Labor Market Effects of the Oxycodone-Heroin Epidemic

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Labor Market Effects of the Oxycodone-Heroin Epidemic David Cho, Daniel I. Garcia, Joshua Montes, and Alison Weingarden 2021-025 Please cite this paper as: Cho, David, Daniel I. Garcia, Joshua Montes, and Alison Weingarden (2021). “Labor Market Effects of the Oxycodone-Heroin Epidemic,” Finance and Economics Discussion Series 2021-025. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2021.025. 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.

Labor Market Effects of the Oxycodone-Heroin ∗ Epidemic † David Cho, Daniel I. Garcia, Joshua Montes, and Alison Weingarden April 12, 2021 Abstract We estimate the causal effects of heroin use on labor market outcomes by proxying for heroin use with prior exposure to oxycodone, the largest of the prescription opioids with a well-documented history of abuse. After a nationwide tightening in the supply of oxycodone in 2010, states with greater prior exposure to oxycodone experienced much larger increases in heroin use and mortality. We find increases in heroin use led to declines in employment and labor force participation rates, particularly for white, young, and less educated groups, consistent with the profile of oxycodone misusers. The results show the importance of extending beyond prescriptions when accounting for the labor market effects of the opioid crisis. Keywords: labor force; participation rate; opioid crisis JEL: J21, I12, I18 ∗TheviewsexpressedinthispaperareourownanddonotreflecttheviewsoftheFederalReserveSystem or its staff. †Cho: Board of Governors of the Federal Reserve System, 20th and C Streets, NW, Washington, DC 20551 (email: david.cho@frb.gov); Garcia: Board of Governors of the Federal Reserve System, 20th and C Streets, NW, Washington, DC 20551 (email: daniel.i.garcia@frb.gov); Montes: Board of Governors of the FederalReserveSystem,20thandCStreets,NW,Washington,DC20551(email: joshua.k.montes@frb.gov); Weingarden: OECD Centre for Entrepreneurship, SMEs, Regions and Cities, 2 rue Andr´e Pascal, 75016 Paris France (email: alison.weingarden@oecd.org). We thank AnneMarie Bryson, Isabel Leigh, Jes Baker, and Paul Tran for excellent research assistance. 1

Since 2000, the opioid crisis has claimed more than 480 thousand lives in the U.S.1 While the tragic consequences on health and human life are salient, the labor market effects of the opioid crisis are stillunderdebate. Muchoftheliteraturetothispointhascenteredontheprescription opioidcrisis. Since use of prescription opioids can have multiple causes, the aggregate effects on labor supply are not immediately obvious. Some prescription use could ameliorate pain and help workers remain in the labor force, while misuse could lead to addiction and major life disruptions including labor force exits. Also, some prescription use is transitory and without long-term consequences. Indeed, the literature that links prescription opioid use to labor supply outcomes finds mixed results. For instance, Harris et al. (2020) and Aliprantis et al. (2019) find large negative effects, while Currie et al. (2019) report modest positive effects for women. We contribute to this literature by estimating the effects on the labor force participation rate (LFPR) and other outcomes of the transition from prescription opioid abuse to illicit opioids, particularly heroin. In contrast to the more ambiguous effects of general prescription opioid use, the labor supply effects of using heroin are likely negative. Still, the aggregate effects could be modest, if the number of heroin users is small and the effects of heroin use on LFPR are also small. However, we find the transition to heroin use has had meaningful, negative effects on the LFPR and the employment-to-population ratio (EPOP), particularly for white, young, and less educated cohorts, who were more likely to abuse prescriptions such as oxycodone. Asastartingpoint,andtomotivateouranalysis,weusedatafromtheNationalSurveyofDrugUse andHealth(NSDUH)toassessrecentincreasesinheroinusersandtheemploymentstatusofheroin users. Accordingtotheself-reportedsurveyestimates,thenumberofindividualshavingusedheroin increasedbyabout1.7millionto5.4millionfrom2007to2017, butthetruefigurescouldhavebeen much larger given the tendency of individuals to underreport illegal drug use in surveys (Harrell 1997; Harrison 1997). Indeed, the number of deaths from heroin overdoses increased fivefold during this period. We also find that heroin users are substantially less likely to be in the labor force. In particular, we estimate a linear probability model of labor force participation and show that heroin users are about 14 percentage points less likely to be in the labor force, a difference comparable to that between those without a high school degree and college graduates. Though suggestive, the NSDUH estimates are not causal, as unobserved individual characteristics could explain both low participation and heroin use. To estimate the effects of the transition to heroin on labor market outcomes, we exploit variation from policy changes in 2010 that limited the supply of oxycodone – the largest of the prescription opioids with a well-documented recent history of abuse (Hays 2004; Cicero et al. 2005; Katz and Hayes 2004). In 2010, Purdue Pharma reformulated OxyContin, a popular brand of oxycodone, which impeded users from crushing or liquefying the pills for consumption (Severtson et al. 2013b; 1See “Overdose Death Rates” (National Institute on Drug Abuse; January 29, 2021) 2

Cicero and Ellis 2015).2 Also in 2010, the DEA and local authorities began shutting down Florida “pill mills,” which had been boosting national supply (Surrat et al. 2013; Davis and Carr 2017;).3 Because of these developments, national manufacturer shipments of oxycodone have declined since peaking in 2010, after more than quadrupling in the 2000s. As supply tightened, many who were addicted to oxycodone turned to heroin (Cicero and Ellis 2015; Alpert et al. 2018 and Evans et al. 2019). To causally estimate the effects of the increase in heroin use on labor market outcomes, we use an event-study design that exploits variation across states prior to 2010 in oxycodone use and the 2010 reductioninoxycodonesupply, followingtheapproachesinAlpertetal.(2018), Evansetal.(2019), andPowellandPacula(2021). Asdocumentedbythisliterature, trendsinheroinmortalitypriorto 2010 were similar across states with high and low oxycodone exposure. However, after 2010, heroin overdoses ramped up more in states with greater prior oxycodone exposure, reflecting an increase in demand for heroin. Also, deaths from fentanyl overdoses (a synthetic opioid), hepatitis-C cases, and other related negative outcomes also increased more in states with greater prior oxycodone exposure (Powell et al. 2019; Powell and Pacula 2021). Given these findings, we proxy for heroin and other illicit drug use after 2010 with oxycodone exposure prior to 2010. We contribute to this literature by estimating meaningful and negative effects on EPOP and LFPR of the oxycodone-heroin epidemic. In our baseline specifications, we track the prime-age cohort in 2010, and account for differences across states in demographics, industry composition, and cyclical performance during the housing boom and Great Recession.4 The baseline estimates show EPOP andLFPRtrendedsimilarlyinstateswithhighandlowoxycodoneexposurepriorto2010. However, after2010,stateswithpreviouslyhigheroxycodoneexposureexperiencedsignificantlylargerrelative declinesinEPOPandLFPR.Inareaswithonestandarddeviationhigheroxycodoneexposureprior, EPOPandLFPRwereonaverageabout30basispointslowerfrom2011to2019relativetothepreperiod. These effects seem driven by behavioral changes related to greater heroin use rather than compositional effects, as we find little evidence of the latter. The unemployment rate estimates are small and not significant, indicating most of the effect of the increase in heroin use on employment occurs through the participation margin. We also find the increase in heroin use led to an increase in the rate of social security disability insurance (SSDI) beneficiaries – consistent with Park and Powell (2021) – and contributed to a modest change in the composition of employment towards part-time work. 2See, for example, “Origins of an Epidemic: Purdue Pharma Knew Its Opioids Were Widely Abused” (New York Times; May 29, 2018) 3Seealso“FedsRaidPain-ClinicsSuspectedofIllegalyDistributingMillionsofPrescriptionDrugs”(ABC News; March 5, 2010) and “Florida Shutting ‘Pill Mill’ Clinics” (New York Times; August 31,2011) 4In the baseline, we focus on people who were prime-age (25-54 year olds) in 2010, as this group faced a plausibly exogenous and unexpected shock to oxycodone supply. That said, we show that the estimates for the group of persons 25-54 years old each year are similar to the cohort approach. 3

We also estimate the employment and participation effects on different demographic groups using the same event-study framework with dependent variables aggregated by race and ethnicity, sex, education, and age cohorts. The EPOP and LFPR estimates are negative and significant for non- Hispanic whites (both male and female), those without a college degree, and younger cohorts. In contrast, the estimates are generally smaller and not significant for nonwhites, those with a college degree or higher, and the cohort of individuals aged 55-64 in 2010. These patterns are consistent with the profile of oxycodone abusers prior to 2010, who according to NSDUH estimates, were predominantly white, less educated, and younger. Despite these differences, we stress that these results do not suggest that nonwhites and those with a college degree have been unaffected. For nonwhites and college graduates, the main estimates are negative and the standard errors are large enough we can neither reject the hypothesis that they are zero nor large and negative. More broadly, death rates related to opioids have recently risen across demographic groups (Drake et al. 2020). Since 2016, mortality rates from synthetic opioids and heroin have increased more rapidly for Blacks and American Indian or Alaska Natives, reaching comparable levels as for non-Hispanic whites. TogainasenseoftheaggregateimplicationsonLFPRfortheprime-agecohortin2010,wecombine the baseline estimates with counterfactual assumptions in which all U.S. states had low oxycodone rates and were thus unaffected by the supply changes in 2010. Using different counterfactual assumptions, we find the oxycodone-heroin shock may have lowered LFPR between 25 and 40 basis points on average from 2010 to 2019. We also relate these magnitudes to the gap between the LFPR and the LFPR predicted by a demographics model. We find the oxycodone-heroin shock could explain between one sixth and one quarter of the gap implied by the demographics LFPR model. In sum, while the drag on LFPR from the increase in heroin use has been sizeable, it is modest compared to other LFPR drivers such as population aging and effects from the Great Recession. We address various concerns with the identifying assumption that states with different exposure to oxycodone would have experienced similar labor market trends in the absence of the 2010 supply shock. Asusual,wecannotdirectlytestforthisassumption. However,LFPRandEPOPpre-trends are flat in the baseline estimates, and the main estimates are similar with and without the various controls. Also, we show that the estimates are robust to alternative specifications, including other ways of controlling for differences across states in industry composition, demographics, and the severity of the Great Recession. We also address the possibility that the cyclicality of drug use could explain some of the results, both by controlling for geographic variation in other painkillers as well as measuring oxycodone exposure in the early 2000s rather than at the end of the decade. In addition, the main findings are not likely related to the secular decline in manufacturing, as we show state-level manufacturing employment shares and oxycodone rates are uncorrelated. Finally, we show the labor market findings are similar when using alternative specifications and sample 4

