The Effect of Prescription Drug Monitoring Programs on Opioid Utilization in Medicare ∗ Thomas C. Buchmueller Ross School of Business University of Michigan [email protected]

Colleen Carey Department of Policy Analysis and Management Cornell University [email protected] April 26, 2016

Abstract The misuse of prescription opioids has become a serious epidemic in the US. In response, states have implemented Prescription Drug Monitoring Programs (PDMPs), which record a patient’s opioid prescribing history. While few providers participated in early systems, states have recently begun to require providers to access the PDMP under certain circumstances. We find that “must access” PDMPs significantly reduce measures of misuse in Medicare Part D. In contrast, we find that PDMPs without such provisions have no effect. We also find that Part D enrollees obtain more opioids from out-of-state pharmacies after a “‘must access” PDMP law has been implemented.

JEL codes: I18 (Government Policy, Regulation, Public Health); I12 (Health Behavior)

∗ This research was supported by the Robert Wood Johnson Foundation Scholars in Health Policy Research Program. Anup Das provided excellent research assistance. The authors thank Jean Abraham and seminar participants at Cornell University and the Midwestern Health Economic Conference.

1

1

Introduction

In recent years, the use of prescription opioids has grown dramatically in the United States; between 1999 and 2010, opioid prescriptions increased by 300% (Kunins et al., 2013). Opioids are an effective analgesic, but bring a high risk of addiction and overdose, and over the same period deaths from opioid poisoning more than quadrupled (Chen et al., 2014). In addition, it became clear that opioids prescribed legitimately by well-meaning medical professionals were being diverted to street sale for non-medical use (Dart et al., 2015). In October 2015, President Obama and health care provider groups announced a set of joint public-private initiatives aiming to improve prescribing practices and expand addiction treatment; the President’s 2017 budget proposal requests more than $1 billion in new funding for related efforts. Most policy activity in response to the opioid problem has taken place at the state level; nearly every state has implemented a Prescription Drug Monitoring Program, or PDMP, that collects data on prescriptions for controlled substances to facilitate detection of suspicious prescribing and utilization. A PDMP allows authorized individuals to view a prescribing history in order to identify those who are misusing or diverting opioids. State programs have evolved over time and continue to vary along several dimensions, including the extent to which medical providers are encouraged to access the data. Early programs were developed by law enforcement agencies and had provisions that made it difficult for providers to access the data. Other programs did not block provider access, but did not actively encourage them to utilize the data. PDMP administrative data show that, when provider access is possible but not mandatory, only a small share of providers create PDMP logins and actually request patient histories (PDMP Center of Excellence, 2014). A goal of the same October 2015 initiative was to “double the number of health care providers registered with their state PDMPs in the next two years”, showing the recent acknowledgment of this problem (White House, 2015). Low provider utilization of these programs may explain why a number of studies find that PDMPs have little or no effect on opioid use and related adverse health outcomes (Paulozzi et al., 2011; Reifler et al., 2012; Jena et al., 2014; Li et al., 2014; Brady et al., 2014; Haegerich et al., 2014). Between 2007 and 2012, six states enacted stronger laws requiring providers to access the PDMP under certain circumstances prior to prescribing. Seventeen other states enacted PDMP legislation that did not include such a “must access” provision. In this paper, we estimate difference-in-differences models using aggregated claims data from Medicare’s prescription drug program (Medicare Part D) to evaluate the effect of these laws on the prescription drug utilization of Medicare beneficiaries. Our large-N microdata allows us to measure rare outcomes in the upper tail of the distribution – the exact outcomes that PDMPs are meant to impede. For example, a key outcome that we analyze is the percentage of opioid users who obtain prescriptions from five or more prescribers, which is a commonly used marker for “doctor shopping”. 2

