The OxyContin Reformulation Revisited: New Evidence From Improved Definitions of Markets and Substitutes
TThe OxyContin Reformulation Revisited: New EvidenceFrom Improved Definitions of Markets and Substitutes
Shiyu Zhang ∗ Daniel GuthJanuary 28, 2021
Abstract
The opioid epidemic began with prescription pain relievers. In 2010 Purdue Pharmareformulated OxyContin to make it more difficult to abuse. OxyContin misuse felldramatically, and concurrently heroin deaths began to rise. Previous research over-looked generic oxycodone and argued that the reformulation induced OxyContin usersto switch directly to heroin. Using a novel and fine-grained source of all oxycodonesales from 2006-2014, we show that the reformulation led users to substitute from Oxy-Contin to generic oxycodone, and the reformulation had no overall impact on opioidor heroin mortality. In fact, generic oxycodone, instead of OxyContin, was the drivingfactor in the transition to heroin. Finally, we show that by omitting generic oxycodonewe recover the results of the literature. These findings highlight the important rolegeneric oxycodone played in the opioid epidemic and the limited effectiveness of apartial supply-side intervention.
Keywords:
Opioids, Drug Overdoses, Heroin, Drug Distributors
JEL Codes:
I11, I12, I18 ∗ Zhang: California Institute of Technology. Corresponding author, email [email protected]. Guth:California Institute of Technology a r X i v : . [ ec on . GN ] J a n Introduction
Since 1999, the opioid epidemic has claimed more than 415,000 American lives (CDC Won-der). What started with fewer than 6,000 opioid-related deaths in 1999 grew steadily everyyear until fatalities reached 47,573 deaths in 2017. Following a small decline in fatal drugoverdoses in 2018, deaths continue to rise. Over the past two decades, millions of Americanshave misused prescription opioids or progressed to more potent opioids, first heroin and laterfentanyl. Many social scientists have tried to understand how this crisis has grown over twodecades despite significant public health efforts to the contrary.Doctors and health economists have long argued that the drug most responsible for prescrip-tion opioid overdose deaths, and the key to understanding the transition from prescriptionopioids to heroin starting in 2010, was OxyContin. Previous research (Van Zee, 2009), courtproceedings (Meier, 2007), and books (Meier, 2003, Macy, 2018) has documented how Pur-due Pharma’s marketing campaign for OxyContin downplayed the risk of addiction startingin 1996. Since then, according to the National Survey on Drug Use and Health (NSDUH),millions of Americans have misused it previously. A key question in this area is whether ornot making prescription opioids, especially OxyContin, more difficult to abuse will reduceoverdose deaths.In this paper, we show that restricting access to OxyContin led many users to switch togeneric oxycodone but had no impact on opioid or heroin mortality. Earlier analyses at-tributing opioid overdose deaths in the late 2000s and the subsequent rise in heroin deathsto OxyContin are incomplete because they omit generic oxycodone. Our analysis shows thatthe misuse of generic oxycodone was prevalent before the reformulation that restricted Oxy-Contin access, and was even more so afterward. We also show that heroin overdose deathsincreased in areas with high generic oxycodone exposure, not high OxyContin exposure,two years after the OxyContin reformulation. In addition, we show that omitting genericoxycodone in our regressions recovers the results of the literature.This analysis was not possible until one year ago when The Washington Post won a courtorder and published the complete Automation of Reports and Consolidated Orders System(ARCOS). The ARCOS tracks the manufacturer, the distributor and the pharmacy of everypain pill sold in the United States. The newly released data allow us to analyze whathappened to sales of generic oxycodone and OxyContin when OxyContin suddenly becamemore difficult to abuse. The previous literature focused on analyzing OxyContin becauseof Purdue’s notorious role in the opioid crisis. However, the new data shows that the sales1f OxyContin was only a small part of the sales of all prescription opioids: in terms of thenumber of pills, OxyContin was 3% of all oxycodone pills sold from 2006 to 2012; in termsof morphine milligram equivalents (MME), OxyContin has closer to 20% market share overthis period. The new transaction-level ARCOS data allows us to track the sales of genericoxycodone and fill in the narrative gaps of how the opioid crisis progressed in the UnitedStates.Following Alpert, Powell, and Pacula (2018), W. Evans, Lieber, and Power (2019) and T.Cicero and Ellis (2015), we treat the introduction of an abuse-deterrent formulation (ADF)of OxyContin as an exogenous shock that should only affect people who seek to bypass theextended-release mechanism for a more immediate high. We construct measures of exposureby combining ARCOS sales and the NSDUH data on drug misuse. The NSDUH is the bestsurvey of people who use drugs at the state level, and by combining it with local sales we cancapture variation in drug use within the state. We leverage this variation in OxyContin andgeneric oxycodone exposure to examine how the reformulation affected OxyContin sales,generic oxycodone sales, opioid mortality, and heroin mortality. Our first contribution isthat we fix the omitted-variable problem by differentiating between OxyContin and genericoxycodone, and we show that this leads to different conclusions than what previous literaturesuggests. Our second contribution is disaggregating the data to metropolitan statistical area(MSA), which allows us to address endogeneity at the state level.To preview our results, we find strong evidence of substitution from OxyContin to genericoxycodone immediately after the reformulation. This substitution was larger in places thathad more OxyContin misuse pre-reform, which is consistent with our hypothesis that userswould switch between oxycodones rather than move on to heroin. Because this substitutionshould be concentrated among people misusing OxyContin, the results imply large changesin consumption at the individual level. Back-of-the-envelope calculation suggests 68% ofthe decline in OxyContin sales was substituted to oxycodone in MSAs with high OxyContinmisuse. The findings are consonant with surveys like Havens et al. (2014), Coplan et al.(2013), and Cassidy et al. (2014) who all document substitution to generic oxycodone afterthe reformulation by people seeking to bypass the ADF. We also find suggestive evidence ofsubstitution from generic oxycodone to OxyContin after the reformulation in places wheregeneric oxycodone misuse was high, a channel that has been unexplored in previous research.Our event study approach also shows that generic oxycodone exposure is predictive of futureheroin overdose deaths whereas OxyContin exposure is not. The results are not contingenton methodology or our construction of exposure measures. Crucially, if we run the same2xact regressions at the state or MSA level and omit generic oxycodone, we recover theresults of the literature where OxyContin misuse appears to be significantly predictive offuture heroin overdose deaths. We find that every standard deviation increase in genericoxycodone exposure pre-reformulation is associated with a 40.8% increase in heroin mortalityin 2012 from the 2009 baseline level. As further evidence against the argument that therewas immediate substitution from OxyContin to heroin after the reformulation, we note thatin all of our regressions the increase in heroin deaths wasn’t statistically significant until2012. As suggested in O’Donnell, Gladden, and Seth (2017), the rise in heroin deaths canbe attributed in part to an increase in the supply of heroin as well as the introduction offentanyl into heroin doses.Our findings highlight the pitfalls of omitting important substitutes to OxyContin in analyz-ing the prescription opioid crisis. Purdue Pharma has received well-deserved attention overthe years for its role in igniting the crisis. The company has been involved in many lawsuitsover the years, but perhaps the most damaging were lesser-known cases that involved losingits patent in 2004 which cleared the way for a rapid increase in generic oxycodone sales in theearly 2000s. While Purdue Pharma was being sued and scrutinized, several manufacturerstook the opportunity to fill in the gaps of OxyContin. By 2006, generic oxycodone outsoldOxyContin by more than 3-to-1 after accounting for pill dosage differences. This paper shedslights on the role generic oxycodone played and continues to play in the opioid crisis andhelps policy makers update their picture of the opioid use disorder (OUD) landscape.The paper also calls attention to the limited effectiveness of a partial supply-side interventionto curb OUD. Purdue Pharma was once a dominant player in the opioid market, but bythe time of the reformulation, that dominance had vanished and it was only one of themany manufacturers whose drugs were actively misused by Americans. Purdue was the firstcompany to include abuse-deterrent formulation (ADF) in their opioids, but it is not untilrecent years that other brands started adding anti-deterrent compounds to their products(Pergolizzi et al., 2018). When substitutions to other abusable opioids are easy, cuttingsupplies of one kind is less effective.The rest of the paper runs as follows. Section 2 gives more background on the opioidcrisis and explains how previous research has characterized the OxyContin reformulation. InSection 3 we describe the new ARCOS sales database, the NSDUH misuse data, the NVSSmortality data, as well as our constructed misuse measure and descriptive statistics. Section4 describes our empirical strategy for testing our hypotheses. Section 5 discusses our results Federal ruling, Risk management plan proposals for generic oxycodone