restrictions. There is a large literature on trends in LFPR and EPOP following the Great Recession. Many causes, both structural and cyclical, have exerted downward pressure on the LFPR (Aaronson et al. (2014), Council of Economic Advisers (2016), Montes (2018), Abraham and Kearney (2020)). The aging of the Baby Boom generation was an important structural factor while the depth of the financialcrisisandfallingdemandforless-skilledworkersalsocontributedtodecliningparticipation rates, particularlyforprime-age, less-educatedmen. AspositedbyKrueger(2017), theopioidcrisis could be an additional contributing factor. Most of the research on this front has estimated LFPR elasticities with respect to prescription opioid use, including Harris et al. (2020), Aliprantis et al. (2019), and Currie et al. (2019). Our main contribution is to instead estimate the effects from the oxycodone-heroin epidemic, similar to concurrent work by Park and Powell (2021). In addition, we present new NSDUH analysis and new findings on the heterogeneous labor market effects. We find the labor supply effects are stronger for white, less educated, and younger cohorts, in line with the profile of oxycodone misusers prior to 2010. Also, we report negative effects for both men and women, with the latter suggesting the Currie et al. (2019) findings of a positive effect on female LFPR from prescriptions may not take into account the oxycodone-heroin link.5 This paper is also related to the literature on policy interventions in the opioids market. Supply restrictions such as the 2010 OxyContin reformulation and some prescription drug monitoring programs (PDMPs) appear successful in reducing prescriptions (Buchmueller and Carey 2018; Grecu et al. 2019).6 These restrictions prevent some new prescription opioid use disorders from starting(Sacksetal.2019), butcanleadtohigherdemandforheroinandotherdrugsforthosewith existing disorders (Cicero and Ellis 2015; Currie and Schwandt 2020). For the latter, treatment via medicationssuchasmethadoneandbuprenorphineishelpful,butthesetreatmentsareunderutilized (Fullerton et al. 2014; NIDA 2020; Jones et al. 2015). Rees et al. (2019) finds expanding access to overdose-reversing drugs such as naloxone lowers mortality rates, though Doleac and Mukherjee (2018) finds the moral hazard effects could outweigh the benefits. 1 Data and Descriptive Analysis We use data from several sources. Our primary dataset that tracks opioid use is the Drug Enforcement Agency’s (DEA) Automation of Reports and Consolidated Orders System (ARCOS) for 5A growing literature examines the effects of the opioid crisis on broad outcomes including consumer finance (Jansen 2019), house prices (D’Lima and Thibodeau 2019), and firm value and investment (Ouimet et al. 2019). 6While some studies find PDMPs are generally effective in reducing opioid abuse, others find the average effects are modest (Meara et al. (2016); Moyo et al. (2017); Davis 2017). In addition, Horwitz et al. (2018) discuss data issues with PDMP studies that may mixed findings in the literature. 5

purchases of prescription opioids by pharmacies and other buyers from manufacturers. These data allowustomeasureopioiddistributiontoretailpharmaciesbystateandovertime,andweusethese shipments of opioids to pharmacies as a proxy for opioid use. To measure deaths related to opioid use, we use of cause-of-death data from the National Vitality Statistics System managed by the Center for Disease Control and Prevention. The Current Population Survey (CPS) is our primary dataset that measures labor market outcomes, such as labor force participation, employment, and part-time work status by various demographic groups. In addition, we use American Community Survey data via the Census to measure geographic variation in demographic characteristics. To supplement our analysis and to provide a descriptive motivation for our work, we use healthsurvey data to look at patterns of heroin use and the labor market attachment of those with and without nonmedical use of prescription pain medications and heroin. To track those patterns, we use the National Survey of Drug Use and Health, NSDUH, which has been conducted by the federal government since 1971 to measure drug use, addiction, and mental health issues. NSDUH is an annual, nationally representative survey of persons aged 12 and over in the civilian noninstitutionalized population of the U.S.7 The survey contains information on various individual characteristics, including labor force status, education, and demographics such as age and race and ethnicity. Among other questions, respondents are asked if they have ever used drugs such as heroin and if they have done so recently. Though individual survey responses are anonymized and confidential, population estimates of illegal drug use from NSDUH are likely undercounts, given the well-known tendency of individuals to underreport drug use especially heroin and other illegal substances (Harrell 1997; Harrison 1997; Col´on et al. 2001; Morral et al. 2000). Moreover, the survey misses some high-risk groups, such as those imprisoned and without a home or shelter. We begin by measuring the number of heroin users based on NSDUH data. As shown in the left paneloffigure1, thenumberofadultswhorecentlyusedheroinincreasedfromabout350thousand to 900 thousand between 2007 and 2017. Over the same period, the number of adults with prior heroin use increased from about 3.7 million to 5.4 million. Although these are notable increases, they are likely conservative, given the noted measurement issues.8 Indeed, as the right panel of figure 1 shows, the dramatic rise in deaths from heroin overdoses – which increased fivefold from 2007 to 2017 – suggests the increase in users could have been much larger.9 That said, some of the rise in mortality could also reflect riskier consumption practices. Our analysis using the NSDUH data suggest that heroin users are much less likely to be in the labor force. Table 1 reports selected coefficient estimates from a linear probability model of labor 7Seehttps://www.samhsa.gov/data/data-we-collect/nsduh-national-survey-drug-use-and-healthformore information. 8See also “The federal government is systematically undercounting heroin users” (Washington Post; August 22, 2017) 9Ruhm (2018) argues deaths from heroin overdoses are also likely understated. 6

Figure 1: Trends in heroin use and mortality Past 12 months Ever 6 5 4 3 1 8. 6. 4. 2. 0 Used heroin ever (millions) Past 12 months (millions) 2003 2007 2011 2015 2019 Year 00051 00001 0005 0 Heroin overdose deaths 2003 2007 2011 2015 2019 Year Source: National Survey of Drug Use and Health (left) and the National Vital Statistics System (right) for heroin overdose deaths (T40.1). forceparticipation, whichcontrolsforvariousdemographic, education, andlocationcharacteristics. The first column estimates a baseline model without considering drug use, and shows that those with a college degree or higher are about 16 percentage points more likely to be in the labor force than those without a high school degree (the omitted group).10 The second column estimates show that those who have abused pain medication within the past year, including prescription opioids, are about one percentage point less likely to be in the labor force. More strikingly, the model estimates in the third column show that recent heroin users are about 14 percentage points less likely to be in the labor force –a magnitude comparable to the coefficient on the college degree or higher indicator. Given the likely undercount of heroin use, the coefficient estimate on heroin use could be attenuated. That said, we stress that these results, while suggestive, are not causal for many reasons, including the possibility that some omitted variables may help explain both heroin use and labor force status. 10Estimatingasimilarlinearprobabilitymodelincolumn1oftable1intheCPSyieldssimilarcoefficients estimates. 7

Table 1: Probability of being in the labor force for people ages 24 to 49, 2002-2014 LF=100 or 0 Baseline Pain Meds Heroin Coef./SE Coef./SE Coef./SE High school 9.93*** 9.92*** 9.89*** (0.41) (0.41) (0.41) Some college 13.48*** 13.46*** 13.43*** (0.41) (0.41) (0.41) College plus 15.93*** 15.89*** 15.84*** (0.41) (0.41) (0.41) Abuse pain meds -1.00** (0.48) Heroin user -14.30*** (2.32) All other controls Yes Yes Yes R-squared 0.06 0.06 0.06 Observations 238267 238267 238267 Note: Theexplanatoryvariablesincludeeducation,raceandethnicity,sex,age,marriage,year,andmetroarea location fixed effects. Source: National Survey of Drug Use and Health and authors’ calculations. To isolate plausibly exogenous variation in heroin use, we exploit 2010 policy changes that limited the supply of oxycodone and which had the unintended consequence of expanding demand for heroin (Cicero and Ellis 2015; Alpert et al. 2018 and Evans et al. 2019). Oxycodone is the powerful primary ingredient in painkillers such as OxyContin, Percocet, and Roxicodone. Before 2010, shipments and use of oxycodone varied significantly by state. To measure oxycodone and other opioids, we use data from the DEA’s ARCOS database, which tracks shipments of controlled substances (by weight of the active ingredient) from manufacturers to buyers such as pharmacies, practitioners, and universities. Figure 2 shows oxycodone is the largest of the prescription opioids and contributed considerably to the 2000s expansion in prescription opioids. 8

