scaled AFQT scores. 126 the 1997 PAY.83 The 1997 norms produce lower AFQT scaled scores than the 1980 norms did. For example, a scaled AFQT score of 31 under the 1980 PAY scale is equivalent to a scaled AFQT score of 26 under the 1997 PAY scale.84 Even though fewer applicants met enlistment test score qualifications after the July 2004 re-norm, the U.S. Army did not adjust AFQT qualification requirements. 3.3 Data and Descriptive Statistics My data combine administrative information for military enlistees with mortality information from the Social Security Administration (SSA) and DC information from the VA. Military enlistment information comes from an extract of the U.S. Military Entrance and Processing Command (USMEPCOM) Examination and Accessions File, which is warehoused at the Defense Manpower Data Center (DMDC). The extract contains information on all persons who applied to join the Active Duty Army between fiscal years 1996 and 2007. I further limit this sample to applicants with no prior military service and who were between the ages of 17 and 34 at the time of their application. 85 The extract includes demographic information, each applicant's three most recent AFQT scores, each veteran's MEPS site, and each veteran's state of residence at the time of his application. 6 The extract also indicates if an applicant tested positive for drugs and, for applicants who took their medical exam after August 2001, whether the individual had any medically disqualifying conditions. For applicants who enlisted, the extract includes which service the applicant entered (e.g. Active Duty Army, Army National Guard, Army Reserves, Active Duty Marine Corps, etc.) and the applicant's date of enlistment. The United States Military Academy's Office of Economic and Manpower Analysis (OEMA) used The 1980 PAY participants were members of the 1979 National Longitudinal Survey of Youth (NLSY) and 1997 PAY participants were members of the 1997 NLSY. 84 See Segall (2004) for a table that translates scaled AFQT scores under the 1997 PAY to scaled AFQT scores under the 1980 PAY (Appendix E). 85 I also exclude approximately 6 percent of applicants who took their ASVAB in high school as part of the ASVAB Career Exploration Program because these applicants observe their AFQT scores before applying. This generates a first order selection problem as most of these applicants only apply if their AFQT exceeds 31. 86 Less than 10 percent of applicants have no reported MEPS site. These individuals most likely took the ASVAB at a Mobile Examination Test (MET) site, a testing option for applicants who do not reside near a MEPS. 83 127 social security numbers to merge military applicant data to VA disability records, which consist of a single cross-section of veterans enrolled in the DC program as of January 2014. VA records include each beneficiary's Combined Disability Rating (CDR) and monthly disability payment. Unfortunately, VA records do not indicate whether a veteran applied for an award but was ultimately denied. They also do not indicate when a veteran first started to receive compensation. OEMA also used social security numbers to merge military applicant data to mortality information found in the SSAs publically available Death Master File (DMF). The DMF indicates the month of death for individuals reported to the SSA as deceased through February 2015. In November 2011, the SSA decided that it could no longer include protected state death records in the public DMF. As a result, the organization removed 4.2 million death records from the public file and reduced the number of annual additions to the DMF by 1 million, or approximately 40 percent of annual deaths in the U.S. (National Technical Information Service, 2011). Appendix Table 3.1, which reports the total number of deaths recorded in the DMF for each month between January 2009 and December 2011, reveals a substantial decline in death records between January and April 2010.87 The table also indicates that reported death rates in 2011 were just over half of the death rates reported in 2009. I therefore predominately rely on DMF records from February 2010 or earlier to measure mortality in my analysis below. Another potential drawback of the DMF is that it tends to under-report mortality for younger individuals. Using the National Death Index (NDI) as the standard for mortality reporting, Hill and Rosenwaike (2001) estimate that the DMF correctly reports 73 percent of deaths among persons aged 25 to 54 who died between 1994 and 1997.88 To investigate the accuracy of DMF data within my sample, I used 2010 life tables from the Center for Disease Control to predict the percentage of applicants expected to be deceased by 2010 based on gender, year of birth, and race. Appendix Table 3.2 compares these predicted mortality percentages, reported in column 1, to actual mortality percentages for all applicants, 87 I observe a similar decline in death records within my sample of U.S. Army applicants. I am currently in the process of merging my data with the National Death Index (NDI), which contains more accurate records on all-cause mortality and includes information on cause of death (Hill and Rosenwaike, 2001). 88 128 applicants who enlisted in the military (veterans), and applicants who did not enlist in the military (nonveterans).8 9 Among white men and white women, both veterans and nonveterans have a slightly higher incidence of reported mortality than predicted mortality, potentially indicating that Army applicants are more prone to risk than the general population. However, among male and female black applicants, predicted mortality percentages are nearly twice as large as actual mortality percentages for both veterans and nonveterans. The DMF might under-report mortality for low-income households; I discuss how this could bias my results in more detail below. Appendix Table 3.2 also reveals a very low incidence of mortality among female applicants. Considering this, I restrict my sample to male applicants in the analysis that follows. Panel A of Table 3.1 reports summary statistics for male applicants in my sample. Column 1 reports statistics for all applicants while column 2 and column 3 report statistics on nonveterans (applicants with no military service) and veterans (applicants with any military service), respectively. Veterans are less likely to be minorities and are more likely to have a high school diploma than nonveterans. On average, veterans also have substantially higher AFQT scores than nonveterans (59 relative to 47), reflecting the Army's goal to recruit high quality soldiers. Notably, 26 percent of applicants were still enrolled in high school at the time of their application.90 Nearly half of all applicants ultimately enlisted in the military. Of those who enlisted, 80 percent joined the Active Duty Army, 9 percent joined the Reserve Component Army (the Army Reserve or the Army National Guard), and another 11 percent enlisted in a separate military service. Slightly more than 1 percent of applicants were deceased by February 2010 and the share deceased is slightly higher among veterans relative to nonveterans (1.25 percent relative to 1.10 percent). Column 3 also indicates that 23 percent of veterans received DC in January 2014. Of note, active duty service members may not receive My definition of "veteran" includes applicants currently serving in the military. This differs from the traditional definition, which indicates someone who previously served in the military but is not currently serving. 89 90 These applicants took their ASVAB at a MEPS site, not as part of the ASVAB Career Exploration Program discussed above. The vast majority of soldiers who apply while they are students complete high school before enlisting. 129 DC until they leave the service. Thus, the percentage of DC beneficiaries will likely grow over time as more soldiers within my sample leave the service and as veteran health diminishes with age. 3.4 Empirical Strategy The plots in Figure 3.1 depict the first stage that underlies my fuzzy RD identification strategy. The left plot shows enlistment rates as a function of AFQT scores for applicants who applied prior to the July 2004 re-norm (pre-July 2004 applicants) and the right plot shows the same relationship for post-July 2004 applicants. As discussed above, July 2004 was the first month that USMEPCOM scaled AFQT scores according to a new PAY sample. Both plots reveal a discontinuous jump in the probability of enlistment when applicants achieve an AFQT score of 31 and again when applicants achieve an AFQT score of 50. The jump at 50 is considerably smaller for post-July 2004 applicants than the same jump for pre-July 2004 applicants, but this is probably due to the high rate of enlistment bonuses offered to applicants with AFQT scores below 50 as the Army grew to respond to operations in Iraq. For the purpose of clarity, I henceforth refer to an initial AFQT score of at least 31 as an "enlistment offer" and an AFQT score of at least 50 as a "bonus offer." In reality, the Army extends enlistment offers to many applicants with AFQT scores below 31. Some of these offers occur when applicants earn higher AFQT scores after retesting, but others are made to applicants with final AFQT scores below 31 but who have otherwise strong applications. Similarly, the Army can offer enlistment bonuses to some applicants with AFQT scores below 50, particularly if demand for new recruits is high. 91 Nevertheless, the discontinuities in enlistment rates around both thresholds (Figure 3.1) suggest that enlistment offers and bonus offers, as just defined, can serve as instruments for military service. A key threat to the validity of any RD strategy is the possibility of precise manipulation of the running variable around the threshold, as discussed in McCrary (2008) and Frandsen (2014). While direct manipulation of exam scores seems unlikely (most exams are computerized adaptive tests), the ability to Twenty percent of active duty Army enlistees with final AFQT scores below 50 received some sort of enlistment bonus. The Army awarded many of these bonuses in fiscal years 2005 and 2006 during the buildup to the surge in Iraq. By contrast, 65 percent of active duty Army enlistees with a final AFQT score above 50 received a bonus. 91 130 retest until an applicant receives an enlistment offer or a bonus offer is potentially problematic. Ideally I would use an applicant's initial AFQT score to avoid this problem, but using the first on file (my data only contains each applicant's three most recent AFQT scores) is probably sufficient for two reasons. First, among applicants whose first AFQT score on file is below 31, 27 percent retested at least once and only 5 percent retested at least twice, suggesting that few retested more than twice. Second, precise manipulation of AFQT scores around either threshold can only occur if an applicant stops retaking the exam after he achieves his desired score. For example, consider an applicant who tested a total of four times. In all likelihood, this applicant did not achieve his desired score on his first exam or his second exam. If this is the case, then using either exam score as his running variable in an RD design would not show evidence of manipulation around a threshold. To visually inspect for manipulation of the running variable around either threshold, Figure 3.2 displays histograms of AFQT scores for pre-July 2004 applicants (left plot) and post-July 2004 applicants (right plot). Both histograms reveal extreme bunching at certain AFQT scores. A comparison of the preJuly 2004 histogram to Table A-3 of Mayberry and Hiatt (1992), which displays the conversion of AFQT raw scores to AFQT scale scores under the 1980 PAY scale, indicates that all of the bunching occurs where two raw AFQT scores equate to a single scaled AFQT score. Likewise, since no raw scores equated to a scaled score of 60 prior to the re-norm, no pre-July 2004 applicants have a scaled AFQT score of 60. The histogram for post-July 2004 applicants also displays extreme bunching at some AFQT scores, but not at the same scores where bunching occurs for pre-July 2004 applicants. This is because the PAY97 scaling adjusted the AFQT scale scores that equate to multiple raw scores (Segall, 2004; Table 2.5). Importantly, there does not appear to be any bunching at AFQT scores adjacent to the thresholds of 31 and 50 in either histogram, arguing against the possibility that applicants can precisely manipulate 131 AFQT scores. 