This section proceeds in chronological order. First, we provide a history of oxycodone and itsmost important formulation, OxyContin. We then describe the OxyContin reformulation in2010 and what it meant for prescription opioid misuse, as well as how the previous literatureanalyzed the reformulation. Next, we present the nascent research on substitution betweendifferent opioids and how our contribution fits in this strain of work. We conclude witha summary of the literature on heroin mortality in the early 2010s and its link with theprescription opioid crisis.Oxycodone was first marketed in the United States as Percodan by DuPont Pharmaceuticalsin 1950. It quickly found to be as addictive as morphine (Bloomquist, 1963), and in 1965California placed it on the triplicate prescription form (Quinn, 1965). Before the 1990s,doctors were hesitant to prescribe oxycodone to non-terminally ill patients due to its highabuse potential (DeWeerdt, 2019). The sales of oxycodone-based pain relievers did not takeoff until the mass marketing of OxyContin, Purdue’s patented oxycodone-based painkiller.OxyContin was first approved by the FDA in 1995. The drug’s innovation was an ‘extended-release’ formula, which allowed the company to pack a higher concentration of oxycodoneinto each OxyContin pill and the patients to take the pills every 12 hours instead of 8 hours.OxyContin’s original label, approved by the FDA, stated that the “delayed absorption, asprovided by OxyContin tablets, is believed to reduce the abuse liability of a drug.” In 2001,the FDA changed OxyContin’s label to include stronger warnings about the potential forabuse and Purdue agreed to implement a Risk Management Program to try and reduceOxyContin misuse. OxyContin was one of the first opioids marketed specifically for non-cancer pain. In the early1990s, pain started to enter the medical discussion as the ‘fifth vital sign’ and somethingto be managed. As described in Meier (2003), Van Zee (2009), and elsewhere, Purdue’ssales representatives pushed OxyContin and were told to downplay the risk of addiction.Quinones (2015) describes how Purdue cited a 1980 short letter published in the New Eng- Triplicate programs required pharmacists to send a copy to the government, and Alpert, Evans, et al.,2019 show that these had a persisting effect on reducing the number of opioid prescriptions. From the FDA Opioid Timeline. $
600 million in fines. Less than six months later, the companyapplied to the FDA for approval of a new reformulated version of OxyContin that includeda chemical to make it more difficult to crush and misuse (Rappaport, 2009). Althoughnot completely effective in reducing misuse, it was approved by the FDA and after August2010 accounted for all OxyContin sales in the United States. Until 2016, with Mallinrockdt’sXtampza ER, Purdue was the only prescription opioid manufacturer to make abuse-deterrentoxycodone pills. The majority of all oxycodone sold over this time was generic oxycodonethat remained abusable. Most research shows that OxyContin misuse fell following the reformulation. As describedin T. Cicero and Ellis (2015), although some users were able to circumvent the abuse-deterrent formulation (ADF) to inject or ingest, the reformulation did reduce misuse. W.Evans, Lieber, and Power (2019) finds that the reformulation coincided almost exactly with astructural break in aggregate oxycodone sales, which had previously been increasing. Shortlyafter the OxyContin reformulation was implemented, researchers began to notice illicit druguse moving towards other drugs such as heroin or generic oxycodone (T. Cicero, Ellis, andSurratt, 2012, Coplan et al., 2013, Alpert, Powell, and Pacula, 2017, W. Evans, Lieber, andPower, 2019, Havens et al., 2014, Cassidy et al., 2014). Our paper extends the analysis ofthe impact of reformulation on opioid use by separately identifying the shifts in OxyContinand generic oxycodone misuse.We build upon a rich literature that studies opioid misuse through surveys or analysis of theaggregated ARCOS reports. Surveys mostly polled either informants or users themselves(for details see Inciardi et al., 2009). The best surveys have been of users in smaller samplesat individual treatment facilities, like in Hays (2004) and Sproule et al. (2009). However,selection bias is a problem for surveying treatment facilities, as that is a specific subset of Many other companies attempted to make abuse-deterrent opioid pills at the same time, as shown inWebster (2009), but Purdue was the first to market. Adler and Mallick-Searle (2018) and Pergolizzi et al.(2018) list other opioids with an ADF.
To estimate the impact of the OxyContin reformulation on opioid use and mortality, wecombine several data sources including sales of OxyContin and non-OxyContin alternativesfrom ARCOS, opioid and heroin mortality from the NVSS, and self-reported OxyContin andPercocet misuse from the NSDUH. Our main regression leverages variations in pre-reformexposure to OxyContin and generic oxycodone to identify the impact of the reformulationon opioid sales and mortality. We define a new measure of exposure by interacting the state-level self-reported opioid misuse and MSA-level opioid sales. In this section, we describe thethree sources of data, the market definition, the construction of the OxyContin and genericoxycodone exposure measure, and present summary statistics of our data.7 .1 Data
As part of the Controlled Substances act, distributors and manufacturers of controlled sub-stances are required to report all transactions to the DEA. This Automation of Reports andConsolidated Orders System (ARCOS) database contains the record of every pain pill soldin the United States. The complete database from 2006 to 2014 was recently released by afederal judge as a result of an ongoing trial in Ohio against opioid manufacturers. The ARCOS database has been used previously to study opioids, but only using the publiclyavailable quarterly aggregate weight of drugs sold (Atluri, Sundarshan, and Manchikanti,2014) or via special request to the DEA (Modarai et al., 2013). The newly released fulldatabase reports the manufacturer and the distributor for every pharmacy order. Thesedata allow us to track different brands of prescription opioids separately, and calculate whatfraction of oxycodone sold is OxyContin at any level of geographic aggregation. We can thusconstruct what we believe is the first public time-series of OxyContin and generic oxycodonesales from 2006-2014.
Note:
We supplemented the 2006 to 2014 data with publicly available aggregate data from 2000to 2005. The publicly available aggregate data does not break down the oxycodone sales bymanufacturer.
Figure 1: Growth of oxycodone and OxyContin sales Link to the ARCOS Data published by the Washington Post.
8s we can see from Figure 1, total oxycodone sales increased substantially from 2000 to2010, with per-person sales nearly quadrupling in the ten years period. From 2010 to 2015,sales of oxycodone declined as a result of aggressive measures taken by the states and thefederal government to counter opioid addiction (Kennedy-Hendricks et al., 2016).The newly available ARCOS data suggests that the commonly held belief about OxyContin’sdominance in the prescription opioid market at the time of reformulation is incorrect. Thelast time OxyContin’s market was estimated was in 2002 by Paulozzi and Ryan (2006), whoacquired from the DEA a year’s worth of ARCOS data aggregated at the state level. In thatyear, OxyContin was 68% of all oxycodone sales by active ingredient weight and scholarshave assumed that Purdue’s market share stayed high until the OxyContin reformulation.However, as Figure 1 shows, by 2006 when our data starts, OxyContin sales only accountedfor 18% of all oxycodone sold by weight and never got above 35% during this period. Theshare is even smaller if we count the number of pills sold, since the average OxyContinactive ingredient weight is 5 to 10 times higher than that of oxycodone from other brands.The share of OxyContin decreased dramatically from 2002 to 2006 because Purdue lost thepatent rights in 2004. As a result, non-OxyContin oxycodone sales grew much faster inthe early 2000s than OxyContin sales. Figure 7 in Appendix presents the market share forall oxycodone manufacturers by dosage strength, and Purdue Pharma is only dominant athigher dosages ( ≥ The second outcome of interest in our main regression is opioid mortality. We use therestricted-use multiple-cause mortality data from the National Vital Statistics System (NVSS)to track opioid and heroin overdose. The dataset covers all deaths in the United States from2006-2014. We follow the literature’s two step procedure to identify opioid-related deaths. We acknowledge some non-OxyContin alternatives are branded and non-generic (i.e. Percocet andPercodan or later Roxicodone), but the majority of them are generic products. Generic oxycodone in thispaper should be interpreted as all non-OxyContin oxycodone products. However, the underestimation wouldnot pose a problem for our regressions. There are variations in how coroners attribute thecause of death across states, but such variation would be captured by the state fixed effects. Specifically, we omit ICD-10 code T50.9 (unspecified poisioning) from our analysis, and some fractionof these deaths are due to opioids or heroin but were not diagnosed or recorded as such.
