Losing insurance and psychiatric hospitalizations

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Losing insurance and psychiatric hospitalizations Johanna Catherine Maclean, Sebastian Tello-Trillo, and Douglas Webber 2022-069 Please cite this paper as: Maclean, J. Catherine, Sebastian Tello-Trillo, and Douglas Webber (2022). “Losing insurance and psychiatric hospitalizations,” Finance and Economics Discussion Series 2022-069. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2022.069. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Losing insurance and psychiatric hospitalizations ∗ Johanna Catherine Maclean, Sebastian Tello-Trillo, Douglas Webber Version: October 17, 2022 Abstract We study the effect of losing insurance on psychiatric – mental health disorder (MHD) and substance use disorder (SUD) – hospital-based care. Psychiatric disorderscosttheU.S.over$1Teachyearandhospitalizationsprovideimportantand valuablecareforpatientswiththesedisorders. Weusevariationinpublicinsurance coverage (Medicaid) eligibility offered by a large-scale and unexpected disenrollment in the state of Tennessee in 2005 that lead to 190,000 individuals losing their insurance. Medicaid enrollees are at elevated risk for psychiatric disorders. Following the disenrollment, hospitalizations for SUDs declined by 15.4%. Findings suggest that MHD hospitalizations declined by 4.2%, but the coefficient estimate is imprecise. The expected financing of hospital care received also changed, with the probability that Medicaid was listed as the expected payer for MHD and SUD hospitalizationsdecliningby27.5%and30.8%respectivelypost-disenrollment. We provide suggestive evidence that psychiatric health declined post-disenrollment. Keywords: healthcare; insurance; mentalhealthdisorders; substanceusedisorders. JEL codes: I1; I11; I18. ∗Maclean (corresponding author): Department of Economics, Temple University, & NBER & IZA, Philadelphia PA, USA, catherine.maclean@temple.edu; Tello-Trillo: Frank Batten School of Public Policy and Leadership, University of Virginia & NBER, Charlottesville, VA, USA. email: sebastian.tello@virginia.edu; Webber: Federal Reserve Board of Govenors & IZA, Washington DC, USA, email: douglas.a.webber@frb.gov. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. All authors contributed equally to this study. HCUP data were obtained by Sebastian Tello-Trillo under the purview of the University of Virginia licenses. The remaining co-authors did not have any unauthorized access to these data while working on this paper. Authors are listed in alphabetical order. We thank Chandler McClellan, Brant Callaway, Ben Cook, Salma Freed, Andrew Goodman-Bacon, Samantha Harris, Steven Hill, Edward Miller, Ryan Mutter, Daniel Rees, and seminar participants at the University of Wisconsin, Southern Economics Association Conference, American Society of Health Economists Conference, and Association for Public Policy and Management for helpful comments, and the Tennessee Department of Health, Division of TennCare for excellent data assistance. Errors are our own 1

1 Introduction In this study we provide the first causal evidence on the effect of losing insurance on psychiatric disorder hospitalizations. Psychiatric disorders − mental health disorders (MHD) and substance use disorders (SUD) − are important to consider as these conditions are common and costly, but are also effectively treated. Each year these disorders cost the United States over $1T (Insel, 2008; Caulkins et al., 2014) and 47.6M and 20.3M Americans have a MHD and SUD respectively (Substance Abuse and Mental Health Services Administration, 2021), with millions more experiencing poor mental health and problematic patterns of substance use and misuse. As a specific example, the U.S. is currently in the middle of an opioid epidemic that has claimed well over 500,000 lives since 1999 (Centers for Disease Control and Prevention, 2022a). We exploit exogenous variation in public insurance coverage eligibility generated by one of the largest disenrollments in the history of the Medicaid program: an August 2005 disenrollment in the state of Tennessee. Medicaid is a public insurance program for low-income people in the U.S. that covers approximately one in five Americans (Kaiser Family Foundation, 2022) and Medicaid enrollees have elevated rates of psychiatric disorders, leading to Medicaid being the single largest payer for associated services in the country (Medicaid and CHIP Payment and Access Commission, 2015). The disenrollment we study resulted in 190,000 enrollees – 10% of those enrolled in Medicaid in the state of Tennessee and 3% of the total state population (Chang and Steinberg, 2009) – unexpectedly losing Medicaid (referred to as ‘TennCare’). The disenrollees were predominantly nondisabled, nonelderly, and childless adults. 24% of this population had serious psychological distress and 12% had an SUD in 2004 (the year prior to the disenrollment). Many more disenrollees likely had less severe psychiatric conditions that could benefit from treatment previously available through TennCare coverage.1 1Authors’ analysis of the public use 2004 National Survey on Drug Use and Health. We calculated the share of adult Medicaid enrollees ages 21 to 64 years of age with no children who (i.e., those likely disenrolled) had serious psychological distress or an SUD in Tenneessee. We expect that this estimate 2

TennCare generously covers a wide range of efficacious psychiatric healthcare services includinghospitalizations,medications,counselingservices,andstandaloneinpatientand outpatient treatment. Additionally, access to primary care, also covered by TennCare, can have important and positive spillover benefits for psychiatric disorders (Maclean et al., 2018) 2 and primary care providers deliver substantial psychiatric care, one recent study found that that over the period 2016-2018, primary care physicians provided most of the care for depression, anxiety, and any mental illness in the U.S. (Jetty et al., 2021) and, for some patients, primary care providers can treat psychiatric health as well as specialists (Gaynes et al., 2009). The disenrollment reduced access to affordable treatment for many low-income individuals who had limited access to other forms of insurance (Farrar et al., 2007). We examine administrative data on psychiatric hospitalizations coupled with difference-in-differences methods to study TennCare effects. Findings from the TennCare experience are useful to both economists and policymakers. Demand theory predicts that insurance, by reducing the out-of-pocket price faced by consumers, should increase the quantity of healthcare demanded (Grossman, 1972). To date, the literature has explored the effect of gaining insurance on psychiatric healthcare use due to available sources of experimental and quasi-experimental variation (Golberstein and Gonzales, 2015; Wen et al., 2015; Baicker et al., 2017; Wen et al., 2017; Maclean et al., 2018; Meinhofer and Witman, 2018; Grooms and Ortega, 2019; Maclean and Saloner, 2019). While there are exceptions, public insurance expansions appear to increase psychiatric healthcare use and shift financing away from patients and tounderstatestheactualnumberdisenrolledindividualswhocouldbenefitfromthepsychiatrichealthcare services offered by TennCare coverage as we consider only serious conditions. For instance, 92% had some level of psychological distress and 20% reported past year illicit drug use (illicit drug use tends to be under-reported in survey settings, suggesting that this is a lower bound on such use). Further, the financial strain and general shock associated with unexpectedly losing insurance (and therefore access to treatment and providers) could potentially exacerbate psychiatric disorders (Maclean et al., 2015), pushing a less serious condition to a serious one. 2Medicaid enrollees have high rates of both physical health conditions and psychiatric disorders. Treating physical health conditions (e.g., diabetes, heart disease) can improve psychiatric disorders through reduced stress and simply improved health overall. 3

wards insurance.3 No study has estimated the impact of losing insurance on psychiatric healthcare or health and this is the contribution of our study. The effects of insurance gains and losses need not be symmetric. Previous work has noted this possibility within the context of physical health and healthcare (Ghosh and Simon, 2015; Tello-Trillo, 2021). For instance, a person who loses insurance can retain accumulated ‘patient education,’ which includes information on one’s health stock, how to manage chronic conditions, how to interact with the healthcare system, and so forth. This education may allow a patient to continue to maintain health after a coverage loss more effectively than a patient with no insurance experience. The nature of psychiatric disorders, patients who receive care, healthcare professionals who deliver care, and the structure of the U.S. healthcare delivery system generally suggests additional reasons to suspect asymmetry. First, a substantial fraction of psychiatric care is provided for free and/or at a heavy discount, with insurance playing a relatively modest rolein the financing of such treatment; this is particularly true for SUD care (Buck, 2011). Charity care may act as a substitute for paid care, thereby muting the effect of any insurance loss. Second, many U.S. healthcare professionals will not accept public insurance patients − in particular Medicaid which offers low reimbursement rates relative to other payers (Decker, 2012), and this practice is particularly common within psychiatric care (Wen et al., 2019).4 Losing insurance that cannot be used to pay for treatment may not lead to substantial reductions in use or shifts in payment source. Finally, there are established psychiatric provider shortages in the U.S. (Buck, 2011; 3We note that Golberstein and Gonzales (2015) document no change in MHD treatment use following a Medicaid expansion. The authors rely on a simulated instrument approach while other studies cited here enter Medicaid policies directly into the regression. This difference in methodology may lead to discordant findings. See Hamersma and Kim (2013) for a thoughtful discussion of alternative approaches to modelling Medicaid generosity. Further, many studies in this literature examine the effects of expansions to non-traditional populations (e.g., low-income childless and non-disabled adults) while Golberstein and Gonzales (2015) examine increases in income generosity for traditional populations (e.g., the disabled and low-income adults with children, often mothers). Differences across studies may be driven by differences in the marginal enrollee who gains coverage following an expansion. 4Nearly 50% of standalone SUD treatment providers did not accept Medicaid in 2004 (authors’ calculations based on the 2004 National Survey of Substance Abuse Treatment Services). 4

