Is There a VA Advantage - Evidence from Dually Eligible Veterans

Is There a VA Advantage? Evidence from Dually Eligible Veterans Abstract Societies face a difficult choice between private and public provision of health care, and there is a shortage of credible evidence to guide this choice. The structure of health care delivery to US veterans provides a distinctive research setting to study this issue. Specifically, veterans aged 65 and older are dually eligible for care in private hospitals (financed by Medicare) or in publicsector hospitals operated by the US Department of Veterans Affairs (VA) in the nation’s largest integrated health care delivery system. We utilize the ambulance design of Doyle et al. (2015) to examine the effect of VA vs. non-VA emergency care on mortality in this high-risk population. We find a VA advantage: a 28-day mortality reduction of 46% (4.5 percentage points, with a 95% confidence interval of 1.1 to 8.0). Survival gains persist for at least a year after the initial ambulance ride, and they accrue despite lower spending in the VA. Evidence suggests that the VA advantage arises in part from some combination of continuity of care and health IT. These results have policy relevance—as the federal government is deciding whether to maintain the existing VA system or to expand financing of private care outside of the VA—and they shed light on sources of inefficiency in private-sector health care in the US. 1 Introduction A key question in the design of health care systems across the world is whether care should be provided by government or by the private sector. In the US, the choice between public and private provision has become a top policy issue for the Department of Veterans Affairs (VA). Seeking to improve veteran access to health care, policymakers have debated whether the VA should expand the capacity of its own system—the Veterans Health Administration—or shift health care delivery to private providers. An extensive descriptive literature (e.g., Reid 2010; Blank et al. 2017) has compared health care outcomes in public vs. private systems. More generally, economists long have debated the appropriate size and role of the public sector in the economy, highlighting theoretical arguments about competitive pressure, ownership structure, and differences in the objectives and constraints in the public vs. private sector (Alchian 1965; Stigler 1965; Hart et al. 1997). Nevertheless, rigorous empirical evaluations of the performance of public vs. private health care providers have been relatively rare, in no small part because public and private providers of health care usually serve different patient populations, either by statute or by patient selection. In this paper, we focus on “dually eligible” veterans aged 65 and older who can receive health care at both VA facilities and private hospitals that accept Medicare. We use the ambulance design proposed by Doyle et al. (2015) to study the causal effect of receiving emergency care at the VA vs. a non-VA facility. Our approach compares veterans sharing key characteristics—zip code of residence, prior VA and non-VA utilization, and location of pick-up (e.g., their home residence vs. a nursing home)—who receive the same dispatched level of ambulance service (i.e., advanced vs. basic life support) from different ambulance companies. Our main analytic sample includes 401, 319 911-dispatched ambulance rides from 2001 to 2014, for veterans with prior attachment to the VA and in a zip code served by at least two ambulance companies. As in Doyle et al. (2015), we show that the leave-out share of dually eligible veterans transported to the VA by the assigned ambulance company is a strong predictor of hospital assignment. Under the plausible assumption that ambulances are quasi-randomly assigned within zip codes and in cells of key characteristics, this design allows us to study the effect of VA vs. non-VA emergency care on health outcomes. We find that in the high-mortality population of elderly veterans with emergencies, there is a VA advantage—a 46% reduction in 28-mortality relative to baseline (4.5 p.p., with a 95% confidence interval of 1.1 to 8.0 p.p). Importantly, we show that our instrumental variables (IV) estimates of the 1 VA effect are robust to the inclusion of a long list of characteristics of the index patient and of other patients transported by the same ambulance company. The latter set of ambulance co-rider controls can account for unobserved patterns of selection across ambulance companies (Altonji and Mansfield 2018). The IV estimates are larger in magnitude than the corresponding OLS estimates, which center around 0.024 p.p., with tight confidence intervals. A possible explanation for this difference is that VA “always takers” (patients who go to the VA even with an ambulance company with a low VA rate) in fact have worse health than VA “never takers,” a pattern suggested by the medical literature and one that we examine in greater detail below (Agha et al. 2000). A critical question for interpreting the survival benefits of VA care is whether these effects fade over longer horizons—as would happen if VA emergency care only temporarily displaces the mortality of fragile patients under “harvesting” (Schwartz 2000). To address this, we use an insight from Abadie (2002) to estimate the weekly potential death rates in the year after the initial ambulance ride among compliers of the quasi-experiment, i.e., patients whose destination hospital is determined by the ambulance company. With this tool, we disentangle the short-term vs. long-term effects of the VA in the setting of competing risks. Despite a high long-term mortality rate (close to one in three veterans will be dead within one year of the ambulance ride), we find that the mortality impact of presenting at the VA is concentrated in the first week, suggesting VA survival gains from care addressing temporary emergency conditions. We find no evidence of harvesting, suggesting that the survival gains are long-lasting. Relying on intuition from Kitagawa (2015), we also use this potential outcomes framework to develop a sharper test of IV validity than the tests typical in the applied literature.1 Finally, we use this framework to document systematically larger long-run mortality hazards for VA always takers than for VA never takers, based on the fact that differences between OLS and IV estimates of the VA advantage consistently widen with longer time horizons. The key potential threat to our research design is that veterans who are taken to the VA are healthier than veterans who are taken to non-VA hospitals. While this runs counter to the selection that we demonstrate above, this could arise if the choice of a specific ambulance company from among those that serve a given zip code is correlated with the risk of death. We present four additional pieces of evidence to address this concern. First, we show balance in characteristics of patients assigned to 1 Specifically, we use the fact that, under IV validity, all indicators for potential outcomes must occur with positive probability among compliers (Balke and Pearl 1997; Kitagawa 2015). In the setting of survival, this implies that the incremental mortality risk must be positive for compliers in every week after the ambulance ride and in both VA and non-VA assignment. This prediction may fail if there are violations of monotonicity that may arise, for example, because ambulance companies with higher VA propensities are less likely to send veterans with certain potential mortality outcomes to the VA. Chan et al. (2019) show that this approach may detect violations in IV validity that remain hidden under standard “judges design” tests of monotonicity (e.g., Arnold et al. 2018). 2 companies with different propensities of taking patients to the VA. Second, we conduct an extensive analysis along the lines suggested by Altonji et al. (2005), evaluating the stability of our estimates as we add controls to the models, including controls that measure the characteristics of other patients transported by the company. Third, in our IV analysis of longer-run survival, we show that aside from the first week, in which there is a large effect on mortality, compliers who go to the VA have mortality hazard rates that are statistically indistinguishable from rates of those who go to non-VA hospitals—ruling out significant differences in underlying health between the two groups. Finally, in heterogeneity analyses, we show that the VA advantage is highly stable across VA and non-VA hospital characteristics that may be related to selection of high-risk patients. In the final section of the paper we then turn to an evaluation of the mechanisms behind the VA advantage. We consider three broad classes of mechanisms. First, the VA might be better suited to treating conditions specific to veterans. Second, along the lines of Doyle et al. (2015), VA hospitals could achieve better outcomes by spending more. Third, better access to patient information and coordination of care may improve the productivity of VA-delivered care, particularly in high-uncertainty and high-stakes environments such as emergency care. This last mechanism is perhaps most consistent with a descriptive literature that highlights integration of care and health information technology (IT) as distinguishing features of the VA relative to the private sector since reforms in the mid-1990s (Mccarthy and Blumenthal 2006; Jha et al. 2009). To evaluate the first explanation—that the VA system is uniquely suited to care for veterans—we note that, although we do not observe non-veterans being treated in the VA, we observe a detailed set of veteran and neighborhood characteristics that predict whether a veteran is more likely to be attached to the VA. We find a VA advantage for patients with very different patterns of prior utilization and comorbidities, although medically needier patients (e.g., those with substance abuse and mental health problems) and those with greater attachment to the VA tend to benefit more. We evaluate the second explanation by examining VA and Medicare spending, using information on actual spending by taxpayers and veterans. Spending following VA care is lower by $2, 548, or about 21%, at 28 days. This suggests that the VA is more productive, achieving better outcomes at lower cost. The third explanation centers on the idea that coordination and continuity of care in an integrated delivery system may improve health outcomes—an explanation consistent with the larger impacts of the VA on medically needy patients and those with greater prior attachment. To provide further evidence on this, we draw on a secondary sample of veterans who have prior attachment to a specific non-VA hospital but no recent use of the VA. As noted, these veterans have little chance of receiving 3 emergency care at a VA hospital but may return to the hospital they visited most in the prior year, where their records may be more easily accessed, or to another facility. Ambulance-based assignment to a patient’s most-visited prior hospital (i.e., their “modal” hospital) indeed reduces 28-day mortality, though only modestly (by about 0.6 p.p.). We infer that the VA survival benefit arises from more than just repeated use of the same facility; instead, it may reflect better care coordination and/or more effective information retrieval in the VA health care system. To probe these channels, we exploit two policy reforms that aimed to improve care coordination and information technology among US hospitals. In 2009 the Health Information Technology for Economic and Clinical Health (HITECH) Act stimulated a large increase in health IT adoption among non-VA hospitals, and in 2011 Medicare began experimenting with alternative payments to “Accountable Care Organizations” (or ACOs) (Blumenthal 2010; Greaney 2011). Consistent with these changes, we find that the modal-hospital survival benefit increases from a negligible effect prior to 2010 to about 1.9 p.p.