The Part D claims also allow us to construct multiple measures capturing the intensity of opioid utilization. Additional data on hospital (Medicare Part A) and outpatient (Medicare Part B) claims provide information on opioid poisoning incidents. We first establish that a PDMP without a “must access” provision is not associated with opioid misuse in Medicare. In contrast, we find that implementing a “must access” PDMP reduces the percentage of Medicare Part D enrollees who obtain prescriptions from five or more prescribers by 12% and reduces the percentage of enrollees who obtain prescriptions from five or more pharmacies by more than one-third. These results suggest that measures that require prescribers to access the PDMP can be an effective way to reduce questionable opioid use patterns. However, we find no effect of “must access” PDMP laws on the percentage of Medicare beneficiaries receiving extremely large volumes of opioids. And even though obtaining opioids from five or more physicians or five or more pharmacies are both strongly correlated with opioid poisoning incidents as observed in medical claims data, we find a negative but statistically insignificant relationship between “must access” PDMP laws and opioid poisoning incidents. Because there is little or no integration of state databases, drug-seeking individuals trying to avoid scrutiny may cross state borders to escape detection by the PDMP in their home state. We test for such avoidance behavior using a novel measure of whether a Part D enrollee disproportionately obtains opioids from out-of-state prescribers or pharmacies (relative to her out-of-state rate for non-opioid prescriptions). Our results indicate that a “must access” PDMP in a given state raises the rate at which that state’s residents obtain opioids from out-of-state pharmacies by roughly a third of a standard deviation. Additional results support a causal interpretation of our findings for “must access” laws. An analysis of leads suggests that states that go on to implement “must access” PDMPs had similar experiences to states that did not. In addition, we estimate models that compare each state with a “must access” PDMP individually to the states without such a policy; the results are broadly similar across all six “treatment” states. As a sort of placebo test, we examine the utilization of statins and antidepressants. It is very uncommon to obtain these drug classes from multiple prescribers or pharmacies, suggesting that our opioid “misuse” measures do not simply reflect poor care coordination among Medicare beneficiaries. We find no effects of a “must access” PDMP on the utilization of either of these two types of drugs. One concern with any strategy aimed at curbing the misuse of prescription drugs is that it will reduce access to pain relief for patients with legitimate clinical needs. To test for such an effect, we conduct the analysis on cancer patients and those near the end of life. We find that the policy has no effect on the utilization of prescription opioids for these two groups. In contrast, we find strong effects for individuals who qualify for Medicare on the basis of a disability, a group that has been identified as having high rates of opioid misuse and abuse (GAO, 2011). In addition, we find no effect of the policy on the percentage of 3

Medicare beneficiaries with any prescription opioid usage or other outcomes at the mean of the utilization distribution; instead, the policy affects only extreme outcomes. These results suggest that the estimated impact of a “must access” PDMP is driven by reducing misuse rather than by reducing appropriate use. Our results help explain why previous analyses concluded that PDMPs have no effect on opioid use or misuse. Firstly, these studies considered early versions of PDMPs that limited access to law enforcement or had very low rates of provider registration and access. We find it plausible that a PDMP in which providers cannot or do not participate does not affect opioid utilization, and indeed we cannot detect any results of these types of PDMPs among Medicare beneficiaries. Secondly, many of the previous studies looked at aggregate consumption outcomes that combine appropriate use and misuse. But if the majority of opioid utilization is appropriate, a successful PDMP may have no effect on mean outcomes. We find that a “must access” PDMP affects the small portion of utilization suggestive of misuse while not affecting average utilization.