Bishop et al., 2014); e.g., 77% of U.S. counties do not have a sufficient mental healthcare workforce to meet local patient demand. If losing insurance curtails a patient’s access to their provider, the patient may be unable to locate alternative care. We have several findings. Psychiatric hospitalizations declined post-disenrollment in Tennessee relative to a comparison group of geographically similar U.S. states, with some suggestive evidence of heterogeneity in effect size across disorders, although coefficient estimates are only precise for SUD hospitalizations. MHD hospitalizations declined by 4.2% and SUD hospitalizations declined by 15.4%. The expected financing of care was also altered, with the probability that Medicaid was the expected payer for hospital care declining 27.5% for MHD and 30.8% for SUD hospitalizations post-disenrollment. MHD disorder hospitalizations were more likely to have private insurance (7.7%) or Medicare (5.8%) as the expected payer post-disenrollment, but there was no change in the probability of either of these insurance forms as the expected payer for SUD hospitalizations. For both types of hospitalizations, the probability that the patient was listedastheexpectedpayerincreased,althougheffectsizesweresomewhatlargerforSUD than for MHD hospitalizations (60.7% and 42.5% respectively). In an extension to the main analysis, we offer suggestive evidence that the severity of psychiatric disorders may have worsened post-disenrollment, potentially through both reduced care and potentially increased financial burden associated with care received. The paper is structured as follows. Section 2 provides background details on psychiatric disorders and associated care, the TennCare disenrollment we exploit, and related literature. Data and methods are described in Section 3. We present our main results in Section 4 and report robustness checking in Section 5. Finally, Section 6 offers a discussion and conclusion. 5

2 Background and prior research 2.1 Psychiatric disorders and associated healthcare The American Psychiatric Association (APA) defines MHDs as ‘...conditions involving changes in thinking, emotion, or behavior (or a combination of these)’ (American Psychiatric Association, 2018). Further, the APA defines SUDs as conditions that occur ‘...when the recurrent use of alcohol and/or drugs causes clinically and functionally significant impairment, such as health problems, disability, and failure to meet major responsibilities at work, school, or home’ (American Psychiatric Association, 2013). These conditions impose substantial burden on affected individuals, families, and communities. Psychiatric disorders are common. In the U.S. 19.1% (47.6M) and 7.4% (20.3M) of adults met diagnostic criteria for a MHD and an SUD in 2020 (Substance Abuse and Mental Health Services Administration, 2021). The U.S. is in the midst of an unprecedented fatal drug overdose epidemic, largely related to opioids, with 93,398 fatal drug overdoses in 2020 (Ahmad et al., 2021). There are over 207 fatal opioid overdoses each day (Centers for Disease Control and Prevention, 2021b) and deaths related to stimulants such as cocaine are rising rapidly (Maclean et al., 2021). In 2020, 45,979 U.S. residents died by suicide (Centers for Disease Control and Prevention, 2022b) and the misuse of alcohol is associated with over 88,000 deaths annually (Centers for Disease Control and Prevention, 2021a). In 2021, MHDs and SUDs cost the U.S. over $1T in healthcare expenditures, disability payments, crime and violence, a less productive work force, and so forth (Insel, 2008; Caulkins et al., 2014).5 While disadvantaged populations (e.g., those with lower income) are at elevated risk for psychiatric disorders (Substance Abuse and Mental Health Services Administration, 2021), no population is immune to these disorders and society as a whole pays the associated costs (financial and otherwise). While they are costly, there are numerous effective treatment options for psychi- 5The original estimates are inflated by the authors to 2021 dollars using the Consumer Price Index. 6

atric disorders, both care provided by specialists and primary care providers (Lu and McGuire, 2002; American Psychiatric Association, 2006; Hunot et al., 2007; Scott et al., 2007; Gaynes et al., 2009; Cuijpers et al., 2011; Popovici and French, 2013; Murphy and Polsky, 2016; Olfson, 2016; Kisely et al., 2017; Krebs et al., 2018; National Institute on Drug Abuse, 2018). For example, individuals with a psychiatric disorder can be prescribed medications or receive therapy from primary care physicians or psychiatrists. Patients can obtain treatment in standalone outpatient and residential facilities, or hospital (community or psychiatric) settings from psychologists, psychiatrists, social workers, counsellors, and other healthcare professionals. Despite effective treatment options, many individuals do not receive psychiatric care – any care or adequate care, where adequate care is care that is of reasonable duration and is matched to patient need (National Institute on Drug Abuse, 2018). In 2020, less than half of U.S. adults who could benefit from mental healthcare did not receive any treatment while approximately one in ten adults meeting diagnostic criteria for an SUD received any care (Substance Abuse and Mental Health Services Administration, 2021). Commonly reported barriers to treatment receipt are inability to pay and lack of insurance coverage (Rowan et al., 2013; Substance Abuse and Mental Health Services Administration, 2021; Federal Reserve Board of Governors, 2022). For example, among adults who needed, but did not receive mental healthcare, in 2020 the most commonly stated reason for failure to receive care was inability to pay (44.9%). Treatment is likely unaffordable for many low-income and uninsured individuals. The reimbursement rates for a single in-network session with a psychiatrist can cost up to $77 (CPT code 99213) to $106 (CPT code 99214) and such treatment typically involves a series of visits (Mark et al., 2018). Food and Drug Administration-approved medications used to treat opioid use disorder (buprenorphine, methadone, and naltrexone) cost up to $16,470 per year (National Institute on Drug Abuse, 2021).6 For comparison, in 2021 the median 6All cost estimates reported in this paragraph inflated by the authors to 2021 dollars using the CPI. 7

maximumincomeatwhichasingleadultwaseligibleforMedicaidintheU.S.was$17,774 (KaiserFamilyFoundation,2022). Finally, thereisevidencethatdelaysinreceivingcare, which could occur following an insurance loss, can have negative effects on psychiatric health (Penttil¨a et al., 2014; Reichert and Jacobs, 2018). In particular, insurance lossinduced treatment discontinuation can be fatal for those with an SUD (Sordo et al., 2017; Connery and Weiss, 2020; Krawczyk et al., 2020). One policy approach to addressing unmet need (quantity and/or quality) for psychiatric healthcare services is the provision of affordable insurance that covers a range of treatment options. TennCare provided such insurance to low-income and uninsurable Tennesseeresidents. Thus, removingthiscoveragefrom190,000peoplewithelevatedrisk for a psychiatric disorders could have lead to a substantial reduction in access to valuable care for chronic conditions. We study the effects of losing TennCare on hospital-based psychiatric treatment. While hospitalizations represent a single modality, they are an important one to consider for several reasons. First, in the U.S., hospitalizations comprise 18% of overall psychiatric healthcare service expenditures and 25% of total payments to providers and are therefore important from a cost perspective (Substance Abuse and Mental Health Services Administration, 2014). In 2004 (the year prior to the TennCare disenrollment), one in four community adult hospitalizations (or 7.6M hospitalizations) involved a psychiatric disorder (Owens et al., 2007). Thus, due to their costs and use by patients, hospitalizations are central in discussions related to containing overall expenditures on healthcare. Second, and perhaps more practically, despite growth in ambulatory facilities and primary care for psychiatric treatment, community hospitals remain a critical facet of treatment for the thousands of Americans with psychiatric disorders (Stensland et al., 2012). Community hospitals provide patient stabilization services, complex drug therapies, coordinated psychotherapy, short-term detoxification, intense observation, and crisis care to psychiatric patients each year. Third, and related to the previous 8

point, hospitalizations are important components of evidence-based psychiatric care: this modality is included in the Addiction Medicine Society of America’s continuum of care and is used to treat patients with relatively severe episodes (Mee-Lee et al., 2013). Fourth, hospitalizations are often an entry point to the healthcare system for individuals with a psychiatric disorder and can lead patients to receipt of a broader set of services (O’Toole et al., 2002; Houry et al., 2018). Finally, through the federal The Emergency Medical Treatment and Labor Act (EMTALA), in the U.S. all emergency departments must accept patients regardless of ability to pay, suggesting that hospitals may be particularly exposed to patients needing care following an insurance loss. 2.2 A brief overview of TennCare Tennesseeoriginallyofferedafee-for-serviceMedicaidprogram. Duetohighcosts,the statetransitionedtoamanagedcareprogramin1994, makingTennesseethefirststatein the U.S. to adopt managed care within Medicaid.7 State legislators anticipated that the transition would reduce overall program costs. The expected savings were allocated to supportalarge-scaleincreaseinMedicaideligibilitytolow-income, non-disabledchildless adults and uninsurable adults, defined as adults with pre-existing conditions that lead to prohibitively high premiums (Farrar et al., 2007). This population has historically had limited access to insurance in the U.S. The program was popular and TennCare covered 22% of the state’s population by 2004. The TennCare program, which used a psychiatric healthcare carve out8 throughout our study period, generously covered a wide range of efficacious psychiatric services (e.g., medications, assessment and evaluation services, and counseling) at low cost-sharing with limited application of utilization management; 7Managed care is a common organizational structure for healthcare provision. Broadly, in managed caresystems,patientsvisitcertainhealthcareprofessionalsandhospitalswhichcontractwithamanaging company that oversees and monitors costs. The goal of managed care is to minimize overuse of services without harming patient outcomes, leading to reduced costs to the insurer. 8The term ‘carve out’ generally is used to describe a specific set of services that is not covered by the primary insurance plan and instead is managed by another company, typically a company that specializes in the specific services ‘carved out’ of the main plan. 9

e.g., prior authorization (Farrar et al., 2007; Chang and Steinberg, 2009; Kaiser Family Foundation, 2021). TennCare coverage for physical health conditions was similarly generous and, through positive spillovers attributable to well-managed co-morbidities (e.g, diabetes, heart disease), psychiatric care delivered by primary care providers (Gaynes et al., 2009; Jetty et al., 2021), and overall improvements in well-being, likely enhanced psychiatric health among enrollees (Maclean et al., 2018). TennCare eventually became financially unsustainable for Tennessee (Bennett, 2014). By 2004 the program accounted for one-third of the state’s budget (Farrar et al., 2007). Between August 2005 and July 2006 Medicaid eligibility was curtailed along several margins,withchangesannouncedinNovember2004. Eligibilityforchildlessnon-disabled and uninsurable adults was terminated. 190,000 enrollees or 10% of the total Medicaid population lost TennCare coverage (Bureau of TennCare, 2005; Chang and Steinberg, 2009).9 Importantly for our design, the disenrollment was essentially unexpected by disenrollees (Farrar et al., 2007). 2.3 TennCare disenrollment literature There is a small literature examining the health insurance, service use, and outcome effects of the TennCare disenrollment using quasi-experimental methods. To date, the literature has focused on general populations and physical health. No studies have examined psychiatric health or healthcare outcomes or insurance among patients with psychiatric disorders. Examining these outcomes is our contribution to this literature. We contend that separate consideration of psychiatric disorders is warranted, which motivates our study. In addition to features of psychiatric disorders and associated care noted in Section, 1, important differences between psychiatric and physical health and healthcare have been highlighted by economists for decades (Frank and McGuire, 2000). 9We emphasize disenrollment effects in our study. We acknowledge that coverage generosity was curtailedtosomeextentamongcontinuingMedicaidenrollees. Nonetheless,coverageremainedrelatively generous and, in line with the broader TennCare literature (reviewed in Section 2.3), we assume that the effects of insurance losses dwarfed the effects of other changes. 10