10n addition, we do not anticipate systematic changes to each state’s practices due to thereformulation.
We use state-level data from the National Survey on Drug Use and Health (NSDUH) tomeasure nonmedical use of opioids. The NSDUH publishes an annual measure of OxyContinmisuse, asking the respondents whether they have ever used OxyContin “only for the ex-perience or feeling they caused” (NSDUH Codebook). As first described in Alpert, Powell,and Pacula (2018), the advantage of the NSDUH misuse measure is that it seperates outmisuse from medical use. However, only OxyContin is reported in the NSDUH and there isno equivalent measure for generic oxycodone.Fortunately, the NSDUH reports PERCTYL2, which asks whether individuals ever mis-used Percocet, Percodan, or Tylox. These drugs are oxycodone hydrochloride with ac-etaminophen and have a maximum dosage of 10mg of oxycodone per pill. The three drugswere popular among users in the pre-OxyContin era (Meier, 2003). In the present day, thePERCTYL2 variable captures misuse of not only the three branded drugs but also othergeneric oxycodone products that are popular on the street.The most direct evidence supporting this claim is the fact that generic oxycodone pills haveoften been referred to as ‘Percs’ colloquially in the last decade. Many news report indicatedthat generic oxycodone has the street name ‘Perc 30’ but is in fact not Percocet. The PatriotLedger reported in a 2011 article that ‘Perc 30s’ were the newest drug of choice in SouthShore of Massachusetts, saying: ‘Perc 30s are not Percocet — the brand name for oxycodone mixed with ac-etaminophen, the main ingredient in Tylenol — but a generic variety of quick-release oxycodone made by a variety of manufacturers. They are sometimes re-ferred to as “roxys” after Roxane Laboratories, the first company to make thedrug, or “blueberries,” because of their color.’ Percocet Drug Information. Tylox was discontinued in 2012 following the FDA regulations limitingacetaminophen. Patriot Ledger Link. Other references to generic non-OxyContin oxycodone as Perc 30s: Pheonix House,Washington State Patrol, The Boston Globe, The Salem News, Massachusetts Court Filing, Cape Cod Times,Pocono Record, Bangor Daily News, Patch, CNN Op-Ed There are also several empirical observations that support this claim. The first is that wecontinue to see increases in the lifetime misuse of Percocet, Percodan, and Tylox even afterthey were replaced by OxyContin as the preferred prescription opioid to misuse. The misuserate of Percocet, Percodan, and Tylox increased 30% from 4.1% to 5.6% from 2002 to 2009(see Figure 8 in Appendix), which would not have been possible if these drugs, or whatpeople believed were ‘Percs’, were not actively misused by new users post-introduction ofOxyContin.The second observation is that, based on the average sales data from 2006 to 2014, a dispro-portionate number of people has reported misusing Percocet, Percodan, or Tylox as com-pared to the actual sales of the three drugs. The sales of Endo Pharma, the manufacturer ofPercocet and Percodan , are orders of magnitude less than the sales of Purdue while morethan twice as many people reported misusing the three drugs as compared to OxyContin (seeFigure 9 in Appendix). A back-of-the-envelope calculation shows that if PERCTYL2 misusecaptures only the misuse of Percocet and Percodan, then the proportion of pills misused outof all pills sold is 29 times higher for Percocet and Percodan than than the same proportionfor OxyContin , an very unlikely situation given the popularity of OxyContin on the street.This deduction is further supported by misuse data reported in the NSDUH. We know thatgeneric oxycodone is commonly misused. If oxycodone has any other drug names, thepopularity of that drug name in the NSDUH surveys should increase to reflect the increasein misuse in recent years. In addition to inquiring about popular brands, the NSDUH surveyasks respondents to list any other prescription oxycodone that they have misused before.Dozens of pain relievers are reported, but in 2010 “oxycodone or unspecified oxycodoneproducts” was only named by 0.10% of the respondents. No other brand of oxycodonepills are reported as commonly misused. We know from the reports in press and documents In the ARCOS dataset these pills are simply listed as ‘Oxycodone Hydrochloride 30mg’ Tylox not included since it was discontinued. In terms of number of pills circulated, OxyContin is 12.1 times Percocet and Percodan from 2006 to2014. In terms of misuse, OxyContin is 41% of Percocet and Percodan in the same period. Law enforcement and journalists have previously identified the 30mg oxycodone pill as the most com-monly trafficked opioid, see DEA Link, ICE Link, and Palm Beach Post Link. NSDUH Codebook variables ANALEWA through ANALEWE list the other pain relievers reported.Even if we assumed all 2.49% of respondents saying they took a prescription pain reliever not listed hadtaken generic oxycodone, it is still less than half of the reported Percocet misuse.
12n court that generic oxycodone is a popular opioid on the street, and we know that Percocetis the only other commonly misused opioid documented in the NSDUH survey. Thus, theonly way to reconcile the discrepancy between these two sources is that people mistakenlyperceive generic oxycodone as Percocet or respond to the NSDUH as if they do. Thus, weuse lifetime OxyContin and lifetime Percocet misuse for the construction of OxyContin andgeneric oxycodone exposure measures in Section 3.3.