92 I use the following reduced form parametric estimating equation to obtain 2SLS estimates of the effects of military service on desired outcomes: (3.1) Yi = a + pZi,3 1 + 6Zi,so + fo(Ai) + Zi. 31 x f(Ai) + Zi.50 x f2 (Ai) + X$fl + Ili Yi is the outcome variable for applicant i, Ai is i's first AFQT score on file, Zi,3 1 indicates an enlistment offer and Zi,50 indicates a bonus offer-that is, Zi,31 = 1(A ! 31) and Zi,50 = 1(Ai 50)-and Xi is a vector of 12 dummy variables indicating the fiscal year that i applied to join the U.S. Army and potentially other individual covariates. I group applicants who took their first AFQT on file between July and September 2004 into fiscal year 2005 to ensure that equation (3.1) only compares the effects of an enlistment offer or a bonus offer for applicants who took the exam under the same scaled norms. I control for the effects of the running variable, Ai, with three second-order polynomials: one for applicants with AFQT scores below 31, another for applicants with AFQT scores in the interval [31, 49], and a third for applicants with AFQT scores above 49. Specifically, (3.2) fj (Ai) = OpjAi + fjA?; j = 0, 1, 2 The results that follow also include estimate equation (3.1) using third and fourth order polynomial controls of the running variable. I also report nonparametric estimates of equation (3.1), which only include linear controls for the running variable but weight observations according to the following variant of a tent-shaped edge kernel: Unfortunately, my data does not contain raw AFQT scores, which would provide a better running variable than scaled AFQT scores. The discrete nature of AFQT scores is also not amenable to the empirical test for manipulation of the running variable suggested by McCrary (2008), which assumes a continuous running variable. Frandsen (2014) devises a test that is similar to McCrary's, but allows for a discrete running variable. However, this test assumes that the running variable is a discretized version of a continuous variable with equally spaced support bins. This is clearly not the case for scaled AFQT scores. 92 132 Kh(AL) =.11 (IA ~1 ~ h 105 1 hi i) <( - 49. 1 )x - 0h if A- 9 40 A hi)if Ag > 40 where h is the bandwidth. I begin with a default bandwidth of 15 before investigating alternative bandwidths for nonparametric estimates. A bandwidth of 15 implies that the sample only contains applicants with AFQT scores in the interval [16 , 64]. Panel B of Table 3.1 reports summary statistics for applicants within this discontinuity window. 93 Appendix Figures 3.1a - 3.1h suggest that applicant characteristics do not vary discontinuously around the score thresholds of 31 and 50, providing further evidence that applicants cannot precisely manipulate their AFQT scores. 94 To complement these figures, Table 3.2 reports estimates of equation (3.1) where the dependent variables are the baseline covariates. The results reveal little evidence of systematic sorting around either threshold. Of note, nonparametric estimates, reported in panel C, suggest statistically significant discontinuities in the probability of high school graduates around both thresholds. However, Appendix Figures 3. ig and 3.1h both reveal a non-linear relationship between AFQT scores and high school graduates within the [16 , 64] interval. Thus, the significant estimates reported in panel C are likely due to misspecification rather than systematic sorting around either threshold. 95 Panel A of Table 3.3 reports first stage estimates of the effect of enlistment offers and bonus offers on military service using AFQT score as the running variable. Columns 1 and 2 report parametric estimates of equation (3.1) with second and third order controls of the running variable, respectively. Column 3 reports nonparametric estimates of the same equation. The point estimate of 0.1015 in column 1 of panel A indicates that an enlistment offer increases the probability of enlistment by 10.15 percentage points. When I examine each cutoff separately, I obtain optimal bandwidths of 1 or smaller for both mortality and DC outcomes using the technique described by Imbens and Kalyanaraman (2012) (IK). These bandwidths are too small for precise inference and are probably the result of extreme bunching at AFQT scaled scores that equate to multiple raw scores. 94 This figure does not include plots of all covariates for post-July 2004 applicants. The excluded plots do not reveal evidence of sorting around either threshold and are available from the author upon request. 93 95 It is also worth noting that an applicant's education status is determined at the time of his most recent ASVAB, not at the time of her first ASVAB on file. Applicants who retake the ASVAB are therefore mechanically more likely to be high school graduates. 133 Results from the same column also indicate that a bonus offer increases the probability of enlistment by an additional 5.67 percentage points. Most of the estimates in Panel A indicate that enlistment offers increase enlistment rates by 10 to 11 percentage points while bonus offers increase enlistment rates by an additional 5 to 6 percentage points. Using any military service as the only endogenous variable in a 2SLS model is potentially problematic because 1) the AFQT threshold of 31 applies to the active Army and the reserve Army, but not to other services, and 2) only the active duty Army extends bonus offers to applicants with AFQT scores of 50 or higher. Subsequent panels of Table 3.3 therefore investigate whether enlistment offers and bonus offers affect enlistment in the active duty Army (panel B), enlistment in a reserve component of the Army (panel C), and enlistment in the Navy, Marines, Air Force, or Coast Guard (panel D). The results indicate that enlistment offers predominately induce applicants to enlist in the active Army, but a few join the reserve Army and some appear less likely to enlist in another military service (enlistment offers decrease the probability of joining a non-Army service by 0.78 percentage points). Likewise, bonus offers slightly decrease the probability of enlistment in the reserve Army or a non-Army service. While these discontinuities potentially muddle the interpretation of 2SLS estimates, their consequences are probably not substantial. The decrease in the probability of reserve service associated with the AFQT threshold of 50, the largest discontinuity associated with a service other than the active Army, is only about one-fifth of the magnitude of the increase in the probability of active Army service at the same threshold. As an additional check, the appendix to this chapter reports estimates from second stage models with two endogenous variables: a dummy for any active duty service and a dummy for any reserve duty service. 3.5 Results A. Effects of Military Service on Mortality and Veterans Disability Compensation Figure 3.3 plots the reduced form effect of enlistment offers and bonus offers on mortality for preJuly 2004 applicants (left plot) and post-July 2004 applicants (right plot). The post-July 2004 plot 134 suggests that enlistment offers are associated with a small, but imprecise, decrease in mortality rates, although there are no substantial jumps in mortality rates around other cutoffs. Panel A of Table 3.4 reports the corresponding 2SLS estimates of the effects of military service on mortality. The 2SLS specifications parallel the reduced-form estimates described by equation (3.1), except that now enlistment offers and bonus offers act as instruments for any military service. Consistent with the plots in Figure 3.3, the estimates in panel A reveal a negative but statistically insignificant effect of military service on mortality. The point estimate of -0.0068 in column 1 suggests that military service decreases mortality by 0.68 percentage points, which is substantial relative to mean mortality rates of 1.25 percentage points. However, the standard error of 0.0080 indicates that I cannot rule out positive effects that are 75 percent larger than mean mortality rates. Subsequent columns in panel A of Table 3.4, which report parametric 2SLS estimates with higher order polynomial controls for the running variable (columns 2 and 3) and nonparametric 2SLS estimates with varying bandwidth (columns 4-6), all indicate a negative but insignificant effect of military service on mortality. Panel A of Table 3.4 and the corresponding reduced form plots in Figure 3.4 indicate that military service increases DC receipt by 24 to 28 percentage points, and all of these estimates are significant. It is not surprising that military service increases DC receipt because only veterans qualify for DC. Still, the magnitudes of these estimates are substantial when compared to the relatively low incidence of federal disability among similarly-aged men. According to the 2013 ACS, 1.5 percent of nonveteran men between 25 and 39 years of age receive SSDI, relative to 2.1 percent of veteran men of the same age. For SSI, these percentages are 1.2 for veterans and 2.2 for nonveterans. A roughly 25 percentage point increase in DC receipt is also large relative to earlier wartime cohorts. By 2000, the causal effect.,of Vietnam-era military service on DC receipt was only 4 percentage points (Angrist et. al, 2010; Table 4). It is also worth noting that estimates for DC receipt are probably a lower bound because some soldiers in my sample were still in the military as of January 2014, and thus not eligible for DC. In an effort to increase precision, panel B of Table 3.4 reports 2SLS estimates using only a single control for the running variable. The reduced form equation is therefore: 135 (3.3) Yi = a + pZi,31 + SZj,5 0 + fo(Ai) + XU3l + m1i Estimates of equation (3.3) using a second order polynomial control for AFQT, reported in panel B (column 1), indicate a negative and marginally significant effect of military service on mortality. However, this significance disappears with 3 rd and 4 th order polynomial controls for the running variable, as seen in columns 2 and 3, respectively. Notably, the upper bound on the 95 percent confidence interval for the estimates reported in column 3 allows me to rule out an effect of enlistment on mortality that is 45 percent larger than mean mortality rates within my sample. 