As noted by Frank and McGuire (2000) in the context of mental vs. physical healthcare: ‘Uncertainty and variation in treatments are greater; the assumption of patient selfinterested behavior is more dubious; response to financial incentives such as insurance is exacerbated; the social consequences and external costs of illness are formidable.’ Further, the risk of an overdose is elevated after periods of abstinence among those with SUDs (Merrall et al., 2010). An abrupt termination in access to treatment or the associated loss of financial resources conferred by insurance (Maclean et al., 2015) could be especially relevant for patients with SUDs. 2.3.1 Coverage effects The TennCare disenrollment reduced Medicaid and overall insurance coverage with some private insurance substitution (Garthwaite et al., 2014; Tarazi et al., 2017; DeLeire, 2019; Tello-Trillo, 2021). For instance, using the Current Population Survey, Garthwaite et al. (2014) find that, post-disenrollment, the probability of having Medicaid declined by 33% among low-income, childless, and non-disabled adults.10 2.3.2 Healthcare service use and health outcome effects Severalstudieshaveexaminedtheeffectofthedisenrollmentonhealthcareandhealth outcomes. Findings from the TennCare literature on physical healthcare suggest that losing insurance reduced service use, increased unmet need, and increased patient financial burden and hospital provision of uncompensated care, although there is some heterogeneity in estimated effects across studies. Tello-Trillo (2021) leverages survey data from the National Health Interview Survey and Behavioral Risk Factor Surveillance Survey (BRFSS). The author finds that the reform lead to a 4% – 5% reduction in mammograms and breast exams, and an increase of 20% in number of days with limited by one’s health and no strong evidence of changes 10There is also work on labor market effects (Garthwaite et al., 2014; Ham and Ueda, 2017). 11

in emergency department visits. Tarazi et al. (2017) use the BRFSS and show that the disenrollment increased cost-related barriers to seeing a doctor but did not change the probability of having a personal doctor. DeLeire (2019) uses the Survey of Income and ProgramParticipationandfindsreportsofunmethealthcareneedandrelianceoncharity care increased and doctor visits and self-reported health declined post-disenrollment. Garthwaite et al. (2018) use American Hospital Association data and document that uncompensated care increased in Tennessee post-disenrollment. Finally, in a working paper, Ghosh and Simon (2015) use the National Inpatient Sample (NIS) and show that, post disenrollment, the share of total non-elderly adult hospitalizations with Medicaid as the expected payer decreased by 21% and hospitalizations with patients as the expected payer increased in Tennessee relative to a comparison group. There was no change in the number of hospitalizations. 3 Data and methods 3.1 Hospitalization data Our primary dataset is the NIS, an administrative database compiled by the Healthcare Cost and Utilization Project (HCUP).11 These data allow us to study hospitalizations and are the largest publicly available U.S. all-payer inpatient healthcare database. The NIS are regularly used by economists to study psychiatric healthcare (Antwi et al., 2015; Golberstein et al., 2015). The sample reflects a 20% stratified sample of U.S. hospitals, withfivetoeightmillionhospitalizationsoccurringatover1,000hospitalseachyear. Hospitals are sampled on region, ownership status, and bed size. In 2007 (the last year of our study) the weighted NIS sample covered 90% of the universe of hospitalizations and 78% of all hospitals (Barrett et al., 2010). The American Hospital Association defines 11The source of the data is the National Inpatient Sample (NIS), Healthcare Cost and Utilization Project (HCUP), Agency for Healthcare Research and Quality. 12

hospitals as ‘all nonfederal, short-term general, and other special hospitals’ (American Hospital Association, 2018). The NIS does not include psychiatric hospitals. We focus on hospitals that have positive hospitalizations for patients 21 to 64 years of age.12 Our study period is January 2002 to December 2007. We choose these years to avoid the early 2000s recession (2001) and the Great Recession (2008-2010). In particular, recessions are linked with insurance (Cawley and Simon, 2005; Cawley et al., 2015), psychiatric disorders (Ruhm, 2015; Carpenter et al., 2017; Hollingsworth et al., 2017), and psychiatric healthcare use (Bradford and Lastrapes, 2014; Maclean et al., 2020).13 TheNISisnotdesignedtobestaterepresentative. Insteadthesedata(overourstudy period) are designed to be representative at the level of the region.14 We are concerned that using the NIS data for Tennessee, which is a state with 5.9M residents (2% of the total U.S. population) in 2004 (University of Kentucky Center for Poverty Research, 2020), could lead to reliance on an unrepresentative sample of hospitals.15 Given the structure of the NIS and our focus on a policy occurring in Tennessee, we propose an alternative empirical approach. We use state-representative data for Tennessee (from administrative sources) instead of the Tennessee observations included in the NIS, and then use the other Southern states (i.e., the comparison group) from NIS. Since we are combining data sets, we perform a simulation to test if our proposed strategy will allow us to estimate more accurate disenrollment effects. The results from this exercise are included in the Data Appendix (Section 8). Our analysis suggests that combining data in this manner allows us to more accurately estimate treatment effects. 12We considered using elderly adults as a within-state comparison group in a triple difference estimator. The elderly have very different trends in psychiatric outcomes than the non-elderly. For example, 8% of adults 18 years and older met diagnostic criteria for an SUD while the share was 2% for elderly adults in 2020 (Substance Abuse and Mental Health Services Administration, 2021). 13Further, there was a major re-certification of all Medicaid enrollees in Tennessee in 2002 (Division ofTennCare,ND)andwedonotwanttoconfoundeffectsbyincludingdatabeforeandafterthisevent. 14In more recent years, the NIS is representative at the level of the nation. 15We are less concerned with this issue in our comparison group as the NIS is designed to be representative at the regional level over our study period and our comparison group covers all other states in the South region. We have confirmed this assumption with an economist at the Agency for Healthcare Quality and Research, the agency that administers the HCUP. Further, we establish this point in a Monte Carlo simulation. Details available on request. 13

We also report results using NIS for Tennessee as a robustness check (see Section 5). We analyze two sets of outcomes to study disenrollment effects. First, we consider the quarterly number of MHD and SUD hospitalizations per hospital (technically, we measure hospital discharges, but we refer to them as hospitalizations for simplicity). We separately classify MHD and SUD hospitalizations based on ICD-9 codes available on the discharge record (Stranges et al., 2011; Heslin et al., 2015). Second, we examine indicators for expected (not actual) payment source: Medicaid, any insurance, private insurance, Medicare, and self-pay (which plausibly includes uninsured patients). We have information on up to two expected payers per hospitalization listed by the hospital and code these variables one if the payer is listed as primary or secondary payer, and zero otherwise. Importantly, self-payment patients are often uninsured and do not necessarily pay the costs of hospitalization themselves, instead these costs are often borne by the hospitals who treat these patients. We are not able to study who eventually paid for the services as we have only expected payer information, which is a limitation of the study. 3.2 Empirical model We estimate the difference-in-differences (DID) model outlined in Equation 1: Y = β +β DID +β pop +α +α +α +ϵ (1) h,s,q,t 0 1 s,q,t 2 s,t h q t h,s,q,t where Y represents a hospitalization or insurance outcome for hospital h in state h,s,q,t s and in quarter q in year t. DID is an indicator variable that takes on a value s,q,t of one in Tennessee after the disenrollment and zero otherwise. pop is the annual s,t state population (University of Kentucky Center for Poverty Research, 2020), we include this variable to capture differences in demand for hospitalizations across states. α is h a vector of hospital fixed-effects (which subsume state fixed-effects), and α and α are q t vectors of quarter and year fixed-effects respectively. ϵ is the error term. We apply h,s,q,t 14