2

Background

2.1

Opioids and Medicare Part D

In recent decades, many physicians have argued that pain, “the fifth vital sign”, was historically undertreated (Pasero and McCaffery, 1997). In response, the medical community began to treat pain more aggressively, often using prescription opioids. The top panel of Figure 2 shows opioid prescriptions becoming more widespread between 2007 and 2012, the time period covered by this analysis. Much of the increase came in in prescriptions for chronic pain, despite the fact that research suggests that opioids are less effective for chronic pain than for acute pain, such as after surgery (Manchikanti et al., 2010). Along with the increase in utilization came significant increases in rates of misuse and overdose; the dotted line in Figure 2 shows a steep increase in emergency department visits for opioid poisoning. According to the National Survey on Drug Use and Health (NSDUH), by 2013 more than 35 million Americans had used pain relievers non-medically at least once in their life (NSDUH, 2014). Given the significant disease burden of the elderly and disabled, it is not surprising that opioid use is quite common in Medicare Part D. The bottom half of Figure 2 depicts opioid utilization and poisonings among Part D enrollees, demonstrating the national trends are reproduced in the sample we consider. In 2010, approximately one-quarter of Part D enrollees took opioids (Jena et al., 2014). Use is especially prevalent among disabled Medicare beneficiaries: in 2011 nearly half filled an opioid prescription (Morden et al., 2014). Over the time period of this analysis, the median days supply among disabled opioid takers rose from 52 days in the first half of 2007 to 90 days in the second half of 2012. A 2011 report by the Government Accountability 4

Office (GAO) found that in 2008 2% of Medicare beneficiaries obtained frequently-abused drugs (opioids as well as some stimulants and benzodiazepines) from five or more prescribers per year, which they deemed indicative of “doctor shopping” (GAO, 2011). Patterns of questionable access are more prevalent among the disabled, who comprise about a third of opioid takers in Medicare Part D. However, opioid misuse among the elderly has been rising in the very recent past. In 2012, the opioid mortality rate among those over sixty years of age surpassed that of those between twenty and fifty-nine (West et al., 2015). For most years since the Medicare prescription drug program was established in 2006, program administrators and plan sponsors have had a limited ability to identify or curtail opioid misuse (GAO, 2011). Efforts to address the problem have largely been limited to encouraging plans to “provide practitioner and beneficiary education as appropriate” when suspicious patterns were identified. While state Medicaid programs can require suspected opioid abusers to receive their drugs from only one prescriber or one pharmacy, the legislation authorizing Medicare Part D did not permit either Part D plans or the Center for Medicare and Medicaid Services (CMS) to restrict access in this way (GAO, 2011). In addition, while beneficiaries can change Part D plans at least yearly, an insurer that has identified suspected drug abuse is not permitted to transmit that information to future insurers (Blum, 2013). In 2013, after our sample period ended, Part D plans gained permission to deny at point-of-sale claims that would result in “unsafe cumulative dosage” (Blum, 2013). To note, opioids are relatively inexpensive, accounting for about 3% of Part D insurers’ total drug costs in 2011, whereas antidepressants, for example, accounted for nearly 10%. If only a small portion of opioid utilization is inappropriate, insurers may not choose to undertake self-financed efforts to interdict it.

2.2

State Prescription Drug Monitoring Programs

The most important policies aimed at curbing opioid abuse are prescription drug monitoring programs, or PDMPs, which are established and operated by states to collect and facilitate the sharing of data on opioid prescriptions. The earliest programs were primarily designed to assist law enforcement in investigations, and were based around carbon copies. In the past decade, PDMPs have migrated online and have become more universal and up-to-date, allowing physicians and pharmacists to easily access a patient’s prescription history. Policymakers have hypothesized that informing medical providers of potential misuse could help impede diversion and recreational use of opioids. A physician or pharmacist might refuse to write or fill a prescription in response to a prescribing history, such as in Figure 1, that showed multiple providers or an unusually large quantity of opioids. However, even as information technology made it easier to obtain the data in a timely fashion, state programs established different rules regarding provider access. In some states, PDMPs began to allow 5