Figure 2: Shipments of oxycodone and other opioids Grams, millions (morphine-equivalents) 90 80 Oxycodone 70 60 50 Methadone 40 30 Morphine Hydrocodone 20 Other 10 0 2000 2005 2010 2015 Source: ARCOS data. Table 2 provides basic characteristics of persons who self-report in NSDUH having used OxyContin for nonmedical reasons within the past year. Recent nonmedical users were much more likely to be white, younger, and without a college degree relative to those without nonmedical use. There are also differences by sex and location, though those are less stark.11 11Note that, in the 2000s, NSDUH did not include an analogous question of recent nonmedical use for broader oxycodone products. 9

Table 2: Characteristics of persons self-reporting nonmedical use of OxyContin, 2006-2009 Nonmedical use within past year? Percent who are: Yes No White (non-Hispanic) 87.15% 68.84% Younger than 50 91.19% 59.34% College degree or higher 10.53% 27.80% High school degree or less 58.04% 46.78% Male 61.27% 48.18% Live in metro area 81.88% 83.30% Live in large metro area 47.97% 53.53% Source: NSDUH, for persons 18 years of age and older. Inthefaceofthegrowingoxycodonecrisis,twokeydevelopmentscurbedthesupplyofoxycodonein 2010. First, PurduePharmareformulatedOxyContin, apopularbrandofoxycodonethathadbeen commonly abused. The new abuse-deterrent formulation successfully impeded many users from crushing and dissolving the pill for non-oral consumption (Severtson et al. 2013b; Severtson et al. 2013a; Cicero and Ellis (2015)). OxyContin had been designed to gradually release large quantities of oxycodone to treat pain over an extended period. However, users crushed or solubilized the pills for a quick and potent release of the oxycodone (Hays 2004; Cicero et al. 2005; Cone et al. 2003). In addition to the reformulation of OxyContin, the DEA and local enforcement began to shut down Florida ‘pill mills’ –clinics with lax standards where physicians prescribed large quantities of oxycodone and other prescription drugs often to out-of-state buyers in 2010. (Surrat et al. 2013; Davis and Carr 2017; Temple 2015). Returning to figure 2, national oxycodone shipments more than quadrupled in the 2000s, and have declined since peaking in 2010. As the supply of oxycodone tightened, many who were addicted to oxycodone turned to heroin as a substitute (Cicero and Ellis 2015; Alpert et al. 2018 and Evans et al. 2019).12 Indeed, deaths from heroin overdoses began to step up shortly after 2010, as documented by Evans et al. (2019) and Alpert et al. (2018). Similar to their approach, we measure state-level oxycodone exposure as the ratio of the oxycodone grams shipped to the state’s chain and retail pharmacies to the state’s 12Cicero et al. (2012) report little substitution specifically on the OxyContin reformulation into other prescriptions such as hydrocodone. At the time, hydrocodone brands including Vicodin contained other ingredients such as acetaminophen, which deters some abuse as high doses of acetaminophen are highly toxic. 10

prime-age population, both averaged over 2006 to 2009.13 Oxycodone2006−2009 OxyRatepre = s (1) s Prime-age population2006−2009 s Figure 3 shows that trends in heroin and synthetic opioids (largely fentanyl) death rates prior to 2010 were similar across states with high and low oxycodone exposure. After 2010, heroin death rates rose nationally, but had a stronger increase in states with high oxycodone exposure (left panel). Throughout the 2010s, deaths from heroin overdoses have been higher in states with higher oxycodone rates, though the gap appears to be closing somewhat in recent years. For synthetic opioids, overdoses trended similarly in both groups and increased more in states with greater oxycodone exposure prior to 2010. Fentanyl deaths began rising a few years after 2010 likely as it took some time for suppliers to innovate (Park and Powell 2021; Pardo et al. 2019). These results reaffirm previous findings by Alpert et al. (2018) and Evans et al. (2019) for heroin overdoses and Powell and Pacula (2021) for fentanyl and cocaine overdoses.14 Figure 3: Overdose death rates by pre-2010 oxycodone exposure Averge of High Oxycodone MME Per Capita States Average of Low Oxycodone MME Per Capita States 61 21 8 4 0 Heroin overdoses per 100,000 Averge of High Oxycodone MME Per Capita States Average of Low Oxycodone MME Per Capita States 2005 2007 2009 2011 2013 2015 2017 2019 04 03 02 01 0 Synthetic opioid overdoses per 100,000 2005 2007 2009 2011 2013 2015 2017 2019 Note: Death rates relative to the prime-age population. Source: ARCOS for state oxycodone rates; National Vital Statistics System for heroin (T40.1) and synthetic opioid (T40.4) overdose deaths. 13Evans et al. (2019) measure is based on the public state-level aggregates of ARCOS data, while our measure is based on the DEA’s ARCOS microdata released by the Washington Post and the Charleston Gazette-Mail of West Virginia. Though in practice the measures are similar, the microdata allow us to exclude manufacturer shipments to buyers such as mail-order pharmacies, practitioners, and universities, which are less likely to use the shipments to fill prescriptions for in-state residents. Alpert et al. (2018) measure is based on state variation in NSDUH estimates of the fraction of OxyContin nonmedical users. They find the NSDUH measure is highly correlated with oxycodone supply measure from ARCOS. 14In addition, Powell et al. (2019) and Beheshti (2019) find evidence of other negative health effects: hepatitis-C infection rates have increased rapidly since 2010, and the increase has been stronger in states with greater pre-2010 exposure to oxycodone. 11

Figure 4: Geographic variation in oxycodone rates, 2006-2009 1 2 3 4 Note: Quartiles of oxycodone rates (mme per capita) with 4 indicating the top quartile. Source: ARCOS. Figure 4 maps quartiles of oxycodone rates prior to 2010. On average, oxycodone rates were highest in the East, comparable in the South and West, and lowest in the Midwest, though there is significant dispersion within regions as well. States in the top quintile of exposure include states with very high heroin mortality rates, such as CT, WV, and NJ. What explains the geographic dispersion in oxycodone rates? Alpert et al. (2019) find that early differences in the supply and marketing of oxycodone have had persistent implications. Indeed, the aggressive marketing tactics used by oxycodone manufacturers such as Purdue Pharma and Mallinckrodtarewell-documented(VanZee2009;Quinones2015;Kolodnyetal.2015).15 Relatedly, differences in physician incentives and quality could explain some of the geographic dispersion (Schnell and Currie (2018)). Anotherpossibilityisthatpooreconomicconditionsmayhavecontributedtopoorhealthoutcomes including both higher deaths (Case and Deaton 2017; Pierce and Schott 2020) and the prescription crisis (Hollingsworth et al. 2017; Charles et al. 2018). Specifically, the positive correlation between manufacturing employment shares and CDC opioid prescription rates suggests lower economic opportunitiescouldpartlyexplaintheprescriptioncrisis. However,oxycodoneratesareweakly(and negatively) correlated with manufacturing employment shares. Hence, the oxycodone-heroin link 15Seealso“PurduePharmaPleadsGuiltytoCriminalChargesforOpioidSales”(NYT;Dec17, 2020)and Mallinckrodt Reaches 1.6 Billion Deal to Settle Opioid Lawsuits (NYT; Feb 25, 2010) 12

is unlikely driven by manufacturing exposure. The positive association between CDC prescription rates and manufacturing seems to be driven by other prescription opioids such as hydrocodone, as shown in 3 and figure 5.16 Table 3: Correlations of state manufacturing shares and prescription opioid indicators Manufacturing CDC Hydrocodone Oxycodone share 1990 RX rate rate rate 2006-2009 2006-2009 2006-2009 Manufacturing Share 1990 1.000 CDC RX Rate 2006-2009 0.402 1.000 Hydrocodone Rate 2006-2009 0.294 0.842 1.000 Oxycodone Rate 2006-2009 -.115 0.301 -.0425 1.000 Source: QCEWformanufacturingemploymentshares;CDCforprescriptionrates;andARCOSforhydrocodone and oxycodone rates. Figure 5 WV KY TN AL LA OK AR MS IN NV WY MTFL AZ DE I U D W T O AM K R GO S MAE NHR M I I OHSC NC ANKM MDCO TV N XA E MA C P T A WI ND SD NJ CAMVTN IL IA HI NY 041 021 001 08 06 04 CDC Prescrption Rate, 2006-2009 DE FL N H V IWY N A MN D K MT MD CO L A N A S Z Y D V W N O N A E VJ K IC U D M A W T M V O AA M T N K R G O C S M A PT I E A A K N Y H R M I I OH T W A N SICLIANR N M C S TX IL .05 .1 .15 .2 .25 .3 Manufacturing Share of Employment, 1990 0051 0001 005 0 Oxycodone MME Per Capita, 2006-2009 .05 .1 .15 .2 .25 .3 Manufacturing Share of Employment, 1990 Source: QCEW for manufacturing employment shares; CDC for prescription rates; and ARCOS for hydrocodone and oxycodone rates. 2 Empirical Framework We estimate the effects of the increase in heroin use on labor market outcomes by proxying for heroin use with exposure to oxycodone prior to 2010. We use this proxy since reliable measures of heroin use are not available, and previous research has shown prior exposure to oxycodone is 16For more discussion on the extent to which changes in drug supply or economic conditions have contributed to the increase in mortality, see Ruhm (2018) and Currie and Schwandt (2020) 13