NIS weights to the (non-Tennessee) NIS data and weight the Tennessee data equally. We follow Garthwaite et al. (2014) and apply a modified block-bootstrap procedure to calculate standard errors.16 We follow, as closely as possible given the states that appear in the NIS, Garthwaite etal.(2014)anduseotherSouthernstatesasacomparisongroup: Arkansas(2004-2007), Florida(2002-2007), Georgia(2002-2007), Kentucky(2002-2007), Maryland(2002-2007), North Carolina (2002-2007), Oklahoma (2005-2007), South Carolina (2002-2007), Texas (2002-2007), Virginia (2002 to 2004 and 2006-2007), and West Virginia (2002-2007). The logic of using this comparison group is that other Southern states are subject to similar economic, social, and political shocks as Tennessee, and therefore offer a suitable counterfactual. As we show in Section 5.2, we observe no evidence of spatial spillovers that could suggest a violation of the stable unit treatment value assumption (STUVA) which is required for identification in DID specifications. The NIS is an unbalanced panel at both the hospital and state level, hence the hospitals and states that appear in the comparison group vary across years. However, all comparison states are observed before and after the TennCare disenrollment. The Tennessee data are the universe, thus changing composition is not likely a concern for our treatment group. A necessary assumption for canonical DID models to recover causal estimates is that the treatment and comparison group would have followed the same trends in outcomes had the treatment group not received treatment (i.e., ‘parallel trends’). The assumption is untestable as the treatment group is treated in the post-period. We provide suggestive evidence on parallel trends through estimation of an event-study. Quarterly hospital-level summary statistics using data from the pre-disenrollment periodinTennesseeandotherSouthernstatesarereportedinTable1. Unadjustedtrends 16In our main specifications, we do not include time-varying covariates, beyond the state population which we believe is important to capture demand for services, as recent work suggests that estimation of a DID method using two-way fixed-effects, as we do here, with time-varying controls can lead to inaccurate estimates of average treatment effects on the treated (Caetano et al., 2022). We report results based on regressions that control for a range of time-varying covariates in a robustness check. 15

in MHD and SUD hospitalizations and expected payment sources for the treatment and comparison groups are reported in Figure 1. While there are some level differences across the two groups of states, trends in outcomes (pre-disenrollment) appear to move in parallel, post-disenrollment there are some differences in trends.17 4 Results 4.1 DID analysis of hospitalizations and expected payment sources Our main DID results for hospitalizations are reported in Table 2. Following the disenrollment, quarterly MHD and SUD hospitalizations declined by 4.9 and 5.8 per hospital operating in Tennessee, although the coefficient estimate in the MHD hospitalizations regression is not precise. Comparing these coefficient estimates to the means in Tennessee pre-disenrollment, our findings suggest a 4.2% reduction in MHD hospitalization (baseline mean = 115.0) and 15.4% in SUD hospitalizations (baseline mean = 37.4). Thus, the disenrollment arguably lead to meaningful reductions in hospitalization care for psychiatric disorders among Tennessee Medicaid enrollees. In Tables 3 and 4 we report the estimated impacts on expected payment sources for MHD and SUD hospitalizations. We note that we interpret these findings cautiously as they are based on samples of patients receiving hospital care for psychiatric conditions, which appear to be impacted by the policy (see Table 2), and hence findings are potentially subject to conditional-on-positive bias. With this caveat in mind, we show that the probability that Medicaid was the expected payer for psychiatric treatment declined by 9.5 percentage points (‘ppts’) and 12.0 ppts for MHD and SUD hospitalizations respectively. Compared to the proportions of hospitalizations expected to be financed by Medicaid in Tennessee pre-disenrollment (34.6% for MHD and 39.0% for SUD), these 17Ideally we would like to examine heterogeneity in our outcomes. Unfortunately our data use agreement with the Tennessee Department of Health does not include this information. 16

coefficient estimates imply that expected Medicaid financing was reduced by 27.5% for MHD hospitalizations and 30.8% for SUD hospitalizations. The probability that any insurance was the expected source of payment for treatment declined more modestly for MHD hospitalizations (3.2 ppts or 3.3%). The decline in any insurance as the expected payer for SUD hospitalizations was nearly as large as the Medicaid decline (10.4 ppts or 12.1%). MHD patients were able to partially offset the insurance loss with private and Medicare coverage (for which expected payment increased by 2.2 ppts or 7.7% and 1.9 or 5.8% respectively), but SUD patients were not able to substitute expected forms of payment in this matter. Expected self-financing increased for both MHD and SUD: 1.7 ppts and 8.2 ppts respectively, which correspond to 42.5% and 60.7% rises. Thus, whilethereductionsintheexpecteduseofMedicaidtopayforcareweresimilar for MHD and SUD patients, the ‘substitute’ expected payment forms were potentially somewhat different, with SUD patients more likely to shoulder the expected costs of care to a larger extent than MHD patients. Given that the average charge for a community hospitalization for psychiatric conditions ranged from $6,432 to $26,986 for an uninsured patient in 2006 (Stensland et al., 2012), this shift in expected financing from Medicaid to patients (in particular SUD patients) was arguably non-trivial. Further, many expected self-pay hospitalizations result in the facility absorbing the cost as patients, ultimately, are not able to (at least fully) pay for treatment (Garthwaite et al., 2018). 4.2 Internal validity We estimate an event study to provide suggestive evidence as to whether our treatment and comparison groups would have followed parallel trends in psychiatric hospitalizations and payment sources post-disenrollment after adjusting for fixed-effects. More specifically, we decompose DID in Equation 1 into a series of interactions between s,q,t an indicator for Tennessee and time leads and lags reflecting periods around the disenrollment. To smooth the data, we use six-month bins to define periods. The omitted 17

category is the six-month period prior to the disenrollment. Event study results are reported graphically in Figures 2 (hospitalizations) and 3 (expected payment source). Therearetwokeyfindingsfromtheeventstudies. First,wedonotobservesubstantial evidence of differential pre-trends in our outcomes between Tennessee and comparison states. Coefficient estimates are small and not statistically distinguishable from zero for the policy lead indicators. Second, examination of the policy lag indicators largely supports our DID findings. 4.3 DID analysis of psychiatric health Next, we explore the impact of the disenrollment on MHD and SUD proxies: deaths by suicide, and fatal alcohol poisoning and drug overdoses. We use the Centers for Disease Control and Prevention’s National Vital Statistics System (NVSS) Underlying Cause of Death public use files. These data record the universe of deaths in the U.S. and classify deaths by cause. We select all related deaths for adults 21 to 64 in each quarter 2002 to 2007 for Tennessee and other Southern states. We report results based on all states but results based on only state/year pairs in the NIS data are available on request and are similar. Deaths are expressed as a quarterly rate per 100,000 adults 21 to 64 and data are weighted by that population.18 We modify Equation 1 to account for the fact that the NVSS data are at the state rather than hospital level, in particular hospital fixed-effects are replaced with state fixed-effects. Results are reported in Table 5. Rates of deaths by suicides increased by 0.479 per 100,000 state non-elderly adult residents or 11.2% relative to the mean in Tennessee predisenrollment. Rates of fatal alcohol and drug poisonings increased by 0.765 or 24.4%. These findings suggest that losing insurance, through reduced service use or increased financial burden associated with healthcare, may lead to worse psychiatric health.19 18All state-month cells in the public use NVSS with less than ten suicides are suppressed for confidentiality reasons. We impute these cells with a value of five. Results are not sensitive to imputing a value of zero (the smallest possible value) or nine (the largest possible value). 19AsnotedbyBlacketal.(2019),wecouldbeunder-poweredtodetecteffects. Tospeaktothisissue, 18

Effects appear to be larger for alcohol poisonings and drug overdoses. This finding is in line with previous research suggests that unexpected treatment discontinuation can be fatal for those with an SUD (Sordo et al., 2017; Connery and Weiss, 2020; Krawczyk et al., 2020). 5 Robustness checks We report a serious of robustness checks. Reassuringly, our findings are stable. For brevity, we report robustness checking for hospitalization outcomes. Results for payment sources are similarly stable and are available on request. 5.1 Alternative specifications, samples, and periods We re-estimate Equation 1 using alternative time periods, comparison groups, and specifications. More specifically, we: (i) include time-varying state-level covariates (see Table 1, demographics from the CPS (Flood et al., 2021) and population from the Census);20 (ii) aggregate to the state-year level; (iii) estimate unweighted regression; (iv) use the NIS data for Tennessee;21 (v) exclude Texas and Georgia from the comparison group – we exclude these states as their Medicaid programs appear, based on available evidence, to potentially cover psychiatric healthcare services less generously than Tennessee over our study period (Kaiser Family Foundation, 2018); (vi) exclude the year of the policy change (2005); (vii) include a separate linear time trend for Tennessee; and we have attempted to assess statistical power in our study. In unreported analyses, we conducted a post-hoc power analysis. Results, available on request, suggest that we are able to detect effect sizes of the magnitudes that we estimate with approximately 80% power. 20Inearlierversionsofthismanuscript,weincludedtime-varyingcovariatesinourmainspecifications, reporting regressions without these controls as a robustness check. We have elected to change our presentation to emphasize the more parsimonious specifications (i.e., not including the time-varying state-level covariates beyondthestatepopulationwhichweviewas important asthisvariable allowsus to account for demand for hospital services within states) based on recent research suggesting that DID regressions are vulnerable to bias from such controls. See Caetano et al. (2022) for a helpful discussion of this issue. 21Weapplythe NIS-providedweights tobothTennesseeand otherSouthernStatesin thisregression. 19

(viii) use a longer study period (2000-2010). Results are reported in Figure 4. Overall, our findings − while not identical across all specifications − appear to be broadly stable. Findings are less precise when we include time-varying covariates, this changemaybeattributabletothevulnerabilitiesoftwo-wayfixed-effectsDIDregressions with such controls included (Caetano et al., 2022). 5.2 Program-induced migration An empirical concern in policy analysis is that the policy under study may have induced individuals to migrate away from or towards the affected locality leading to biased estimates of regression coefficients due to violation of STUVA (Moffitt, 1992). To explore this possibility, we draw micro-level data from the Annual and Social Economic Supplement (ASEC) to the CPS between 2003 and 2008 and model past-year acrossstate migration among respondents ages 21 to 64 years as a function of the disenrollment using a modified version of Equation 1. ASEC data over the period 2003 to 2008 pertains to migration 2002 to 2007. We exclude those respondents with family income > 400% FPL as such respondents are unlikely to be eligible for Medicaid coverage in any state in our analysis sample. We include all Southern states but results based on only state/year pairsthatappearintheNISdataareverysimilar.22 Dataareweightedbysurveyweights provided by IPUMS administrators. We estimate an augmented version of Equation 1 as outlined in Section 4.3, although due to the structure of the ASEC, we have annual not quarterly data and thus we cannot include quarter fixed-effects in the regression. We find no evidence that the disenrollment lead to changes in the propensity to move towards or away from Tennessee among non-elderly adults of lower income levels. The coefficient estimate is small in size and not statistically distinguishable from zero. 22We do not control for the state population as we do in other regressions given the close connection betweenmigrationandpopulation,e.g.,out-migrationallelseequalwillmechanicallylowerpopulation. However, we have included this covaraite in the regression in unreported analyses and the results are very similar and available on request. 20