providers to query the PDMP, while in others it was explicitly prohibited or allowed only for current (but not prospective) patients (Davis et al., 2014). Similarly, in some states, the PDMP actively generated reports to prescribers and pharmacies when a patient appeared to be misusing opioids; in other states, it was illegal for the PDMP to do so (Davis et al., 2014). Despite a growing recognition of the dangers of opioids, physician groups have not generally endorsed PDMP access as a solution. Physicians complained about interference with clinical practice, new paperwork burdens, and the difficulty of contextualizing and interpreting a prescribing history (Islam and McRae, 2014; Gourlay, 2013). When prescribers were allowed access, they were sometimes granted immunity from prosecution for not checking the PDMP (Davis et al., 2014). In this context, it is not surprising that when not mandated, only the most conscientious providers actively used the PDMP to inform prescribing decisions (Haffajee et al., 2015). For example, in the first year after a voluntary PDMP was established in Florida, a state with a well-publicized opioid misuse problem, fewer than 1 in 10 physicians had even created a login for the system (Poston, 2012). Similarly, two years after Rhode Island allowed prescribers to query its PDMP, officials estimated it was used in only 10% of prescriptions (Arditi, 2014). Levy et al. (2015) show that nearly half of prescriptions for opioids from the United States are written by primary care specialists, suggesting that even if the heaviest prescribers are enrolling, these low rates of uptake will limit the effectiveness of PDMPs. In recent years, a number of states have strengthened their PDMPs by requiring providers to access the data under certain circumstances. These “must access” provisions appear to have substantially increased provider take-up. When New York implemented a “must access” provision in 2013 (after the period of our analysis), the number of registrants increased fourteenfold, and the number of daily queries rose from fewer than 400 to more than 40,000. Similarly, in Kentucky, Tennessee and Ohio, implementing a “must access” provision increased by several orders of magnitude the number of providers registered and the number of queries received per day (PDMP Center of Excellence, 2014). Table 1 summarizes state PDMP laws by providing the date that the first PDMP law went into effect and when a “must access” provision went into effect (where applicable). Our main source of information on state laws is the detailed database of PDMPs collected by the Network for Public Health Law (Davis et al., 2014). This database was created by multiple legally-trained researchers independently reviewing state laws relevant to PDMP operations for content and dates of implementation. This database covers 1998 to 2011; we independently researched whether any other states introduced “must access” PDMPs after 2011. During the period of our analysis (2007 to 2012), six states implemented such laws. Note that in all cases but one (Delaware) the “must access” provision strengthened an existing PDMP. Previous studies on PDMPs have generally not attempted to distinguish among different types of pro6

grams. For example, Paulozzi et al. (2011) examine the relationship between whether a state has any kind of PDMP between 1999 and 2005 and opioid poisoning incidents and mortality. They find no effect. Similarly, using cross-sectional data from 2010, Jena et al. (2014) find no relationship between the presence of a state PDMP and the number of Medicare beneficiaries obtaining prescriptions from multiple providers. Reifler et al. (2012), considering 1999-2003, find that the growth in opioid use slows when states pass PDMP laws. Brady et al. (2014) and Li et al. (2014) find no effect on opioids dispensed or drug overdose mortality from PDMPs between 1999 and 2008. In a review, Haegerich et al. (2014) conclude that 14 studies of PDMPs “have not clearly established significant effects on total opioid prescribing or health outcomes with PDMPs.” In this paper we focus on the states that require prescribers (under certain circumstances) to access the PDMP prior to prescribing opioids. Although we combine these six states in a single category, their laws differ slightly in their scope (Davis et al., 2014). Kentucky’s law is the strongest, requiring access before every prescription and at three-month intervals thereafter. In Delaware, Ohio, and Nevada, prescribers must access the PDMP when he or she “has a reasonable belief that the patient may be seeking the controlled substance...for any reason other than the treatment of an existing medical condition.” In Louisiana and Oklahoma, the requirement is limited to pain and drug abuse clinics, respectively.