96 A closer inspection of both plots in Figure 3.2 reveals a substantial decrease in the first stage slope immediately to the right of the enlistment offer threshold and another, less substantial, slope change to the right of the bonus offer threshold. These slope changes offer the possibility of adding Regression Kink (RK) instruments to the RD identification strategy to increase power even further. Specifically, the interaction of Zj, 31 and a linear term for AFQT forms one instrument while the interaction of Zj,50 and a linear term for AFQT forms the other instrument. Nonparametric first stage results from a model that includes two RD instruments and two RK instruments, reported in column 3 of Appendix Table 3.3, reveal statistically significant slope changes around both cutoffs. First stage RK estimates under parametric specifications are substantially less precise than corresponding nonparametric estimates. However, even parametric estimates incorporating second order polynomial controls for AFQT indicate a significant slope change at the 31 threshold and a marginally significant slope change at the 50 threshold. Panel C of Table 3.4 displays 2SLS estimates of the effects of military service using RD and RK instruments (henceforth referred to as 2SLS RD-RK estimates). The nonparametric estimates reported in column 4 are substantially more precise than 2SLS estimates using only RD instruments (2SLS RD estimates): the standard errors for both mortality and disability outcomes decrease by nearly 60 percent with the inclusion of RK instruments. However, mortality estimates derived from a nonparametric 2SLS Although reported in panel B of Table 3.4, nonparametric estimates of (3.3) are likely to be biased because the underlying relationship between AFQT scores and military enlistment is clearly not linear, as seen in Figure 3.2. 96 136 RD-RK model with a bandwidth of 15 are the opposite direction of the corresponding 2SLS RD estimates. This result, reported in column 4 of panel C, indicates that military service increases the probability of mortality by 0.11 percentage points (approximately one-tenth the size of mean mortality rates), although this estimate is still indistinguishable from 0. Figure 3.3 suggests that the positive mortality estimate from the nonparametric 2SLS RD-RK model is predominately driven by low mortality rates among applicants with AFQT scores below 23. On one hand, this could reflect a positive effect of military service on mortality, a result that 2SLS RD estimates cannot rule out. On the other hand, this could also be the result of poor mortality data among low-scoring applicants. Several studies have documented a negative relationship between cognitive ability and mortality. 97 Yet within my data, both veterans and nonveterans with relatively low AFQT scores have substantially lower mortality rates than applicants with higher AFQT scores. This can be seen in Appendix Figure 3.2, which plots mortality rates against AFQT scores separately for veterans and nonveterans (to facilitate interpretation, I group AFQT score into bins of 3). The figure also indicates that veterans have lower mortality rates than nonveterans over nearly all AFQT scores. If the causal effect of military service on mortality is positive, then two conditions must hold: 1) soldiers must be more riskaverse than applicants who do not enlist 2) applicants with particularly low AFQT scores must be more risk-averse than applicants with higher AFQT scores. 98 I cannot rule out both of these possibilities, but there are other reasons to believe the positive estimate of military service on mortality is biased upwards. First, this estimate is not robust to different bandwidths. Nonparametric 2SLS RD-RK estimates derived from a bandwidth of 8 suggest that military service decreases mortality by 0.51 percentage points. Second, parametric 2SLS RD-RK estimates using second order polynomial controls for AFQT suggest that military service decreases mortality by 0.54 Jokela, Elovainio, Singh-Manoux, and Kivimaki (2009) document a negative association between AFQT score and early mortality among 1979 NLSY participants. See Pearce, Deary, Young, and Parker (2006), Batty, Deary, and Gottfredson (2007), and Batty, Shipley, Mortensen, Boyle et. al. (2008) for systematic reviews on the link between IQ and mortality. 98 By "risk-averse," I mean an applicant with a low probability of death absent military service. 97 137 percentage points. Thus, nonparametric 2SLS RD-RK estimates derived from a bandwidth of 15 probably reflect a non-linear relationship between AFQT and mortality rather than a discrete change in the slope of the mortality-AFQT relationship at the enlistment offer threshold. It is also worth mentioning that 2SLS RD-RK models estimating DC receipt produce very similar point estimates, but substantially smaller standard errors, than the corresponding 2SLS RD models. In particular, nonparametric 2SLS RD-RK estimates (column 4, panel C) indicate that military service increases DC receipt by 25.5 percentage points, which is near the 2SLS RD point estimate of 26.6 percentage points (column 4, panel A). Furthermore, the standard error for the 2SLS RD-RK estimate is small enough to detect an effect of 1.4 percentage points. This offers reassurances that the 2SLS RD-RK identification strategy has enough power to investigate the effect of military service on earnings, employment, education, and other outcomes. B. OLS and Logit Estimates To complement 2SLS estimates of the effects of military service on mortality and DC receipt, Table 3.5 reports OLS estimates of the following equation: (3.4) Yj = a + yDi + Xfl + ej, where Y is the outcome variable for applicant i, Di is and indicator for enlistment in the military, and Xi is a vector of controls that includes dummy variables for each application fiscal year, each AFQT score, year of birth, race (white or nonwhite), 4 education levels (still in high school, GED or high school dropout, high school diploma, some college or more), and state of birth. Similar to the identification strategy used in Angrist (1998), OLS estimates of (3.4) are unbiased if factors that influence outcomes are uncorrelated with military service after controlling for the individual characteristics captured in Xi. 99 99 Angrist (1998) includes a saturated set of controls for each combination of application year, AFQT category, education level, and year of birth to construct a covariate-specific comparison of veterans to nonveteran applicants. I do not estimate a saturated model because this leads to many combinations of the covariates where no applicants are deceased as of February 2010. These individuals would then be excluded from the logit estimates that follow. 138 Column 1 of Table 3.5 reports estimates of (3.4), but where Xi only includes indicator variables for fiscal year of application. The statistically significant point estimate of -0.0020 implies that mortality rates among veterans are 0.20 percentage points lower than mortality rates among nonveteran applicants after accounting only for an applicant's cohort. Estimates that include the full set of controls for Xi, reported in column 4, indicate that veterans are 0.24 percentage points less likely to be deceased as of February 2010 than similar nonveteran applicants, a point estimate that is 18 percent as large as the incidence of mortality among nonveterans. Notably, average marginal effects from logit estimates of equation (3.4), reported in column 5, are nearly identical to OLS estimates for both mortality and DC outcomes. A key advantage of OLS estimates is that they are substantially more precise than 2SLS RD and 2SLS RD-RK estimates. This is particularly important for outcomes that have little variation, such as mortality among young adults. To facilitate comparison, columns 6 and 7 of Table 3.5 reproduce parametric and nonparametric 2SLS RD-RK estimates. Importantly, OLS estimates are near the midpoint of parametric and nonparametric 2SLS RD-RK estimates, offering reassurances that any unobservable characteristics that influence selection into the military are probably not correlated with mortality.10 0 A key concern with this interpretation, however, is that the military selects only applicants who meet rigorous physical and health standards, which would cause OLS estimates of (3.4) to be biased 0 downwards (i.e. more negative than the true effect of military service on mortality).1 1 To investigate whether military health standards explain the negative OLS estimates of military service on mortality, Table 3.6 reports estimates of equation (3.4) with additional controls for drug test and medical exam results. Testing positive for drugs or having a medically disqualifying condition substantially reduces an applicant's chances, of enlisting in the military. Among applicants in my sample who tested positive for drugs-mostly marijuana-fewer than one-third enlisted, relative to nearly 79 '0 Note that even if this is true, OLS estimates might differ from 2SLS RD-RK estimates if applicants induced to enter the military as a result of an enlistment offer or a bonus offer are not representative of typical volunteers, as discussed in Angrist and Pischke (2009). 1 The means reported in Table 3.5 indicate that a few nonveterans (0.28 percent) receive DC. Most of these applicants are probably veterans who were incorrectly coded as not enlisting in the military. 139 percent of applicants who took a medical exam and did not test positive for drugs. Similarly, only 59 percent of applicants with at least one medically disqualifying condition enlisted. 02 The military only administers drug tests and medical exams to applicants who attend the second day of the military application process. 0 3 Column 1 of Table 3.6 therefore reports estimates of (3.4) using the full RD sample while column 2 reports estimates using only applicants with drug test information on file. The estimates reported in column 3 include an additional control that indicates whether an applicant tested positive for drugs. Notably, testing positive for drugs is a strong predictor of early-life mortality: applicants who tested positive for drugs are 0.35 percentage points more likely to be deceased than applicants who tested negative for drugs. Importantly, however, the estimated effect of military service on mortality remains negative and significant with the inclusion of a positive drug-test indicator (column 3), suggesting that drug-screening does not fully explain the negative association between military service and mortality. As mentioned above, a further limitation of my data is that it only contains information on medically disqualifying conditions, other than positive drug tests, for individuals who applied in September 2001 or later. Subsequent columns of Table 3.6 therefore report estimates where the sample is restricted individuals who applied after August 2001. The negative relationship between military service and mortality is substantially smaller in magnitude for individuals who applied after August 2001, a result I discuss in greater detail below. Still, the estimates reported in columns 7 and 8 suggest that the association between military service and mortality hardly changes with the inclusion of a medical failure indicator. Furthermore, the estimated effect of any medical failure on mortality is not statistically different from zero. The military may grant enlistment waivers to applicants with medically disqualifying conditions. Thus, the presence of a disqualifying condition does not necessarily preclude military service. 