strongly predictive of increases in mortality from opioids such as heroin and fentanyl. Moreover, oxycodoneexposureisalsopredictiveofotherrelatedpublichealthissues,includinghigherincidence of hepatitis-C (Powell et al. 2019) and drug overdoses from cocaine, often fentanyl-laced (Powell and Pacula 2021). Thus, exposure to oxycodone proxies for both increases in heroin use and other related negative public health outcomes caused by the decline in oxycodone supply. Specifically, we employ a difference-in-difference framework exploiting variation in states’ exposure to oxycodone prior to 2010 interacted with year effects, according to equation 2. Our estimates show differences in how labor markets evolved as explained by variation in their pre-2010 exposure to oxycodone. This approach is similar to that of Alpert et al. (2018), Evans et al. (2019), Powell and Pacula (2021), and Park and Powell (2021) who show the 2010 shock to oxycodone supply led to higher demand for heroin and heroin overdoses. Y = α +γ +δ ×OxyRatepre+ωX +(cid:15) (2) st t s t s st st State-level dependent variables Y include EPOP, LFPR, and other indicators including death st rates from heroin and synthetic opioids (mainly fentanyl). We interact OxyRatepre, the average s oxycodone rate over 2006-2009, scaled to have mean zero and standard deviation of 1, with a full set of year fixed effects. The model includes year and state fixed effects, α and γ , respectively, to t s account for both national trends and fixed differences across states. X include different cyclical st and structural labor market characteristics: the 2010 white, male, college graduate, and foreignborn population shares, each interacted with a full set of year fixed effects; the lagged employment share in the service industry; and house price and unemployment rate changes from 2003 to 2006 and2006to2009,eachinteractedwithafullsetofyearfixedeffects.17 Standarderrorsareclustered at the state-level to account for serial correlation. In the baseline, we include the 50 U.S. states and the District of Columbia and use regressions weighted by population. The main coefficients of interest are estimates of δ , which are shown graphically with the 2009 t coefficient normalized to zero. These estimates show the causal effect of the oxycodone-heroin shift on the labor market, conditional on the identifying assumption that labor markets would have evolved similarly across states absent the oxycodone supply change in 2010. As usual, this assumption is not testable. However, pre-trends in our baseline estimates are flat, and the main estimates with and without the various controls are similar. In addition, the estimates are robust to different specifications including other ways of accounting for state-level variation in the severity of the Great Recession, industry composition, other drug use, region trends, and other controls. 17Data come from the American Community Survey via the Census Bureau; the Quarterly Census of Employment and Wages via Bureau of Labor Statistics for industry composition; CoreLogic, Inc., Private- Label Loan, Home Equity Servicing and HPI Data for house prices; and the Current Population Survey for the labor market indicators used throughout the paper. 14

Also, we show throughout the paper that the labor market estimates are similar in different specifications, including unweighted regressions and sample restrictions. Furthermore, as noted above, manufacturingemploymentsharesandoxycodoneratespre-2010areessentiallyuncorrelated,hence our labor market estimates are unlikely to reflect manufacturing exposure. We also estimate the following related specification, where β estimates the conditional difference in the level of the outcome variable by oxycodone exposure between the pre and post-periods: Y = α +γ +β (cid:2) Post ×OxyRatepre(cid:3) +ωX +(cid:15) (3) st t s t s st st Post is an indicator that is equal to 1 for years 2011 to 2019 and is zero for years 2005-2010. t Similarly to Alpert et al. (2018) and Park and Powell (2021), we conservatively treat 2010 as part of the pre-period. We rely on both specifications, as the model in equation 2 is more flexible, but the model in equation 3 allows for more concise reporting. Forourlabormarketdependentvariables, weconstructlocalareaaggregatesfromtheCPS.Forthe baselinespecifications, wefocusonpeoplewhowereofprime-age(25-54yearolds)in2010, theyear that the supply of oxycodone was tightened significantly and the shift to heroin began. We choose to track this specific prime-age cohort, as this group faced a plausibly exogenous and unexpected shock to oxycodone supply. The cohort approach is a tractable way of omitting prime-age workers who were teenagers in 2010 and thus less attached to the labor force at that time. However, in the robustness section we also include estimates for persons who were 25-54 years old each year and find the results are similar. 15

Figure 6: LFPR for the prime-age cohort in 2010 LFPR for the prime-age cohort in 2010 83 82 Age-adjusted LFPR 81 80 79 LFPR 78 2005 2007 2009 2011 2013 2015 2017 2019 Source: CPS and authors’ calculations Figure6plotstheaggregateLFPRforthiscohortofindividualswhowere25to54yearsoldin2010. As the cohort has aged, participation has trended down, since this cohort started to age into its retirement ages after 2010, and those ages are associated with significantly lower LFPRs compared to the prime-age. Of course, we want to avoid picking up the effects of aging for this cohort in our estimates of heroin use on labor market outcomes. Thus, in all of our baseline estimates, we age-adjust each of our labor market outcome variables. Specifically, we use the CPS sample from 1995 on to estimate equation 4 at the individual i level, and aggregate the residuals: Y = θ +υ (4) ist age(i) ist where θ are a full set of age fixed effects. Figure 6 plots the aggregate age-adjusted LFPR, age(i) normalized to start in 2005 at the same level as the unadjusted series. The age-adjustment leads to an appreciable difference in the aggregate LFPR time series. After declining after the Great Recession,theage-adjustedLFPRhasrecoveredgraduallyasthelabormarkethasstrengthened. In our baseline estimates, we use the age-adjusted measures. That said, we show the age-adjustment has little effect on our labor market estimates, which are cross-sectional. More broadly, we find little evidence of a change post-2010 in the age and other demographic structure of states with greater oxycodone exposure. 16

Table 4: Effects on heroin and synthetic opioids deaths per 100,000 Dependent variables: Heroin Heroin Heroin Heroin Synth. Synth. Synth. Synth. Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 1.717** 2.120*** 0.771 2.312** 4.978*** 4.477*** 3.138** 5.274*** (0.81) (0.78) (0.96) (1.11) (1.31) (1.11) (1.33) (1.65) Other controls No Yes Yes Yes No Yes Yes Yes Weighted No No Yes Yes No No Yes Yes FL included Yes Yes Yes No Yes Yes Yes No R-squared 0.36 0.70 0.78 0.80 0.18 0.79 0.84 0.84 Observations 765 765 765 750 765 765 765 750 The table shows estimates of β from equation 3. The models in columns 1 and 5 include only the levels and interaction of Post and OxyRatepre. The remaining columns include all baseline controls. Standard errors t s clustered by state. *** p<0.01, ** p<0.05, * p<0.10. 3 Results First, we briefly review and update the evidence that states with greater exposure to oxycodone prior to 2010 experienced larger increases in death rates from heroin and other illicit opioids after 2010. Second, we show that the oxycodone-heroin shock led to persistently lower EPOP and LFPR since 2010 in more affected states, particularly for non-Hispanic whites, both male and female, younger cohorts, and persons without a college degree. These effects are likely driven by behavioralchangesratherthandemographicchangesincompositioncausedbytheincreaseinheroin demand and mortality. Next, we discuss the aggregate implications of the baseline findings, before concluding with a discussion about the sensitivity of these estimates to alternative specifications. 3.1 Effects on heroin and synthetic opioid death rates First, we update and review the evidence presented in Alpert et al. (2018), Evans et al. (2019), Park and Powell (2021), Powell et al. (2019), Powell and Pacula (2021), and others, that areas with greater exposure to oxycodone pre-2010 experienced subsequent relative increases in heroin and synthetic opioid (mainly fentanyl) mortality. Relative to the cited papers, the data we use (through 2019) is more recent.18 Table4reportsestimatesfromequation3fortheheroinandsyntheticopioiddeathratemodelswith 18Alpertetal.(2018)reportdatathrough2013; Evansetal.(2019)through2014; ParkandPowell(2021) and Powell et al. (2019) through 2015; and Powell and Pacula (2021) through 2017. 17

and without all the baseline controls, population weights, and the inclusion of Florida. Generally, across the models, states with one standard deviation higher oxycodone rates pre-2010 experienced on average a relative increase of 2 more deaths from heroin overdoses and 5 more deaths from synthetic opioids per 100 thousand of the prime-age population in the 2011-2019 period relative to the 2005-2010 pre-period. Though we find the results are typically robust, we note the estimates aresmallerintheheroinmodelwhenincludingFloridaandweightingbypopulation. Inoursample, Florida has a high oxycodone rate pre-2010, but some oxycodone prescriptions there were likely diverted to other states as the ‘pill mills’ boomed. In line with the cited literature, we keep Florida andweightbypopulationinourbaselinespecifications. However,weincluderesultswithoutFlorida and without population weights throughout the paper, and find the labor market effects are quite similar. Figure 7 plots point estimates and 95 percent confidence intervals from equation 2 for heroin and synthetic opioid death rates for regressions without population weights (top row), with population weights (middle row), and with population weights but without Florida (bottom row). In line with the findings in the literature, the pre-trends are flat and heroin deaths rates appreciably step up shortly after 2010. Fentanyl death rates begin rising a few years later, as it took some time for suppliers to introduce fentanyl (Powell and Pacula 2021). In recent years, the coefficient estimates in the heroin mortality models have plateaued or declined. Recently, heroin death have declined somewhat, particularly in states with higher oxycodone exposure pre-2010 (see also figure 3). Thus, we might expect the labor market effects to also plateau or decline in recent years. That said, heroin mortality rates are a noisy proxy for heroin use, as mortality rates reflect risky consumption practices. 3.2 Effects on the labor market Table 5 presents estimates from equation 3, which quantify the effect of the oxycodone-heroin shift over the 2011-2019 period on EPOP and LFPR for the baseline cohort. Columns 1 and 5 show estimates for the EPOP and LFPR models without other controls, while the remaining columns report estimates including all the baseline controls. With or without controls, the coefficient estimates are quite similar, even as the R-squared is much higher when including controls. The table also reports estimates with and without population weights and with and without Florida. Overall, the coefficient estimates are quite similar in the different specifications. In the model with population weights and Florida, the coefficient estimate on EPOP and LFPR is around -0.30, indicating states with one standard deviation higher oxycodone exposure before 2010 had a 30-basis point relative decline in LFPR and EPOP in the 2011-2019 period relative to the 2005-2010 pre-period. As the estimates on EPOP and LFPR are very close, the effects on the unemployment rate are very small 18