Previous studies of the OxyContin reformulation depend on state-level variation to causallyidentify the impact of the reformulation. Treating OxyContin reformulation as an exogenousshock at the state level is potentially problematic. Although the timing of the reformulationis exogenous, each state’s exposure to it is a result of a combination of the state’s regulatoryenvironment and Purdue’s initial marketing strategy (Alpert, Evans, et al., 2019). Thesefactors have substantial impact on how people in a state respond to the reformulation,creating a hidden link between exposure to the reformulation, the identifying variation, andsubsequent drug use, the outcome variable.One can limit the impact of endogenous regulation by disaggregation, but only if there issubstantial intra-state variation in exposure to the reformulation. Both the ARCOS databaseand the NVSS mortality data have great geographic detail. Conducting our analysis onmetropolitan statistical areas (MSAs), we find large variation in both OxyContin use andopioid mortality across MSAs in the same state. At the aggregate level in 2009, the averageOxyContin market share in a state is 35.6%. 65 of the 379 MSAs (17.1%) in our samplehave an OxyContin market share that is 10% greater or smaller than their state average.The average opioid mortality is 0.343 deaths per 100,000 population in 2008. The variationin death is even more significant. More than 310 (83%) MSAs have a mortality rate 20%higher or lower than their state average, and more than 192 (51%) have a mortality rate50% higher or lower than their state average . We present the full distribution of deviationsof OxyContin market share and opioid mortality from state average in Figure 10 and Figure11 in the Appendix.Disaggregating to the MSA-level allows us to control for the state’s regulatory environmentand hence eliminate the most problematic source of endogeneity. We use intra-state variationin exposure to the reformulation for identification. Intra-state heterogeneity in opioid use isassociated with past economic conditions (Carpenter, McClellan, and Rees, 2017), location13f hospitals and treatment centers (Swensen, 2015), preferences of local physicians (Schnell,2017), and local policy, some of which could still be correlated with the locality’s responseto the reformulation. Analysis at the MSA level clearly allows us to make a much strongerclaim than analysis at the state level.In addition, as we will show in the next sections, the disaggregation increases the statisticalpower of our regressions beyond the impact of the tripled sample size. Our results indicatethat defining the market at the MSA level better captures the interaction between druguse and mortality than the state level. The important variations in drug use, for examplebetween Los Angeles-Long Beach-Santa Ana at 4.4% of nonmedical use of pain relievers andSan Francisco-Oakland-Fremont at 5.6%, disappears when they’re aggregated to the statelevel (
Since the OxyContin reformulation was a national event independent of local conditions, wecan estimate its impact by comparing the outcomes in areas of high prior exposure to opioidswith outcomes in areas of low exposure. Ideally, we want to quantify exposure using thevolume of OxyContin misused in each region pre-reform while controlling for the volume ofgeneric oxycodone misused. In practice, we do not observe these quantities. The best proxyin the literature is the self-reported misuse rate from the NSDUH.Based on the NSDUH misuse, we create a new measure of OxyContin and non-OxyContinoxycodone exposure by combining the NSDUH state-level misuse rate with ARCOS MSA-level sales. Specifically, for each drug, we calculate:Exposure pre-reform m = Lifetime Misuse − s × Sales m (1)Our measure is the interaction term of sales of OxyContin/generic oxycodone in an MSA andthe lifetime misuse rate of that drug in the corresponding state. This new measure has twoadvantages over the conventional misuse rate from NSDUH: it captures intra-state variationin misuse and it more accurately reflects the current misuse of both OxyContin and genericoxycodone.The NSDUH surveys approximately 70,000 respondents every year and uses sophisticatedreweighting techniques to get accurate state level estimates. Once we get to the MSA level,the number of people surveyed as well as the number of positive responses to questions on14pioid misuse are extremely small. As a result, most of the outcomes at the MSA level arecensored by the NSDUH to protect individual privacy. Using only the survey data meansthat we would use same state misuse value for all MSAs and therefore forgo any intra-statevariation in drug use. In comparison, our proposed measure relies on deviations from normalsales patterns to generate variations in exposure rates for the MSAs. Our definition assumesthat the percentage of people reported misusing a particular drug in a state is equivalent tothe proportion of sales that are being misused. In a state where all the MSAs have identicalsales, all the MSAs will have identical exposure rates by definition. However, if one MSAhas higher sales of OxyContin compared with the rest of the state, our OxyContin exposuremeasure in that MSA will be higher than the rest of the state. This construction of exposuremirrors our intuitive understanding that the misuse of a drug in a locality is a function ofthe overall misuse and the availability of that particular drug in the area.The NSDUH survey reports past-year misuse of OxyContin but only lifetime misuse ofgeneric oxycodone. Previous studies did not focus on generic oxycodone misuse, so thesestudies rely on past-year OxyContin misuse rate. In our case, to disentangle substitutionamong prescription opioids, we have to make the comparison between OxyContin and genericoxycodone equal. Resorting to lifetime misuse rates for both series sacrifices the timelynature of the NSDUH misuse rates. By combining the lifetime misuse rates with sales inthe year before reformulation, we capture recent changes in use of both drugs. To makeour results comparable with previous studies, in the Appendix section, we repeat our entireanalysis with OxyContin last-year misuse and generic oxycodone lifetime misuse. Most ofour conclusions stand despite giving OxyContin a more favorable treatment.To construct our measure, we follow the precedent set in the literature by using a six-yearsaverage state level lifetime misuse rate pre-reform (2004 - 2009) and sales in 2009. The goalof the time average is to reduce the variance of the state-level misuse rates. We check thevalidity of our measure by regressing opioid death on it and compare the results with the sameregressions on either only ARCOS sales or only NSDUH misuse. Results are summarizedin Table 2 in Appendix. The fit of the generic oxycodone regression is much improvedwith the interacted variable ( R = 0 . R = 0 . R = 0 . R = 0 . R = 0 . R = 0 . In all surveys prior to 2014.
All MSAs MSAs withlowOxyContinexposure MSAs withhighOxyContinexposure MSAs withlowoxycodoneexposure MSAs withhighoxycodoneexposure
NSDUH lifetime misuse rates (2004-2009)
OxyContin misuse rate (%) 2.22 1.88 2.56 1.87 2.56Oxycodone misuse rate (%) 5.19 4.22 6.17 3.75 6.64
Annual ARCOS sales (all sample period)
Oxycontin sales per person 65.71 43.47 88.06 50.70 80.79Oxycodone sales per person 181.84 112.50 251.55 99.24 264.88
Annual death per 100,000 (all sample period)
Opioid 0.32 0.23 0.41 0.23 0.42Heroin 0.13 0.09 0.16 0.10 0.16
Census Demographics (2009)
Number of MSAs 379 190 189 190 189Population 679878 745327 614082 663740 696101Age 36.13 34.68 37.59 34.84 37.43Male (%) 49.24 49.35 49.13 49.40 49.08Separated (%) 18.83 18.24 19.42 18.32 19.34High school and above (%) 84.20 82.79 85.61 83.68 84.72Bachelor and above (%) 25.36 24.77 25.96 24.85 25.87Mean income 64213 63414 65016 63058 65374Low income (%) 35.38 35.79 34.98 35.90 34.86White (%) 82.17 79.99 84.36 81.22 83.12Black (%) 11.20 13.09 9.30 11.80 10.60Asian (%) 3.03 3.47 2.60 3.52 2.54Native American (%) 0.18 0.20 0.17 0.20 0.17
Note:
Simple average, not weighted by population.
Table 1 reports summary statistics for five groups of MSAs: All MSAs, MSAs with high Oxy-Contin exposure, MSAs with low OxyContin exposure, MSAs with high generic oxycodoneexposure, and MSAs with low generic oxycodone exposure. MSAs with high OxyContinexposure and MSAs with high generic oxycodone exposure have similar demographic sum-mary statistics. These two groups of MSAs also are not different statistically in their heroinmortality. Disentangling the impact of various opioids on the rise in heroin mortality isimpossible with nationally aggregated or state level data due to the high correlation in mis-use. The high correlation also implies that regressing heroin death on OxyContin withoutcontrolling for generic oxycodone use will likely lead to an overestimation of OxyContin’simpact. 16SAs with high misuse differ from MSAs with low misuse. High misuse states have highersales of both types of prescription opioids (twice as much for both types of opioids), highermortality rate (twice as much for both opioid and heroin overdose), smaller population,higher average age, higher median income, higher percentage of white population, and lowerpercentage of black population. The differences in racial composition repeat well establishedfindings in the literature: prescription opioid misuse was originally concentrated among whiteusers, and by 2010 new heroin users were almost entirely white (T. J. Cicero et al., 2014).These differences in demographic variables motivate the inclusion of control variables in ourmain regressions.