—approximately one-half of the VA survival benefit—after 2010. We also find tentative evidence linking the increase in the modal-hospital survival benefit to hospital-specific health IT adoption and, to a lesser degree, ACO participation. Our findings contribute to three sets of related literature. First, the public vs. private provision of health care is a central question for the field of comparative health policy, which compares health care systems across the world to inform the design of health care systems (Blank et al. 2017). The literature in this field has been mainly descriptive.2 Comparing the performance of health care systems across countries is intrinsically difficult for obvious reasons: Countries differ in their populations and in other determinants of health. To our knowledge, the only quasi-experimental exception in this literature is a recent working paper by Frakes et al. (2020). Studying military mothers who give birth in two different hospitals due to a move in between births, they find higher spending but lower rates of complications in private hospitals. Second, an important literature has sought to measure the quality of care in the VA, which budgeted $84 billion for medical care in 2020 (Department of Veterans Affairs 2020).3 Following a well-known reorganization and investment in health IT in the mid-1990s (Mccarthy and Blumenthal 2006), this literature has documented favorable VA quality, compared to care outside of the VA, in 2 As an example of the amount of material devoted to such comparative studies, the European Observatory on Health Systems and Policies (www.euro.who.int) produces policy commentary and “health system reviews” on the health care systems of individual countries. 3 Spending continues to grow. The 2019 enacted budget allocated $77 billion for VA medical care, and the 2021 proposed budget requests $94 billion for medical care. For the last ten years, spending on medical care has nearly doubled (Department of Veterans Affairs 2020). 4 terms of process measures and health outcomes (e.g., Jha et al. 2003). The question of performance in the VA health care system has become particularly relevant in recent years, as the Department considers ways to improve access to care for veterans and as Congress has sought to increase private health care delivery for veterans (113th Congress 2014; 115th Congress 2018). So far, however, this literature has mainly compared outcomes of veterans receiving care in the VA system to those of non-veterans outside of the VA. A third and very large literature studies why health care in the US appears to be a low-productivity outlier among developed countries, spending more as a percentage of GDP than any country but with poor outcomes relative to this spending (Garber and Skinner 2008). Experts have drawn attention to fragmentation in the US health care system, which could increase spending and potentially worsen outcomes (Cebul et al. 2008; Cutler 2010; Agha et al. 2019). Policymakers have responded by incentivizing adoption of health IT and coordination of care, but whether such policies improve health outcomes remains an open empirical question.4 Our results are consistent with a productivity advantage (better outcomes at lower cost) at the VA, the nation’s largest integrated health care system. We also find suggestive evidence that government regulations to incentivize private hospitals to adopt health IT and integrate care may achieve better outcomes, at least in our setting for patients with continuity of care at these hospitals. The remainder of this paper proceeds as follows. Section 2 describes the setting and data. Section 3 presents our main analysis of the VA survival benefit. Section 4 discusses our survival analysis over time. Section 5 presents evidence on mechanisms driving the VA survival benefit. Section 6 discusses policy implications and concludes. 2 Setting and Data 2.1 The Veterans Health Administration and US Health Care The Veterans Health Administration (VHA) of the US Department of Veterans Affairs (VA) provides health care for 9 million veteran enrollees every year. The VHA is the nation’s largest integrated health care delivery system, including 170 medical centers and more than 1,000 outpatient sites of 4 A recent empirical literature documents modest reductions in spending and improvements in patient satisfaction among provider forming ACOs (McWilliams et al. 2014, 2016; Trombley et al. 2019). An older literature on health maintenance organizations (HMOs) documents impacts on spending and technology adoption in the 1990s (Baker 2001; Cutler et al. 2000), although policy analysts have noted that HMOs were primarily insurance products, focused more on limiting utilization than on improving quality (Luft 2010). Finally, a mixed literature on health IT adoption has shown health improvements in some cases (e.g., Miller and Tucker 2011) but null results in general (e.g., Agha 2014). To our knowledge, our paper is the first to assess the complementarity between health IT and continuity of care. 5 care, with a budget of $84 billion in 2020 for medical care (Department of Veterans Affairs 2020). Two features of the VHA distinguish it from the rest of the US health care delivery system. First, the VHA is owned and administered by the government, while the rest of US health care system is largely run by private parties. Second, health care is integrated at the VHA. The VHA directly employs all of its physicians and health care workers. In contrast, physicians outside of the VA are mostly independent of the hospitals at which they work and can affiliate at will with multiple hospitals. Health care in the VHA is organized by region and coordinated across inpatient, emergency department, and outpatient locations, as well as across different services and specialties of care. In the mid-1990s, the VHA implemented one of the first and most widely used electronic health record (EHR) systems in the US. VHA spending on health information technology (IT) infrastructure, currently at $5.7 billion yearly, has grown dramatically; the VHA also spends around $800 million yearly on research and development in disease-specific areas (e.g., substance abuse, chronic disease), coordination of care, quality measurement and improvement, access to care, and veteran-specific concerns (e.g., suicide, homelessness) (Department of Veterans Affairs 2020). Outside of the VA, the US health care system is marked by a high level of complexity involving multiple private and public (federal, state, and local) parties. The US spends more on health care per capita than any other country—50% greater than the second-highest country, Norway—but has lower life expectancy than most other high-income countries (Rice et al. 2013). Compared to other highincome countries, the private sector plays a greater role in the US health care system, and health care financing in the US is largely fee-for-service. Despite a large and well-trained workforce, as well as advanced institutions providing secondary and tertiary care, experts have noted poor coordination of care and strikingly low adoption of health IT (Cebul et al. 2008; Cutler 2010). Prior to the Affordable Care Act (ACA), only 1.5% of US hospitals maintained a comprehensive EHR system (Jha et al. 2009). In the wake of the ACA, federal policies have attempted to spur care coordination and health IT adoption in the private sector (Blumenthal 2010; Greaney 2011). 2.2 Comparing VA and Non-VA Care Over the past decade, lawmakers have enacted major reforms that allow veterans to receive VA-funded care at private facilities (113th Congress 2014; 115th Congress 2018).5 These reforms broaden the 5 There have been additional well-funded efforts to shift care further into the private sector (Rein et al. 2018; Kefe 2018; Shulkin 2018; Gordon 2019). According to an official recommendation to the congressionally established Commission on Care, some have even proposed that “if veteran choice dictates it over time, the long term goal of the transformation is the total transition to community care” (Blom 2016). 6 role of the VA to that of an insurer of care for veterans (similar to the role of Medicare for the elderly), with concomitant functions such as authorizing care, processing claims, and ex post monitoring of claims for waste and fraud. Related to these initiatives, the quality of VA vs. non-VA care has been a longstanding subject of interest to policymakers and researchers. The health services literature has documented that the VA provides care that is of the same or higher quality than that of the private sector, as measured by a wide variety of process measures and health outcomes.6 However, these comparisons are potentially confounded by differences, due to eligibility and self-selection, between the populations that utilize care in the VA and in non-VA facilities. Indeed, the vast majority of existing research has compared the care of veterans in the VA with the care of non-veterans in non-VA facilities.7 We use two key ideas to extend the literature on comparisons between VA and non-VA care. First, we focus on dually eligible veterans who are aged 65 and older. These veterans may receive care in the VHA and at non-VA hospitals using Medicare (Hynes et al. 2007). Second, we build on the ambulance design strategy of Doyle et al. (2015) to sidestep concerns about the endogenous selection of where to obtain care. Specifically, we study veterans who arrive at a hospital via a 911dispatched ambulance, comparing veterans from the same zip code who could have obtained services from different ambulance companies with different propensities to transport patients to a VA hospital. Importantly, Doyle et al. (2015) document that the company dispatched to serve a given patient may be chosen independently of the patient’s characteristics, due to rotational assignment, direct competition between available providers, or software that may consider the placement of available ambulance units at the time of the 911 call (Chiang et al. 2006; Ragone 2012). Ambulance companies also exhibit different tendencies to transport patients to different hospitals, based on their ownership, headquarter location, and other characteristics (Skura 2001). We further describe our quasi-experimental design and assess its assumptions in Section 3. 2.3 Data We use data from two main sources—Medicare claims and VHA administrative data—for the universe of enrolled veterans in the VHA from the years 2000 to 2014. We observe all Medicare claims 6 See Shekelle et al. (2010), Trivedi et al. (2011), and O’Hanlon et al. (2017) for systematic reviews. The literature includes dozens of studies on hundreds of quality of care process measures, as well as several studies on health outcomes. 7 Two studies are noteworthy for having better identification. Nuti et al. (2016) compare outcomes for veterans in VA hospitals with outcomes for non-veterans in non-VA hospitals but restrict comparisons between VA and non-VA hospitals in the same metropolitan statistical areas. In an older study, Wright et al. (1999) look at 47,598 dually eligible veterans with a myocardial infarction. These studies find no difference or slightly better mortality outcomes in VA hospitals. Of note, a related literature suggests that veterans generally have poorer health than non-veterans (e.g., Agha et al. 2000). 