6 Discussion We provide the first evidence on the effect of losing public insurance on psychiatric hospitalizations. Our findings are relevant from both an economic and a policy perspective. First, we extend the economic literature that has estimated the insurance-elasticity of demand for psychiatric healthcare by leveraging plausibly exogenous variation offered by a large and unexpected Medicaid disenrollment. We are the first study to document this elasticity. Second, many countries grapple with providing quality healthcare to individuals and curtailing high healthcare costs to governments, which are in turn borne by taxpayers. For example, in the United States there have been recent policy debates regarding repealing or re-designing the Affordable Care Act (ACA) – a massive transformation of the U.S. healthcare system which substantially increased both public and private coverage and required insurers to cover psychiatric healthcare services. These findings speak to the insurance elasticity of demand for psychiatric healthcare services and can inform policy decisions how best to provide coverage for such care. Our findings suggest that the disenrollment reduced MHD hospitalizations by 4.2% and SUD hospitalizations by 15.4%, although findings for MHD hospitalizations are not precise. In additional, the expected financing of psychiatric care changed, MHD hospitalization expected financing was partially shifted to private and Medicare insurance while SUD expected financing was transferred entirely to patients. These findings suggest that losing insurance reduces access to valuable psychiatric care and likely imposes financial burden on patients. Given that self-payments by patients often translates into bad debt to hospitals (Garthwaite et al., 2018), this shift in financing may also spillover to healthcare providers who treat patients with psychiatric disorders. Further, while both MHD and SUD patients were arguably negatively impacted by the disenrollment, effects were disproportionately larger for those with SUDs. While we lack the data to fully explore why findings differed for MHD and SUD patients, we can propose some hypotheses. SUD patients may be more disadvantaged than 21

MHD patients. This disadvantage may manifest in being less able to secure employment that offers insurance, and insurance that covers SUD treatment, following a large-scale disenrollment. That is, patients with MHD may be able to find a job that will offer coverage for MHD or offer sufficient income so that these patients can purchase coverage individually post-disenrollment. Over our study period, while private coverage for MHD and SUD services was limited, more states required private plans to cover MHD than did SUD services, thus obtaining private coverage that included services was likely less difficult for MHD than SUD patients. Providers may have been less willing to treat patients with SUDs in the hospital due to stigma, misunderstanding of addiction, and so forth. For example, an SUD patient appearing in the emergency department may have been allowed to detox in the emergency department and then discharged, while a patient with a MHD appearing in the emergency department may have been admitted to the hospital. PatientswithMHDswerepotentiallybetterabletosecureotherformsofpublic coverage than patients with SUDs following the disenrollment. For example, MHDs are a qualifying condition for Supplemental Security Income (SSI), but SUD was removed as a qualifying condition for this program in the 1990s, thus those with SUD may be less able to qualify for other forms of public insurance. We note that our explanations are not likely fully satisfactory and we encourage more work on understanding how those with SUDs and MHDs are able to respond to policy shocks such as a large-scale insurance disenrollment and associated loss of coverage. We can compare our findings to studies that have examined the effect of pre- and post-ACA state Medicaid expansions to ‘non-traditional’ populations similar to the population impacted by the TennCare disenrollment (Wen et al., 2015, 2017; Meinhofer and Witman, 2018; Maclean et al., 2018; Cher et al., 2019; Ghosh et al., 2019; Grooms and Ortega, 2019; Maclean and Saloner, 2019; Saloner and Maclean, 2020) and the Oregon Medicaid experiment (Baicker et al., 2017). Non-traditional populations are low-income non-disabled childless adults that historically been categorically ineligible for Medicaid 22

coverage. To date, this literature has examined the effect of insurance gains attributable to Medicaid expansions. We examine insurance losses following the TennCare disenrollment. While there are numerous reasons to expect asymmetry in the effects of insurance gains and losses on psychiatric healthcare, we do not find that to be the case. We observe suggestive evidence that psychiatric health may worsen following an insurance loss but there is less of improvements in this health metric following an insurance gain (Abouk et al., 2019; Averett et al., 2019), although Wettstein (2019) shows that private insurance gains among young adults appear to improve psychiatric health. OurfindingsdepartfromtheTennCareliteratureonall-causehospitalizations. Ghosh and Simon (2015) show that there was no change in the number of total hospitalizations post-disenrollment. These differing findings may imply differences in the insuranceelasticity of demand across services. Previous work demonstrates differences in the priceelasticity of demand for healthcare services (Ellis et al., 2017). However, findings for shifts in the expected financing of hospitalizations post-disenrollment across total and psychiatric hospitalizations: for example Ghosh and Simon (2015) find a 21% reduction in the expected use of Medicaid to pay for hospitalizations while we document a 27.5% and 30.8% reduction for MHD and SUD hospitalizations respectively. In sum, we provide the first evidence on the impact of losing insurance on psychiatric hospitalizations and financing, and measures of psychiatric health. Given the ongoing opioid and mental illness crisis in the U.S. (Maclean et al., 2021), alongside an emerging stimulant crisis, understanding how insurance impacts these outcomes is critical. Establishing and understanding differences in responsiveness to policy changes across services is critical to optimal insurance design. Our findings shed light on these questions. 23

References Abouk, R., L. Helmchen, A. Moghtaderi, and J. Pines (2019). The aca medicaid expansions and opioid mortality: Is there a link? Medical Care Research and Review, 1077558720967227. Ahmad, F., L. Rossen, and P. Sutton (2021). Provisional drug overdose death counts: 2021. Technical report, National Center for Health Statistics. American Hospital Association (2018). Fast facts on u.s. hospitals, 2018. Technical report, American Hospital Association. American Psychiatric Association (2006). American Psychiatric Association practice guidelines for the treatment of psychiatric disorders: Compendium 2006. American Psychiatric Publications. American Psychiatric Association (2013). Diagnostic and Statistical Manual of Mental Disorders, 5th Edition. American Psychiatric Association. American Psychiatric Association (2018). What is mental illness? Antwi, Y. A., A. S. Moriya, and K. I. Simon (2015). Access to health insurance and the use of inpatient medical care: Evidence from the affordable care act young adult mandate. Journal of Health Economics 39, 171–187. Averett, S. L., J. K. Smith, and Y. Wang (2019). Medicaid expansion and opioid deaths. Health Economics 28(12), 1491–1496. Baicker, K., H. L. Allen, B. J. Wright, and A. N. Finkelstein (2017). The effect of medicaid on medication use among poor adults: evidence from oregon. Health Affairs 36(12), 2110–2114. Barrett, M., E. Wilson, D. Whalen, et al. (2010). Summary 2007 hcup nationwide inpatient sample (nis) comparison report: Hcup methods series report no. 2010-03. Technical report, Agency for Healthcare Research and Quality. Bennett, C. (2014). TennCare, one state’s experiment with Medicaid expansion. Vanderbilt University Press. Bishop, T. F., M. J. Press, S. Keyhani, and H. A. Pincus (2014). Acceptance of insurance by psychiatrists and the implications for access to mental health care. JAMA Psychiatry 71(2), 176–181. Black, B. S., A. Hollingsworth, L. Nunes, and K. I. Simon (2019). The effect of health insurance on mortality: power analysis and what we can learn from the affordable care act coverage expansions. NBER working paper (w25568). Bradford, W. D. and W. D. Lastrapes (2014). A prescription for unemployment? recessions and the demand for mental health drugs. Health Economics 23(11), 1301–1325. 24

Buck, J. A. (2011). The looming expansion and transformation of public substance abuse treatment under the affordable care act. Health Affairs 30(8), 1402–1410. Bureau of TennCare (2005). Fiscal year 2005-2006 annual report: Bureau of tenncare. Technical report, Bureau of TennCare. Caetano, C., B. Callaway, S. Payne, and H. S. Rodrigues (2022). Difference in differences with time-varying covariates. arXiv preprint arXiv:2202.02903. Carpenter,C.S.,C.B.McClellan,andD.I.Rees(2017). Economicconditions,illicitdrug use, and substance use disorders in the united states. Journal of Health Economics 52, 63–73. Caulkins, J., A. Kasunic, and M. A. Lee (2014). Societal burden of substance abuse. International Public Health Journal 6(3), 269. Cawley, J., A. S. Moriya, and K. Simon (2015). The impact of the macroeconomy on healthinsurancecoverage: Evidencefromthegreatrecession. Health Economics 24(2), 206–223. Cawley, J. and K. I. Simon (2005). Health insurance coverage and the macroeconomy. Journal of Health Economics 24(2), 299–315. Centers for Disease Control and Prevention (2021a). Alcohol and public health: Frequently asked questions. Centers for Disease Control and Prevention (2021b). Opioid overdose: Understanding the epidemic. Centers for Disease Control and Prevention (2022a). Drug overdose deaths drug overdose deaths remain high. Technical report, Centers for Disease Control and Prevention. Centers for Disease Control and Prevention (2022b). Suicide and self-harm injury. Chang, C.F.andS.C.Steinberg(2009). Tenncaretimeline: Majoreventsandmilestones from 1992 to 2009. Technical report, Methodist Le Bonheur Center for Healthcare Economics, University of Memphis. Cher, B. A., N. E. Morden, and E. Meara (2019). Medicaid expansion and prescription trends: opioids, addiction therapies, and other drugs. Medical Care 57(3), 208. Connery, H. S. and R. D. Weiss (2020). Discontinuing buprenorphine treatment of opioid use disorder: What do we (not) know? American Journal of Psychiatry 177(2), 104– 106. Cuijpers, P., F. Clignet, B. van Meijel, A. van Straten, J. Li, and G. Andersson (2011). Psychological treatment of depression in inpatients: A systematic review and metaanalysis. Clinical Psychology Review 31(3), 353–360. 25