Our goal is to investigate two questions. First, what was the reformulation’s immediateimpact on OxyContin and generic oxycodone use? Second, what was the reformulation’slong-run effect on opioid mortality, heroin mortality, and on the progression of opioid addic-tion?We follow the event study framework from Alpert, Powell, and Pacula (2018) to estimate thecausal impact of the OxyContin reformulation on OxyContin and generic oxycodone sales andopioid and heroin mortality. We exploit variation in MSAs’ exposure to the reformulationdue to the differences in their pre-reform OxyContin use while controlling for pre-reformgeneric oxycodone use. Our approach is similar to Finkelstein (2007), where the OxyContinreformulation has more “bite”, or more of an effect, in areas where OxyContin misuse washigher than in places where generic oxycodone was the preferred drug. The approach allowsus to measure whether MSAs with higher exposure to OxyContin experienced larger declinesin OxyContin sales, larger increases in alternative oxycodone, or larger increases in opioidand heroin mortality. The empirical framework is: Y mt = α s + δ t + (cid:88) i =2006 { i = t } β i × OxyContin Exp
P rem + (cid:88) i =2006 { i = t } β i × Oxycodone Exp
P rem + X (cid:48) mt γ + (cid:15) mt (2)17here Y mt are the outcome variables of interest in MSA m at year t ; OxyContin Exp P rem and Oxycodone Exp
P rem are time-invariant measures of OxyContin and oxycodone exposurebefore the reformulation (see Section 3.5 for construction), and are interacted with a set of β t and β t for each year. We include state fixed effects to control for regulatory differences amongstates and year fixed effects to control for national changes in drug use. We also include afull set of MSA-level demographic variables. We weight the regression by population andexclude Florida. We show the full set of β t estimates graphically, normalizing by the 2009coefficient. The β t identifies the differences in sales and death across MSAs due to theirhigher or lower pre-reform OxyContin or oxycodone exposure. Standard errors are clusteredat the MSA level to account for serial correlation. In the Appendix section, we present betaestimations from variations of our base model, which include (1) using a MSA fixed effectinstead of state fixed effect, (2) replacing OxyContin lifetime misuse rate with OxyContinlast-year misuse rate, (3) regressing at the state level, and show that our conclusion areinsensitive to most of these variations.To complement our results, we also use a strict difference-in-difference framework to esti-mate effect of the reformulation conditioning on OxyContin and non-OxyContin oxycodoneexposure levels. Our specification is: Y mt = α s + γ t + δ { t > } + δ { m ∈ HighOxyContin } + δ { m ∈ HighOxycodone } + δ { t > } × { m ∈ HighOxyContin } + δ { t > } × { m ∈ HighOxycodone } + X (cid:48) mt β + (cid:15) mt (3)where HighOxyContin and
HighOxycodone are the set of MSAs with higher than medianpre-reform exposure to OxyContin and oxycodone respectively. We restrict the regressionto include only the three years prior (2008 to 2010) and the three years after (2011 to2013) the reformulation. The advantage of this specification is that it does not assume thatOxyContin or oxycodone exposure affects the outcome variable linearly. Instead of havinga flexible δ for each year, we have only one δ for each of the pre- or post-reform period. Inthis specification, we simply test whether higher exposure MSAs reacted differently to thereformulation as compared to lower exposure MSAs (if δ and δ are significant). We includestate fixed effects to control for state-level heterogeneity, year fixed effects for national trend,and a set of time-varying MSAs level covariates. Again, standard errors are clustered at theMSA level. The literature excludes Florida because it underwent massive increases in oxycodone sales over thisperiod, some of which was trafficked to other states. Results
We proceed in two steps. First, we provide direct evidence that the OxyContin reformulationcaused OxyContin sales to decrease and generic oxycodone sales to increase, and that thechanges in sales are proportional to the pre-reformulation level of OxyContin exposure.Second, we estimate the impact of the reformulation on opioid and heroin mortality. Wefind that high pre-reformulation levels of OxyContin exposure were not associated with highopioid deaths, but there was a strong positive effect from generic oxycodone exposure in boththe pre- and post-reform period. We find that higher pre-reform OxyContin and pre-reformoxycodone exposure were both positively but not significantly associated with later heroindeaths, but the oxycodone coefficient is larger. If we run the heroin regression separatelywith only OxyContin exposure we recover the results of the literature, but running the heroinregression with only oxycodone exposure better fits the data.
We begin by showing graphically that OxyContin sales decreased and generic oxycodonesales increased in high OxyContin misuse MSAs immediately after reformulation. Figure 3and Figure 4 present the full set of coefficients from estimating the event study frameworkon OxyContin and generic oxycodone sales. Each data point in the figure is the coefficientof the interactive term of misuse and sale, which we call exposure, for OxyContin or genericoxycodone in a specific year, and it captures any additional change in sales in that yeardriven by high OxyContin or oxycodone exposure. In Figure 3, we observe a larger decreasein OxyContin sales post-reform in MSAs with higher pre-reform OxyContin exposure. AsFigure 4 shows, higher OxyContin exposure MSAs saw greater increases in generic oxycodonesales post-reform. The effects are well identified at 95% confidence level. An one standarddeviation increase in OxyContin exposure translates into an additional 21.2 MME decreasein per person OxyContin sales and 11.8 MME increase in per person oxycodone sales in2011. These changes represents a 24% decrease in OxyContin sales and a 8.8% increase inoxycodone sales from the 2009 level. The effects are economically significant especially giventhat the reformulation should only affect the population abusing OxyContin, so this dropin sales is driven by a fraction of all users. The two observations combined support thehypothesis that reformulation caused substantial substitution from OxyContin to genericoxycodone. 19igure 3: Main regression on OxyContin sales. Shaded regions are the 95 percent confidenceintervals with standard errors clustered at the MSA level.Figure 4: Main regression on generic oxycodone sales. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.Figure 3 also documents that high pre-reform oxycodone misuse MSAs saw large increasesin OxyContin sales right after the reformulation. This phenomenon has been unreportedpreviously, but would be consistent with Schnell (2017)’s physician benevolence hypothesiswhere good physicians switch patients from oxycodone to reformulated OxyContin to lowerthe future risk of abuse. Although the switch toward OxyContin is smaller in magnitudethan the switch from OxyContin, this increase is the first documented positive impact of theOxyContin reformulation in the literature. It seems both physicians and users saw the twotypes of drugs as substitutes. Unfortunately, there are not enough MSAs where the switch20oward OxyContin is significant enough that it cancels out the switch away from OxyContinto examine the possible substitution channel in the other direction.Because we include both OxyContin and generic oxycodone misuse in the same regression,we can separate out the increases in oxycodone sales due to its own popularity from the in-creases due to spillover effects from the OxyContin reformulation. Figure 4 shows increasinggrowth in oxycodone sales in MSAs with higher oxycodone misuse until 2011, and the growthrate declined after. The smoothness of the oxycodone curve indicates that the OxyContinreformulation had no impact on how oxycodone misuse affected oxycodone sales. This trendcorresponds well with many states tightening control over opioid prescription policies in 2011and 2012 in response to rising sales and and increased awareness of opioid misuse.Another way of estimating the impact of the reformulation is through difference-in-differenceregressions. Column (1) of Table 3 in Appendix shows the regression on OxyContin sales.OxyContin sales in all MSAs decreased by 8.05 MME post-reform, a 9.4% decrease withrespect to the average per person sales of 85.6 MME in 2009. High OxyContin misuse MSAshad a higher level of OxyContin sales to start with, but experienced an additional 15.1 MMEdrop (an additional 17% decrease) post-reform. Given that only 2.46% of the populationever misused OxyContin and the reformulation only affected the people misusing it, a 17%additional decrease in all OxyContin sales would translate into a very significant decrease insales to the population that misuses it. The negative and significant Post × High OxyCon-tin coefficient confirms previous findings that high OxyContin exposure MSAs saw largerdecreases in OxyContin sales post-reform.Column (2) of the same table reports the regression on generic oxycodone sales. Genericoxycodone sales per person increased 41.7 MME in the post period, a 31.2% increase withrespect to the average per person alternative oxycodone sales of 133.5 MME in 2009. HighOxyContin misuse MSAs experienced an additional 10.3 MME increase, which translatesto a 68% conversion from OxyContin to generic oxycodone in those areas. Combining thefindings from columns (1) and (2), we see direct substitution from OxyContin to genericoxycodone in local sales immediately after reformulation, and the substitution pattern ismore pronounced in MSAs with high OxyContin exposure as expected.To help our readers visualize the trend of OxyContin and alternative oxycodone sales, inFigure 12 in the Appendix, we break all MSAs into three bins by the magnitude of theobserved drop in OxyContin sales due to the reform. Then, we plot the per person OxyContin NSDUH, 2010.