Figure 7: Effects on heroin and synthetic opioid death rates tneiciffeoC detamitsE 8 6 4 2 0 Heroin : unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 02 51 01 5 0 Synthetic opioids: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 6 4 2 0 2- Heroin : weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 51 01 5 0 5- Synthetic opioids: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 8 6 4 2 0 2- Heroin : weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 02 51 01 5 0 Synthetic opioids: weighted, no FL 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) for overdose deaths per hundred thousand of the prime-age population 2009 from heroin and synthetic opioids overdoses. Regressions are unweighted (top row), weighted (middle row), and weighted without Florida (bottom). Baseline controls included. Standard errors clustered by state. 19

Table 5: Effects on labor market indicators Dependent variables: (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) EPOP EPOP EPOP EPOP LFPR LFPR LFPR LFPR SSDI PT Coef/SECoef/SE Coef/SE Coef/SE Coef/SECoef/SECoef/SECoef/SECoef/SECoef/SE Oxy Rate pre X 2011−2019 -0.274** -0.241**-0.328***-0.369*** -0.226* -0.199 -0.287** -0.346* 0.056 0.203 (0.12) (0.11) (0.11) (0.14) (0.13) (0.13) (0.13) (0.18) (0.05) (0.13) Other controls No Yes Yes Yes No Yes Yes Yes Yes Yes Weighted No No Yes Yes No No Yes Yes Yes Yes FL included Yes Yes Yes No Yes Yes Yes No Yes Yes R-squared 0.02 0.86 0.90 0.90 0.15 0.62 0.67 0.65 0.95 0.85 Observations 765 765 765 750 765 765 765 750 707 765 The table shows estimates of β from equation 3. The models in columns 1 and 5 include only the levels and interaction of Post and OxyRatepre. The remaining columns include all baseline controls. Standard errors t s clustered by state. *** p<0.01, ** p<0.05, * p<0.10. (not shown). Hence, the results suggest the effects on EPOP are entirely from the participation margin. Figure 8 plots point estimates and 95 percent confidence intervals from equation 2 for the baseline EPOP and LFPR models for regressions without population weights (top row), with population weights(middlerow),andwithpopulationweightsbutwithoutFlorida(bottomrow). Priorto2010, the coefficients are close to zero and not statistically significant. After 2010, however, states with previously larger oxycodone exposure experienced relative declines in EPOP and LFPR. Though the negative effects have persisted through the 2010s, they have grown smaller in recent years, perhaps as differences in heroin use across states have grown smaller. Indeed, the gap in heroin death rates explained by oxycodone exposure has fallen recently (see figure 7). Overall, the patterns and magnitudes of the coefficient estimates across specifications are quite similar. Hence, from now on we use specifications with population weights and Florida as the baseline. That said, we include additional results in the robustness section for the alternative specifications. Figure 9 plots point estimates and 95 percent confidence intervals for models of the rate of social security disability insurance (SSDI) beneficiaries and the prime-age part-time employment rate.19 Consistent with Park and Powell (2021), we find substitution towards heroin has caused some increase in the rate of disability beneficiaries. Within employment, we also find some evidence of an increase in the share of part-time workers. Columns 9 and 10 of Table 5 show a relative increase 19Total SSDI beneficiaries data come from the Social Security Administration. The microdata is not available so we could not construct measures for the prime-age cohort in 2010 as we did for CPS variables. 20

Figure 8: EPOP and LFPR effects tneiciffeoC detamitsE 5. 0 5.- 1- 5.1- Prime-age EPOP: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 5. 0 5.- 1- 5.1- Prime-age LFPR: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 5. 0 5.- 1- 5.1- Prime-age EPOP: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 5. 0 5.- 1- 5.1- Prime-age LFPR: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 5. 0 5.- 1- 5.1- Prime-age EPOP: weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 5. 0 5.- 1- 5.1- Prime-age LFPR: weighted, no FL 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) for the rates of employment to population (left) and labor force par- 2009 ticipation (right). Regressions are unweighted (top row), weighted (middle row), and weighted without Florida (bottom). Baseline controls included. Standard errors clustered by state. 21

Figure 9: SSDI and part-time employment rate effects tneiciffeoC detamitsE 2. 1. 0 1.- SSDI Rate: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 5.1 1 5. 0 5.- Part-time employment rate: weighted 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) (with δ normalized to zero) for the rates of social security disabil- 2009 2009 ity insurance beneficiaries (left) and part-time employment (right). Baseline controls included. Standard errors clustered by state. Observations weighted by population. in SSDI and the part-time employment rate in states with greater oxycodone exposure though the increases are modest and not statistically significant. Next, we explore the heterogenous effects of the oxycodone-heroin epidemic. We estimate separate models of equation 2 for local area aggregates by demographic groups. Figure 10 plots point estimates and 95 percent confidence intervals for LFPR models by age, race and ethnicity, and education. Starting with the age results in the top row, the left panel shows estimates for the baseline prime-age cohort, and the right panel shows the estimates for the cohort aged 55-64 in 2010. For the older cohort, the point estimates are generally not significant and tend to be positive, in contrast with the negative and significant estimates for the younger cohort. The middle row shows estimates for non-Hispanic whites (left) and nonwhites (right), while the bottomrowshowsestimatesforthosewithoutacollegedegree(left)andthosewithacollegedegree or higher (right). For non-Hispanic whites and those without a college degree, the estimates show a negativeandsignificanteffectoftheincreaseinheroinuse. Fornonwhitesandcollegegraduates,the estimates are not generally significant and are sometimes positive (for college graduates). Tables 6 and 7 summarize the heterogeneous results for both LFPR and EPOP, respectively, based on estimates from equation 3, showing analysis consistent with figure 10. Figure 11 plots estimates for LFPR by sex and ethnicity. For white males and females (top row), the estimates are generally negative and significant. In contrast, for nonwhites (bottom row) the effects seem largely flat though they are noisier particularly for nonwhite females. Though the negative effects on the white, younger, and less educated cohorts are stark, we stress we do not interpret these results as showing the oxycodone-heroin epidemic as having a negligible 22

Figure 10: LFPR effects by age, race and ethnicity, and education tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Ages 55-64 LFPR: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: White: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Nonwhite: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Less than college: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: College or higher: weighted 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalizedtozero)for: theprime-agecohortin2010(topleft);thecohortaged55-64in2010 2009 (top right); prime-age non-Hispanic whites (middle left); nonwhites (middle right); persons withoutacollegedegree(bottomleft);personswithacollegedegreeorhigher(bottomright). Baseline controls included. Standard errors clustered by state. Observations weighted by population. 23

Figure 11: LFPR effects by sex and race and ethnicity tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: White male 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: White female 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Nonwhite male 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Nonwhite female 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) for prime-age in 2010 non-Hispanic white males (top left) and females 2009 (top right); and nonwhite males (bottom left) and females (bottom right). Baseline controls included. Standard errors clustered by state. Observations weighted by population. effect on other groups. The results in table show that for nonwhites and college graduates the estimates are negative, and the standard errors are large enough that we can neither reject the hypothesis that they are zero nor that they are large and negative. Moreover, since 2010, heroin and fentanyl mortality have risen across demographic groups (Drake et al. 2020). Next, we examine whether these effects could be explained by composition rather than behavioral changes associated with the oxycodone-heroin shock. For example, changes in death rates or net migrationacrossstatesduetotheshifttoheroinusewouldcausechangesincompositionthatcould affect the LFPR and EPOP. To explore this possibility, we decompose EPOP and LFPR into age, education, and demographic components. Specifically, we estimate the following equation akin to equation 4 and aggregate the predicted values to the state and year levels.20 Y = θ +θ +θ +θ +υ (5) ist age(i) edu(i) sex(i) race(i) ist 20Foreducation,thecategoriesusedare: somehighschoolorless;highschool;somecollege;collegedegree; andmorethancollege. Forraceandethnicity, thecategoriesusedare: non-Hispanicwhite; Black; Hispanic; and other. The fitted values are aggregated using the person weights provided in the CPS. 24