103 A disproportionate share of applicants who do not show for the medical exam have AFQT scores that qualify for neither an enlistment offer nor a bonus offer, as seen in Appendix Figure 3.3. As a result, restricting the sample to applicants with medical exam information would invalidate RD estimates because medical examinees appear to sort around both AFQT thresholds. It is also worth noting that about 10 percent of applicants who enlisted do not have a medical exam on file, thus indicating that medical exam data is not complete 102 140 These results reported in Table 3.6 weigh against the hypothesis that health screening standards fully explain the negative relationship between military service and mortality, at least within samples restricted to applicants with valid medical exam information. Of course, even if positive selection on health factors is not a major contributor to the negative association between military service and mortality, OLS estimates of equation (3.4) provide less than conclusive evidence of a causal effect because veterans may differ from nonveterans on unobservable dimensions. C. Estimates by Subgroup Table 3.7 explores OLS estimates of the effects of military service by race (white and non-white), application timeframe (pre-September 2001 applicants and post-September 2001 applicants), and AFQT score ranges. Since mortality outcomes to this point have defined mortality as being deceased as of February 2010, Table 3.7 includes additional estimates where mortality is defined as being deceased as of February 2015. A comparison of results for both of these outcomes generally indicates that the negative relationship between military service and mortality persists through 2015. However, results for mortality as of February 2015 should be interpreted with caution given the substantial reduction in Death Master File Records after 2010, as discussed above. The effects of military service on mortality and DC receipt appear similar for whites and nonwhites, as suggested by the estimates reported in columns 2 and 3. On the other hand, the negative relationship between military service and mortality is much stronger for pre-September 2001 applicants than the same relationship for post-September 2001 applicants. While differential selection into the military could partially explain these results, a more plausible explanation is that post-September 2001 applicants were more likely to deploy to combat zones in Iraq and Afghanistan than pre-September 2001 applicants. Consistent with mortality results by application timeframe, Table 3.7 also indicates that military service has a larger effect on DC receipt for later applicants than for earlier applicants. Columns 6 and 7 of Table 3.7 report estimates for two AFQT score ranges within the RD sample: applicants with AFQT scores in the interval [16, 42] and applicants with AFQT scores in the interval [43, 141 64]. To further investigate the possibility of heterogeneous treatment effects by AFQT score ranges, column 8 reports results for applicants with AFQT scores of 65 and higher. Notably, the effect of military service on mortality appears to be much closer to zero for applicants within the highest AFQT category. Thus, even if military service is protective for some individuals, it is not necessarily protective for applicants with relatively high AFQT scores. On the other hand, the negative point estimates in column 8 still argue against the hypothesis that military service increases mortality. Parametric and nonparametric 2SLS RD-RK estimates by race and application timeframe, as reported in Appendix Table 3.4, are broadly consistent with OLS estimates reported in Table 3.7. One exception to this is 2SLS estimates of the effect of military service on DC receipt for post-September 2001 applicants, which are noticeably larger than OLS estimates. 2SLS estimates indicate that military service increases DC receipt for post-September 2001 applicants by 31 to 35 percentage points, while OLS estimates only suggest a 26 percentage point increase. Some of the difference between 2SLS and OLS estimates for later applicants occurs because 2SLS "compliers"-applicants induced to enlist as a result of an enlistment offer or a bonus offer-exit the military faster than typical enlistees. Although not reported, OLS estimates of equation (3.4) where the outcome is a dummy variable that indicates if an applicant is still serving in the active duty Army (as of September 2013) produces a point estimate of 0.175, implying that 17.5 percent of post-September 2001 veterans are still in the active duty Army and therefore ineligible for DC benefits. This is significantly larger than the corresponding 2SLS estimates, which indicate that 12 to 13 percent of veteran 2SLS compliers are still in the active duty Army. 3.6 Mortality Trends over Time by Application Cohorts The results up to this point suggest that military service does not increase mortality, but reveal little about whether a veteran's time in the military or his post-service years are potentially protective or dangerous in nature. To offer some descriptive evidence on this matter, Table 3.8 reports OLS estimates of (3.4) where the outcomes indicate whether an applicant is deceased within a specified number of years since applying for the military. To account for differences in service experience and data availability over 142 time, the table reports results by four groups of application cohorts: FY1996-FY1998 applicants, FY1 999-FY2001 applicants, FY2002-FY2004 applicants, and FY2005-FY2007 applicants.1 04 The results for all four applicant cohorts suggest that soldiers experience substantial reductions in mortality, relative to similar nonveteran applicants, during their first year of military service.1 05 For the FY1996-FY1998 application cohort, military service is associated with a 0.12 percentage point reduction in mortality one year after applying, a difference that is equal to 80 percent of the one-year mortality rate for nonveterans in FY1996-FY1998 applicant cohort. Subsequent estimates for the FY1996-FY1998 cohort suggest that mortality among veterans declines by an additional 0.12 percentage points, relative to nonveterans, during a soldier's second and third years of service. Within 11 years of applying, mortality among veterans is 0.36 percentage points lower than mortality among nonenlistees. FY1999-FY2001 applicants (panel B) follow a similar trend to FY1996-FY1998 applicants. Within this application cohort, veterans witness a sharp drop in mortality, relative to nonveterans, during their first year of service, with further reductions during the second and third years of service. After three years, the overall difference in mortality between veterans and nonveterans from the FY1999-FY2001 applicant cohort remains relatively stable. Similar to other application cohorts, veterans from the FY2002-FY2004 cohort are substantially less likely to die than their nonveteran counterparts within a year of applying. However, unlike earlier cohorts, this trend reverses over the next three years. After four years of service, the net effect of military service on mortality is nearly zero. This period of relative danger associated with military service, which is unique to the FY2002-FY2004 cohort during their second through fourth years of military service (roughly 2003-2005 for FY2002 applicants, and 2005-2007 for FY2004 applicants), coincided with the most lethal years in Iraq (U.S. Department of Defense, 2015). Interestingly, veterans and nonveterans Differences in data availability occur because I am unable to determine whether relatively recent applicants are deceased several years after applying. For example, I cannot determine if an FY2006 applicant is deceased within ten years of applying because I only have valid mortality data through February 2010. 104 105 By years of service, I actually mean years since applying. While most service members enlist within a month or two of applying, some enlist several months later. 143 have similar mortality rates between four and five years after applying, suggesting that the relative danger associated with the military after the first year of service had leveled off by that time. This could be attributed to a decline in hostilities in Iraq around 2007 and to soldiers leaving the service after completing their initial four-year enlistment contract. The results from this subsection are purely descriptive in nature. Many factors could cause the effects of military service to vary by application cohort. Still, the results reported in Table 3.8 suggest that military service is protective during a soldier's first year on duty. Furthermore, the gap in mortality between veterans and nonveterans seems to persist, or even grow, during times of relative peace. On the other hand, the protective effects of military service appear to diminish during times of relative hostility. 3.7 Conclusion Any death of a soldier is a tragedy. Public concern over soldier deaths, whether by suicide, enemy fire, or other causes, is therefore warranted. But the results presented here suggest that military service does not increase mortality for individuals who applied to the active duty U.S. Army between 1996 and 2007. Although relatively imprecise, 2SLS estimates do not indicate that military service increases a soldier's likelihood of death. OLS estimates that compare nonveteran applicants to soldiers with similar characteristics (test scores, age, race, education, state of birth) provide additional evidence that military service does not increase mortality. My findings also suggest that soldier's experience substantial reductions in mortality during their first year in the Army. This is not terribly surprising considering that a soldier's first year of service is characterized by close supervision from drill instructors, first line supervisors, and other non-commissioned officers. Importantly, the gap in mortality between soldiers and nonveteran applicants diminishes during times of relative hostility. However, the net effect of military service on mortality appears to close to zero even for veteran cohorts that were most likely to deploy to a combat zone. Active duty Army soldiers and Marines bore the brunt of hostile casualties in Afghanistan and Iraq. Thus, the results presented here could potentially reflect an upper bound on the mortality consequences of 144 military service, at least for enlisted personnel. Of course, a negative relationship between military service and all-cause mortality could mask effects for specific causes of death, as suggested by (Conley and Heerwig, 2012). For example, military service could conceivably increase the probability of suicide while reducing the probability of all-cause mortality. My results also reveal little about the long-term effects of military service because I only have valid mortality data through early 2010. My ongoing efforts to merge military applicant data with the National Death Index, which includes all-cause mortality and causespecific mortality information through 2013, should resolve some of these questions. While I find little evidence that military service increases mortality, my results indicate that military service during the past two decades increases DC receipt by 24 to 28 percentage points. By comparison, Angrist et. al (2010) estimate that Vietnam-era military service increased DC receipt by only 4 percentage points nearly 30 years after the conclusion of the Vietnam War. The substantial, positive effect of recent military service on DC receipt complicates the interpretation of mortality results. One possible explanation is that military service reduces overall health and quality of life in manners that do not lead to early-life mortality. Another interpretation is that large gains in DC receipt could potentially reflect health problems not explicitly caused by military service. Alternatively, temporary reductions in earnings that typically occur when soldiers depart the service, as documented by Angrist (1990) and Angrist (1998), could induce some veterans to turn to DC to offset such reductions. The last possibility is particularly concerning considering recent evidence that DC receipt reduces labor force participation among Vietnam-era veterans (Autor, Duggan, Greenberg, and Lyle, 2015). Future research on the causal effects of military service during recent decades is warranted, and the fuzzy RD identification strategy presented in this chapter offers a new method for obtaining such estimates. I am currently working with the Social Security Administration (SSA) to merge military applicant data with employment, earnings, SSDI, and SSI outcomes. This and other research will hopefully shed more light on the complicated effects of military service during the post-9/l1 era. 145 3.8 Chapter 3 Appendix: Multiple Endogenous Variables Appendix Table 3.5 reports OLS and 2SLS estimates of models with two endogenous variables: an indicator for active duty military service and an indicator for reserve duty military service.' 