Next, we estimate the impact of the reformulation on overdose mortality. In Figure 5, wereport the full set of coefficients from estimating the event study framework on opioid mor-tality. Each data point in the figure is the coefficient of the interactive term of misuse andsale for OxyContin or generic oxycodone in a specific year, and it captures any additionalchange in opioid mortality in that year driven by high OxyContin or oxycodone exposure.The OxyContin coefficients are never significant, suggesting higher pre-reform OxyContinmisuse is not predictive of either higher or lower opioid death post-reform. The lack of anytrend indicates that any benefit of the OxyContin reformulation on reducing OxyContin con-sumption is offset by the substitution to generic oxycodone. In aggregate, the reformulationhad no impact on non-heroin opioid deaths. 22igure 5: Main regression on opioid mortality. Shaded regions are the 95 percent confidenceintervals with standard errors clustered at the MSA level.Figure 6: Main regression on heroin mortality. Shaded regions are the 95 percent confidenceintervals with standard errors clustered at the MSA level.In Figure 6, we report the event study coefficients on heroin mortality. Again the OxyContincoefficients are tiny and insignificant, while the oxycodone coefficients grow over time butnever reach statistical significance at conventional levels. The lack of statistical significanceis due to the small number of heroin moralities in the whole sample and high correlations be-tween OxyContin and oxycodone exposure. If we were to run the OxyContin and oxycodoneregression separately (See Figure 34 and Figure 38 in Appendix), oxycodone exposure had amuch larger and more significant impact on heroin mortality. The results provide tentativeevidence that the higher the generic oxycodone exposure in an MSA, the higher the increases23n heroin mortality. However, the results do not support the alternative hypothesis that theOxyContin reformulation was solely responsible for the increase in heroin mortality.The difference-in-difference results mirror our finding from the event study framework. Col-umn (3) of Table 3 in Appendix suggests that opioid deaths are 0.08 higher in high oxycodoneexposure MSAs, which is equivalent to 27% of the average opioid overdose of 0.29 per 10,000people in 2009. Opioid mortality is 0.05 lower (17% of the 2009 average) in higher OxyContinexposure MSAs after controlling for oxycodone use. Higher OxyContin exposure does notlead to higher or lower opioid overdose post-reform, while higher generic oxycodone exposureis associated with 0.06 (20.6% of 2009 average) more opioid death in the post period.Column (4) of the same table reports the difference-in-difference regression on heroin death.Heroin mortality has increased by 0.14 in the post period in all MSAs, which is equivalentto a 111% increase from the average 2009 level of 0.126 heroin death per 10,000 population.High OxyContin exposure MSAs did not experience additional jumps in heroin mortality,while high oxycodone exposure MSAs did experience an additional 0.07 (56% with respectto 2009 average) increase in death. Again, the evidence from the difference-in-differenceregressions indicate that OxyContin was not responsible for the rise in heroin mortality.In Figure 13 in the Appendix, we show the average trend of the opioid and heroin mortalityfor groups with high, medium and low observed drop in Oxycontin sales. If the reformulationwas responsible for the subsequent heroin epidemic, then the MSAs mostly likely to haveadditional jumps in heroin mortality would be the MSAs with the largest OxyContin drop.As shown in the figure, the three groups went through the same explosive growth in heroinmortality (around 38% from 2009 to 2011, and similar rate afterward), indicating the rise inheroin was independent of the decrease in OxyContin sales. This evidence conclusively rejectsthe hypothesis that the OxyContin reformulation is solely responsible for the subsequentheroin epidemic. (A) The Reformulation’s Impact on Opioid Mortality
Until now, the literature has found mixed results for the effects of the OxyContin reformu-lation on opioid mortality. In contrast to previous work, we find no statistically significantimpact of the reformulation on opioid mortality as a result of substantial substitutions fromOxyContin to generic oxycodone post-reform. Increases in generic oxycodone sales com-24ensated for 55% of the drop in OxyContin sales in high OxyContin misuse MSAs by ourevent study framework, and 68% by our different-in-difference estimation. Opioid mortal-ity continued to increase in the post-reform period, but not was driven by high OxyContinexposure. (B) The Reformulation’s Impact on Heroin Mortality
Our results stand in direct contrast to the findings of the literature. Instead of being theevent that precipitated the heroin epidemic, the OxyContin reformulation shifted misuse toother opioids, of which heroin was only one. We cannot refute the hypothesis that someOxyContin users switched to heroin due to the reformulation. Our analysis refutes thehypothesis that the reformulation was the sole cause of the heroin epidemic. Instead ofOxyContin misuse, we identified generic oxycodone misuse as a much more powerful driverof increases in heroin mortality post-2011. What prompted the increases in heroin use is stillan unresolved question. Previous research has suggested an increase in the supply of heroin(O’Donnell, Gladden, and Seth, 2017) around this time, as well as crackdowns in Florida onpill-mills reducing the supply of oxycodone (Kennedy-Hendricks et al., 2016). (C) Bridging the Differences between our Findings and the Literature
One of the innovations we’ve made in this paper is to shed light on a hidden source of opioidmisuse: the misuse of generic oxycodone. This segment of prescription opioids was overlookedby other scholars because of OxyContin’s dominance in opioid misuse in the early years aswell as, we argue, the lack of identifiable brand names for the generic products. Empiricalstudies based on market data or interviews of opioid users noted that many people misusedgeneric oxycodone products (Paulozzi and Ryan, 2006, Inciardi et al., 2009). Leaving outoxycodone misuse, an important driver of opioid and heroin mortality that is positivelycorrelated with OxyContin misuse, would produce spurious regression results.To show that the difference in findings is not driven by our constructed misuse measure,or our choice of framework, we test whether we can reproduce findings in the literature byrunning all of our regressions using only OxyContin (see Section 7.4.4 in the Appendix). OurOxyContin misuse exposure individually predicts an increase in opioid and heroin mortalitypost-reform as the literature claims. This finding is the basis of previous studies supportingthe claim that the OxyContin reformulation is the main cause of the subsequent heroinepidemic. However, if we run the same set of regressions using only generic oxycodone (seeSection 7.4.5), we were able to produce the same findings. The only way to differentiatethe impact of OxyContin from that of generic oxycodone is to include both in the same25egressions. Variations in local OxyContin and oxycodone exposure allow us to identify theimpact of both series, if any exist. As we’ve shown in our main regressions, the impact ofOxyContin on heroin disappears after controlling for the effect of generic oxycodone. (D) Market Definition
Another innovation we’ve made in this paper is a finer definition of the opioid market. It isimportant to consider what we gain from disaggregating to the MSA level. The specific Oxy-Contin market share in a state is endogenous to a great many things, including advertising(Van Zee, 2009) and triplicate status (Alpert, Evans, et al., 2019). Although the OxyContinreformulation was an exogenous shock, its interpretation is made very complicated becauseits impact depended on each state’s regulatory history and prescribing environment. We doour regressions at the MSA level, where there are unobserved local conditions that affectedsales of OxyContin and generic oxycodone, while controlling for state-level laws and restric-tions. By comparing two different MSAs with the same regulatory environment but differentexposures to the reformulation, we can get at the marginal effects of OxyContin and genericoxycodone exposure. Contrasting the state-level regression estimates (see Section 7.4.3) withour main results, our main results are larger in magnitude and more statistically significant.The MSA level estimation of the effect of exposure on mortality is more stable. (E) Definition of OxyContin Misuse
The literature relies on NSDUH’s OxyContin past-year misuse. To make our findings com-parable with previous studies and robust to the choice of misuse measure, we repeat ourentire analysis with OxyContin last-year misuse and generic oxycodone lifetime misuse (seeSection 7.4.2 for results.) As noted in Section 3.3, using last-year OxyContin misuse givesan unfair advantage to OxyContin due to the timeliness of the measure. If our findings onoxycodone persist despite the unequal treatment of the two misuse measures, then it is astronger indication of the essential role generic oxycodone played in the opioid and heroinepidemic.Comparing the two sets of results, we observe the same decline in OxyContin sales andincrease in generic oxycodone sales, although smaller in magnitude. Both sets of coefficientson opioid mortality become positive but insignificant. Finally, comparing the heroin result,at the state level we do detect a positive effect on heroin mortality from OxyContin. Inaggregate, our results lose some significance when we replace lifetime OxyContin misuse withlast-year OxyContin misuse. The loss of significance, however, is in the direction predictedby the unfair advantage given to OxyContin. This exercise highlights the importance of26reating the two misuse measures equally. When we use measures that more accuratelycapture recent OxyContin misuse than recent generic oxycodone misuse, we could mistakenlyattribute effects of generic oxycodone to OxyContin.