Decker, S. L. (2012). In 2011 nearly one-third of physicians said they would not accept new medicaid patients, but rising fees may help. Health Affairs 31(8), 1673–1679. DeLeire, T. (2019). The Effect of Disenrollment from Medicaid on Employment, Insurance Coverage, and Health and Health Care Utilization. Emerald Publishing Limited. Division of TennCare (N/D). Tenncare timeline. Technical report, Division of TennCare. Ellis, R. P., B. Martins, and W. Zhu (2017). Health care demand elasticities by type of service. Journal of Health Economics 55, 232–243. Farrar, I., D. Eichenthal, B. Coleman, and C. Reese (2007). Tenncare reform, one year later. Technical report, Robert Wood Johnson Foundation. Federal Reserve Board of Governors (2022). Report on the economic well-being of u.s. households in 2021. Technical report. Flood, S., M. King, R. Rodgers, S. Ruggles, J. R. Warren, and M. Westberry (2021). Integrated public use microdata series, current population survey: Version 9.0 [dataset]. Technical report, IPUMS: University of Minnesota. Frank, R. G. and T. G. McGuire (2000). Economics and mental health. Handbook of Health Economics 1, 893–954. Garthwaite, C., T. Gross, and M. J. Notowidigdo (2014). Public health insurance, labor supply, and employment lock. The Quarterly Journal of Economics 129(2), 653–696. Garthwaite, C., T. Gross, and M. J. Notowidigdo (2018). Hospitals as insurers of last resort. American Economic Journal: Applied Economics 10(1), 1–39. Gaynes, B. N., D. Warden, M. H. Trivedi, S. R. Wisniewski, M. Fava, and A. J. Rush (2009). What did star* d teach us? results from a large-scale, practical, clinical trial for patients with depression. Psychiatric Services 60(11), 1439–1445. Ghosh, A. and K. Simon (2015). The effect of medicaid on adult hospitalizations: Evidence from tennessee’s medicaid contraction. Technical report, National Bureau of Economic Research. Ghosh, A., K. Simon, and B. D. Sommers (2019). The effect of health insurance on prescription drug use among low-income adults: evidence from recent medicaid expansions. Journal of Health Economics 63, 64–80. Golberstein, E., S. H. Busch, R. Zaha, S. F. Greenfield, W. R. Beardslee, and E. Meara (2015). Effectoftheaffordablecareact’syoungadultinsuranceexpansionsonhospitalbased mental health care. American Journal of Psychiatry 172(2), 182–189. Golberstein, E. and G. Gonzales (2015). The effects of medicaid eligibility on mental health services and out-of-pocket spending for mental health services. Health Services Research 50(6), 1734–1750. 26

Grooms, J. and A. Ortega (2019). Examining medicaid expansion and the treatment of substance use disorders. AEA Papers and Proceedings 109, 187–92. Grossman, M. (1972). On the concept of health capital and the demand for health. Journal of Political Economy 223(10.2307), 1830580223. Ham, J. C. and K. Ueda (2017). The perils of relying solely on the march cps: The case of estimating the effect on employment of the tenncare public insurance contraction. Unpublished Working Paper. Hamersma, S. and M. Kim (2013). Participation and crowd out: Assessing the effects of parental medicaid expansions. Journal of Health Economics 32(1), 160–171. Heslin, K., A. Elixhauser, and C. Steiner (2015). Hospitalizations involving mental and substance use disorders among adults, 2012. HCUP Statistical Brief 191. Hollingsworth, A., C. J. Ruhm, and K. Simon (2017). Macroeconomic conditions and opioid abuse. Journal of Health Economics 56, 222–233. Houry, D. E., T. M. Haegerich, and A. Vivolo-Kantor (2018). Opportunities for prevention and intervention of opioid overdose in the emergency department. Annals of Emergency Medicine 71(6), 688–690. Hunot, V., R. Churchill, V. Teixeira, and M. S. de Lima (2007). Psychological therapies for generalised anxiety disorder. Cochrane Database of Systematic Reviews (1). Insel, T. R. (2008). Assessing the economic costs of serious mental illness. American Journal of Psychiatry 165(6), 663–665. Jetty, A., S.Petterson, J.M.Westfall, andY.Jabbarpour(2021). Assessingprimarycare contributions to behavioral health: a cross-sectional study using medical expenditure panel survey. Journal of Primary Care & Community Health 12, 21501327211023871. Kaiser Family Foundation (2018). Medicaid benefits: Rehabilitation services – mental health and substance use. Kaiser Family Foundation (2021). Medicaid income eligibility limits for adults as a percent of the federal poverty level. Kaiser Family Foundation (2022). Medicaid state fact sheets. Kisely, S. R., L. A. Campbell, and R. O’Reilly (2017). Compulsory community and involuntary outpatient treatment for people with severe mental disorders. Cochrane Database of Systematic Reviews 3. Krawczyk, N., R. Mojtabai, E. A. Stuart, M. Fingerhood, D. Agus, B. C. Lyons, J. P. Weiner, and B. Saloner (2020). Opioid agonist treatment and fatal overdose risk in a state-wide us population receiving opioid use disorder services. Addiction 115(9), 1683–1694. 27

Krebs, E., B. Enns, E. Evans, D. Urada, M. D. Anglin, R. A. Rawson, Y.-I. Hser, and B. Nosyk (2018). Cost-effectiveness of publicly funded treatment of opioid use disorder in california. Annals of Internal Medicine 168(1), 10–19. Lu, M. and T. G. McGuire (2002). The productivity of outpatient treatment for substance abuse. Journal of Human Resources, 309–335. Maclean, J. C., B. L. Cook, N. Carson, and M. F. Pesko (2018). Public insurance and psychotropic prescription medications for mental illness. BE Journal of Economics and Policy Analysis 19(1), Online. Maclean, J. C., B. P. Horn, and J. H. Cantor (2020). Business cycles and admissions to substance abuse treatment. Contemporary Economic Policy 38(1), 139–154. Maclean, J. C., J. Mallatt, C. J. Ruhm, and K. Simon (2021). Economic studies on the opioid crisis: Costs, causes, and policy responses. In Oxford Research Encyclopedia of Economics and Finance. Maclean, J. C., C. McClellan, M. F. Pesko, and D. Polsky (2018). Reimbursement rates for primary care services: Evidence of spillover effects to behavioral health. Technical report, National Bureau of Economic Research. Maclean, J. C. and B. Saloner (2019). The effect of public insurance expansions on substance use disorder treatment: evidence from the affordable care act. Journal of Policy Analysis and Management 38(2), 366–393. Maclean, J. C., D. Webber, and M. T. French (2015). Workplace problems, mental health and substance use. Applied Economics 47(9), 883–905. Mark, T. L., W. Olesiuk, M. M. Ali, L. J. Sherman, R. Mutter, and J. L. Teich (2018). Differential reimbursement of psychiatric services by psychiatrists and other medical providers. Psychiatric Services 69(3), 281–285. Medicaid and CHIP Payment and Access Commission (2015). Behavioral health in the medicaid program—people, use, and expenditures. Technical report, Medicaid and CHIP Payment and Access Commission. Mee-Lee, D., G. Shulman, M. Fishman, et al. (2013). The asam criteria. American Society of Addiction Medicine. Meinhofer, A. and A. E. Witman (2018). The role of health insurance on treatment for opioid use disorders: Evidence from the affordable care act medicaid expansion. Journal of Health Economics 60, 177–197. Merrall, E. L., A. Kariminia, I. A. Binswanger, M. S. Hobbs, M. Farrell, J. Marsden, S. J. Hutchinson, and S. M. Bird (2010). Meta-analysis of drug-related deaths soon after release from prison. Addiction 105(9), 1545–1554. 28

Moffitt, R. (1992). Incentive effects of the us welfare system: A review. Journal of Economic Literature 30(1), 1–61. Murphy, S. M. and D. Polsky (2016). Economic evaluations of opioid use disorder interventions. Pharmacoeconomics 34(9), 863–887. National Institute on Drug Abuse (2018). Principles of drug addiction treatment: A research-based guide. Technical report, National Institutes on Drug Abuse. National Institute on Drug Abuse (2021). Medications to treat opioid use disorder. Technical report, National Institutes on Drug Abuse. Olfson, M. (2016). Building the mental health workforce capacity needed to treat adults with serious mental illnesses. Health Affairs 35(6), 983–990. Owens, P., M. Myers, A. Elixhauser, and C. Brach (2007). Care of adults with mental health and substance abuse disorders in us community hospitals, 2004. HCUP fact book (10), 07–0008. O’Toole, T. P., E. C. Strain, G. Wand, M. E. McCaul, and M. Barnhart (2002). Outpatient treatment entry and health care utilization after a combined medical/substance abuse intervention for hospitalized medical patients. Journal of General Internal Medicine 17(5), 334–340. Penttil¨a, M., E. J¨aa¨skela¨inen, N. Hirvonen, M. Isohanni, and J. Miettunen (2014). Duration of untreated psychosis as predictor of long-term outcome in schizophrenia: systematic review and meta-analysis. The British Journal of Psychiatry 205(2), 88–94. Popovici, I. and M. French (2013). Economic evaluation of substance abuse interventions: Overview of recent research findings and policy implications. In B. McCrady and E. Epstein (Eds.), Addictions: A comprehensive guidebook, Chapter 2. Oxford University Press. Reichert, A. and R. Jacobs (2018). The impact of waiting time on patient outcomes: Evidence from early intervention in psychosis services in e ngland. Health Economics 27(11), 1772–1787. Rowan, K., D.D.McAlpine, andL.A.Blewett(2013). Accessandcostbarrierstomental health care, by insurance status, 1999–2010. Health Affairs 32(10), 1723–1730. Ruhm, C. J. (2015). Recessions, healthy no more? Journal of Health Economics 42, 17–28. Saloner, B. and J. C. Maclean (2020). Specialty substance use disorder treatment admissions steadily increased in the four years after medicaid expansion: Study looks at whether medicaid expansion led to more low-income adults with substance use disorders receiving treatment. Health Affairs 39(3), 453–461. 29