Table 6: Effects on LFPR by demographic groups LFPR: White WM WF No Col Age 25−34 Nonwhite Col Age 55−64 Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.375** -0.375** -0.358 -0.297* -0.305** -0.133 -0.125 0.237 (0.16) (0.17) (0.23) (0.16) (0.15) (0.18) (0.12) (0.19) Other controls Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.59 0.56 0.48 0.75 0.57 0.50 0.53 0.60 Observations 765 765 765 765 765 765 765 765 The table shows estimates of β from equation 3. Estimates are for the prime-age cohort in 2010 unless otherwise stated. Results shown for: non-Hispanic whites (column 1), white males (column 2), and white females (column 3); persons without a college degree (column 4); the cohort aged 20-32 in 2010 (column 5); nonwhites(column6); personswithacollegedegreeorhigher(column7); andthecohortaged55-64in2010 (column 8). All baseline controls included. Standard errors clustered by state. Observations weighted by population. *** p<0.01, ** p<0.05, * p<0.10. Table 7: Effects on EPOP by demographic groups EPOP: White WM WF No Col Age 25−34Nonwhite Col Age 55−64 Coef/SE Coef/SECoef/SECoef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.409***-0.398** -0.404* -0.311** -0.266 -0.221 -0.157 0.201 (0.15) (0.17) (0.20) (0.13) (0.16) (0.15) (0.13) (0.19) Other controls Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.82 0.81 0.68 0.89 0.84 0.84 0.68 0.65 Observations 765 765 765 765 765 765 765 765 The table shows estimates of β from equation 3. Estimates are for the prime-age cohort in 2010 unless otherwise stated. Results shown for: non-Hispanic whites (column 1), white males (column 2), and white females (column 3); persons without a college degree (column 4); the cohort aged 20-32 in 2010 (column 5); nonwhites (column 6); persons with a college degree or higher (column 7); and the cohort aged 55-64 in 2010 (column 8). All baseline controls included. Standard errors clustered by state. Observations weighted by population. *** p<0.01, ** p<0.05, * p<0.10. 25

Table 8: Effects on composition EPOP predicted by: LFPR predicted by: Age Educ. Demographics Age Educ. Demographics Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.038 0.017 -0.047 -0.031 0.018 -0.034 (0.03) (0.06) (0.05) (0.02) (0.05) (0.04) Other controls Yes Yes Yes Yes Yes Yes Observations 765 765 765 765 765 765 The table shows estimates of β from equation 3 for EPOP and LFPR predictions based on age, education, and broader set of demographics (see equation 5). All baseline controls included. Standard errors clustered by state. Observations weighted by population. *** p<0.01, ** p<0.05, * p<0.10. We include the predicted values for EPOP and LFPR as dependent variables, for decompositions fromageeffectsonly,educationeffectsonly,andbroaddemographiceffectsincludingage,education, sex, and race and ethnicity. Table 8 reports the baseline estimates from equation 3. Across the models,thecoefficientestimateonoxycodoneexposureisclosetozeroandnotsignificant,indicating states with greater oxycodone exposure did not experience a relative change in their demographics structure after 2010 in a way that majorly impacted EPOP or LFPR. Thus, we interpret the main findings as likely reflecting behavioral changes from greater heroin use rather than compositional effects. 26

3.3 Aggregate Implications To gain a sense of the aggregate implications on LFPR, we combine the baseline estimates with a counterfactual assumption in which all states have low oxycodone rates and are unaffected by the oxycodone-heroin shock. As is generally the case, this counterfactual is not observed. We begin by assuming states in the bottom 5 percent of the oxycodone rate distribution were unaffected, and reduce oxycodone rates for the remaining states to the 5th percentile.21 Next, we calculate the counterfactual LFPR by year using the baseline estimates from equation 2. This main exercise suggeststheoxycodone-heroinshockcontributedtoanaverage40basispointdeclineinLFPRfrom 2010 to 2019. In an alternative counterfactual, we assume the lowest quartile of the distribution by oxycodone rates was unaffected, and reduce oxycodone rates for the remaining states to the 25th percentile. In this exercise, the oxycodone-heroin shock explains an average 25 basis point decline in LFPR from 2010 to 2019. We relate these magnitudes to the gap between the unadjusted LFPR and the LFPR predicted by demographics(age,educationalattainment,sex,andraceandethnicity)fromequation5. Figure12 plots the LFPR, the demographics-predicted LFPR, and the counterfactual LFPR under the main aggregation exercise, with the latter two normalized to start at the same level as the unadjusted series in 2005. On average, LFPR was about 1.5 percentage points below the demographicspredictedLFPRfrom2010to2019,likelymainlyreflectingcyclicaleffectsfromtheGreatRecession (Cajner et al. (2020)). The main counterfactual explains about one-quarter of the gap during this period, whereas the alternative counterfactual explans about one-sixth of the gap. Theaccuracyoftheseaggregationexercisesdependonotherfactorsinadditiontothecounterfactual assumptions. This approach treats states as independent units, but trade and other equilibrium effects could be important.22 Also, there is noise in the baseline estimates. When repeating the aggregation exercise using the 90 percent confidence intervals from equation 2, the oxycodoneheroin shock could explain between 3 and 76 basis point decline in LFPR for the period between 2010 and 2019. Given these caveats, we interpret these exercises as a useful robustness check. That said, there is much uncertainty around the aggregate effect. 21Alternatively, we could have assumed that all states had zero oxycodone rates prior to 2010. However, this would require extrapolating out of the sample as all states had positive oxycodone rates. See Alpert et al. (2018) for a related approach and discussion. 22See Chodorow-Reich (2020) for a discussion on translating regional estimates to aggregate implications. 27

Figure 12: Counterfactual LFPR for the prime-age cohort in 2010 83 82 Counterfactual Demographic-predicted 81 LFPR 80 79 LFPR 78 2005 2007 2009 2011 2013 2015 2017 2019 The figure plots the LFPR, the LFPR predicted from equation 5, and the LFPR from the main counterfactual where states had low enough oxycodone rates prior to 2010 they were unaffected by the decline in oxycodone supply. The latter two are normalized to equal the level of LFPR in 2005. See the text for details. 3.4 Robustness In this section we examine the sensitivity of the baseline EPOP and LFPR estimates to alternative specifications. First, we report additional results for specifications without Florida and without population weights. Next, we report estimates for the prime-age cohort each year rather than for the prime-age cohort in 2010. Also, we address the extent to which the findings could be explained by: the cyclicality of drug use; unobserved differences between states in cyclical or demographic characteristics; industry composition; the age-adjustment to LFPR and EPOP; and using metropolitan statistical area (MSA)-level rather than state-level data. In the main text, we provided results when dropping Florida or estimating regressions without population weights for the main LFPR and EPOP specifications. Here, we provide additional results for different demographic groups. The additional estimates from equation 3 are reported in tables 9 and 10 for EPOP and LFPR, respectively. Also, figures 13 and 14 plot estimates from equation 2 for LFPR for the different demographic groups. Overall, the results without Florida and without population weights are quite similar to the baseline. In our baseline specification, we track the labor market outcomes for the prime-age cohort in 2010, as this cohort faced a plausibly exogenous and unexpected shock to the supply of oxycodone. 28

Table 9: Effects on EPOP without Florida or population weights NoFlorida Unweighted All White No Col Age 55−64 All White No Col Age 55−64 Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.369*** -0.457** -0.336* 0.309 -0.241** -0.310* -0.233 0.171 (0.14) (0.18) (0.18) (0.26) (0.11) (0.15) (0.16) (0.33) Other controls Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.90 0.83 0.89 0.65 0.86 0.78 0.85 0.50 Observations 750 750 750 750 765 765 765 765 The table shows estimates of β from equation 3 for the baseline prime-age cohort in 2010, non-Hispanic whites, persons without a college degree, and the cohort that was 55-64 in 2010. In columns 1-4, weighted estimates are shown without Florida. In columns 5-8, results are unweigthed. Standard errors clustered by state. *** p<0.01, ** p<0.05, * p<0.10. Table 10: Effects on LFPR without Florida or population weights NoFlorida Unweighted All White No Col Age 55−64 All White No Col Age 55−64 Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.346* -0.452** -0.340 0.366 -0.199 -0.274* -0.209 0.201 (0.18) (0.20) (0.24) (0.27) (0.13) (0.16) (0.19) (0.33) Other controls Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.65 0.58 0.73 0.61 0.62 0.57 0.69 0.46 Observations 750 750 750 750 765 765 765 765 The table shows estimates of β from equation 3 for the baseline prime-age cohort in 2010, non-Hispanic whites, persons without a college degree, and the cohort that was 55-64 in 2010. In columns 1-4, weighted estimates are shown without Florida. In columns 5-8, results are unweigthed. Standard errors clustered by state. *** p<0.01, ** p<0.05, * p<0.10. 29

Figure 13: LFPR effects by demographic groups, without FL tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Ages 55-64 LFPR: weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: White: weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Nonwhite: weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Less than college: weighted, no FL 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: College or higher: weighted, no FL 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) for non-Hispanic white males (top left) and females (top right); and 2009 nonwhite males (bottom left) and females (bottom right). Baseline controls included. Standard errors clustered by state. Observations weighted by population and Florida excluded. 30

Figure 14: LFPR effects by demographic groups, unweighted tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Ages 55-64 LFPR: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: White: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Nonwhite: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: Less than college: unweighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 2 1 0 1- 2- Prime-age LFPR: College or higher: unweighted 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) for non-Hispanic white males (top left) and females (top right); and 2009 nonwhite males (bottom left) and females (bottom right). Baseline controls included. Standard errors clustered by state. Figure 15: EPOP and LFPR effects: Ages 25-54 tneiciffeoC detamitsE 5. 0 5.- 1- EPOP: Ages 25-54: weighted 2005 2010 2015 2020 year tneiciffeoC detamitsE 5. 0 5.- 1- LFPR: Ages 25-54: weighted 2005 2010 2015 2020 year The figure shows point estimates and 95 percent confidence intervals of δ from equation 2 (with t δ normalized to zero) for EPOP and LFPR for persons aged 25-54 each year. Standard errors 2009 clustered by state. Observations weighted by population. 31