Researchers have attributed the prescription opioid opioid crisis and recent increase in heroinuse to OxyContin. Previous studies have documented how Purdue Pharma’s marketingdownplayed the risks of OxyContin’s abuse potential, which fomented the prescription opioidcrisis; recent studies identified the OxyContin reformulation as the event that pushed usersto switch to heroin, which precipitated recent increase in heroin use. This paper revisitsthe roles OxyContin and the Oxycontin reformulation played in the opioid crisis with fine-grained sales data that includes OxyContin’s most immediate substitute, generic oxycodone.We have three main findings.First, we find direct evidence of substitution to from OxyContin to generic oxycodone post-reformulation. Our difference-in-difference estimation indicates a 68% substitution fromOxyContin to generic oxycodone due to the reform. Looking at the decline in OxyContinsales and rise in generic oxycodone sales from 2002-2006, we believe this substitution (fordifferent reasons, namely Purdue’s loss of its patent) also happened years before the refor-mulation. The size of this substitution, and indeed the size of the generic oxycodone marketpre-reform, may come as a surprise to researchers. Paulozzi and Ryan, 2006 estimate that in2002 OxyContin’s market share was 68%. By the time of the reformulation in 2010, it hadfallen by more than half. OxyContin played an essential part in igniting the prescriptionopioid crisis but, after losing its patent in 2004, other companies took up the torch andsurpassed Purdue by selling generic oxycodone.Our second main finding is that the OxyContin reformulation had no overall effect on opioidmortality. In our estimation, the OxyContin coefficients are not significant in the entiresample period, suggesting that higher OxyContin exposure is not predictive of either higheror lower opioid death. The lack of any trend indicates that the benefits of the OxyContinreformulation, if exist, are offset by substitution to oxycodone. In addition, we do find thathigh oxycodone exposure is predictive of rise in opioid mortality from 2011, confirming theincreasingly important role of generic oxycodone in the recent prescription opioid crisis.Third and most importantly, we show that the heroin overdose deaths after 2010 were pre-27icted by generic oxycodone exposure, not OxyContin exposure. Our main event-studymodel shows positive and significant effects from oxycodone exposure on heroin deaths after2012, but OxyContin exposure is not predictive of heroin deaths once we control for oxy-codone. The difference-in-difference results are similar, showing that oxycodone exposure waspredictive of heroin deaths before or after the reformulation, and OxyContin exposure afterthe reformulation is weakly positive but not statistically significant. We also do not observean additional rise in heroin deaths immediately after reformulation in areas where Oxy-Contin sales declined the most post-reformulation. In particular, without including genericoxycodone in the analysis, we recover the same results from the literature that OxyContinwas responsible for the rise in heroin deaths. The evidence shows that omitting oxycodone,an important substitute to OxyContin, produces erroneous results. This paper demonstratesthe pernicious effects of generic oxycodone, which had thus far escaped scrutiny until theWashington Post acquired data and reported on it.28 eferences [1] . 2012. url : .[2] Jeremy A Adler and Theresa Mallick-Searle. “An overview of abuse-deterrent opi-oids and recommendations for practical patient care”. In: Journal of multidisciplinaryhealthcare
11 (2018), p. 323.[3] Abby Alpert, Evans, et al. “Origins of the Opioid Crisis and Its Enduring Impacts”.In: Working Paper Series 26500 (Nov. 2019). doi :
10 . 3386 / w26500 . url : .[4] Abby Alpert, David Powell, and Rosalie Liccardo Pacula. “Supply-Side Drug Policyin the Presence of Substitutes: Evidence from the Introduction of Abuse-DeterrentOpioids”. In: Working Paper Series 23031 (Jan. 2017). doi : . url : .[5] Abby Alpert, David Powell, and Rosalie Liccardo Pacula. “Supply-side drug policy inthe presence of substitutes: Evidence from the introduction of abuse-deterrent opioids”.In: American Economic Journal: Economic Policy
ASIPP ® )”. In: Cal-ifornia medicine
Journalof Health Economics
52 (2017), pp. 63–73.[9] Theresa A Cassidy et al. “Changes in prevalence of prescription opioid abuse afterintroduction of an abuse-deterrent opioid formulation”. In:
Pain Medicine
JAMA psychiatry
JAMA Psychiatry issn : 2168-622X. doi : . eprint: https://jamanetwork.com/journals/jamapsychiatry/articlepdf/2174541/yoi140121.pdf . url : https://doi.org/10.1001/jamapsychiatry.2014.3043 .[12] Theodore Cicero, Matthew Ellis, and Hilary Surratt. “Effect of abuse-deterrent for-mulation of OxyContin”. In: New England Journal of Medicine
New England Journalof Medicine
Pharmacoepidemiology and drug safety
Drug and alcohol dependence
145 (2014), pp. 238–241.[16] Sarah DeWeerdt. “Tracing the US opioid crisis to its roots.” In:
Nature
Review of Economics and Statistics
The quarterly journal of economics
Drug and alcoholdependence
139 (2014), pp. 9–17.[20] Lon R Hays. “A profile of OxyContin addiction”. In:
Journal of Addictive Diseases
Journal ofaddictive diseases
American journal of public health
NewEngland Journal of Medicine
Dopesick: Dealers, doctors, and the drug company that addicted America .Little, Brown, 2018.[25] Justine Mallatt. “The effect of prescription drug monitoring programs on opioid pre-scriptions and heroin crime rates”. In:
Available at SSRN 3050692 (2018).[26] Sarah G Mars et al. ““Every ‘never’I ever said came true”: transitions from opioid pillsto heroin injecting”. In:
International Journal of Drug Policy $
600 million”. In:
New YorkTimes
10 (2007).[28] Barry Meier.
Pain killer: A” wonder” drug’s trail of addiction and death . Rodale, 2003.[29] F Modarai et al. “Relationship of opioid prescription sales and overdoses, North Car-olina”. In:
Drug and alcohol dependence
MMWR. Morbidity and mortalityweekly report
American journal of preventive medicine
Current medical research and opinion
Cali-fornia medicine
Dreamland: The true tale of America’s opiate epidemic . BloomsburyPublishing USA, 2015.[35] Bob Rappaport.
APPLICATION NUMBER: 22-272 - MEDICAL REVIEW(S) . FDACENTER FOR DRUG EVALUATION and RESEARCH, 2009. url : .3136] Christopher J Ruhm. “Geographic variation in opioid and heroin involved drug poison-ing mortality rates”. In: American journal of preventive medicine
American family physician
Canadian Family Physician
Journal of PublicEconomics
122 (2015), pp. 13–30.[41] Art Van Zee. “The promotion and marketing of oxycontin: commercial triumph, publichealth tragedy”. In:
American journal of public health
Pain Medicine
Appendix
Note:
We compute market share based on the average of 2006-2014 sales data. We kept only thetop twenty manufacturers for better readability of the table. The rest of the 35 manufacturerscombined contribute 0.18% of total sales. During this sample period, Purdue Pharma was thedominant manufacturer of high dosage oxycodone pills ( ≥ Figure 7: Market share of different opioid manufacturers33 ote:
The figure shows the misuse rate of OxyContin (OXYFLAG or OXYCONT2)and the misuse rate of Percocet, Percodan and Tylox (PERCTYL2). Data obtainedfrom annual NSDUH. Percocet was a popular prescription oxycodone to misuse inthe pre-OxyContin period. We see in this graph that the PERCTYL2 misuse rateincreased 30% from 2002 to 2009, suggesting that the lifetime misuse rate capturesmore than historical Percocet, Percodan and Tylox misuse.