06 The OLS estimates reported in panel A suggest that reserve duty is more protective than active duty service. Still the point estimate of -0.0020 is very similar to OLS estimate of the effect of any military service on mortality (-.0024, Table 5), reflecting how most individuals who apply to join the active duty Army ultimately enlist in that service. 2SLS estimates of the effects of active duty service on mortality are also similar to the corresponding estimates using only one endogenous variable, but estimates of the effects of reserve duty service are very imprecise. Panel B of Appendix Table 3.5 indicates that the effect of active duty service on DC receipt is stronger than the effect of reserve duty service on DC receipt. OLS estimates reported in column 1 suggest that the association between active duty service and DC receipt is twice as large as the association between reserve duty service and DC receipt. Nonparametric 2SLS RD-RK estimates of the effects of reserve duty service on DC receipt, reported in column 3, are precise enough to confirm that this association is likely a causal effect. Reserve duty military service implies enlisting in the Army National Guard, the Army Reserves, or similar reserve duty for the Air Force, Navy, or Marines. Applicants who enlist in the reserves and are called up for Active Duty service are still considered as performing reserve duty within my data. 106 146 Percentage who Enlist in Any Service - Pre July 2004 Renorm (D (N CO Percentage who Enlist in Any Service - Post July 2004 Renorm N- (0 0. 0 0 (N C6LO) CO N 0 6r- III 20 20 25 25 30 30 354I111511- 5] 35 40 60iIi 45 rt1111 First AFQT on File 50 -- 55 60 20 25 30 35 40 45 50 55 60 First AFQT on File Figure 3.1. This figure plots the percentage of applicants who enlist in any military service by AFQT score for applicants who applied before the July 2004 ASVAB renorm (left) and after the renorm (right). Distribution of Initial AFQT Score - Post July 2004 Renorm Distribution of Initial AFQT Score - Pre July 2004 Renorm N J- 0 0 C6 C6) CdJ 0 LO 0 0 -'- 20 25 30 35 40 45 50 First AFQT on File 55 60 -~- 20TTrTTTi i3 35m 4i1i41111111 11111111 20 25 30 35 40 45 50 First AFQT on File Figure 3.2. This figure plots the distribution of the first AFQT scores on file for individuals who applied before the July 2004 ASVAB renorm (left) and after the renorm (right). 55 60 Percentage Deceased - Post July 2004 Renorm Percentage Deceased - Pre July 2004 Renorm C0 0 4 0 * 0 00 .0 C~) * @0 0 0 0 4 0 0 0 .0 C~J 0 .0 * 0 0 * 0 0 .0 0 * 0 w 0 0 0 0 R OR0- I 3 401 45111 5011 551 601111111111111 20 25 30 l 35 40 45 First AFQT on File 50 55 60 20 251111 25 3 30 35 40 45 50 55 60 First AFQT on File Figure 3.3. This figure plots mortality percentages, as of February 2010, by AFQT score for pre-July 2004 applicants (left) and post-July 2004 applicants (right). Percentage Receiving DC - Post July 2004 Renorm Percentage Receiving DC - Pre July 2004 Renorm CO mV 0 0 I 0 0 ~w.. @0 0 0)_ * 0* * 0 0 C) e 0 M0 0 UO 0 0 6 I I I I IIII I I I I I I I I I I I I I I I I I I I I I I I I I5I I I I 5I I I I I45 20 25 30 35 40 45 First AFQT on File 50 55 60 20 25 30 35 40 45 50 55 60 First AFQT on File Figure 3.4. This figure plots DC receipt percentages, as of January 2014, by AFQT score for pre-July 2004 applicants (left) and post-July 2004 applicants (right). Table 3.1. Summary Statistics Age at application AFQT White A. Full Sample No Mil All Service Applicants (2) (1) 20.3 20.5 [3.7] [3.5] 46.7 53.0 [23.9] [25.4] 0.74 0.70 Any Mil Service (3) 20.1 [3.2] 58.7 [20.8] 0.77 B. RD Window All No Mil Any Mil Applicants Service Service (2) (1) (3) 20.1 20.4 19.8 [3.3] [3.6] [3.1] 41.7 38.4 44.9 [13.3] [13.8] [11.9] 0.70 0.68 0.73 Black 0.18 0.21 0.15 0.22 0.24 0.20 Enrolled in HS 0.26 0.27 0.25 0.26 0.27 0.26 HS Graduate 0.55 0.52 0.57 0.50 0.47 0.52 Single (no dependents) 0.90 0.91 0.89 0.90 0.90 0.89 No MEPS Indicated 0.09 0.09 0.09 0.09 0.09 0.09 Any Military Service 0.52 0 1 0.51 0 I Active Army Service 0.42 0 0.80 0.40 0 0.79 Reserve/Guard Army 0.05 0 0.09 0.05 0 0.10 Other Military Service 0.06 0 0.11 0.06 0 0.11 Deceased by FEB2010 0.0118 0.0125 0.0110 0.0125 0.0136 0.0115 0.12 0 0.23 0.12 0 0.24 1,101,703 529,947 571,756 668,149 330,233 337,916 DC Receipt by JAN2014 Observations Notes: This table reports descriptive statistics for men who applied to join the U.S. Army during fiscal years 1996 through 2007. Columns 1, 2, and 3 report descriptive statistics for all applicants, nonveterans, and veterans, respectively. AFQT scores indicate an applicant's first AFQT score on file (military records include AFQT scores from the three most recent exams). An applicant's age is determined on the day of his first AFQT on file while his education and marital status are determined on the day of his most recent exam. Other military service includes service in the Marines, the Air Force, the Navy, and the Coast Guard. Panel B reports descriptive statistics for applicants whose first AFQT scores on file are in the interval [16, 64]. Standard deviations are reported in brackets. 151 Baseline Characteristic Mean (1) Age White Black 20.09 [3.33] 0.70 0.22 Table 3.2. Covariate Balance Parametric Estimates 2nd Order Poly 3rd Order Poly 1 (AFQT 3 1) 1(AFQT 50) 1 (AFQT23 1) 1(AFQT250) (2) (3) (4) (5) -0.0141 (0.0304) -0.0026 (0.0044) 0.0015 (0.0041) 0.0004 (0.0040) 0.0006 (0.0046) -0.0284 (0.0284) -0.0005 (0.0037) 0.0030 Still In High School 0.26 High School Diploma 0.50 Single 0.90 0.0013 0.09 (0.0027) 0.0010 (0.0027) (0.0032) -0.0013 (0.0037) -0.0022 (0.0042) -0.0014 (0.0026) -0.0009 (0.0025) 0.996 668,149 0.606 668,149 No MEPS Indicated P Val (Joint x2 Test) Observations 668,149 -0.0410 (0.0424) -0.0063 (0.0061) 0.0017 (0.0057) -0.0036 (0.0056) -0.0031 (0.0064) 0.0010 (0.0038) 0.0000 (0.0037) -0.0026 (0.0384) -0.0064 0.569 668,149 0.402 668,149 (0.0050) 0.0080* (0.0043) -0.0046 (0.0050) 0.0058 (0.0057) -0.0022 (0.0035) -0.0044 (0.0033) Nonparametric Estimates BW= 15 1 (AFQT 3 1) 1 (AFQT 50) (6) (7) 0.0031 -0.0196 (0.0213) 0.0060* (0.0194) -0.0009 (0.0031) -0.0040 (0.0025) 0.0027 (0.0029) -0.0028 (0.0022) -0.0009 (0.0028) -0.0102*** (0.0025) 0.0076*** (0.0032) -0.0012 (0.0019) 0.0007 (0.0019) (0.0029) -0.0011 (0.0018) -0.0002 0.000 668,149 0.039 668,149 (0.0017) Notes: This table reports discontinuities in covariates estimated using models like the reduced-form specification described by equation (1). Columns 2, 4, and 6 report discontinuities at the AFQT = 31 threshold while columns 3, 5, and 7 report discontinuities at the AFQT = 50 threshold. Columns 2 and 3 report estimates from a parametric equation with three 2nd order polynomial controls for the running variable: one for AFQT scores less than 31, another for AFQT scores between 31 and 49, and another for AFQT scores greater than 49. Columns 4 and 5 report estimates from a parametric equation with three 3rd order polynomial controls for the running variable. Columns 6 and 7 report estimates from a nonparametric equation with three linear controls for the running variable. Nonparametric estimates weight observations according to the variant of the edge kernel described in the text using a bandwidth of 15. All estimates include dummy controls for fiscal year of application. Joint x 2 tests are for the null hypothesis that discontinuities in all regressions are equal to zero. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 152 Table 3.3. First Stage Estimates Parametric Estimates 3rd Order Poly 2nd Order Poly (2) (1) Instrument 1(AFQT 31) l(AFQT>50) A. Endogenous Variable: Any Military Service 0.1015*** 0.1080*** (0.0045) (0.0062) 0.0567*** 0.0509*** (0.0042) (0.0056) First Stage F-Statistic Observations 1(AFQT 31) 1(AFQT 50) 407.3 668,149 174.1 668,149 B. Endogenous Variable: Active Duty Army Service 0.1019*** 0.1027*** (0.0042) (0.0059) 0.0740*** 0.0674*** (0.0042) (0.0057) First Stage F-Statistic Observations 523.4 668,149 198.4 668,149 Nonparametric BW = 15 (3) 0.1049*** (0.0032) 0.0506*** (0.0029) 615.0 668,149 0.1027*** (0.0030) 0.0711*** (0.0029) 752.5 668,149 C. Endogenous Variable: Reserve Component Army Service I(AFQT 31) 0.0078*** 0.0089*** 0.0086*** (0.0021) (0.0030) (0.0015) 1(AFQT150) -0.0127*** -0.0131*** -0.0131*** (0.0019) (0.0025) (0.0013) First Stage F-Statistic Observations I(AFQT>31) 1 (AFQT>50) 26.1 668,149 20.3 668,149 D. Endogenous Variable: Other Military Service -0.0078*** -0.0033 (0.0021) (0.0029) -0.0039 -0.0049** (0.0028) (0.0020) First Stage F-Statistic Observations 11.6 668,149 1.5 668,149 85.4 668,149 -0.0060*** (0.0015) -0.0076*** (0.0014) 19.5 668,149 Notes: This table reports parametric and nonparametric.first stage estimates of the effects of enlistment offers (AFQT 3 1) and bonus offers (AFQT50) on enlistment in any military service (panel A), enlistment in the active duty Army (panel B), enlistment in a reserve component of the Army (panel C), and enlistment in a non-Army service (panel D). All estimates include dummy controls for fiscal year of application and three polynomial (parametric estimates) or linear (nonparametric estimates) controls for the running variable: one for AFQT scores less than 31, another for AFQT scores between 31 and 49, and another for AFQT scores greater than 49. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 153 Dependent Variable Deceased by FEB10 E[Y] DC Receipt by JAN14 E[Yj Deceased by FEB10 E[Yj DC Receipt by JAN14 E[Yj Deceased by FEB10 E[Y,] DC Receipt by JAN14 E[Yj Table 3.4. 2SLS Estimates Parametric Estimates 2nd Order Poly 3rd Order Poly 4th Order Poly (3) (1) (2) -0.0068 (0.0080) 0.0125 -0.0097 (0.0122) 0.0125 0.2814*** (0.0214) 0.1220 0.2417*** (0.0322) 0.1220 -0.0118* (0.0064) 0.0125 A. 2SLS RD Estimates -0.0083 (0.0143) 0.0125 0.2544*** (0.0380) 0.1220 Nonparametric Estimates (Edge Kernel) BW=15 BW=12 BW=8 (6) (5) (4) -0.0094 (0.0065) 0.0125 -0.0084 (0.0069) 0.0127 -0.0066 (0.0078) 0.0128 0.2664*** (0.0171) 0.1220 0.2656*** (0.0181) 0.1234 0.2607*** (0.0205) 0.1267 -0.0034 (0.0039) 0.0127 -0.0057 (0.0048) 0.0128 0.2503*** (0.0091) 0.1220 0.2531*** (0.0103) 0.1234 0.2562*** (0.0127) 0.1267 0.0011 (0.0026) 0.0125 -0.0013 (0.0031) 0.0127 -0.0051 (0.0042) 0.0128 0.2554*** (0.0071) 0.1220 0.2576*** (0.0085) 0.1234 0.2585*** (0.0114) 0.1267 B. 2SLS RD Estimates, Single Control for Running Variable -0.0080 -0.0062 -0.0020 (0.0057) (0.0058) (0.0034) 0.0125 0.0125 0.0125 0.2655*** (0.0166) 0.1220 0.2729*** (0.0154) 0.1220 -0.0054 (0.0053) 0.0125 -0.0123 (0.0114) 0.0125 0.2764*** (0.0148) 0.1220 0.2388*** (0.0314) 0.1220 0.2742*** (0.0161) 0.1220 C. 2SLS RD-RK Estimates -0.0123 (0.0126) 0.0125 0.2553*** (0.0349) 0.1220 Observations 668,149 668,149 668,149 668,149 599,218 473,276 Notes: This table reports 2SLS estimates of the effects of any military service on mortality and DC receipt where enlistment offers (AFQT>31) and bonus offers (AFQT>50) serve as instruments for any military service. Estimates in Panel A include three polynomial or linear controls for the running variable: one for AFQT scores less than 31, another for AFQT scores between 31 and 49, and another for AFQT scores greater than 49. Estimates in panel B include a single polynomial or linear control for the running variable. Estimates in panel C include two additional instruments: enlistment offers interacted with a linear term for the running variable and bonus offers interacted with a linear term for the running variable. All regressions include dummy controls for fiscal year of application. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 154 Table 3.5. OLS, Logit, and 2SLS Estimates 2SLS RD-RK Logit OLS Dependent Variable 2nd Order Poly (1) Deceased by FEB10 E[Y 1 ID,=0] DC Receipt by JAN14 E[Y, ID, =0] App FY Controls YOB Controls AFQT Controls Additional Controls Observations -0.0020*** (0.0003) 0.0136 0.2341*** (0.0007) 0.0031 (2) -0.0019*** (0.0003) 0.0136 0.2372*** (0.0007) 0.0031 (4) (3) -0.0025*** (0.0003) 0.0136 0.0136 0.2368*** (0.0008) 0.0031 X X X X X X 668,149 668,149 668,149 -0.0024*** (0.0003) (5) -0.0024*** (0.0003) 0.0136 0.2357*** (0.0008) 0.0031 0.2356*** (0.0008) 0.0031 X X X X 668,149 X X X X 668,149 (6) NP (BW=15) (7) -0.0054 (0.0053) 0.0011 (0.0026) 0.0136 0.0136 0.2770*** (0.0148) 0.0031 0.2553*** (0.0070) 0.0031 X X 668,149 668,149 * Notes: This table reports OLS, Logit, and 2SLS estimates of the effects of military service on mortality and DC receipt. Column I reports mean outcomes for persons who applied but did not join the military. Column 2 reports estimates from a regression of the dependent variable on an indicator for military enlistment while controlling for application fiscal year. Estimates in column 3 include additional indicators for each possible AFQT scores and estimates in columns 4 and 5 include additional controls for year of birth, race (white or nonwhite), education level (still in high school, high school dropout or GED, high school diploma, some college), and state of birth. Columns 6 and 7 report parametric and nonparametric 2SLS RD-RK estimates. 2SLS estimates include three polynomial controls (column 6) or linear controls (column 7) for the running variable: one for AFQT scores less than 31, another for AFQT scores between 31 and 49, and another for AFQT scores greater than 49. Robust standard errors are reported in parentheses. ***,**, and denote significance at the 1%, 5%, and 10% level, respecitvely. 155 Table 3.6. Mortality Estimates with Drug and Medical Test Controls Dependent Variable: Deceased as of FEB10 (1) Any Military Service Positive Drug Test Other Medical Failure -0.0024*** (0.0003) -0.0030*** (0.0004) (3) -0.0025*** (0.0005) 0.0036*** (0.0009) (4) -0.0009** (0.0004) (5) -0.0007 (0.0006) (6) -0.0003 (0.0006) 0.0028** (0.0013) (7) -0.0008 (0.0006) -0.0009 (0.0006) Drug Tested Applicants Post-AUGO I Applicants E[YID,=0] Observations (2) 0.0136 668,149 X X 0.0152 381,863 0.0152 381,863 0.0103 310,629 (8) -0.0005 (0.0006) 0.0027** (0.0013) -0.0008 (0.0006) X X X X 0.0108 179,232 0.0108 179,232 0.0108 179,227 0.0108 179,227 Notes: This table reports OLS estimates of the effects of military service on mortality as of February 2010 for applicants with AFQT scores in the interval [16, 64]. Column 1 reports estimates from the full RD sample while columns 2, 3, and 5 through 8 limit the RD sample to applicants with a valid drug test on file. Columns 4 through 8 further limit the sample to individuals who applied in September 2001 or later since medical exam information is not available for earlier applicants. "Other medical failure" indicates if an applicant failed any portion of his medical exam other than the drug test. All regressions include indicator variables for each application fiscal year, each possible AFQT score, year of birth, race (white or nonwhite), education level (still in high school, high school dropout or GED, high school diploma, some college), and state of birth. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 156 Dependent Variable Full RD Sample (1) Table 3.7. OLS Estimates by Race, Application Timeframe, and AFQT Score FY02-FY07 FY96-FY01 16 AFQT 42 Applicants Applicants Whites Non-whites (6) (5) (4) (3) (2) 435AFQT64 (7) AFQT>65 (8) Deceased by FEB10 E[Y1 1D1 =0] -0.0024*** (0.0003) 0.0136 -0.0015*** (0.0005) 0.0113 -0.0027*** (0.0004) 0.0148 -0.0037*** (0.0004) 0.0163 -0.0008** (0.0004) 0.0102 -0.0025*** (0.0004) 0.0131 -0.0023*** (0.0004) 0.0145 -0.0010*** (0.0004) 0.0114 Deceased by FEB15 E[Y,1D 1 =0] -0.0029*** (0.0003) 0.0177 -0.0024*** (0.0006) 0.0149 -0.0030*** (0.0004) 0.0189 -0.0042*** (0.0005) 0.0203 -0.0012*** (0.0004) 0.0143 -0.0029*** (0.0005) 0.0169 -0.0028*** (0.0005) 0.0187 -0.0007* (0.0004) 0.0144 DC Receipt by JAN14 0.2357*** (0.0008) 0.0031 0.2219*** (0.0014) 0.0033 0.2409*** (0.0009) 0.0030 0.2119*** (0.0010) 0.0047 0.2640*** (0.0011) 0.0010 0.2367*** (0.0012) 0.0021 0.2349*** (0.0010) 0.0046 0.2241*** (0.0009) 0.0055 668,149 197,748 470,401 363,319 304,830 340,006 328,143 373,707 Observations Notes: This table reports OLS estimates of the effects of military service on mortality and DC receipt. Column 1 reports results for the full RD sample, which includes applicants with AFQT scores in the interval [16, 64]. Columns 2 through 7 limit the RD sample to the subgroups identified in the column header. Column 8 reports estimates for applicants with AFQT scores in the interval [65, 99]. All regressions include indicator variables for each application fiscal year, each possible AFQT score, year of birth, race (white or nonwhite), education level (still in high school, high school dropout or GED, high school diploma, some college), and state of birth. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 157 Table 3.8. Effects of Military Service on Mortality Over Time by Application Cohort Years Since Application FY96 - FY98 Applicants E[Y, ID =0] OLS (1) (2) 1 0.0015 2 0.0029 3 0.0042 4 0.0054 5 0.0065 6 0.0076 7 0.0086 8 0.0100 9 0.0113 10 0.0127 11 0.0143 Observations -0.0012*** (0.0002) -0.0019*** (0.0002) -0.0024*** (0.0003) -0.0027*** (0.0003) -0.0028*** (0.0003) -0.0031*** (0.0004) -0.0030*** (0.0004) -0.0033*** (0.0004) -0.0035*** (0.0005) -0.0034*** (0.0005) -0.0036*** (0.0005) 183,380 FY99 - FY01 Applicants E[Y; ID =0] (3) 0.0019 0.0032 0.0046 0.0061 0.0072 0.0086 0.0104 0.0120 OLS (4) -0.0017*** (0.0002) -0.0024*** (0.0002) -0.0027*** (0.0003) -0.0027*** (0.0004) -0.0024*** (0.0004) -0.0022*** (0.0004) -0.0022*** (0.0005) -0.0024*** (0.0005) 179,939 FY02 - FY04 Applicants E[Y, ID =0] (1) 0.0019 0.0033 0.0048 0.0062 0.0079 OLS (2) -0.0015*** (0.0002) -0.0013*** (0.0003) -0.0007* (0.0004) -0.0002 (0.0004) -0.0002 (0.0005) 161,464 FY05 - FY07 Applicants E[Y, ID =0] (3) 0.0029 0.0052 OLS (4) -0.0018*** (0.0002) -0.0019*** (0.0004) 143,366 Notes: This table reports OLS estimates of the effects of military service on mortality within the number of years specified at left. Column I reports mean outcomes for persons who applied but did not join the military. Column 2 reports estimates from a regression of the dependent variable on an indicator for military enlistment while including indicator variables for each application fiscal year, each possible AFQT score, year of birth, race (white or nonwhite), education level (still in high school, high school dropout or GED, high school diploma, some college), and state of birth. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 158 a) Age at Application - Pre July 2004 Renorm b) Percentage White - Pre July 2004 Renorm CSJ 00 I S 48jj% 0 IW~0" *00' OWSO I,- CO 004* N- 0)- (O 0)i rU) I 2III0 20 11125111111111111111 I 35 40 45 50 55 60 25 30 35 40 45 50 N-i LO 4 55 se 60 20 25 30 35 40 50 45 55 60 First AFQT on File First AFQT on File c) Percentage Black - Pre July 2004 Renorm d) Percentage without Dependents - Pre July 2004 Renorm N.- Oi - m~ C. *e 0 0@ I *-- 0. (6 0) 0 - O- *0 0 (N 201111111111 I 20 25 0" I 1111111111111II I I II 25 I3 I3 4 4 5 30 35 40 45 First AFQT on File 50 55 60 0 20 25 30 35 40 45 50 0 55 60 First AFQT on File Appendix Figure 3.1. This figure plots covariate percentages by AFQT score for pre-July 2004 applicants. f) Percentage Still in High School - Pre July 2004 Renorm e) Percentage with no MEPS Indicated - Pre July 2004 Renorm .0 * 0 S 0 0 0 * a 0 0 *.0 I 0 0 0 0 S N 4 * 0 0 6 0* C') .0 S S S 0 * 0 0 0 0 0 0)6 0 0 6 0 = 35 30 25 20 40 45 60 55 50 First AFQT on File First AFQT on File g) Percentage High School Diploma - Pre July 2004 Renorm U) h) Percentage High School Diploma - Post July 2004 Renorm CO L6 LO - 0) 0. 0 .0 0 I I I I I I I I I I I I I I IN I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I 60 55 50 45 40 35 30 25 20 .000 0 00 C6 M6- UO * 0 0* 0 0. 0 00. C') LOj 0)O M~J 6 00 0 0 0 "b 0) 0 U)- *1....... L T111U111II1 1111111111111111111111111 _________________________ v~1 20 25 30 35 40 45 First AFQT on File 50 55 60 20 25 30 35 40 45 50 55 60 First AFQT on File Appendix Figure 3.1 (Continued). Plots (e) - (g) are for pre-July 2004 applicants. Plot (h) is for post-July 2004 applicants. Percentage Deceased - Post July 2004 Renorm Percentage Deceased - Pre July 2004 Renorm CO O LO) LO) 0000 0 0 0 0) 0 0 (0- ) (- 0 i 0 0 0 0 0 0 0 . 0 Cv)11111111111 20 25 11 ii 30 35 il il illill 40 45 ill 50 ill ill1 55 First AFQT on File * Nonveterans e Veterans 60 l I1 I I||I11111I 11 20 25 30 1 35 1 II 40 1I I1 ill |I1 45 50 55 i ii 60 First AFQT on File l Nonveterans * Veterans Appendix Figure 3.2. This fig ure plots mortality percentages, as of February 2010, for soldiers (red) and nonveteran applicants (blue) by AFQT score. Distribution of Initial AFQT Score - Pre July 2004 (with Med Exam) Distribution of Initial AFQT Score - Post July 2004 (with Med Exam) 0 0 0 0 0U, 04 0 0 0 0 0~ 0 LO 0-J1--'111 0-*'- 20 25 30 35 40 45 50 First AFQT on File 55 60 20 25 30 35 40 45 50 55 60 First AFQT on File Appendix Figure 3.3. This plots in this figure show the distribution of the first AFQT scores on file for applicant. with valid medical exam results. Appendix Table 3.1. Monthly SSA DMF Records Month January February March April May June July August September October November December January February March April May June July August September October November December January February March April May June July August September October November December Number of Deaths 220,803 198,570 213,890 200,708 198,417 188,936 192,675 190,134 187,656 205,111 199,449 215,643 218,633 184,776 124,982 111,554 113,166 107,505 115,191 110,469 99,617 108,992 109,174 117,647 120,953 109,699 114,022 106,075 105,134 100,188 101,406 102,071 99,918 106,300 105,547 115,348 Year 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2011 2011 2011 2011 2011 2011 2011 2011 2011 2011 2011 2011 Notes: This table indicates the number of deaths recorded in the Social Security Administration's Public Death Master File for each month between January 2009 and December 2011. 163 Appendix Table 3.2. Predicted vs. Documented Mortality Death Master File Merge Rates Predicted Mortality Nonveterans Veterans Percentage All Applicants (4) (3) (1) (2) 0.0104 A. White Men 0.0124 0.0135 0.0115 0.0187 B. Black Men 0.0101 0.0106 0.0094 0.0045 C. White Women 0.0049 0.0053 0.0043 0.0072 D. Black Women 0.0039 0.0042 0.0033 Notes: Column 1 of this table reports predicted mortality percentages, by race and gender, based on 2010 U.S. Life Tables, available at http://www.cdc.gov/nchs/data/nvsr/nvsr63/nvsr63_07.pdf. Columns 2, 3, and 4 report merge rates with the public SSA Death Master File (DMF) for all applicants (column 2), applicants who enlist (column 3), and applicants who do not enlist (column 4). A merge with the DMF is defined as a matching Social Security Number and a matching month of birth. 