7 for any dually enrolled veteran. These claims data include the beneficiary’s zip code and demographic information (age, race, and gender), as well as a record of medical services, each defined by an encounter date, Current Procedural Terminology (CPT) code(s), diagnostic (International Classification of Diseases, Ninth Revision, or ICD-9) codes, and provider identity. On the VHA side, we have a complete record of clinical encounters in the electronic health record system that we transform into a corresponding set of encounter dates, CPT codes, ICD-9 codes, and provider identities.8 We begin by selecting ambulance ride events for dually eligible veterans, as recorded in the Medicare claims.9 We restrict attention to “lights and sirens” emergency ambulance rides that originate from 911 dispatch calls.10 As in Doyle et al. (2015), we extract the date of the ambulance ride and the identity of the ambulance company, based on its tax identification number (TIN). We use the ambulance company identity to develop our instrumental variable for the propensity of the ambulance company to deliver patients to the VA or to non-VA hospitals. We also extract information on interventions provided by the ambulance (e.g., intravenous fluids, intubation), the level of care (advanced life support or basic life support), the pick-up location (e.g., private residence, nursing home, skilled nursing facility, accident), and the ambulance diagnosis (ICD-9) codes assigned by the ambulance personnel. We then link these ambulance rides to emergency department (ED) visits at VA and non-VA hospitals. This constitutes our main treatment of interest. For each patient we collect information on medical conditions and outpatient, ED, and inpatient utilization over the prior year, as recorded in the Medicare claims and VHA records. We use the ICD-codes for past medical conditions to identify 31 Elixhauser indices (Elixhauser et al. 1998) of comorbidities, noting the source of each condition (i.e., from visits to the VA, to non-VA facilities, or both). These comorbidities range from common conditions such as hypertension to rarer conditions such as lymphoma. Our main outcome measure is mortality. We obtain information on exact date of death from three sources: records of inpatient deaths from the VA and Medicare claims; records of death from the Veterans Benefits Administration (VBA), and records of death from the Social Security Administration (SSA). The latter two sources are particularly reliable, as they determine whether the veteran will re8 The VHA system includes patient home address information. However, we use the zip code information from the Medicare claims records as our source of home location, since this information is updated frequently and has been widely used in previous studies, including Doyle et al. (2015). 9 VHA policy is that patients with outside insurance should have ambulance services paid for by that insurance. In our dually eligible population, therefore, ambulance rides will be recorded in the Medicare claims. 10 We select ambulance rides with Healthcare Common Procedure Coding System (HCPCS) codes A0322, A0328, A0330, A0362, A0368, A0370, A0427, A0429, A0433, or Q3019. We restrict to modifiers “RH”, “SH”, “NH”, and “EH”, corresponding to rides to a hospital from a residential location, a scene of an accident or acute event, a skilled nursing facility, and an extended care facility, respectively. 8 ceive payments from either the VBA or the SSA, and draw on reports from family, funeral directors, post offices, financial institutions, other federal agencies, and state vital records agencies. To construct our main analytical sample of 401, 319 ambulance rides, we make the following restrictions (see Appendix Table A.1). First, we remove patients who live in zip codes more than 20 miles away from either the nearest VA hospital or the nearest non-VA hospital. We also drop patients who traveled more than 50 miles from their zip code to the hospital. Second, we require that patients live in zip codes served by at least two ambulance companies with at least 20 rides, at least 5% of rides transported to a VA hospital, and at least 5% transported to a non-VA hospital. Finally, for our baseline analysis of VA vs. non-VA care, we drop veterans with no VA primary, ED, or inpatient care in the prior year, since fewer than 1% of these veterans are ever transported to the VA. In our secondary analysis of veterans who may be transported to modal or non-modal hospitals outside of the VA, in Section 5, we study an analogous sample of 1, 414, 217 ambulance rides of veterans who did not use VA care in the previous year and who live in zip codes with at least two non-VA hospitals within 20 miles. Appendix Table A.14 describes the selection process for this sample. Table 1 describes the characteristics of the veterans and the episode associated with their ambulance ride at different steps in the creation of the main analytical sample. The average 28-day mortality rate is stable across steps and relatively high, between 9.7 and 11.5 p.p., reflecting the illness acuity of elderly veterans who arrive by 911-dispatched ambulance. Similarly, the weekend rate, which is the proportion of ambulance rides arriving on a weekend day, is remarkably stable and close to two-sevenths, which reflects the unplanned nature of these health events (Card et al. 2009). The main impact of our sample restrictions is to increase the share of rides going to a VA hospital. In some of the steps, such as the step imposing zip-code distance restrictions, the sample becomes more concentrated in urban areas with shorter distances to nearby VA and non-VA hospitals. Black veterans also comprise a larger share of the sample. Patient characteristics otherwise remain stable across sample restriction steps. 3 3.1 Benchmark Analysis Quasi-Experiment Following Doyle et al. (2015), our main empirical strategy relies on the assignment of ambulances to patients in emergencies and the role of ambulance companies in determining the hospital that a patient is taken to for emergency care. As detailed in Doyle et al. (2015), several companies typically 9 serve the same narrow geographic area. The assignment of a particular company may be quasiexperimentally determined such that the identity of the assigned ambulance company is plausibly unrelated to patient characteristics. Furthermore, ambulance companies exhibit “preferences” for delivering patients to certain hospitals, due to their ownership or the location of their operations. We define conditioning sets within which ambulance assignment may be as good as random. First, we condition on the origin zip code 𝑧 (𝑖) of ambulance ride 𝑖, so that we compare patients from the same zip code who are picked up by different ambulance companies. Second, we categorize the ambulance by whether it offers advanced life support (ALS) or basic life support (BLS) based on ambulance Healthcare Common Procedure Coding System (HCPCS) codes. We further categorize rides by the pickup site category (e.g., residential address, nursing home, scene of an accident), the day of the week, and month-year interactions (e.g., January 2010). Finally we condition on measures of the patient’s primary care, ED, and inpatient utilization over the past year at VA and non-VA facilities.11 For simplicity we refer to the joint set of controls for the type of ambulance, pickup site, date of pickup, and patient prior utilization as X0𝑖 . Unlike Doyle et al. (2015), we do not include patient demographics, prior medical conditions, or ambulance diagnoses in the set of baseline controls. Instead, we “hold out” these variables—many of which are highly predictive of mortality—and show that they are balanced across local ambulance
companies with differing propensities to send patients to the VA, conditional on 𝑧(𝑖), X0𝑖 . Our treatment of interest is delivery to a VA hospital, which we denote by the indicator 𝐷 𝑖 ∈ {0, 1} for ambulance ride 𝑖. Transfers are rare in our sample.12 Ride 𝑖 is provided by company 𝑗 (𝑖) ∈ J𝑧 (𝑖) ,where J𝑧 is the set of companies that serve zip code 𝑧.13 Associated with each ride and company is a potential treatment indicator 𝐷 𝑖 ( 𝑗); thus 𝐷 𝑖 = 𝐷 𝑖 ( 𝑗 (𝑖)). Our main outcome is 28-day mortality of the patient, denoted by 𝑌𝑖 ∈ {0, 1}. The associated potential outcomes, 𝑌𝑖 (𝑑), depend on whether the patient was transported to a VA hospital (𝑑 = 1) or not (𝑑 = 0), with 𝑌𝑖 = 𝑌𝑖 (𝐷 𝑖 ). Under the assumptions that different ambulance companies have systematically different tendencies to transport patients to the VA, and that the assignment of 𝑗 (𝑖) is as good as random, conditional
on 𝑧 (𝑖) , X0𝑖 , the identity of the ambulance company can be used to construct a a valid instrumental 11 The latter set of prior utilization measures may capture ambulance service areas within large zip codes, which may in turn account for correlations between prior use of VA vs. non-VA care and the identity of ambulance companies. 12 Of the 132,535 rides that go to the VA, 828 (0.6%) have a non-VA hospital ED visit on the subsequent day. Of the 268,784 rides to a non-VA hospital, 2,191 (0.8%) have an ED visit at a VA hospital the next day. Of 79,684 VA admissions, 418 (0.5%) were transferred to a non-VA hospital within 7 days of the ambulance ride. Of 157,682 non-VA admissions, 1,774 (1.1%) were transferred to a VA facility within 7 days. 13 We define an “ambulance company” as the interaction between an ambulance company tax identification number (TIN) and the health referral region (HRR) of the ride. This accounts for a few large corporations with a single TIN that serve multiple regions. 10 variable for 𝐷 𝑖 . More formally, we consider the following conditions for IV validity (Imbens and Angrist 1994): Condition 1 (IV Validity). For a random sample of ambulance rides 𝑖 provided by ambulance companies 𝑗, the following conditions hold:   (i) Relevance: 𝐸 𝐷 𝑖 ( 𝑗) 𝑧 (𝑖) , X0𝑖 is a nontrivial function of 𝑗 ∈ J𝑧 (𝑖) . (ii) Independence and Exclusion: The vector of potential outcomes, (𝑌𝑖 (0) ,𝑌𝑖 (1) , 𝐷 𝑖 ( 𝑗)), is inde
pendent of the assigned ambulance company, 𝑗 (𝑖) ∈ J𝑧 (𝑖) , conditional on 𝑧 (𝑖) , X0𝑖 .
(iii) Monotonicity: Conditional on 𝑧 (𝑖) , X0𝑖 , for any 𝑗 and 𝑗 0, 𝐷 𝑖 ( 𝑗) ≥ 𝐷 𝑖 ( 𝑗 0) for all 𝑖, or 𝐷 𝑖 ( 𝑗) ≤ 𝐷 𝑖 ( 𝑗 0) for all 𝑖. Because only veterans are typically treated in VA hospitals, the compliers in this quasi-experiment (i.e., the set of rides 𝑖 such that 𝐷 𝑖 ( 𝑗) > 𝐷 𝑖 ( 𝑗 0) for some 𝑗, 𝑗 0 ∈ J𝑧 (𝑖) ) will involve patients who can state to an ambulance that they are a veteran and who would be open to care at the VA. We estimate and report complier characteristics in Section 5.1. As is standard in the judges-design literature (e.g., Kling 2006, Dahl et al. 2014), to deal with finite samples, we construct a leave-out (or jackknife) instrumental variable that reflects the propensity of the ambulance company 𝑗 (𝑖) assigned to ride 𝑖 to transport other patients to the VA. We compute this as the average fraction of other patient who were picked up by company 𝑗 (𝑖) and went to the VA. Specifically, for ambulance ride 𝑖 transporting patient 𝑘 (𝑖) we define the leave-out probability 𝑍𝑖 of transport to the VA: 𝑍𝑖 = ∑︁ 1 (𝑘 (𝑖 0) ≠ 𝑘 (𝑖)) 𝐷 𝑖0 , 𝐾 𝑗 (𝑖) − 1 0 𝑁 𝑘 (𝑖0), 𝑗 (𝑖) 1 (1) 𝑖 ∈I𝑗 (𝑖) where 𝐾 𝑗 is the total number of patients transported by company 𝑗 , 𝑁 𝑘, 𝑗 is the total number of rides taken by patient 𝑘 with company 𝑗, and I𝑗 is the set of rides transported by ambulance company 𝑗. We estimate 𝑍𝑖 using the sample of dually eligible veteran ambulance rides (Column 1 of Table 1).