Scott, J., F.Colom, andE.Vieta(2007). Ameta-analysisofrelapserateswithadjunctive psychological therapies compared to usual psychiatric treatment for bipolar disorders. International Journal of Neuropsychopharmacology 10(1), 123–129. Sordo, L., G. Barrio, M. J. Bravo, B. I. Indave, L. Degenhardt, L. Wiessing, M. Ferri, and R. Pastor-Barriuso (2017). Mortality risk during and after opioid substitution treatment: systematic review and meta-analysis of cohort studies. British Medical Journal 357. Stensland, M., P.R.Watson, andK.L.Grazier(2012). Anexaminationofcosts, charges, and payments for inpatient psychiatric treatment in community hospitals. Psychiatric Services 63(7), 666–671. Stranges, E., K. Levit, C. Stocks, and P. Santora (2011). State variation in inpatient hospitalizations for mental health and substance abuse conditions, 2002–2008: Statistical brief# 117. Technical report, Agency for Healthcare Quality and Research. Substance Abuse and Mental Health Services Administration (2014). Projections of national expenditures for treatment of mental and substance use disorders, 2010–2020. Technical report, Substance Abuse and Mental Health Services Administration. Substance Abuse and Mental Health Services Administration (2021). Key substance use and mental health indicators in the united states: Results from the 2020 national survey on drug use and health. Technical report, Substance Abuse and Mental Health Services Administration. Tarazi, W. W., T. L. Green, and L. M. Sabik (2017). Medicaid disenrollment and disparities in access to care: evidence from tennessee. Health Services Research 52(3), 1156. Tello-Trillo, D. S. (2021). Effects of losing public health insurance on preventative care, health, and emergency department use: Evidence from the tenncare disenrollment. Southern Economic Journal 88(1), 322–366. University of Kentucky Center for Poverty Research (2020). State level data of economic, political, andtransferprograminformationfor1980-2018. Technicalreport, University of Kentucky Center for Poverty Research. Wen, H., B. G. Druss, and J. R. Cummings (2015). Effect of medicaid expansions on health insurance coverage and access to care among low-income adults with behavioral health conditions. Health Services Research 50(6), 1787–1809. Wen, H., J. M. Hockenberry, and J. R. Cummings (2017). The effect of medicaid expansion on crime reduction: Evidence from hifa-waiver expansions. Journal of Public Economics 154, 67–94. Wen, H., A. S. Wilk, B. G. Druss, and J. R. Cummings (2019). Medicaid acceptance by psychiatrists before and after medicaid expansion. JAMA Psychiatry 76(9), 981–983. 30

Wettstein, G. (2019). Health insurance and opioid deaths: Evidence from the Affordable Care Act young adult provision. Health Economics 28(5), 666–677. 31

7 Tables and figures Table 1: Summary statistics: Pre-disenrollment Sample: Tennessee Other Southern States MHD hospitalizations 115.018 118.703 SUD hospitalizations 37.483 38.395 Age (years) 36.317 35.696 Female 0.515 0.512 Black 0.167 0.167 Other race 0.025 0.044 Hispanic 0.043 0.162 High school 0.267 0.240 Some college 0.191 0.195 College 0.158 0.166 Population 21-64 3,564,215 7,272,9032 Observations 2172 5712 Hospitals 142 911 Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a hospital in a state in quarter in a year. TN observations are unweighted and non-TN observations are weighted by NIS sample weights. Table 2: Effect of TennCare disenrollment on psychiatric hospitalizations among adults 21-64 years per hospital-quarter: Hospitalization data 2002-2007 Outcome: MHD hospitalizations SUD hospitalizations TN × post -4.885 -5.769** (6.079) (2.763) Observations 11,894 11,894 TN mean, pre-policy 115.018 37.483 Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, andWV.Theunitofobservationisahospitalinastateinquarterinayear. TNobservationsare unweighted and non-TN observations are weighted by NIS sample weights. Regressions are estimated with least squares and control for hospital, quarter, and year fixed-effects, and state population as a covariate. Standard errors are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported in parentheses. ***, **, & * = statistically different from zero at the 1%, 5%, and 10% level of confidence. 32

Table 3: Effect of the TennCare disenrollment on expected source of payment for MHD hospitalizations among adults 21-64 years per hospital-quarter: Hospitalization data 2002-2007 Expected payment source: Medicaid Any insurance Private Medicare Self-pay TN × post -0.095*** -0.032*** 0.022** 0.019** 0.017* (0.0129) (0.0077) (0.0088) (0.0081) (0.0094) Observations 11,871 11,871 11,871 11,871 11,871 TN mean, pre-policy 0.346 0.958 0.287 0.325 0.040 Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a hospital in a state in quarter in a year. TN observations are unweighted and non-TN observations are weighted by NIS sample weights. Regressions are estimated with least squares and control for hospital, quarter, and year fixed-effects, and state population as a covariate. Standard errors are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported in parentheses. ***, **, & * = statistically different from zero at the 1%, 5%, and 10% level of confidence. Table 4: Effect of the TennCare disenrollment on expected source of payment for SUD hospitalizations among adults 21-64 years per hospital-quarter: Hospitalization data 2002-2007 Expected payment source: Medicaid Any insurance Private Medicare Self-pay TN × post -0.120*** -0.104*** 0.005 -0.002 0.082*** (0.0117) (0.0129) (0.0110) (0.0090) (0.0119) Observations 11,664 11,664 11,664 11,664 11,664 TN mean, pre-policy 0.390 0.860 0.224 0.246 0.135 Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a hospital in a state in quarter in a year. TN observations are unweighted and non-TN observations are weighted by NIS sample weights. Regressions are estimated with least squares and control for hospital, quarter, and year fixed-effects, and state population as a covariate. Standard errors are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported in parentheses. ***, **, & * = statistically different from zero at the 1%, 5%, and 10% level of confidence. 33

Table 5: Effect of the TennCare disenrollment on rates of suicides and fatal alcohol poisonings and drug overdoses among adults 21-64 years per quarter: NVSS 2002-2007 Cause of death: Suicide Alcohol poisoning or drug overdose TN × post 0.479*** 0.765*** (0.122) (0.167) Observations 408 408 TN mean, pre-policy 4.288 3.141 Notes: Other Southern states include DE, DC, AR, AL, KY, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a state in a quarter in a year. Data are weighted by the state non-elderly adult population. Regressions are estimated with least squares and control for state, quarter, and year fixed-effects. Standard errors are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported in parentheses. ***, **, & * = statistically different from zero at the 1%, 5%, and 10% level of confidence. 34

Table6: EffectoftheTennCaredisenrollmentonmigrationamongadults 21-64 years and population: ASEC-CPS 2002-2007 Outcome: Move across state lines in the past year TN × post -0.004 (0.006) Observations 129,938 TN mean, pre-policy 0.036 Notes: Other Southern states include DE, DC, AR, AL, KY, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a respondent in a state in a year. Data are weighted by ASEC survey weights. Regressions are estimatedwithleastsquaresandcontrolforstateandyearfixed-effects. Standarderrors are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported in parentheses. ***, **, & * = statistically different from zero at the 1%, 5%, and 10% level of confidence. 35

Figure 1: Trends in psychiatric hospitalizations among adults 21-64 years per hospitalquarter: Hospitalization data 2002-2007 MHD hospitalizations SUD hospitalizations Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. Data are aggregated to the treatment-time period level. 36

Figure 2: Effect of TennCare disenrollment on psychiatric hospitalizations among adults 21-64 years per hospital-quarter using an event study: Hospitalization data 2002-2007. MHD hospitalizations SUD hospitalizations Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a hospital in a state in a year. TN observations are unweighted and non-TN observations are weighted by NIS sample weights. Regressions are estimated with least squares and control for hospital, quarter, and year fixed-effects, and state population as a covariate. 95% confidence intervals are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported with vertical lines. 37

Figure 3: Effect of TennCare disenrollment on psychiatric hospitalization expected payment sources among adults 21-64 years per hospital-quarter using an event study: Hospitalization data 2002-2007. MHD hospitalizations SUD hospitalizations Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a hospital in a state in a year. TN observations are unweighted and non-TN observations are weighted by NIS sample weights. Regressions are estimated with least squares and control for hospital, quarter, and year fixed-effects, and state population as a covariate. 95% confidence intervals are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported with vertical lines. 38

Figure 4: Effect of TennCare disenrollment on psychiatric hospitalizations among adults 21-64 years per hospital-quarter using alternative samples, specifications, and time periods: Hospitalization data 2002-2007 Notes: Other Southern states include AK, FL, GA, KY, LA, MD, MS, NC, OK, SC, TX, VA, and WV. The unit of observation is a hospital in a state in a year unless otherwise noted. TN observations are unweighted and non-TN observations are weighted by NIS sample weights unless otherwise noted. Regressions are estimated with least squares and control for hospital, quarter, and year fixed-effects and state population as a coviarate unless otherwise noted. 95% confidence intervals are estimated with a modified block-bootstrap procedure following Garthwaite et al. (2014) and are reported with vertical lines. 39

8 Data appendix The NIS is not designed to be state representative. Instead, the dataset is representative at the national and regional level over our study period. HCUP administrators strongly advise researchers against using the NIS for single-state estimates.23 We study a subset of the population (ages 21 to 64) and specific types of hospitalizations (i.e, psychiatric). There are important differences across hospitals that provide different service lines. For instance, psychiatric healthcare is a low profit service line, and for-profit hospitals are less likely to offer this type of care. We are concerned that using the NIS data for Tennessee could lead to reliance on an unrepresentative sample of hospitals. We propose an alternative empirical approach, in which we use a state-representative data for Tennessee (from administrative sources) instead of the Tennessee included in the NIS and then use the other southern states (i.e., the comparison group) from NIS. Since we are combining data, we perform a simulation to test if our proposed strategy will allow us to estimate more accurate disenrollment effects. We conduct this analysis for MHD and SUD hospitalizations (which are the focus of our study) and, for completeness, we report results for total hospitalizations. We initially provide suggestive evidence on the implications of relying on a regionally representative dataset to study TennCare through a Monto Carlo simulation. First, we use data from the Bureau of Labor Statistics Quarterly Census of Employment and Wages (QCEW) database to determine the total number of community hospitals in each Southern state in each year over the period 2002 to 2007 (we use the six digit North American Industry Classification System code 622110). This information provides us with the universe of community hospitals. We construct variables that take on three values (1, 2, and 3) for (i) ownership and (ii) bed size, to mimic the variables used by 23See: ’...strongly advises researchers against using the NIS to estimate state-specific statistics. ... However, these NIS samples were not designed to yield a representative sample of hospitals at the State level’: https://www.hcup-us.ahrq.gov/db/nation/nis/nis_statelevelestimates.jsp (last accessed September 8th, 2021). 40