Table 11: Effects on EPOP and LFPR by age groups EPOP: LFPR: All White No Col Age 55−64 All White No Col Age 55−64 Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.291*** -0.401*** -0.206* 0.206 -0.272*** -0.368*** -0.237** 0.258 (0.08) (0.11) (0.11) (0.22) (0.07) (0.10) (0.10) (0.22) Other controls Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.91 0.84 0.90 0.46 0.70 0.61 0.76 0.33 Observations 765 765 765 765 765 765 765 765 The table shows estimates of β from equation 3 for the group each year that was prime-age, white, without a college degree, and aged 55-64. Standard errors clustered by state. Observations weighted by population. *** p<0.01, ** p<0.05, * p<0.10. Instead of tracking this cohort, we now report estimates for the prime-age group each year. Figure 15 shows estimates from equation 2 and table 11 shows estimates from equation 3. The results are similar to the baseline. As drug use could vary with economic conditions (Hollingsworth et al. 2017), it is plausible that geographic variation in oxycodone rates measured over 2006 to 2009 could have been partly caused by differences across states in the severity of the Great Recession. We address this possibility in two ways – by controlling for exposure to other prescription drugs, and by measuring oxycodone exposure in 2000 rather than the 2006-2009 period – and do not find this channel could have an important effect on the estimates. First, we include as additional controls hydrocodone rates measured over 2006 to 2009 with a full set of year effects; these estimates are reported in columns 1-2 of table 12. Hydrocodone is the second largest of the prescription opioids and state-level hydrocodone rates are correlated with the manufacturing share of employment, as shown in figure 5. Second, we repeat the baseline estimates, measuring average oxycodone exposure from 2000 to 2001 instead of from 2006 to 2009. The correlation coefficient between these two measures is 0.85, consistent with Alpert et al. (2019) who find that early differences in oxycodone supply have had persistent implications. Columns 3-4 of table 12 show these estimates. In both cases, the coefficients of interest are very similar to the baseline estimates. Another possibility is that omitted variable bias may influence the results if there are unobserved factorscorrelatedwithoxycodoneuseandtheperformanceoflabormarkets. Thebaselineestimates already flexibly control for differences in the severity of the Great Recession, demographics, and industry composition, as discussed in section 2. Indeed, the LFPR and EPOP pre-trends are flat as shown in figure 8. Also, employment for nonwhites tends to covary more with the cycle than for whites (Cajner et al. (2017)), but our findings are stronger for the latter. Nonetheless, we estimate 32

Table 12: Robustness checks Hydrocodone Oxy2000 URandHPI Manufshares EPOP LFPR EPOP LFPR EPOP LFPR EPOP LFPR Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.341*** -0.309** -0.400*** -0.372*** -0.304** -0.301** (0.11) (0.12) (0.07) (0.08) (0.12) (0.14) Oxy Rate 2000 X 2011−2019 -0.227* -0.248* (0.12) (0.14) Other controls Yes Yes Yes Yes Yes Yes Yes R-squared 0.91 0.69 0.90 0.66 0.88 0.62 0.90 0.67 Observations 765 765 765 765 765 765 765 765 The table shows estimates of β from equation 3 for alternative models. In columns 1-2, hydrocodone rates from 2006 to 2009 are interacted with year effects. In columns 3-4, oxycodone rates are measured from 2000 to 2001. In columns 5-6, three lags of the unemployment rate and house price index are included. In colums 7-8, threelagsoftheservicesandmanufacturingemploymentsharesareincluded. Standarderrorsclustered by state. Observations weighted by population. *** p<0.01, ** p<0.05, * p<0.10. alternative specifications. First, we include three lags of the unemployment rate and house price index as controls, instead of the approach in the baseline where we interact changes from 2003 to 2006 and 2006 to 2009 with year fixed effects. Columns 5-6 of table 12 show these estimates. Second, we include three lags of both the manufacturing share of employment and the service industry share of employment, instead of only controlling for one lag of the service share as in the baseline (columns 7-8). The estimates are close to the baseline results. In addition, when constructing the EPOP and LFPR aggregates, we strip out a broad set of demographic effects beyond age (see equation 5) and repeat the analysis.23 Columns 1-2 of table 13 show these estimates. We also report estimates for the unadjusted LFPR and EPOP in columns 3-4 of Table 13. In both cases, the results are similar to the baseline. Finally, we estimate additional specifications. We report estimates when including division fixed effects interacted with year effects to account for common regional trends (columns 5-6).24 Lastly, we repeat the analysis using MSA-level data rather than state-level data (columns 7-8). We prefer state-level specifications as estimates based on smaller geographies are likely noisier from both greater oxycodone diversion and smaller CPS samples. We are able to aggregate the LFPR and EPOP into about 200 distinct MSAs using the CPS microfiles. In these cases, the coefficient estimates are again similar to the baseline. 23Specifically, we separately regress LFPR and EPOP on age, education, race and ethnicity, and sex fixed effects and aggregate the residuals to the state-year level. 24The U.S. Census divides the country into 9 divisions. See here for more info. 33

Table 13: Additional robustness checks Demographics Unadjusted Division MSA EPOP LFPR EPOP LFPR EPOP LFPR EPOP LFPR Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Oxy Rate pre X 2011−2019 -0.319** -0.284** -0.366*** -0.318** -0.359** -0.334** -0.326** -0.292* (0.12) (0.14) (0.11) (0.13) (0.14) (0.15) (0.15) (0.16) Other controls Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.90 0.71 0.86 0.90 0.94 0.79 0.53 0.23 Observations 765 765 765 765 765 765 2910 2910 The table shows estimates of β from equation 3 for alternative models. In columns 1-2, EPOP and LFPR are adjusted for age, education, race, and sex (as per equation 5). In columns 3-4, Florida is dropped. In columns 5-6, controls include the interactions of year and U.S. division fixed effects. Columns 7-8 use MSA-level data with outliers dropped (1 percent of each tail by oxycodone rates). Standard errors clustered by state. Observations weighted by population. *** p<0.01, ** p<0.05, * p<0.10. 34

4 Conclusion Previous work, including Alpert et al. (2018), Evans et al. (2019), and Powell and Pacula (2021), have shown areas with greater oxycodone exposure prior to 2010 experienced relative increases in heroin and fentanyl death rates after prescription supply tightened in 2010. We contribute to this literature by showing more exposed areas also experienced relative declines in EPOP and LFPR, particularly for white, less educated, and younger cohorts, consistent with the profile of oxycodone nonmedical users prior to 2010. These results indicate that, when accounting for the effects of the opioid crisis, extending beyond prescriptions is important. In particular, while tighter oxycodone supply may have deterred some new addictions, many with existing opioid use disorders turned to heroin and other drugs including fentanyl. In recent years, deaths from heroin and fentanyl overdoses have surpassed deaths from prescription opioid overdoses. We view research on the health and economic effects of medication-assisted and other treatments for opioid use disorder as an important area of further work. 35

References Aaronson, S., Cajner, T., Fallick, B., Galbis-Reig, F., Smith, C., and Wascher, W. (2014). Labor force participation: recent developments and future prospects. Brookings Papers on Economic Activity, 2014(2):197–275. Abraham, K. G. and Kearney, M. S. (2020). Explaining the decline in the us employment-topopulation ratio: A review of the evidence. Journal of Economic Literature, 58(3):585–643. Aliprantis, D., Fee, K., and Schweitzer, M. E. (2019). Opioids and the Labor Market. Working Papers 201807R2, Federal Reserve Bank of Cleveland. Alpert, A., Powell, D., and Pacula, R. L. (2018). Supply-side drug policy in the presence of substitutes: Evidence from the introduction of abuse-deterrent opioids. American Economic Journal: Economic Policy, 10(4):1–35. Alpert, A. E., Evans, W. N., Lieber, E. M., and Powell, D. (2019). Origins of the opioid crisis and its enduring impacts. Technical report, National Bureau of Economic Research. Beheshti, D. (2019). Adverse health effects of abuse-deterrent opioids: Evidence from the reformulation of oxycontin. Health economics, 28(12):1449–1461. Buchmueller, T. C. and Carey, C. (2018). The Effect of Prescription Drug Monitoring Programs on Opioid Utilization in Medicare. American Economic Journal: Economic Policy, 10(1):77–112. Cajner, T., Coglianese, J., and Montes, J. (2020). The Long-Lived Cyclicality of the Labor Force Participation Rate. Unpublished working paper, Federal Reserve Board of Governors. Cajner,T.,Radler,T.,Ratner,D.,andVidangos,I.(2017). RacialGapsinLaborMarketOutcomes in the Last Four Decades and over the Business Cycle. Finance and Economics Discussion Series 2017-071, Board of Governors of the Federal Reserve System (U.S.). Case, A. and Deaton, A. (2017). Mortality and Morbidity in the 21st Century. Brookings Papers on Economic Activity, 48(1 (Spring):397–476. Charles, K. K., Hurst, E., and Schwartz, M. (2018). The Transformation of Manufacturing and the Decline in U.S. Employment. Working Papers 2018-028, Human Capital and Economic Opportunity Working Group. Chodorow-Reich, G. (2020). Regional data in macroeconomics: Some advice for practitioners. Journal of Economic Dynamics and Control, 115(C). 36