Under Condition 1, an IV estimate based on 𝑍𝑖 , conditioning on 𝑧 (𝑖) , X0𝑖 , recovers a local average treatment effect (LATE) of the VA on mortality among compliers. For comparison, we also consider the observational “treatment effect” of going to the VA on mortality of patients who arrive
at hospital in a 911-dispatched ambulance, controlling for 𝑧 (𝑖) , X0𝑖 : 𝑌𝑖 = 𝛽𝐷 𝑖 + X0𝑖 𝛿0 + 𝜁0,𝑧 (𝑖) + 𝜀0,𝑖 11 (2) where 𝜁0,𝑧 represents an unrestricted fixed effect for rides originating in zip code 𝑧. Estimating Equation (2) by OLS yields 𝛽ˆ𝑂𝐿𝑆 , while instrumenting 𝐷 𝑖 with 𝑍𝑖 yields 𝛽ˆ𝐼 𝑉 . The gap between 𝛽ˆ𝑂𝐿𝑆 and 𝛽ˆ𝐼 𝑉 will depend on differences in the potential outcomes between never takers (i.e., patients who go to a non-VA facility regardless of the ambulance company) and always takers (i.e., patients who go to the VA regardless of the ambulance company), as well as on differences in treatment effects between compliers and non-compliers. We explore this gap more directly in Section 4. 3.2 First Stage, Balance, and Reduced Form We begin our empirical analysis by demonstrating instrument relevance, Condition 1(i), with the following first-stage regression: 𝐷 𝑖 = 𝜋1 𝑍𝑖 + X0𝑖 𝛿1 + 𝜁1,𝑧 (𝑖) + 𝜀 1,𝑖 . (3) The coefficient 𝜋1 reflects the impact of ambulance company preferences on the probability that the ride goes to the VA, conditional on our baseline controls for ambulance type, pickup site, zip code, date categories, and veteran prior utilization. Figure 1, Panel A, shows a binned scatterplot of residualized 𝐷 𝑖 on the y-axis with respect to residualized 𝑍𝑖 on the x-axis and reports 𝜋ˆ 1 = 0.882 (s.e. 0.034). The first-stage relationship between 𝐷 𝑖 and 𝑍𝑖 is very predictive and close to linear. To assess independence, Condition 1(ii), we test whether 𝑍𝑖 is correlated with patient characteristics that are correlated with mortality. Specifically, we construct an estimate of predicted mortality 𝑌ˆ𝑖 using “hold-out” patient characteristics of patient demographics and 31 Elixhauser indices for prior medical conditions.14 We then fit models for 𝑌ˆ𝑖 based on the same right-hand-side specification as in Equation (3). Panel B of Figure 1 shows (with hollow dots) that there is no relationship between
𝑌ˆ𝑖 and 𝑍𝑖 , controlling for 𝑧 (𝑖) , X0𝑖 . In contrast, the same panel shows (with solid dots) that the reduced-form relationship between actual mortality, 𝑌𝑖 , and 𝑍𝑖 is significantly negative, under the same controls. Specifically, for the reduced-form relationship, 𝑌𝑖 = 𝜋2 𝑍𝑖 + X0𝑖 𝛿2 + 𝜁2,𝑧 (𝑖) + 𝜀2,𝑖 , 14 Patient (4) demographics include age, gender, and race and ethnicity. Age is captured by two-year age bins from 65 years to 100 years. Race and ethnicity is captured with three dummies for white, Black, and Hispanic; the omitted category is Asian/other. We use the 31 Elixhauser indices as described in Elixhauser et al. (1998), interacting each index with the source of the record indicating the comorbidity. There are three possible sources: VA only, Medicare claims only, and VA and Medicare claims. This results in 3 × 31 = 93 dummies. These hold-out patient characteristics are described in Appendix Table A.3. 12 we find 𝜋ˆ 2 = −0.040 (s.e. 0.016). This suggests that quasi-random assignment to an ambulance company more likely to transport to the VA results in an intention-to-treat reduction in mortality. The exclusion condition in Condition 1(ii) asserts that ambulance companies do not affect outcomes other than through their effect on whether a patient arrives at a VA or non-VA hospital. Implicitly, however, our notation also assumes that each complier has a well-defined non-VA hospital that is stable across ambulance companies. In Appendix A.1.1, we evaluate the robustness of our results to potential violations of the exclusion condition. Specifically, we assess and find no evidence of any correlation between 𝑍𝑖 and ambulance treatments captured in summary charges or between 𝑍𝑖 and ambulance propensities to deliver patients to different non-VA hospitals.15 To assess the monotonicity assumption given by Condition 1(iii), we follow the standard practice in the judges-design literature to show that the first-stage relationship between 𝐷 𝑖 and 𝑍𝑖 remains positive for subgroups of patients defined by different observable characteristics (e.g., Arnold et al. 2018; Bhuller et al. 2020). We detail these analyses in Appendix A.1.2. In Section 4, we present a stronger test of monotonicity (and IV validity) based on potential outcomes. Following the reasoning in Kitagawa (2015), this test amounts to showing a positive density for the potential outcome of death in a given week among compliers. 3.3 Mortality Effect With this background, we now move to our main results on patient mortality. In Table 2, we show both OLS and IV estimation results for Equation (2). Panel A of the table shows 𝛽ˆ𝑂𝐿𝑆 from Equation (2), while Panel B shows 𝛽ˆ𝐼 𝑉 = 𝜋ˆ 2 /𝜋ˆ 1 from the first-stage and reduced-form regressions in Equations (3) and (4). Column 1 shows our baseline specification, controlling for zip code and the variables in X0𝑖 . The OLS estimate is 𝛽ˆ𝑂𝐿𝑆 = −0.024 (s.e. 0.001), while the IV estimate is 𝛽ˆ𝐼 𝑉 = −0.045 (s.e. 0.018).16 Relative to the mean 28-day mortality of 9.7 p.p., both estimates imply a sizeable reduction in mortality for compliers who are taken to the VA. The other columns in Table 2 show OLS and IV estimates as we include additional controls to the models: (i) patient demographics (age, race, gender), (ii) ambulance diagnostic (ICD-9) codes, (iii) 15 Following Kolesar et al. (2015), these analyses correspond to the weaker assumption that there are no systematic correlations between our instrument and other ambulance-specific treatments that impact our outcome of interest. Specifically,  under this weaker version of exclusion, we require that ambulance companies with higher values of 𝐸 𝐷 𝑖 ( 𝑗) 𝑧 (𝑖) , X0𝑖 do not also systematically apply treatments during the ambulance ride that affect mortality, or systematically deliver patients to higher- or lower-quality non-VA alternatives. 16 Appendix Figure A.1 shows the IV estimate visually, by plotting the predicted first-stage probability of treatment from Equation (3) on the x-axis and predicted reduced-form effect on mortality from Equation (4) on the y-axis. The slope of this visual IV relationship corresponds to 𝛽ˆ 𝐼 𝑉 = −0.045. 13 Elixhauser comorbidity indicators, and (iv) ambulance and co-rider controls, which are all described in Appendix Table A.3. The latter controls are meant to capture any unobservable patient selection at the ambulance company level by using characteristics of other rides and patients under the same ambulance company, following the reasoning in Altonji and Mansfield (2018). Specifically, these controls address the potential concern that sicker patients tend to be allocated to certain ambulance companies that may be more or less likely to take patients to the VA. Reassuringly, both 𝛽ˆ𝑂𝐿𝑆 and 𝛽ˆ𝐼 𝑉 remain stable as we add additional controls. Figure 2 illustrates this stability as controls are added in a more granular fashion, and Appendix Figure A.2 shows stability of the IV estimates as we permute the order in which the extra controls are added. The stability of both the OLS and IV estimates suggests lack of selection on observable characteristics; under the reasoning of Altonji et al. (2005), this stability suggests limited scope for selection on unobservable characteristics that predict potential 28-day mortality. However, IV estimates are larger than OLS estimates, suggesting either that never takers are healthier than always takers (i.e., selection runs counter to treatment effects on mortality) or that the LATE is larger than the unconditional average treatment effect (ATE).17 We investigate these possibilities in the next section and in Section A.4. 4 Survival Analysis In this section we develop and apply a survival analysis framework to understand the dynamics of potential survival outcomes following the ambulance ride. We use this framework to draw several insights. First, we determine the time course of VA effects on mortality. Second, we use the empirical results of this framework to provide further validation of Condition 1, beyond the standard benchmark analysis in Section 3.2. Third, we investigate the implications of heterogeneity in mortality risks between compliers and non-compliers of our ambulance quasi-experiment. 4.1 Approach Consider a set of potential survival outcomes 𝑆𝑖 (𝑡; 𝑑) ∈ {0, 1} under VA care (𝑑 = 1) and non-VA care (𝑑 = 0) for each week 𝑡 ∈ {1, . . . , 52} following the ambulance ride.18 By definition, if 𝑆𝑖 (𝑡; 𝑑) < 17 We note that a Hausman test for equality of the two estimates has a t-statistic of only 1.0, so based on this evidence alone, the gap between OLS and IV could be simply due to sampling error. In the next section, however, we show a dynamic pattern of IV and OLS estimates, over the year after the initial ambulance ride, that points more definitively to systematic differences. That is, using additional data over time, we infer with high confidence that the causal VA advantage is larger than the (precisely estimated) OLS effect. 18 We adopt the convention that a death within the first 7 days is a death in week 1. Thus a death within 28 days is a death by the end of week 4. 