Figure 8: NSDUH national lifetime misuse rate34 ote:
This graph shows the difference in oxycodone sales between Purdue and Endo Pharma.The small market share of Endo Pharma leads us to believe that individuals misreport the drugsthey consume on the NSDUH.
Figure 9: Comparison of Sales of Purdue and Endo Pharma
Note:
Left is the absolute difference in market share (0.1 means that MSA share is 10% higherthan the state average) and right is percentage difference (10% means that MSA share is 1.1times the state average).
Figure 10: Within-state variation in OxyContin Market Share35 ote:
Left is the absolute difference in opioid mortality (0.1 means that MSA mortality per10,000 people is 0.1 higher than the state average) and right is percentage difference (10% meansthat MSA mortality per 10,000 people is 1.1 times the state average)
Figure 11: Within-state variation in opioid mortality
Note:
We categorized all MSAs into high, mid, and low by the drop in the observed per personOxyContin sales from 2009 to 2011. The series are population weighted and Florida is excluded.The high group saw a 30% drop in OxyContin sales, mid group a 3.9% drop, and low group a15% increase. The high group experienced a 46% increase in generic oxycodone sales, mid groupa 34% increase, and low group a 29% increase. The three groups share similar oxycodone growthtrends until the reformulation.
Figure 12: Opioid sales by empirical OxyContin drop36able 2: Testing Constructed Exposure Measure Against Opioid Mortality
Opioid overdose deaths per 100,000OxyContin Generic Oxycodone(1) (2) (3) (4) (5) (6)NSDUH misuse 10.235 2.909(1.719) (0.570)ARCOS sales 0.001 0.001(0.0002) (0.0001)Combined exposure 0.093 0.087(0.012) (0.009)Number of observations 379 379 379 379 379 379R-square 0.086 0.089 0.130 0.065 0.178 0.189Adjusted R-square 0.084 0.086 0.128 0.062 0.176 0.187
Notes:
Standard errors are in parentheses. We report coefficients from OLS regressions ofopioid mortality on misuse, sales or exposure. NSDUH misuse rates is the 6-year averageOxyContin or Percocet lifetime misuse rate from pre-reform period (2004-2009). ACROS sales isOxycontin or generic oxycodone sales per person from 2009. Combined exposure is the productof the previous two measures normalized (see equation 1). Overdose from 2009. Regressionsare weighted by MSA population.
Note:
Similarly to the previous figure, we categorized all MSAs into high, mid, and low by thedrop in the observed per person OxyContin sales from 2009 to 2011. The series are populationweighted and Florida is excluded. No trend break in opioid mortality in the high drop group.The high group saw an 35% increase in heroin mortality, the mid group 38%, and the low group37%. The similar increases in heroin mortality post-reform indicates that drops in OxyContinuse post-reform did not lead to additional increase in heroin use.
Figure 13: Opioid mortality by empirical OxyContin drop37able 3: Difference in difference regression results
Opioid sales per person Overdose per 10,000OxyContin Oxycodone Opioid Heroin(1) (2) (3) (4)Post -8.05 41.74 0.01 0.14(2.86) (4.92) (0.02) (0.02)High OxyContin 47.24 56.46 -0.05 -0.07(5.78) (13.36) (0.03) (0.02)High Oxycodone 26.84 95.90 0.14 0.08(6.66) (15.47) (0.04) (0.05)Post x High OxyContin -15.14 10.30 0.02 0.03(6.39) (8.90) (0.02) (0.02)Post x High Oxycodone -2.33 33.99 0.06 0.07(6.37) (8.80) (0.02) (0.02)Number of observations 2148 2148 2148 2148R-square 0.665 0.737 0.517 0.469Adjusted R-square 0.654 0.728 0.501 0.452
Notes:
We report coefficients from the difference-in-difference estimation (see equation (3)).All MSAs in Florida are excluded. In all specifications, we include MSA-level control variables,state fixed effects and year fixed effects. Standard errors are clustered at the MSA level. .2 Tables7.3 Map Note:
Data from 2004-2009 NSDUH lifetime OxyContin misuse rate (NSDUH ticker OXXYR). 0.01 isinterpreted as 1% of the state population have ever misused OxyContin.
Figure 14: OxyContin lifetime misuse rate at state level39 ote:
Data from 2004-2009 NSDUH lifetime Percocet, Percodan, Tylox misuse rate (NSDUH tickerPERCTYL2). 0.01 is interpreted as 1% of the state population have ever misused one of the three drugs.Percocet lifetime misuse rate on average is much higher than OxyContin lifetime misuse rate.
Figure 15: Percocet lifetime misuse rate at state level40 ote:
The figure plots the absolute difference in percentile ranking of the two state level lifetime misuserate. A 0.1 should be interpreted as a 10% difference in percentile ranking between OxyContin lifetimemisuse rate and Percocet lifetime misuse rate. For example, Colorado’s OxyContin misuse rate is 0.0063(42 percentile) and it’s Percocet misuse rate is 0.092 (97 percentile), which is a 55% difference in percentileranking. We rely on the difference between two misuse rate to separately identify the impact of OxyContinand oxycodone.
Figure 16: Difference in state level misuse rates41 ote:
This figure shows OxyContin exposure by MSA. We show Florida here, which had very low OxyContinexposure/sales, but omit it from analysis because it had abnormally high generic oxycodone sales with largeamounts being trafficked to other states.
Figure 17: OxyContin exposure at MSA level42 ote:
Florida is excluded in this analysis. MSAs grouped by high vs low OxyContin exposure and high vslow generic oxycodone exposure.
Figure 18: Diff-in-diff regression categories43 .4 Alternative Regression Specifications
Figure 19: Regression on OxyContin sales with MSA FE. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.Figure 20: Regression on oxycodone sales with MSA FE. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.44igure 21: Regression on opioid mortality with MSA FE. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.Figure 22: Regression on heroin mortality with MSA FE. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.45 .4.2 Last Year OxyContin Misuse
Figure 23: Regression on OxyContin sales with last-year OxyContin. Shaded regions are the95 percent confidence intervals with standard errors clustered at the MSA level.Figure 24: Regression on oxycodone sales with last-year OxyContin. Shaded regions are the95 percent confidence intervals with standard errors clustered at the MSA level.46igure 25: Regression on opioid mortality with last-year OxyContin. Shaded regions are the95 percent confidence intervals with standard errors clustered at the MSA level.Figure 26: Regression on heroin mortality with last-year OxyContin. Shaded regions are the95 percent confidence intervals with standard errors clustered at the MSA level.47 .4.3 State Level Regression
Figure 27: Regression on OxyContin sales at state level. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.Figure 28: Regression on oxycodone sales at state level. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.48igure 29: Regression on opioid mortality at state levelFigure 30: Regression on heroin mortality at state level. Shaded regions are the 95 percentconfidence intervals with standard errors clustered at the MSA level.49 .4.4 OxyContin Only
Figure 31: Regression on OxyContin sales with OxyContin only. Shaded regions are the 95percent confidence intervals with standard errors clustered at the MSA level.Figure 32: Regression on oxycodone sales with OxyContin only. Shaded regions are the 95percent confidence intervals with standard errors clustered at the MSA level.50igure 33: Regression on opioid mortality with OxyContin only. Shaded regions are the 95percent confidence intervals with standard errors clustered at the MSA level.Figure 34: Regression on heroin mortality with OxyContin only. Shaded regions are the 95percent confidence intervals with standard errors clustered at the MSA level.51 .4.5 Oxycodone Only.4.5 Oxycodone Only