AHRQ administrators to select hospitals for inclusion in the NIS; we note that our simulation does not exactly replicate the NIS sampling method, but our approach is very similar. Second, we generate an outcome variable where the data generating process emulates a standard DID functional form. In particular, we generate a variable Y that is determined by a DID estimate (Tennessee interacted with a post-disenrollment period indicator), hospital fixed-effects, year fixed-effects, and a random error term. We used Southern states that appear in the NIS to form the comparison group. Third, we perform various draws from the universe of hospitals: (1) a 100% sample; (2) a 20% sample by year, ownership, and bed size and retaining only hospitals that provide psychiatric healthcare and that treat non-elderly adult patients;24 and (3) a 100% sample in Tennessee and a 20% sample by year, ownership, and bed size in other Southern states that are included in the NIS data set for hospitals that provide psychiatric healthcare and that treat non-elderly adult patients. We select the data in this manner to reflect the universe of hospitals (1), the NIS sample after making exclusions necessary for our research question (2), and the value of our proposed combination of NIS and other administrative data which we describe in more detail later (3). We then estimate a DID specification in each sampling scheme across 1,000 simulated populations. Full details are available on request. ThesimulationresultsarereportedinDataAppendixFigureA1,andtheimplications of reliance on the NIS are immediately apparent. First, all distributions of beta hat are centered on the true parameter value (constructed to have a value of two). However, the sampling framework employed has consequential implications for the likelihood that the estimated DID parameter will be significantly over- or under-estimated. The difference between region- and state-representativeness matters in a DID context, particularly one in which only a single state is treated. In a regionally representative dataset such as the 24WeconstructedthesevariablestohavedifferentdistributionsinTennesseeandotherSouthernstates includedintheNISdatasetandtheyaredesignedtomimic,albeitimperfectly,ourfocusonpsychiatric hospitalizations among non-elderly adults in our main analysis. 41

NIS, a state like Tennessee may contribute a very small number of observations within any given conditional cell (e.g., ownership, bed size, and year). Intuitively, when the identifying variation is small number of observations which by design are not necessarily representative of the treatment group, the variability in coefficient estimates increases substantially. Using the 100% sample, the distribution of beta hats is tightly centered around the true value, as we would expect. The 20% sample with exclusions required to form our analysis sample produces a very wide distribution of estimates, with many estimates differingsubstantiallyfromthetruetreatmenteffect. Usingacombinationoftheuniverse of Tennessee hospitals and a 20% sample of hospitals in other Southern states, while not fullyalleviatingtheincreaseindistributionspread,substantiallytightensthedistribution of beta hats and nearly offsets the increase in the width of the distribution induced by sampling. In sum, this simulation exercise implies that sampling used to construct the NIS, while not leading to bias, substantially increases the chance that the researcher will have an unrepresentative sample which can lead to inaccurate estimates of treatment effects. This outcome is what NIS administrators state will occur when the NIS is used for state-level analyses, but researchers (including ourselves in previous work) often ignore this caution. A high-variance estimator, like the one arising from the typical state-level analysis administrators warn researchers against applying when using the NIS data, is underpowered to detect small and modest effect sizes. When using an underpowered estimator, the distribution of statistically significant coefficient estimates will be biased away from zero. This situation illustrates the real danger of using such estimators: researchers who find statistically significant evidence are potentially overstating the true effect size. Using the simulation evidence above to help us identify the best possible estimation strategy, weproposeacombinationoftheNISandadministrativedataforTennessee. We replace the NIS Tennessee observations with the universe of hospitalizations at commu- 42

nity hospitals for this state that we obtained from the Tennessee Department of Health (‘DOH’ data).25 Of note, the NIS Tennessee is drawn from the DOH data that we use, thus the datasets are entirely compatible. Put differently, we are not introducing bias into the analysis attributable to using different data sources as the NIS uses a sub-set of the DOH data to form the Tennessee sample. This combination mimics (3) in our simulation and, we hypothesize, will allow more accurate estimates of treatment effects than reliance on the NIS data alone.26 We further investigate the value of our combined dataset by comparing NIS and DOH data for Tennessee over our study period. We exclude non-community hospitals from the DOH data to match the NIS sample frame; we confirmed our definition of community hospitals with administrators at the Tennessee DOH (details available on request). We plot trends in the average number of hospitalizations among non-elderly adults per hospital in each quarter of our study period in Tennessee in the NIS and the DOH data. Trends in MHD, SUD, and total hospitalizations (Data Appendix Figure A2) display more period-to-period variation in the NIS data than the DOH data, and the variation in the NIS occurs around the disenrollment period. Differences in trend between NIS hospitals and the universe of hospitals in the DOH are arguably more pronounced for total hospitalizations. Of note, based on our analysis of the data, review of Tennessee DOH hospital directories, and discussions with DOH staff, the deviations apparent in the NIS data for Tennessee do indeed reflect the unrepresentative nature of the hospitals sampled for the NIS, which further underscores our concern that reliance on the NIS for a single-state analysis may be problematic. The correlations between the DOH and NIS time series are 0.46 (MHD), 0.56 (SUD), and 0.39 (total). Thus, while the 25Because the Tennessee NIS data is drawn from the DOH data the measures, we use in our analysis are standardized across the two datasets (i.e., NIS Tennessee sample and DOH). However, we note that theparametersofourdatauseagreementwiththeTennesseeDepartmentofHealthdonotpermitusto linkdataintheDOHtotheTennesseeNISdatatocomparehospitals. Fulldetailsavailableonrequest. 26The State Inpatient Database or the State Emergency Department Database for Tennessee are not availabletothepublic(https://www.hcup-us.ahrq.gov/db/availability_public.jsp;lastaccessed September 8th, 2021). 43

DOH and NIS time series generally follow similar trends, there are clear and non-trivial differences, in particular at pivotal periods relative to the TennCare disenrollment. Data Appendix Tables A1 and A2 report the shares of MHD and SUD hospitalizations, and total hospitalizations respectively that appear in the NIS in each year of our study period. We include only hospitals that have positive MHD or SUD hospitalizations in at least one quarter during our study period in our calculations.27 There are substantial differences across the DOH and the NIS data suggesting that the NIS data are not representative of all community hospitals in Tennessee, which is perhaps not unexpected based on our simulation exercise and cautions from NIS administrators. For instance, in 2000 22% of all community hospitals in Tennessee appeared in the NIS while in 2004 the share had increased to 38%. In 2005 this share declined to 32%, and by 2007 the share fell to 20%. Changes in sample that are concurrent with the policy under study can lead to inaccurate estimates. We refer to the combined NIS (for non-Tennessee states) and DOH (for Tennessee) dataset as the ‘hospitalizations dataset.’ We view our large sample size for Tennessee, we have the universe of community hospitals, as an advantage over previous TennCare studies that have relied on smaller, non-state representative datasets for Tennessee. Other studies seeking to use NIS, or other datasets, to investigate single-state treatments may consider such a combination. Or, if data are not available, the potential limitation could be noted in the study. 27Somehospitalshavezerohospitalizationsinagivenyear-quarter. Whencreatingapercentagemeasure, we could not divide by zero as this value is undefined. Thus, to avoid losing these observations, weaddedavalueofonetoeachhospitalinoursample. Theminimumforyear-quartersisthereforeone as opposed to zero. This change shifts the distribution but does not affect the coefficient estimates. We test whether the disenrollment influenced the probability that a hospital has any MHD or SUD hospitalizations using the empirical model outlined in Equation (1). The coefficient estimate and associated standard error for are -0.004 and 0.008. Thus, we observe no statistically significant evidence that the disenrollment influenced this probability. 44

Table A1: Share of all hospitalizations captured by NIS: Hospitals with hospitalizations patients ages 21−64 years Year Share of TN hospitals with > 0 psychiatric hospitalizations appearing in NIS 2002 25.56 2003 27.56 2004 27.11 2005 26.33 2006 21.99 2007 22.26 Notes: Denominator is the number of community hospitalizations in the Tennessee Department of Health data. Numerator is the number of community hospitalizations in the NIS. The numbers are very similar when MHD and SUD hospitalizations are measured separately. Table A2: Share of all hospitalizations captured by NIS: Hospitals with hospitalizations among patients ages 21−64 years Year Share of all TN hospitalizations appearing in NIS 2002 29.61 2003 34.79 2004 38.20 2005 31.71 2006 32.91 2007 20.20 Notes: The denominator is the number of community hospitalizations in the Tennessee Department of Health data. The numerator is the number of community hospitalizations in the NIS. 45

Figure A1: Monte Carlo simulation of estimated treatment effects Notes: Each of the four simulations are conducted using 1,000 repetitions. Data is generated such that the true value of the treatment effect is two. See text for full details. 46

Figure A2: Effect of TennCare disenrollment on psychiatric hospitalization payment sources among adults 21-64 years per hospital-quarter using an event study: Hospitalization data 2002-2007. Total hospitalizations MHD hospitalizations SUD hospitalizations Notes: Outcomesarequarterlyaveragesofhospitalizationsamongpatients21 to 64 years. DOH = Tennessee Department of Health data. 47