14 𝑆𝑖 (𝑡 − 1; 𝑑), then the patient in ambulance ride 𝑖 would die in the 𝑡 th week following the ambulance ride if exposed to treatment 𝑑. Of course, potential survival outcomes must weakly decrease over time, i.e., 𝑆𝑖 (𝑡; 𝑑) ≤ 𝑆𝑖 (𝑡 − 1; 𝑑) for all 𝑖, 𝑑, and 𝑡. As with mortality outcomes, for each ambulance ride 𝑖, we can only observe the set of survival outcomes corresponding to 𝑑 = 𝐷 𝑖 : 𝑆𝑖 (𝑡) = 𝐷 𝑖 𝑆𝑖 (𝑡; 1) + (1 − 𝐷 𝑖 ) 𝑆𝑖 (𝑡; 0). However, appealing to Abadie (2002), we can recover the expected survival outcomes for the set of compliers C whose hospital choice depends on which ambulance company picks them up. In particular, under Condition 1, we can estimate 𝑠 𝐼 𝑉 (𝑡; 1) ≡ 𝐸 [ 𝑆𝑖 (𝑡; 1)| 𝑖 ∈ C] by two-stage least squares using the first-stage Equation (3) and a reduced-form equation similar to Equation (4) but with dependent variable 𝑆𝑖 (𝑡) 𝐷 𝑖 . Similarly, we can estimate 𝑠 𝐼 𝑉 (𝑡; 0) ≡ 𝐸 [ 𝑆𝑖 (𝑡; 0)| 𝑖 ∈ C] using the same first stage model but replacing the reduced-form outcome variable in Equation (4) with 𝑆𝑖 (𝑡) (𝐷 𝑖 − 1). Note that by construction, the IV estimand of the VA treatment effect on 28-day mortality in Section 3, satisfies 𝛽 𝐼 𝑉 = 𝑠 𝐼 𝑉 (4; 1) − 𝑠 𝐼 𝑉 (4; 0) . Given the potential survival outcomes, we can then estimate potential hazard rates for mortality, under VA and non-VA assignment: ℎ 𝐼 𝑉 (𝑡; 𝑑) ≡ 𝐸 [ 1 − 𝑆𝑖 (𝑡 + 1; 𝑑)| 𝑆𝑖 (𝑡; 𝑑) = 1, 𝑖 ∈ C] 𝑠 𝐼 𝑉 (𝑡; 𝑑) − 𝑠 𝐼 𝑉 (𝑡 + 1; 𝑑) , = 𝑠 𝐼 𝑉 (𝑡; 𝑑) (5) for 𝑑 ∈ {0, 1} and 𝑡 ∈ {1, . . . , 52} , corresponding to weekly mortality hazard rates up to one year after the initial ambulance ride. Under Condition 1, differences between {ℎ 𝐼 𝑉 (𝑡; 1)}𝑡 and {ℎ 𝐼 𝑉 (𝑡; 0)}𝑡 can be interpreted as the causal effect of VA assignment, among compliers, on the set of mortality hazard rates.19 As in Section 3, we calculate risk-adjusted OLS survival functions and mortality hazard rates, conditional on 𝐷 𝑖 . We estimate 𝑠𝑂𝐿𝑆 (𝑡; 𝑑) ≡ 𝐸 [ 𝑆𝑖 (𝑡; 𝑑)| 𝐷 𝑖 = 𝑑] = 𝐸 [ 𝑆𝑖 (𝑡)| 𝐷 𝑖 = 𝑑] by OLS, replacing the outcome variable in Equation (2) with 𝑆𝑖 (𝑡) 𝐷 𝑖 for 𝑠𝑂𝐿𝑆 (𝑡; 1) and with 𝑆𝑖 (𝑡) (𝐷 𝑖 − 1) for 𝑠𝑂𝐿𝑆 (𝑡; 0). Our OLS estimand of the VA effect on 28-mortality, 𝛽𝑂𝐿𝑆 , is similarly equal to 𝑠𝑂𝐿𝑆 (4; 1) − 𝑠𝑂𝐿𝑆 (4; 0). Corresponding mortality hazard rates can also be calculated based on ob19 We emphasize that any gap between ℎ 𝐼 𝑉 (𝑡, 1) and ℎ 𝐼 𝑉 (𝑡, 0) at some later time horizon (e.g., 𝑡 = 12) could arise because treatments at the VA affected the population of compliers who survive to week 𝑡 − 1 and are therefore at risk of death in week 𝑡, or because of a treatment effect on the week 𝑡 hazard, holding the population fixed. 15 served risk-adjusted survival: ℎ𝑂𝐿𝑆 (𝑡; 𝑑) ≡ 𝐸 [ 1 − 𝑆𝑖 (𝑡 + 1; 𝑑)| 𝑆𝑖 (𝑡; 𝑑) , 𝐷 𝑖 = 𝑑] 𝑠𝑂𝐿𝑆 (𝑡; 𝑑) − 𝑠𝑂𝐿𝑆 (𝑡 + 1; 𝑑) = . 𝑠𝑂𝐿𝑆 (𝑡; 𝑑) (6) Compared to the potential survival functions and mortality hazards, the OLS analogues also incorporate outcomes for the always takers and never takers whose choice of hospital is unaffected by the specific ambulance company that picked them up. Specifically, 𝑠𝑂𝐿𝑆 (𝑡; 1) and ℎ𝑂𝐿𝑆 (𝑡; 1) reflect survival outcomes for a combination of always takers and compliers, while 𝑠𝑂𝐿𝑆 (𝑡; 0) and ℎ𝑂𝐿𝑆 (𝑡; 0) reflect survival outcomes for a combination of never takers and compliers. 4.2 Time Course of Mortality Effects Since we examine potential survival outcomes one year after an ambulance ride, for the analysis in this section we restrict analysis to ambulance rides of patients with no prior ride within one year.20 Figure 3 shows the estimated potential survival curves and potential hazard rates in weeks 0 to 52 for compliers assigned to the VA and those assigned to a non-VA hospital. The potential survival curves, shown in Panel A, reveal a high risk of mortality among compliers. Mortality at 28 days among compliers assigned to a non-VA hospital is greater than the sample mean of 9.7 p.p., and cumulative mortality at one year is approximately 30 p.p. However, despite the substantial mortality risk over the subsequent year, the gap in survival between VA- and non-VA-assigned compliers (i.e., the mortality treatment effect) is fully realized at 28 days and remains stable for the rest of the year. In Panel B, we examine the implied hazard rates and show that the differences in mortality are concentrated in the first week following the ambulance ride. Thereafter, though the hazard rates for both VA- and non-VA-assigned compliers remain relatively high, they are indistinguishable from each other. This similarity suggests that the VA advantage results entirely from events within the first week following the ambulance ride. The potential hazard profiles in Figure 3 suggest that mortality risks for the compliers in our 20 This restriction attributes survival for a given patient in a given week to the “upstream” ambulance ride, rather than attributing the survival event to both upstream and downstream ambulance rides. This changes (decreases) the sample in Appendix Table A.1 to 254, 782 rides and 188, 299 patients. In Appendix Figure A.9, we show that this restriction (or any other restriction on prior rides) does not lead to qualitative differences in our estimated OLS or IV treatment effects on mortality over time. Qualitatively, regardless of the number of days within which we require no prior ride, the IV estimates are larger than 4 p.p. at 28 days and remain mostly stable within the year following the ambulance ride; the OLS estimates are between 2.0 and 2.5 p.p. at 28 days and essentially disappear by one year following the ambulance ride. We evaluate the implications of the long-term difference between IV and OLS treatment effects in Section 4.4. 16 analysis comprise two separate risks: (i) a relatively high short-term risk component that is affected by VA vs. non-VA assignment, and (ii) a relatively stable long-term risk component that is the same between compliers who go to the VA and those who go to a non-VA hospital. If the latter risk reflects underlying patient health and is independent of the risk that led to the ambulance call, then we would expect the long-run weekly mortality rate (after, e.g., three months) to be the same for veterans who were quasi-randomly assigned to VA and non-VA hospitals; we formalize this as a test in Section 4.3. The potential hazard rates also allow us to assess whether excess mortality at non-VA hospitals involves “harvesting,” or mortality displacement, in which deaths for patients at the VA are simply delayed (Schwartz 2000; Honore and Lleras-Muney 2006). Under this hypothesis, survival gains from VA care observed at 28 days are temporary and will fade in the long-term. Such mortality displacement would imply that the hazard of dying increases among VA-assigned compliers after a time. We find no evidence of this in the potential hazard rates in Panel B of Figure 3. In Appendix A.2, we formally test that ℎ 𝐼 𝑉 (𝑡; 1) ≤ ℎ 𝐼 𝑉 (𝑡; 0) for all 𝑡 and cannot reject this null hypothesis of no harvesting.21 This suggests that the VA prevents rather than displaces deaths, leading to a persistent survival benefit, as shown by the stable gap between potential survival curves in Panel A. 4.3 Extended IV Validity We can also use the estimated potential survival outcomes to test the validity of our IV strategy based on ambulance assignment. Under Condition 1, the density of any characteristic, including characteristics defined by potential outcomes, must be positive among compliers of the quasi-experiment (Balke and Pearl 1997; Imbens and Rubin 1997): Pr ( 𝑋𝑖 = 𝑥,𝑌𝑖 = 𝑦| 𝑖 ∈ C) ≥ 0, (7) for all possible characteristics 𝑥 ∈ X and all possible potential outcomes 𝑦 ∈ Y. Kitagawa (2015) proposes a formal test of this implication, and Chan et al. (2019) show that applying this test to potential outcomes can provide a stronger test of the conditions for IV validity, particularly the monotonicity assumption in Condition 1(iii).22 21 Our test builds on the suggestion of Wolak (1987) to form a test statistic based on a quadratic form that represents the deviations of the data from the predictions of a constrained model that imposes the inequality restrictions. We use a simple bootstrap procedure to derive critical values of the test. 22 Specifically, testing Equation (7) with respect to potential outcomes 𝑦 ∈ Y may be more likely to detect violations of Condition 1 than standard tests of monotonicity, focusing on patient characteristics, that we employ in Appendix A.1.2. The intuition behind this is that testing Equation (7) with respect to potential outcomes will reveal violations in Condition 1 that relate not only to observed patient characteristics but also to unobserved patient characteristics correlated with potential 17 In our setting, we partition survival potential outcomes into weeks of potential mortality for 52 weeks following the ambulance ride, for both VA- and non-VA-assigned compliers. Since survival can only decrease over time, the potential mortality hazard rates for any week must be positive (i.e., ℎ 𝐼 𝑉 (𝑡; 𝑑) ≥ 0 for all 𝑡 ∈ {0, . . . , 51}, 𝑑 ∈ {0, 1}). This prediction may be violated if patient risk of death in some week 𝑡 is correlated with the assigned ambulance’s propensity to take patients to the VA (a violation of independence), or if there exist “defiers” (i.e., patients that are less likely to go to the VA when assigned to ambulances that transport more often to the VA overall) with higher risk of death in some week 𝑡 (a violation in monotonicity). In Appendix A.2, we formally test the joint inequality constraint that ℎ 𝐼 𝑉 (𝑡; 𝑑) ≥ 0 for all 𝑡 ∈ {0, . . . , 51}, 𝑑 ∈ {0, 1}, and cannot reject this null hypothesis, with a bootstrap-based p-value of 1.00. If the short-term and longer-term mortality risks facing veterans are independent (as is typically assumed in a competing risks model) and treatment at the VA only affects the short-term risk component, then Condition 1 also implies that ℎ 𝐼 𝑉 (𝑡; 1) = ℎ 𝐼 𝑉 (𝑡; 0) for 𝑡 ≥ 𝑡, for some 𝑡 after the acute ambulance episode. Specifically, if the short-term risk component disappears after some time 𝑡, and if the assignment of compliers to VA and non-VA hospitals is as good as random, then the death rates of the two groups of compliers should be the same after 𝑡. Visually, it appears that the potential hazard rates of the compliers are very similar in weeks 𝑡 ∈ {1, . . . , 51}. Consistent with this impression, in Appendix A.2, we show that we cannot reject that ℎ 𝐼 𝑉 (𝑡; 1) = ℎ 𝐼 𝑉 (𝑡; 0) for all weeks 𝑡 ≥ 1, with a bootstrap-based p-value of 0.31. 