164 Instrument I (AFQT>3 1) Stage Appendix Table 3.3. RD and RK First Appendix Table 3.3. RD and RK First Stage Parametric Estimates 3rd Order Poly 2nd Order Poly (2) (1) (3) (4) 0.1015*** (0.0045) I (AFQT>50) 0.0567*** 0.0567*** (0.0042) (0.0042) -0.0157*** (0.0012) -0.0022* (0.0012) 407.3 668,149 446.4 668,149 I (AFQT>3 1) x AFQT 1 (AFQT>50) x AFQT First Stage F-Statistic Observations 0.1015*** (0.0045) 0.1080*** (0.0062) 0.0509*** (0.0056) 174.1 668,149 0.1080*** (0.0062) 0.0509*** (0.0056) -0.0079** (0.0033) 0.0001 (0.0030) 96.9 668,149 Nonparametric Estimates BW = 15 (5) 0.1049*** (0.0032) 0.0506*** (0.0029) 615.0 668,149 (6) 0.1049*** (0.0032) 0.0506*** (0.0029) -0.0172*** (0.0004) -0.0024*** (0.0004) 2122.6 668,149 Notes: This table reports parametric and nonparametric first stage estimates from parametric and nonparametric 2SLS RD and 2SLS RDRK specifications. The endogenous (LHS) variable for all estimates is an indicator for any military service. The RD instruments are enlistment offers (AFQT>3 1) and bonus offers (AFQT>50) while the RK instruments interact enlistment offers and bonus offers with a linear AFQT term. All estimates include dummy controls for fiscal year of application and a linear control for AFQT (the running variable). Parametric estimates using second order polynomial controls for the running variable also include the following controls: (AFQT 2) X I (AFQT<3 1), (AFQT 2) x 1 (AFQT>3 1), and (AFQT 2) x 1 (AFQT>50). Parametric estimates using third order polynomial controls include these additional controls: (AFQT 3) x 1(AFQT<3 1), (AFQT ) x 1 (AFQT>3 1), and (AFQT 3) x 1 (AFQT>50)for Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 165 Dependent Variable Appendix Table 3.4. 2SLS Estimates by Race and Application Timeframe FY96-FYO 1 FY02-FY07 Full RD Sample Non-whites Whites Applicants Applicants (1) (2) (3) (4) (5) A. Parametric 2SLS RD-RK Estimates: 2nd Order Polynomial Control for RV Deceased by -0.0054 -0.0108 -0.0046 -0.0055 -0.0049 FEB10 (0.0053) (0.0080) (0.0071) (0.0073) (0.0075) E[Y,] 0.0125 0.0106 0.0134 0.0147 0.0099 Deceased by FEB15 E[Y, ID1 =0] -0.0026 (0.0061) 0.0163 -0.0093 (0.0093) 0.0139 -0.0015 (0.0080) 0.0173 -0.0050 (0.0082) 0.0184 0.0000 (0.0088) 0.0138 Observations 0.2764*** (0.0148) 0.1220 668,149 0.3020*** (0.0241) 0.1061 197,748 0.2645*** (0.0190) 0.1287 470,401 0.2207*** (0.0178) 0.1075 363,319 0.3499*** (0.0249) 0.1394 304,830 Deceased by FEB10 E[Yj] B. Nonparametric 2SLS RD-RK Estimates: Bandwidth = 15 0.0011 -0.0003 0.0009 0.0005 (0.0026) (0.0038) (0.0036) (0.0037) 0.0125 0.0106 0.0134 0.0147 0.0019 (0.0036) 0.0099 Deceased by FEB15 E[Y, ID1 =0] 0.0033 (0.0030) 0.0163 0.0026 (0.0045) 0.0139 0.0029 (0.0040) 0.0173 0.0023 (0.0042) 0.0184 0.0046 (0.0042) 0.0138 0.2554*** (0.0071) 0.1220 668,149 0.2501*** (0.0113) 0.1061 197,748 0.2592*** (0.0092) 0.1287 470,401 0.2119*** (0.0088) 0.1075 363,319 0.3144*** (0.0115) 0.1394 304,830 DC Receipt by JAN14 DC Receipt by JAN14 Observations Notes: This table reports 2SLS estimates of the effects of military service on mortality and DC receipt. Column 1 reports results for the full RD sample, which includes applicants with AFQT scores in the interval [16, 64]. Columns 2 through 5 limit the RD sample to the subgroups identified in the column header. All regressions include indicator variables for each application fiscal year. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 166 Endogenous Variable Appendix Table 3.5. 2SLS Estimates with 2 Endogenous Variables B. DV: DC Receipt by JAN14 A. DV: Deceased by FEB10 2SLS RD-RK OLS 2SLS RD-RK OLS 2nd Order Poly NP (BW=15) 2nd Order Poly NP (BW=15) (3) (2) (1) (3) (2) (1) 0.2846*** 0.2794*** Any Active Duty Service -0.0020*** -0.0051 -0.0021 0.2498*** Any Reserve Service (0.0003) -0.0054*** (0.0005) (0.0057) -0.0152 (0.0632) (0.0053) 0.0209 (0.0269) (0.0008) 0.1290*** (0.0018) (0.0159) -0.0112 (0.1841) (0.0151) 0.1037 (0.0772) 668,149 First Stage F-Statistics 1,598.2 428.1 263.9 16.2 668,149 668,149 668,149 428.1 16.2 668,149 1,598.2 263.9 668,149 Any Active Duty Service Any Reserve Service Observations Notes: This table reports OLS and 2SLS RD-RK estimates of the effects of active duty and reserve duty military service on mortality and DC receipt. All regressions include dummy controls for fiscal year of application. OLS estimates include additional indicators for each possible AFQT scores and estimates in columns 4 and 5 include additional controls for year of birth, race (white or nonwhite), education level (still in high school, high school dropout or GED, high school diploma, some college), and state of birth. Robust standard errors are reported in parentheses. ***,**, and * denote significance at the 1%, 5%, and 10% level, respecitvely. 167 References Abbott, Michael and Orley Ashenfelter. 1976. "Labour Supply, Commodity Demand and the Allocation of Time," Review ofEconomic Studies, 43(3), 389-411. Air Force Health Study. 2000. "Air Force Health Study: An Epidemiological Investigation of Health Effects in Air Force Personnel Following Exposure to Herbicides. 1997 Follow-up Examination Results." Brook Air Force Base: Air Force Research Laboratory. American Psychiatric Association. 1994. Diagnostic and StatisticalManual ofMental Disorders:DSM-4. ManMag. Angrist, Josh. 1990. "Lifetime Earnings and the Vietnam-Era Draft Lottery: Evidence from Social Security Administrative Records," American Economic Review, 80(3), 313-336. Angrist, Josh. 1998. "Estimating the Labor Market Impact of Voluntary Military Service Using Social Security Data on Military Applicants." Econometrica, 66(2), 249-288. Angrist, Josh. 2001. "Estimation of Limited Dependent Variable Models with Dummy Endogenous Regressors." JournalofBusiness & Economic Statistics, 19(1). Angrist, Josh and Jorn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An Empiricist's Companion. Princeton: Princeton University Press. Angrist, Josh, Stacey Chen, and Brigham Frandsen. 2010. "Did Vietnam Veterans Get Sicker in the 1990s? The Complicated Effects of Military Service on Self-Reported Health." JournalofPublic Economics, 94(11), 824-837. Ashenfelter, Orley and James J. Heckman. 1974. "The Estimation of Income and Substitution Effects in a Model of Family Labor Supply." Econometrica, 42(1), 73-85. Autor, David H., Mark Duggan, Kyle Greenberg, and David S. Lyle. 2015. "The Impact of Disability Benefits on Labor Supply: Evidence from the VA's Disability Compensation Program." NBER Working Paper. Autor, David H. and Mark G. Duggan. 2003. "The Rise in the Disability Rolls and the Decline in Unemployment." QuarterlyJournalofEconomics, 118(1), 157-206. Autor, David H. and Mark G. Duggan. 2007. "Distinguishing Income from Substitution Effects in Disability Insurance." American Economic Review Papersand Proceedings, 97, 119-124. Autor, David H., Mark G. Duggan, and David Lyle. 2011. "Battle Scars? The Puzzling Decline in Employment and Rise in Disability Receipt among Vietnam Era Veterans." American Economic Review Papersand Proceedings, 101, 339-344. Autor, David H., Nicole Maestas, Kathleen J. Mullen and Alexander Strand. 2015. "Does Delay Cause Decay? The Effect of Administrative Decision Time on the Labor Force Participation and Earnings of Disability Applicants." NBER Working Paper. Batty, G. David, Ian J. Deary, and Linda S. Gottfredson. 2007. "Premorbid (early life) IQ and Later Mortality Risk: Systematic Review." Annals ofEpidemiology, 17(4), 278-288. 168 Batty, G. David, M.J. Shipley, L.H. Mortensen, S.H. Boyle, J. Barefoot, M. Gronbaek, C.R. Gale, and I.J. Deary. 2008. "IQ in Late Adolescence/Early Adulthood, Risk Factors in Middle Age and Later All-Cause Mortality in Men: The Vietnam Experience Study." Journal of Epidemiology and Community Health, 62(6), 522-531. Benitez-Silva, Hugo, Moshe Buchinsky, Hiu Man Chan, Sofia Cheidvasser, and John Rust. 2004. "How Large is the Bias in Self-Reported Disability?" JournalofApplied Econometrics, 19(6) 649-670. Bilmes, Linda J. and Joseph E. Stiglitz. 2008. The Three Trillion Dollar War: The True Cost of the Iraq Conflict. New York: W.W. Norton and Company. Black, Dan, Kermit Daniel, and Seth Sanders. 2002. "The Impact of Economic Conditions on Participation in Disability Programs: Evidence from the Coal Boom and Bust."American Economic Review, 92, 27-50. Bollinger, Mary J., Susanne Schmidt, Jacqueline A. Pugh, Helen M. Parsons, Laurel A. Copeland, and Mary Jo Pugh. 2015. "Erosion of the Health Soldier Effect in Veterans of U.S. Military Service in Iraq and Afghanistan." PopulationHealth Metrics, 13(8). Bonds, Timothy M., Dave Baiocchi, and Laurie L. McDonald. 2010. Army Deployments to OIF and OEF. RAND Arroyo Center Santa Monica CA. Borghans, Lex, Anne Gielen, and Erzo Luttmer. 2014. "Social Support Substitution and the Earnings Rebound: Evidence from a Regression Discontinuity in Disability Insurance Reform." American Economic Journal:Economic Policy, 6(4), 34-70. Bound, John. 1989. "The Health and Earnings of Rejected Disability Insurance Applicants." American Economic Review, 79(3), 482-503. Bound, John and Richard Burkhauser. 1999. "Economic Analysis of Transfer Programs Targeted on People with Disabilities." Handbook of Labor Economics, Volume 3 (eds. Orley Ashenfelter and David Card), Amsterdam, North-Holland. Bound, John and Timothy Waidmann. 1992. "Disability Transfers, Self-Reported Health, and the Labor Force Attachment of Older Men: Evidence from the Historical Record." Quarterly Journal of Economics, 1393-1419. Boyle, Melissa A. and Johanna Lahey. 2010. "Health Insurance and the Labor Supply Decisions of Older Workers: Evidence from a US Department of Veterans Affairs Expansion." Journal of Public Economics, 94(7-8), 467-478. Bradley, Rebekah, Jamelle Greene, Eric Russ, Lissa Dutra, and Drew Westen. 2005. "A Multidimensional Meta-Analysis of Psychotherapy for PTSD." American JournalofPsychiatry, 162(2), 214-227. Brewin, Chris R, Ruth A Lanius, Andrei Novac, Ulrich Schnyder, and Sandro Galea. 2009. "Reformulating PTSD for DSM-V: Life after Criterion A." Journal of Traumatic Stress, 22(5), 366-373. 169 Calvert, G., Sweeney, M., Deddens, J. and D. Wall. 1999. "An Evaluation of Diabetes Mellitus, Serum Glucose, and Thyroid Function among U.S. Workers Exposed to 2, 3, 7, 8-tetrachlorodibenzo-pdioxin." Occupationaland EnvironmentalMedicine, 56(4), 270-276. CDC, NIH, DoD, and VA Leadership Panel. 2013. Report to Congress on Traumatic Brain Injury in the United States: Understanding the Public Health Problem Among Current and Former Military Personnel. Centers for Disease Control and Prevention (CDC), the National Institutes of Health (NIH), the Department of Defense (DoD), and the Department of Veterans Affairs (VA). Cesur, Resul, Joseph J Sabia, and Erdal Tekin. 2013. "The Psychological Costs of War: Military Combat and Mental Health." JournalofHealth Economics, 32(1), 51-65. Chen, Susan and Wilbert van der Klaauw. (2008) "The Work Disincentive Effects of the Disability Insurance Program in the 1990s." JournalofEconometrics, 142(2), 757-784. Chetty, Raj. 2012. "Bounds on Elasticities with Optimization Frictions: A Synthesis of Micro and Macro Evidence on Labor Supply." Econometrica, 80(3), 969-1018. Congressional Budget Office. 2014. Veterans Disability Compensation: Trends and Policy Options. Congressional Budget Office, August 2014. Coile, Courtney, and Jonathan Gruber. 