Cicero, T. J. and Ellis, M. S. (2015). Abuse-Deterrent Formulations and the Prescription Opioid Abuse Epidemic in the United States: Lessons Learned From OxyContin. JAMA Psychiatry, 72(5):424–430. Cicero, T. J., Ellis, M. S., and Surratt, H. L. (2012). Effect of abuse-deterrent formulation of oxycontin. New England Journal of Medicine, 367(2):187–189. Cicero, T. J., Inciardi, J. A., and Mun˜oz, A. (2005). Trends in abuse of oxycontin® and other opioid analgesics in the united states: 2002-2004. The Journal of Pain, 6(10):662–672. Col´on,H.M.,Robles,R.R.,andSahai,H.(2001). Thevalidityofdruguseresponsesinahousehold survey in Puerto Rico: comparison of survey responses of cocaine and heroin use with hair tests. International Journal of Epidemiology, 30(5):1042–1049. Cone, E. J., Fant, R. V., Rohay, J. M., Caplan, Y. H., Ballina, M., Reder, R. F., Spyker, D., and Haddox, J. D. (2003). Oxycodone involvement in drug abuse deaths: a dawn-based classification scheme applied to an oxycodone postmortem database containing over 1000 cases. Journal of Analytical Toxicology, 27(2):57–67. Council of Economic Advisers (2016). The Long-term Decline in Prime-age Male Labor Force Participation. Technical report, Executive Office of the President. Currie, J., Jin, J., and Schnell, M. (2019). US Employment and Opioids: Is There a Connection? InPolachek, S.W.andTatsiramos, K., editors, Health and Labor Markets, volume47ofResearch in Labor Economics, pages 253–280. Emerald Publishing Ltd. Currie, J. and Schwandt, H. (2020). The Opioid Epidemic Was Not Caused by Economic Distress But by Factors that Could be More Rapidly Addressed. NBER Working Papers 27544, National Bureau of Economic Research, Inc. Davis, C. S. and Carr, D. H. (2017). Self-regulating profession? administrative discipline of “pill mill” physicians in florida. Substance Abuse, 38(3):265–268. PMID: 28394740. D’Lima, W. and Thibodeau, M. (2019). Health Crisis and Housing Market Effects - Evidence from the U.S. Opioid Epidemic. Technical report, Working Paper. Doleac, J. and Mukherjee, A. (2018). The Moral Hazard of Lifesaving Innovations: Naloxone Access, Opioid Abuse, and Crime. IZA Discussion Papers 11489, Institute of Labor Economics (IZA). Drake, J., Charles, C., Bourgeois, J. W., Daniel, E. S., and Kwende, M. (2020). Exploring the impact of the opioid epidemic in black and hispanic communities in the united states. Drug Science, Policy and Law, 6:2050324520940428. 37

Evans, W.N., Lieber, E.M.J., andPower, P.(2019). HowtheReformulationofOxyContinIgnited the Heroin Epidemic. The Review of Economics and Statistics, 101(1):1–15. Fullerton, C. A., Kim, M., Thomas, C. P., Lyman, D. R., Montejano, L. B., Dougherty, R. H., Daniels, A. S., Ghose, S. S., and Delphin-Rittmon, M. E. (2014). Medication-assisted treatment with methadone: assessing the evidence. Psychiatric services (Washington, D.C., 65(2):146–157. Grecu, A. M., Dave, D. M., and Saffer, H. (2019). Mandatory Access Prescription Drug Monitoring ProgramsandPrescriptionDrugAbuse. Journal of Policy Analysis and Management,38(1):181– 209. Harrell, A. (1997). The validity of self-reported drug use data: the accuracy of responses on confidential self-administered answered sheets. NIDA research monograph, 167:37–58. Harris, M. C., Kessler, L. M., Murray, M. N., and Glenn, B. (2020). Prescription Opioids and Labor Market Pains: The Effect of Schedule II Opioids on Labor Force Participation and Unemployment. Journal of Human Resources, 55(4):1319–1364. Harrison, L. (1997). The validity of self-reported drug use in survey research: an overview and critique of research methods. NIDA research monograph, 167:17–36. Hays, L. R. (2004). A profile of oxycontin addiction. Journal of Addictive Diseases, 23(4):1–9. PMID: 15339710. Hollingsworth, A., Ruhm, C. J., and Simon, K. (2017). Macroeconomic conditions and opioid abuse. Journal of Health Economics, 56(C):222–233. Horwitz, J., Davis, C. S., McClelland, L. S., Fordon, R. S., and Meara, E. (2018). The Problem of Data Quality in Analyses of Opioid Regulation: The Case of Prescription Drug Monitoring Programs. NBER Working Papers 24947, National Bureau of Economic Research, Inc. Jansen, M. (2019). Spillover Effects of the Opioid Epidemic on Consumer Finance. Technical report, Western Finance Association Annual Meeting Paper. Jones, C. M., Campopiano, M., Baldwin, G., and McCance-Katz, E. (2015). National and state treatmentneedandcapacityforopioidagonistmedication-assistedtreatment. American Journal of Public Health, 105(8):e55–e63. PMID: 26066931. Katz, D. and Hayes, L. (2004). Adolescent OxyContin Abuse. Journal of the American Academy of Child and Adolescent Psychiatry, 43(2):231–234. Kolodny, A., Courtwright, D. T., Hwang, C. S., Kreiner, P., Eadie, J. L., Clark, T. W., and Alexander, G. C. (2015). The prescription opioid and heroin crisis: a public health approach to an epidemic of addiction. Annual review of public health, 36:559–574. 38

Krueger, A. B. (2017). Where Have All the Workers Gone? An Inquiry into the Decline of the U.S. Labor Force Participation Rate. Brookings Papers on Economic Activity, 48(2 (Fall)):1–87. Meara, E., Horwitz, J. R., Powell, W., McClelland, L., Zhou, W., O’malley, A. J., and Morden, N. E. (2016). State legal restrictions and prescription-opioid use among disabled adults. New England Journal of Medicine, 375(1):44–53. Montes, J. (2018). Cbo’s projection of labor force participation rates. CBO WP 2018-04, www.cbo.gov/publication/53616. Morral, A. R., McCaffrey, D., and Iguchi, M. Y. (2000). Hardcore drug users claim to be occasional users: drug use frequency underreporting. Drug and Alcohol Dependence, 57(3):193–202. Moyo, P., Simoni-Wastila, L., Griffin, B. A., Onukwugha, E., Harrington, D., Alexander, G. C., and Palumbo, F. (2017). Impact of prescription drug monitoring programs (pdmps) on opioid utilization among medicare beneficiaries in 10 us states. Addiction, 112(10):1784–1796. NIDA (2020). Medications to Treat Opioid Use Disorder Research Report. Ouimet, P., Simintzi, E., and Ye, K. (2019). The Impact of the Opioid Crisis on Firm Value and Investment. CEPR Discussion Papers 14236, C.E.P.R. Discussion Papers. Pardo, B., Caulkins, J. P., Kilmer, B., Pacula, R. L., Reuter, P., and Stein, B. D. (2019). The Synthetic Opioid Surge in the United States: Insights from Mortality and Seizure Data. RAND Corporation, Santa Monica, CA. Park, S. and Powell, D. (2021). Is the Rise in Illicit Opioids Affecting Labor Supply and Disability Claiming Rates? Journal of Health Economics, 76. Pierce,J.R.andSchott,P.K.(2020). Tradeliberalizationandmortality: evidencefromuscounties. American Economic Review: Insights, 2(1):47–64. Powell, D., Alpert, A., and Pacula, R. L. (2019). A transitioning epidemic: how the opioid crisis is driving the rise in hepatitis c. Health Affairs, 38(2):287–294. Powell, D. and Pacula, R. L. (2021). The Evolving Consequences of OxyContin Reformulation on Drug Overdoses. American Journal of Health Economics, 7(1):41–67. Quinones, S. (2015). Dreamland: The true tale of America’s opiate epidemic. Bloomsbury Publishing USA. Rees, D. I., Sabia, J. J., Argys, L. M., Dave, D., and Latshaw, J. (2019). With a Little Help from My Friends: The Effects of Good Samaritan and Naloxone Access Laws on Opioid-Related Deaths. Journal of Law and Economics, 62(1):1–27. 39

Ruhm, C. J. (2018). Corrected us opioid-involved drug poisoning deaths and mortality rates, 1999–2015. Addiction, 113(7):1339–1344. Sacks, D. W., Hollingsworth, A., Nguyen, T. D., and Simon, K. I. (2019). Can Policy Affect Initiation of Addictive Substance Use? Evidence from Opioid Prescribing. NBER Working Papers 25974, National Bureau of Economic Research, Inc. Schnell, M. and Currie, J. (2018). Addressing the Opioid Epidemic: Is There a Role for Physician Education? American Journal of Health Economics, 4(3):383–410. Severtson, S. G., Bartelson, B. B., Davis, J. M., Mun˜oz, A., Schneider, M. F., Chilcoat, H., Coplan, P. M., Surratt, H., and Dart, R. C. (2013a). Reduced abuse, therapeutic errors, and diversion following reformulation of extended-release oxycodone in 2010. The Journal of Pain, 14(10):1122–1130. Severtson, S. G., Bartelson, B. B., Davis, J. M., Mun˜oz, A., Schneider, M. F., Chilcoat, H., Coplan, P. M., Surratt, H., and Dart, R. C. (2013b). Reduced abuse, therapeutic errors, and diversion following reformulation of extended-release oxycodone in 2010. The journal of pain, 14(10):1122–1130. Surrat, H. L., O’Grady, C., Kurtz, S. P., Stivers, Y., Cicero, T. J., Dart, R. C., and Chen, M. (2013). Reductions in prescription opioid diversion following recent legislative interventions in Florida. Pharmacoepidemeology and Drug Safety, 23(3):314–320. Temple, J. (2015). American pain: How a young felon and his ring of doctors unleashed America’s deadliest drug epidemic. Rowman & Littlefield. VanZee,A.(2009). Thepromotionandmarketingofoxycontin: commercialtriumph,publichealth tragedy. American journal of public health, 99(2):221–227. 40