4.4 Heterogeneity in Mortality Risks Finally, we take a closer look at death rates during the year after the ambulance ride to better understand the differences between our main OLS and IV estimates of the VA advantage. As shown in Panel A of Figure 4, we find that, remarkably, OLS survival curves cross about nine to ten months after the ambulance ride. This reflects a reversal in the sign of the OLS-estimated VA treatment effect: While patients arriving at the VA experience an immediate survival benefit that peaks at 14 days after the ambulance ride, the survival benefit eventually reverses, such that patients arriving at the VA are more likely to die within a year. Consistent with this observed survival pattern, Panel B of Figure 4 reveals a cross-over in the observed hazard rates of death for patients who are taken to VA and non-VA hospitals. In the first outcomes. Violations in quasi-random assignment or monotonicity may be more likely to occur along potential outcomes if agents act according to an objective function based on potential outcomes. 18 week after the ambulance ride, the death rate is lower for patients at the VA, though the gap between the VA and non-VA hazards is smaller than the corresponding potential-outcomes gap for compliers shown in Figure 3. Thereafter, the hazard rate for patients at the VA, ℎ𝑂𝐿𝑆 (𝑡; 1), is consistently higher than the hazard rate for those who went to a non-VA hospital, ℎ𝑂𝐿𝑆 (𝑡; 0). This gap suggests differences in baseline risk between always takers and never takers that are initially offset by the shortterm VA advantage but reemerge, soon after the first week. These differences in baseline mortality hazards accumulate over time to generate large differences in long-term survival. To identify differences in the baseline mortality risk between VA-assigned compliers and always takers, we compare ℎ 𝐼 𝑉 (𝑡; 1) and ℎ𝑂𝐿𝑆 (𝑡; 1); to identify differences between non-VA-assigned compliers and never takers, we compare ℎ 𝐼 𝑉 (𝑡; 0) and ℎ𝑂𝐿𝑆 (𝑡; 0). In Appendix A.2 we show that we cannot reject the null hypothesis that ℎ 𝐼 𝑉 (𝑡; 1) = ℎ𝑂𝐿𝑆 (𝑡; 1) for 𝑡 ≥ 1. However, we can strongly reject the null hypothesis that ℎ 𝐼 𝑉 (𝑡; 0) = ℎ𝑂𝐿𝑆 (𝑡; 0) for 𝑡 ≥ 1. The average value of ℎ 𝐼 𝑉 (𝑡; 0), for 𝑡 ≥ 1, is significantly larger than the corresponding average value of ℎ𝑂𝐿𝑆 (𝑡; 0), for 𝑡 ≥ 1, which implies that never takers are healthier than compliers. This survival analysis shows, with substantially more precision than that afforded by the benchmark analysis in Section 3.3, that the VA advantage is in fact larger than the (precisely estimated) OLS effect would imply. 5 Mechanisms This section probes further into the mechanisms behind the large VA mortality advantage. We first examine characteristics of compliers in our quasi-experiment. Second, we use a simple Olsen (1980) control function approach to estimate the average treatment effect (ATE) and compare it with the LATE estimated in Section 3. Third, we examine treatment effect heterogeneity by hospital and patient characteristics. Fourth, we ask whether the VA produces superior health outcomes by spending more; spending less would imply mechanisms that improve productivity. Fifth, we perform an analysis of health IT and integrated care among veterans using non-VA care, as potential mechanisms that may set the VA apart from non-VA care. 5.1 Complier Characteristics We perform a standard complier analysis examining characteristics of compliers relative to the overall sample.23 Table 4 shows results for various characteristics. Compliers are more likely to be Black, 23 Specifically, we employ the same approach from Abadie (2002) that we introduced in Section 4.1. Under IV validity in Condition 1, we can estimate 𝐸 [ 𝑋𝑖 | 𝑖 ∈ C] for some characteristic 𝑋𝑖 by two-stage least squares, involving the first-stage 19 to have lower income, to have a prior VA ED visit, and to suffer from mental illness and substance abuse. Compliers have slightly fewer recorded Elixhauser comorbidities and are less likely to receive Advanced Life Support (ALS). In Appendix Table A.7, we show similar patterns comparing always takers and never takers, following an approach from Dahl et al. (2014) that we describe in Appendix A.3. Consistent with our analysis in Section 4.4, we find that VA always takers have higher predicted mortality, based on observable characteristics, than either compliers or never takers.24 Researchers and policymakers have noted greater incidence of mental health and substance abuse issues among veterans (Adamson et al. 2008). Recognizing this need, Congress allocated $152 million for increasing mental health care programming in 1999; in the following two decades, VHA stations expanded mental health services and hired thousands of mental health providers (106th Congress 1999; U.S. Government Accountability Office 2015). This capacity to treat mental health disorders contrasts with the non-VA health care sector, where mental health services have long been underfunded and underprovided (Huskamp and Iglehart 2016). 5.2 Selection Model We consider a structural model of selection, both to assess how VA treatment effects vary with a veteran’s propensity to go to the VA and to infer the ATE. Following the “marginal treatment effects” (MTE) literature (see, e.g., Heckman and Vytlacil (2007) for a review), we exploit our multivalued ambulance instrument in order to characterize the relationship between treatment effects and veterans who are induced to go to the VA. Specifically, we allow for flexibility in the returns to VA care among compliers who are induced into VA care, ranging from ambulances with low propensities to deliver patients to the VA to those with high propensities. Using a control function model, we also extrapolate this relationship in treatment effects to always takers and never takers, thereby imputing the ATE from the semiparametric structure of the model. We provide further details of our approach in Appendix A.4. We find evidence of moderate “selection on gains,” in which veterans with larger mortality reductions from going to the VA are more likely to go to the VA. In Appendix Figure A.7, we show the MTE function ranging from veterans who are most likely to use the VA to those who are least likely to use the VA. Veterans induced to go to the VA by lower-propensity ambulances have higher returns Equation (3) and a reduced-form equation replacing the outcome variable in Equation (4) with 𝑋𝑖 𝐷 𝑖 . 24 For Table 4 and Appendix Table A.7, we predict mortality using the same regression of mortality on baseline and hold-out characteristics described in Section 3.2 applied only on rides going to non-VA hospitals, in order to separate the VA advantage from coefficients used in the prediction. The intuition for this is described in Chetty et al. (2014, p. 2598). Adopting their approach of estimating coefficients on predictors while controlling for 𝐷 𝑖 yields nearly identical results. 20 to VA care than veterans who are induced by high-propensity ambulances. In Appendix Table A.8, we find a substantial ATE, only marginally smaller than the LATE, across a variety of specifications. 5.3 Heterogeneity by Hospitals and Patients We next assess heterogeneity in the VA mortality effect by hospitals in a patient’s choice set and by patient characteristics. We consider a wide range of characteristics, in three categories: (i) characteristics of non-VA hospitals serving a given zip code, weighting the hospitals by volume of rides from the zip code; (ii) characteristics of the VA hospital serving a given zip code; and (iii) patient characteristics. For each of these characteristics 𝑥, we construct a binary indicator variable, 𝐼 𝑥,𝑖 ∈ {0, 1}. For example, for the non-VA hospital characteristic of number of staffed beds, we create a binary indicator variable for whether the volume-weighted average number of staffed beds across non-VA hospitals in   a zip code is above or below the median. We include a demeaned 𝐼˜𝑥,𝑖 ≡ 𝐼 𝑥,𝑖 − 𝐸ˆ 𝑖 𝐼 𝑥,𝑖 in the following linear control function regression: 𝑌𝑖 = 𝛽 𝑥 𝐷 𝑖 + 𝜌 𝑥 𝐷 𝑖 𝐼˜𝑥,𝑖 + 𝜋 𝑥 𝐼˜𝑥,𝑖 + 𝛾 𝑥 𝜀ˆ1,𝑖 + X0𝑖 𝛿 𝑥 + 𝜁 𝑥,𝑧 (𝑖) + 𝜖 𝑥,𝑖 , (8) where 𝜀ˆ1,𝑖 is the first-stage error from Equation (3). Controlling for the endogeneity of selection, this approach yields estimates of binary heterogeneous treatment effects along several dimensions.25 Since 𝐼˜𝑥,𝑖 has a mean of 0, we can interpret 𝛽 𝑥 as the LATE, controlling for 𝐼˜𝑥,𝑖 ; 𝜌 𝑥 is the difference in the VA effect on mortality between 𝐼 𝑥,𝑖 = 1 and 𝐼 𝑥,𝑖 = 0. We calculate standard errors by bootstrap, drawing blocks of data by zip code. Table 5 shows differences in hospital characteristics between VA and non-VA hospitals. For example, VA hospitals have fewer ED visits and admissions per bed and are more likely to be teaching hospitals.26 However, we find at most modest treatment heterogeneity with respect to any of these hospital characteristics (Appendix Tables A.9 to A.11). Heterogeneity along any of the VA or nonVA hospital characteristics across zip codes is less than 20% of the main VA advantage, suggesting that the VA advantage pertains across the spectrum of VA and non-VA alternatives. Non-VA hospital characteristics in a zip code that are associated with a modestly smaller VA advantage include trauma center status and number of nurses per 1,000 patient-days. As for heterogeneity by VA hospital 25 For a discussion of this general approach, see Wooldridge (2015), Section III. 26 VA hospitals appear to have more long-term care admissions, which explains a higher average length of stay (i.e., fewer admissions for slightly larger average daily census). As shown in Table 3, the difference in length of stay is not borne out in our sample; the IV estimate of the effect on length of stay suggests that the VA reduces length of stay. 21 characteristics, the VA advantage is perhaps slightly greater for larger VA hospitals. In Appendix A.6, we describe complementary results from an empirical Bayes approach to heterogeneity in stationspecific OLS estimates of the VA advantage; in that approach, we fail to demonstrate meaningful heterogeneity in the VA advantage across VA stations. Across patients (Appendix Table A.12), the VA advantage is likely as large for minority veterans (Black and Hispanic) as for non-minority veterans. The VA survival benefit appears greater for veterans who suffer from mental illness or substance abuse, who have more prior visits at the VA, or who have higher predicted mortality. However, the differences in the VA survival benefit across all of these dimensions are still smaller than the overall scale of the survival benefit. Importantly, the VA survival benefit is not limited to select medical conditions that stereotypical users of the VA might have; even patients who are less than likely to use VA care experience a similar VA survival benefit. 