2007. "Future Social Security Entitlements and the Retirement Decision." Review of Economics and Statistics, 89(2), 234-246. Coile, Courtney, Mark Duggan and Andrew Guo. (2015) "Veterans' Labor Force Participation: What Role Does the VA's Disability Compensation Program Play?" forthcoming in American Economic Review Papers and Proceedings. Conley, Dalton, and Jennifer Heerwig. 2012. "The Long-Term Effects of Military Conscription on Mortality: Estimates from the Vietnam-Era Draft Lottery." Demography, 49(3), 841-855. Duggan, Mark, Perry Singleton and Jae Song. 2007. "Aching to Retire? The Rise in the Full Retirement Age and its Impact on the Disability Rolls." JournalofPublic Economics, 91(7), 1327-1350. Duggan, Mark, Robert Rosenheck and Perry Singleton. 2006. "Federal Policy and the Rise in Disability Enrollment: Evidence for the VA's Disability Compensation Program." NBER Working Paper. Duggan, Mark, Robert Rosenheck and Perry Singleton. 2010. "Federal Policy and the Rise in Disability Enrollment: Evidence for the VA's Disability Compensation Program." Journal of Law and Economics, 53, 379-398. Eriksson, Stefan and Dan-Olof Rooth. 2014. "Do Employers Use Unemployment as a Sorting Criterion When Hiring? Evidence from a Field Experiment." The American Economic Review, 104(3), 1014-1039. Frandsen, Brigham R. 2014. "Party Bias in Union Representation Elections: Testing for Manipulation in the Regression Discontinuity Design When the Running Variable is Discrete." BYU Working Paper. French, Eric and Jae Song. 2014. "The Effect of Disability Insurance Receipt on Labor Supply." American Economic Journal:Economic Policy, 6(2), 291-337. 170 Frisch, Ragnar. 1959. "A Complete Scheme for Computing All Direct and Cross Demand Elasticities in a Model with Many Sectors," Econometrica, 27(2), 177-196. Frueh, B Christopher, Anouk L Grubaugh, Jon D Elhai, and Todd C Buckley. 2007. "US Department of Veterans Affairs Disability Policies for Posttraumatic Stress Disorder: Administrative Trends and Implications for Treatment, Rehabilitation, and Research." American Journal of Public Health, 97(12), 2143-2145 Gates, Robert M. 2014. Duty: Memoirs of a Secretary at War. Random House LLC. Gros, Daniel F., Matthew Price, Erica K. Yuen, and Ron Aciemo. 2013. "Predictors of Completion of Exposure Therapy in OEF/OIF Veterans with Posttraumatic Stress Disorder." Depression and Anxiety, 30(11), 1107-1113. Gruber, Jon. 2000. "Disability Insurance Benefits and Labor Supply," JournalofPoliticalEconomy, 108, 1162-1183. Hagel, Charles. 2014. "MEMORANDUM: Supplemental Guidance to Military Boards For Correction of Military/Naval Records Considering Discharge Upgrade Requests by Veterans Claiming Post Traumatic Stress Disorder." Hearst, N., T.B. Newman, and S. B. Hulley. 1986. "Delayed Effects of the Military Draft on Mortality: A Randomized Natural Experiment." The New EnglandJournalofMedicine, 314, 620-624. Hill, Mark E. and Ira Rosenwaike. 2001. "The Social Security Administration's Death Master File: The Completeness of Death Reporting at Older Ages." Social Security Bulletin, 64(1), 45-51. Hoge, Charles W., Carl A. Castro, Stephen C. Messer, Dennis McGurk, Dave I. Cotting, and Robert L. Koman. 2004. "Combat Duty in Iraq and Afghanistan, Mental Health Problems, and Barriers to Care." New EnglandJournalofMedicine, 351(1), 13-22. Hoge, Charles W., Jennifer L. Auchterlonie, and Charles S. Milliken. 2006. "Mental Health Problems, Use of Mental Health Services, and Attrition from Military Service After Returning from Deployment to Iraq or Afghanistan" Journal of American Medical Association, 295(9), 10231032. Home, David K. "The Impact of Soldier Quality on Army Performance." 1987. Armed Forces & Society, 13(3), 443-455. Imbens, Guido and Karthik Kalyanaraman. 2012. "Optimal Bandwidth Choice for the Regression Discontinuity Estimator." Review ofEconomics and Statistics, 79(4), 933-959. Institute of Medicine. 2000. "Veterans and Agent Orange: Herbicide/Dioxin Exposure and Type 2 Diabetes." Washington, D.C.: The National Academies Press. Institute of Medicine. 2010. "Returning Home from Iraq and Afghanistan: Preliminary Assessment of Readjustment Needs of Veterans, Service Members, and Their Families." 171 Institute of Medicine. 2011. "Long-term Health Consequences of Exposure to Burn Pits in Iraq and Afghanistan." Washington, D.C.: The National Academies Press. Jokela, Markus, Marko Elovainio, Archana Singh-Manoux, and Mika Kivimaki. 2009. "IQ, Socioeconomic Status, and Early Death: The US National Longitudinal Survey of Youth." PsychosomaticMedicine, 71(4), 322-328. Kang, H.K. and Bullman, T.A. 2009. "Is There an Epidemic of Suicides Among Current and Former U.S. Military Personnel?" Annals ofEpidemiology, 19(10), 757-760. Kaplan, M.S., B.H. McFarland, N. Huguet, and M. Valenstein. 2012. "Suicide Risk and Precipitating Circumstances Among Young, Middle-aged, and Older Male Veterans." American Journal of Public Health, 102(Supplement 1), S131-S137. Kolkow, Tonya T., James L. Spira, Jennifer S. Morse, and Thomas A. Grieger. 2007. "Post-Traumatic Stress Disorder and Depression in Health Care Providers Returning from Deployment to Iraq and Afghanistan." Military Medicine, 172(5), 451-455. Kostol, Andreas and Magne Mogstad. 2014. "How Financial Incentives Induce Disability Insurance Recipients to Return to Work." American Economic Review 104(2): 624-55. Kroft, Kory, Fabian Lange, and Matthew J. Notowidigdo. 2013. "Duration Dependence and Labor Market Conditions: Evidence from a Field Experiment." QuarterlyJournalof Economics, 128(3), 11231167. Macklin, Michael L., Linda J. Metzger, Brett T. Litz, Richard J. McNally, Natasha B. Lasko, Scott P. Orr, and Roger K. Pitman. 1998. "Lower Precombat Intelligence Is a Risk Factor for Posttraumatic Stress Disorder." Journalof Consulting and ClinicalPsychology, 66(2), 323. Maercker, Andreas, Chris R. Brewin, Richard A. Bryant, Marylene Cloitre, Geoffrey M. Reed, Mark van Ommeren, Asma Humayun, Lynne M. Jones, Ashraf Kagee, Augusto E. Llosa et al. 2013. "Proposals for Mental Disorders Specifically Associated with Stress in the International Classification of Diseases- I1." The Lancet, 381(9878), 1683-1685. Maestas, Nicole, Kathleen Mullen, and Alexei Strand. 2013. "Does Disability Insurance Receipt Discourage Work? Using Examiner Assignment to Estimate Causal Effects of SSDI Receipt." American Economic Review, 103(5), 1797-1829. Mayberry, Paul W. and Hiatt, Catherine M. 1992. "Computing AFQT Scores from Historical Data." Center for Naval Analysis. McClelland, Robert and Shannon Mok. 2012. "A Review of Recent Research on Labor Supply Elasticities." Congressional Budget Office Working Paper #2012-12, October. McCrary, Justin. 2008. "Manipulation of the Running Variable in the Regression Discontinuity Design." Journalof LaborEconomics, 13(2), 289-308. McNally, Richard J. and B. Christopher Frueh. 2013. "Why are Iraq and Afghanistan War Veterans Seeking PTSD Disability Compensation at Unprecedented Rates?" Journal ofAnxiety Disorders, 27(5), 520-526. 172 McNally, Richard J. and Lisa M. Shin. 1995. "Association of Intelligence with Severity of Posttraumatic Stress Disorder Symptoms in Vietnam Combat Veterans." The American Journal of Psychiatry, 152(6), 936. National Technical Information Service. 2011. "Important Notice: Change in Public Death Master File Records." U.S. Department of Commerce. Oberholzer-Gee, Felix. 2008. "Nonemployment Stigma as Rational Herding: A Field Experiment." Journal ofEconomic Behavior & Organization,65(1), 30-40. Parsons, Donald. 1980. "The Decline in Male Labor Force Participation." The Journal of Political Economy 88: 117-134. Pearce, M.S., I.J. Deary, A.H. Young, and L. Parker. 2006. "Childhood IQ and Deaths Up to Middle Age: The Newcastle Thousand Families Study." PublicHealth, 120(11), 1020-1026. Ruggles, Steven J., Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder, and Matthew Sobek. 2010. Integrated Public Use Microdata Series: Version 5.0. Minneapolis: University of Minnesota. Schiller, Jeannine, Jacqueline Lucas, Brian Ward, Jennifer Peregoy. 2010. "Summary Health Statistics for U.S. Adults: National Health Interview Survey, 2010." Vital and Health Statistics, 10(252). Schnurr, Paula P., Carole A. Lunney, Anjana Sengupta, and Avron Spiro III. 2005. "A Longitudinal Study of Retirement in Older Male Veterans." Journal of Consulting and Clinical Psychology, 73(3), 561. Scribner, Barry L., D. Alton Smith, Robert H. Baldwin, and Robert L. Phillips. 1986. "Are Smart Tankers Better? AFQT and Military A Productivity." Armed Forces & Society, 12(2) 193-206. Segall, Donald 0. 2004. "Development and Evaluation of the 1997 ASVAB Scale Score." Defense Manpower Data Center. Seltzer, Carl C., and Seymour Jablon. 1974. "Effects of Selection on Mortality." American Journal of Epidemiology, 100(5), 367-372. Siminski, Peter, and Simon Ville. 2011. "Long-Run Mortality Effects of Vietnam-Era Army Service: Evidence from Australia's Conscription Lotteries." American Economic Review Papers and Proceedings, 101, 345-349. Siminski, Peter. 2013. "Employment Effects of Army Service and Veterans' Compensation: Evidence from the Australian Vietnam-Era Conscription Lotteries" Review of Economics and Statistics, 95(1), 87-97. Tanielian, Terri L. and Lisa Jaycox. 2008. Invisible Wounds of War: Psychological and Cognitive Injuries, Their Consequences, and Services to Assist Recovery. Santa Monica, CA: RAND Corporation. U.S. Army. 2010. "Army Health Promotion, Risk Reduction, and Suicide Prevention Report." Washington, D.C.: United States Army. 173 U.S. Army Recruiting Command. 2012. "USAREC Regulation 601-96: Enlistment, Accessions, and Processing Procedures." Fort Knox, KY. U.S. Department of Defense, 2004. PopulationRepresentation in the Military Services: Fiscal Year 2002. Washington D.C.: Office of the Under Secretary of Defense (Personnel and Readiness). U.S. Department of Defense. 2015. U.S. Military Casualties- Operation Iraqi Freedom (QIF): Casualty Summary by Month and Service. Defense Manpower Data Center (Defense Casualty Analysis System). U.S. Department of Veterans Affairs. 2003. "Agent Orange Brief." Environmental Agents Service, (131) Al. U.S. Department of Veterans Affairs: Office of the Inspector General. 2005. "Review of State Variances in VA Disability Compensation Payments." U.S. General Accounting Office. 2006. "VA Should Improve Its Management of Individual Unemployability Benefits by Strengthening Criteria, Guidance, and Procedures." Report GAO06-309. U.S. General Accounting Office. 2009. "Social Security Disability: Additional Outreach and Collaboration on Sharing Medical Records Would Improve Wounded Warriors' Access to Benefits." Report GAO-09-762. U.S. Veterans Benefits Administration. Selected Years. "Veterans Benefits Administration Annual Benefits Report." Available online at http://www.vba.va.gov/reports.htm. U.S. Veterans Benefits Administration. 2013. "Trends in the Utilization of VA Programs and Services." Von Wachter, Till, Joyce Manchester, and Jae Song. 2011. "Trends in Employment and Earnings of Allowed and Rejected Applicants to the SSDI Program." American Economic Review, 33083329. White House. 2010. "Weekly Address: President Obama Announces Changes to Help Veterans with PTSD Receive the Benefits They Need." 10 July 2010. Winkler, John D. 1999. "Are Smart Communicators Better? Soldier Aptitude and Team Performance." Military Psychology, 11(4), 405-422. Worthen, Mark D. and Robert G. Moering. 2011. "A Practical Guide to Conducting VA Compensation and Pension Exams for PTSD and Other Mental Disorders," PsychologicalInjury and Law, 4(34), 187-216. Yankovich, Michael. 2011. "The Effects of Combat Exposure on Military Experience." Working Paper. 174