5.4 Effect on Spending and Utilization In light of the important literature on the returns to spending in health care (e.g., Garber and Skinner 2008), we examine the causal effect of VA vs. non-VA care on spending. The motivation behind this analysis is similar to that in Doyle et al. (2015), who sought to understand whether higher-spending hospitals achieve better health outcomes. To perform this analysis, we rely on both internal VA cost data and Medicare payment data from claims. Internal VA cost accounting apportions costs by VA utilization data and scales the cost of each encounter so that total spending matches actual budgeted spending within each VHA station.27 On the Medicare side, we include payments made both by the veteran (i.e., coinsurance and deductible) and by the government. Therefore, we measure the cost of both VA and non-VA health care in terms of dollars spent by the government and the veteran. Using the same instrumental variables approach as in our benchmark analysis, we study the effect of VA vs. non-VA care on daily VA and non-VA spending over time since the ambulance ride. Specifically, we combine VA and Medicare spending in various weekly intervals since the ambulance ride. Table 3 further shows that the VA reduces 28-day combined spending by $2, 548, or by 21% of the mean 28-day spending. The reduction in spending reflects a lower probability of inpatient admission and fewer hospital days associated with VA care, although VA care results in slightly more outpatient visits in the following 28 days.28 We find similar results when we use an alternative 27 The apportioning is based on inputs such as Relative Value Units (RVUs) associated with CPT codes, DiagnosisRelated Group (DRG) weights, patient characteristics, and admission lengths of stay. This methodology is detailed in Wagner et al. (2003) and in Phibbs et al. (2019). 28 Although the average Medicare outpatient visit costs less than the average VA outpatient visit, the average Medicare 22 measure of spending that fixes prices for utilization in VA and non-VA settings.29 In Figure 5, we show potential cumulative spending curves during the first year and implied weekly spending rates conditional on survival, for compliers transported to a VA hospital and those transported to a non-VA hospital. Differences in cumulative spending remain throughout the end of the year. Unlike survival outcomes, differences in spending accrue until at least three months after the ambulance ride. The result that the VA saves lives while reducing spending is significant for two reasons. First, the result speaks directly to the policy question of whether the VA should privatize its care in a Medicaretype arrangement. The potential role of the VA as an insurer of private care has featured heavily in recent policy proposals and laws (113th Congress 2014; 115th Congress 2018). We show that, at least for the patients in our design, this privatization arrangement would be dominated by the status quo, as it would lead to both higher spending and worse health outcomes. Second, this joint finding suggests that the general mechanism behind the VA survival benefit is not higher spending but higher productivity. Our evidence points to productive inefficiency, rather than “flat of the curve” spending that underlies the relatively low-returns to US health care. This implication complements a growing literature on productivity differences across personnel (Chan et al. 2019; Silver 2020) and hospitals (Chandra and Staiger 2007, 2020) by showing an important productivity difference between health care systems. 5.5 Health IT and Integrated Care Our final analysis investigates the role of health IT and integrated care in explaining the VA survival advantage. A substantial literature has reported the qualitative importance of these mechanisms in the VA’s “transformation” into a high-quality health care organization in the mid-1990s (e.g., Jha et al. 2003).30 The lack of information flow and the high degree of fragmentation across providers in the US private health care sector have long been highlighted as potential roots of inefficiency (Cebul et al. inpatient day costs more than the average VA inpatient day. In our main analytic sample described in Appendix Table A.1, the average VA outpatient visit costs $181.35, and the average VA inpatient day costs $1,577.38. In the same sample, the average Medicare outpatient visit costs $108.30 (lower than the average VA cost), but the average Medicare inpatient day costs $1,811.99 (higher than the average VA cost). 29 We follow the methodology laid out in Finkelstein et al. (2016), where we impute spending based on Relative Value Units (RVUs) for procedures with CPT codes and Diagnosis-Related Group (DRG) weights for inpatient stays. We scale prices by a constant so that imputed total Medicare spending equals actual total Medicare spending. According to this spending measure, the VA reduces 28-day combined spending by $4, 036. 30 For recent qualitative research that illuminates this mechanism between VA and non-VA care, see, e.g., Nevedal et al. (2019) and Rinne et al. (2019). Outside of the VA, qualitative research also suggests important complementarities between IT and organization for producing quality care (e.g., Haque et al. 2019). 23 2008; Jha et al. 2009; Cutler 2010). These information-based mechanisms would be consistent with greater health care productivity of the VA, particularly for regular users and patients whose conditions are more responsive to informed management (e.g., substance abuse). Effective coordination of care also seems necessary to explain the large body of evidence that documents superior VA performance on a wide range process measures characterizing timeliness and appropriateness of care (e.g., O’Hanlon et al. 2017). Ideally, we would study these mechanisms by observing the VA’s implementation of health IT and its reorganization into more integrated care in the mid-1990s. However, these reforms predate our data.31 Similarly, it is not possible to examine the VA’s effect on mortality among veterans who have no prior utilization at the VA, since it is exceedingly rare for these veterans to be transported to the VA by ambulance, as shown in Appendix Figure A.9. Nevertheless, this figure also shows that veterans may utilize more than one non-VA hospital system and that veterans may or may not be transported to their modal non-VA hospital, defined as the hospital system at which they had the largest number of utilization days in the prior year. Our analytic strategy thus centers on a sample of veterans with no prior VA utilization but with some prior non-VA utilization.32 While these veterans will almost certainly be transported to a non-VA hospital, we assess mortality outcomes as a function of whether they are quasi-randomly assigned—via a similar ambulance instrument as the one we use in our benchmark analysis—to their modal non-VA hospital. This modal-hospital effect on mortality arguably captures at least some of the potential effect of continuity of care in the private sector. In order to more explicitly investigate the role of health IT and integrated care, we further exploit two changes induced by incentives in federal laws and payment policies during our study period. First, the HITECH Act of 2009 dramatically increased the share of hospitals using health IT (Blumenthal 2010).33 Second, in 2011, Medicare began to incentivize care integration via alternative payment arrangements to “Accountable 31 Indeed, the VHA’s adoption of a common health IT platform (VistA) in the mid-1990s paved the way for research on health services within the VHA system, including this study. 32 We detail the sample selection process for this analysis in Appendix Table A.13 and present patient and ride characteristics in this sample Appendix Table A.14. Since no veteran in this sample has prior VA utilization, the sample is disjoint from our benchmark sample (Appendix Table A.1). We only include zip codes with at least two non-VA hospitals within 20 miles, but we make no requirement on proximity to a VA hospital. The probability of transport to a VA hospital in this sample is 0% (as opposed to 33% in the benchmark sample), but rates of weekend transport and 28-day mortality are remarkably similar. 33 Jha et al. (2009) report in 2009 that 1.5% of US non-federal hospitals had an electronic health record (EHR) system in all clinical units and an additional 7.6% had an EHR system in at least one clinical unit. By 2014, 97% of such hospitals had possessed an EHR technology meeting requirements of the Department of Health and Human Services, and 76% of hospitals had implemented the EHR system in at least one clinical unit (Charles et al. 2015). However, interoperability (i.e., the ability to share electronic medical records) across private hospitals has remained low (Holmgren et al. 2017). To this day, multiple EHR platforms exist in the private sector, and they do not communicate with each other. 24 Care Organizations” (ACOs) (Greaney 2011). As an analog to our benchmark VA instrument in Equation (1), we construct an instrument that reflects a given ambulance company’s leave-out propensity to deliver patients to the index patient’s modal non-VA hospital. Let ℎ (𝑖) denote the hospital that ambulance ride 𝑖 is transported to, and let ℎ 𝑚 (𝑖) represent the modal non-VA hospital used by patient 𝑘 (𝑖) in ride 𝑖. Our treatment of interest is 𝐷 𝑖𝑚 ≡ 1 (ℎ (𝑖) = ℎ 𝑚 (𝑖)), which indicates whether ambulance ride 𝑖 transports its patient 𝑘 (𝑖) to his modal hospital. Our instrumental variable for this treatment is 𝑍𝑖𝑚 = 1 ∑︁ 𝐾 𝑗 (𝑖),𝑧 (𝑖) − 1 𝑖0 ∈I𝑗 (𝑖) ,𝑧 (𝑖) 1 (𝑘 (𝑖 0) ≠ 𝑘 (𝑖)) 𝐷 𝑖𝑚0 . 𝑁 𝑘 (𝑖0),𝑧 (𝑖0), 𝑗 (𝑖0) (9) where 𝐾 𝑗,𝑧 is the number of patients transported by company 𝑗 from zip code 𝑧, 𝑁 𝑘,𝑧, 𝑗 is the number of rides taken by patient 𝑘 originating in zip code 𝑧 with company 𝑗, and I𝑗,𝑧 is the set of rides transported by ambulance company 𝑗 from zip code 𝑧. This is the leave-out probability that ambulance company 𝑗 (𝑖) transports other patients from the same zip code to the modal hospital ℎ 𝑚 (𝑖) of patient 𝑘 (𝑖).34 We use the following first-stage and reduced-form equations, similar to Equations (4) and (3): 𝑚 𝑚 𝐷 𝑖𝑚 = 𝜋1𝑚 𝑍𝑖𝑚 + 𝛾1𝑚 𝑍 𝑖 + X0𝑖 𝛿1𝑚 + 𝜁1,𝑧 (𝑖) + 𝜀 1,𝑖 ; 𝑚 (10) 𝑚 (11) 𝑚 𝑚 𝑌𝑖 = 𝜋2𝑚 𝑍𝑖𝑚 + 𝛾2𝑚 𝑍 𝑖 + X0𝑖 𝛿2𝑚 + 𝜁2,𝑧 (𝑖) + 𝜀 2,𝑖 . We include in these equations an additional control variable: 𝑚 𝑍𝑖 = ∑︁ 1 (𝑘 (𝑖 0) ≠ 𝑘 (𝑖)) 𝐷 𝑚0 𝑖 , 𝐾 𝑧 (𝑖) − 1 0 𝑁 𝑘 (𝑖0),𝑧 (𝑖0) 1 𝑖 ∈I𝑧 (𝑖) where 𝐾 𝑧 is the number of patients from zip code 𝑧, 𝑁 𝑘,𝑧 is the number of rides taken by patient 𝑘 originating in zip code 𝑧, and I𝑧 is the set of rides originating in zip code 𝑧. This is the leave-out probability that patients from the same zip code 𝑧 (𝑖) are transported to hospital ℎ 𝑚 (𝑖), unconditional on the ambulance company. The modal-hospital effect may also capture hospital quality or hospitalpatient match effects. We further assess the modal hospital effect both (i) while including hospital fixed effects in Equations (10) and (11) and (ii) while splitting rides 𝑖 into samples based on whether the ride was before or after the hospital ℎ (𝑖) adopted health IT or joined an ACO. 34 As with the benchmark instrument, we construct this instrument from data in the overall sample of ambulance rides with dually eligible veterans (Column 1, Table 1 and A.14). For patients with multiple hospitals that tie for highest utilization in the prior year, we designate the set of these highest-use hospitals as the “modal hospital.” 25 In the sample of veterans with only non-VA prior utilization (Panel B of Appendix Table A.13), we demonstrate in Appendix Figure A.10 a well-behaved first-stage relationship between 𝐷 𝑖𝑚 and 𝑍𝑖𝑚 and  𝑚  balance between predicted mortality, 𝑌ˆ𝑖 , and 𝑍𝑖𝑚 , conditional on 𝑍 𝑖 , X0𝑖 , 𝑧 (𝑖) .35 The IV estimate of the modal-hospital effect on mortality is −0.006 (s.e. 0.004), which is less than 20% of the VA effect on mortality. The visual IV graph in Appendix Figure A.11 shows that the overall relationship between the reduced form and first stage is not particularly striking.36 However, computing the same IV estimate separately by years, we show in Figure 6 a stronger modal-hospital effect emerges after the passage of the HITECH Act of 2009, which led to a rapid rise in electronic medical record systems. The modal-hospital effect is close to 0 and stable prior to 2009; following 2009, the modal-hospital effect grows to about half the size of the VA effect on mortality. In Table 6, we examine how the modal-hospital effect relates to dates of hospital health IT or ACO adoption.37 We estimate the modal-hospital effect in four subsamples defined by whether or not each veteran’s modal hospital, at the time of his ride, had adopted health IT and similarly by whether or not the hospital had joined an ACO. We also use a control function approach to estimate separate modal-hospital effects, depending on health IT or ACO adoption, in the overall sample and with hospital group fixed effects (see Appendix A.7 for details). The results in the table provide suggestive evidence that the growth in the modal-hospital effect is associated with health IT adoption, holding hospitals fixed; the relationship with ACO adoption appears similar but is imprecise. While estimates in the table control for hospital or hospital group fixed effects, we find that results are essentially unchanged regardless of their inclusion. 6 Conclusion The structure of health care delivery to US veterans provides a distinctive research opportunity, allowing us to study fundamentally different systems of health care that coexist for a large patient population. Specifically, millions of older veterans (those at least aged 65) are dually eligible for care 35 Analogously to Figure 1, this figure presents binned scatterplots of the first-stage regression in Equation (10), the reduced-form regression in Equation (11), and a balance regression with predicted mortality as the outcome variable and the same design matrix. 36 Analogously to Figure 2 and Appendix Figure A.2 in the benchmark analysis, Appendix Figure A.12 shows stability in OLS and two-stage least squares estimates with increasing controls, and Appendix Figure A.13 shows robustness of two-stage least squares estimates under an exhaustive set of control combinations. 37 We measure health IT adoption from a dataset from the Office of the National Coordinator of Health Information Technology (ONC), coding any certified product on the Certified Health IT Product List (CHPL) as health IT adoption. We measure ACO participation from the Medicare Shared Savings Program (MSSP) Accountable Care Organizations (ACO) dataset. Consistent with other research, only 11% of non-VA hospitals participated in ACOs by the last year of our sample (Colla et al. 2016). 26 in a public system, operated by the Veterans Health Administration, or in private-sector hospitals, financed by Medicare. The ambulance setting provides plausible quasi-experimental assignment of veterans to these health care systems. Our work has current policy relevance, as the Department of Veterans Affairs is now considering whether to bolster its existing public delivery system or to replace it, either partially or fully, with a system of financing private care. Our work has implications more broadly for understanding the impact of public vs. private health care on mortality and spending. We find a significant VA advantage: Our preferred instrumental variables estimate, based on veterans who are induced by their ambulance company to use the Veterans Health Administration (VHA), shows a 4.5 p.p. survival gain at 28 days (95% confidence interval 1.1 to 8.0 p.p.), implying about a 46% reduction in mortality relative to the overall average. In a novel survival analysis of this quasi-experiment, we show that these survival gains occur in the first week following the ambulance ride and appear to be long-lasting. We further use this survival analysis framework to validate our IV quasi-experiment and to demonstrate differences in long-term mortality hazards between VA and non-VA users who are non-compliers. Our analysis of long-term hazards provides a compelling explanation for the difference in magnitude between IV and OLS estimates of the VA effect on 28day mortality. Although we find some intuitive margins of heterogeneity in the VA advantage, the VA outperforms the non-VA alternative in a wide variety of locations with different types of non-VA hospitals and for all types of patients we consider, not only for patients with stereotypical medical conditions. Importantly, we also find that the VA reduces total spending by 21% relative to non-VA providers, which points to higher productivity in the VHA than in the private sector. Using our quasi-experiment, we shed light on mechanisms, many of which have been raised more descriptively or qualitatively (e.g., Jha et al. 2003). We interpret our findings as consistent with the idea that the VA advantage arises from continuity of care, health IT, and organization. For example, we find that compliers are more likely to have prior VA care and have larger survival gains from VA assignment than average; in a selection model that rationalizes this finding, we show veterans who are more likely to use the VA also have larger survival gains. Interestingly, we tentatively find a similar effect in the private sector for veterans who primarily use a private hospital system. These veterans also experience reduced mortality when quasi-experimentally assigned to their modal private hospital, but only in a period following adoption of health IT due to government incentives, and even then on a smaller scale (approximately half the size) than the overall VA advantage. This suggests a sizeable residual VA advantage, even when accounting for private efforts to adopt health IT and reorganize care. 27 Our results contribute more broadly to two streams of literature on the efficiency of production. First, we contribute to the descriptive analysis that compares the performance of the US health care system to systems in other developed countries (Blank et al. 2017). By almost all accounts, comparisons of US health outcomes and health care spending are unfavorable with those of other developed countries (Garber and Skinner 2008; Rice et al. 2013). Our analysis points to a potentially large source of inefficiency in the US context: its version of private provision of health care, in which well-known information and coordination gaps may be fatal, at least for veterans in emergencies. Although several developed countries that outperform the US also feature private provision of care, the US system arguably has the most complex configuration of financing and delivery, with high levels of uninsurance, administrative costs, and fragmentation (Cebul et al. 2008). Second, we provide empirical support in the context of health care for the general idea of production complementarities among three innovations in production: information technology (IT), workplace reorganization, and products and services (Bresnahan et al. 2002). The VHA adopted a comprehensive health IT system almost two decades before the vast majority of private hospitals in the US. This reform was accompanied by an integration of care involving both reorganizing the delivery system and redefining services involved in patient care. For private hospitals, redefining health care products and services is limited by fee-for-service payment systems and the difficulty of measuring quality (Cutler 2010). Hospitals without a broad network of clinics and an overarching mandate for a population’s health may find it difficult to reorganize and redefine its services. Our result that health IT in private hospitals may improve survival—but to a muted extent and only for patients that the hospitals have previously treated—is consistent with these production complementarities. Complementarities in health care production may pose barriers for replicating the VA advantage in the fragmented private landscape of US health care. 28