3

Data and Methods

3.1

Measures of Opioid Use, Misuse, and Poisoning

Our primary dataset is the prescription drug and medical claims of a random 5% subsample of Medicare beneficiaries enrolled in Part D and fee-for-service Medicare (not Medicare Advantage) in any year between 2007 and 2012. Because several of the legislative changes during this period occurred in the middle of the calendar year, to better match the claims data to the appropriate policy regime, we divide the data into six-month periods.1 Table 2 provides basic summary statistics for the full sample of FFS/Part D enrollees (first column) and the subsample with at least one opioid claim during each half-year (second column). For each of twelve half-years, the sample size is approximately 950,000; individuals appear an average of ten times. A quarter of the sample is entitled to Medicare due to disability and more than two in five are dually-eligible for Medicaid due to low wealth and income. In any half-year, about 17.5% FFS/Part D enrollees fill at least one prescription for opioids. Opioid takers are more likely than the average Part D enrollee to be eligible for Medicaid and more likely to be disabled. Not surprisingly, opioid users also have higher rates of cancer and are more likely to be near the end of life.2 We consider the full population of opioid takers, even though 1 We

assign laws to half-years on the basis of the regime in the state for the majority of the half-year. identified by the ICD9 codes that comprise CMS Hierarchical Condition Codes 7, 8, 9, and 10 in any position in

2 Cancer

7

more than 40% of these individuals fill only a single prescription (often as part of post-operative care). Table 3 describes the measures of opioid use and misuse that we construct using the claims data. Pharmacy and prescriber identifiers are not included on our 2007 Part D claims file; therefore, variables using those data are reported for the 2008-2012 period. The prescription drug claims denote the exact drug purchased, days supply, purchase date, prescriber identifier and pharmacy identifier. Opioids are identified using the United States Pharmacopeia 2011 Medicare Model Guideline Formulary Reference File.3 Even with this rich detail, it is difficult to distinguish appropriate and inappropriate use of prescription opioids. We draw on prior research to construct several proxy measures of misuse. One of the main ways that a PDMP can reduce opioid abuse is by curtailing “doctor shopping” behavior. In the Part D claims, we can count the number of distinct providers from whom an enrollee received a prescription for opioids during the half-year and the number of different pharmacies at which opioid prescriptions were filled. For both variables, the modal value is one, though there is a long right tail, especially for the number of physicians. We construct a measure of doctor-shopping by identifying all Part D enrollees who received opioid prescriptions from 5 or more providers in a single half-year. In our data, 1.6% of opioid takers meet this criterion. We construct an analogous measure for individuals who fill opioid prescriptions at five or more pharmacies in a half-year. Such “pharmacy shopping” is less common: approximately half of a percent of opioid users visit five or more pharmacies in a half-year. Figures 3 and 4 present the distribution of these variables, with the right panel zooming in on the distribution above the cutoff. The right panels show the frequency of observations, revealing thousands of observations of even these rare phenomena. Although receiving prescriptions from multiple physicians or pharmacies is likely to be indicative of questionable use, results from the NSDUH indicate that most opioids consumed non-medically are actually obtained from a single prescriber rather than from multiple (NSDUH, 2014). Therefore, we develop three quantity-based measures reflecting the upper tail of the utilization distribution. The goal is to capture utilization patterns that previous research suggests are either indicative of misuse or dangerous per se. First, we create an indicator variable that equals one for all individuals who obtained 211 or more days supply (more than seven thirty-day prescriptions) of opioids in a half-year. The distribution of days supply is given in Figure 5, again showing the distribution above the threshold in the right panel. Roughly 8% of opioid takers fill more than 210 days supply of opioid prescriptions in a half-year; high days supply could indicate abuse or opioid diversion, but could also result from patients experimenting with different ingredients or strengths, or combinations of long-acting and short-acting opioids. To address this second (innocuous) cause of high days supply, we convert each prescription to morphine-equivalent dosage. The morphine-equivalent the inpatient, outpatient, and carrier (physician) medical claims. Cause of death is not noted in Medicare claims data. 3 We do not distinguish between Schedule II (most opioids) and Schedule III (hydrocodone combinations and buprenorphine) claims. Over our time period, all PDMPs covered Schedule II and nearly all covered Schedule III drugs (Davis et al., 2014).

8

dosage allows opioids of different ingredients, strengths, and form (routes of administration) to be converted into equivalent milligrams of morphine.4 We divide the total MED obtained in a six-month period by 180 days to create a “daily MED”.

5

Figure 6 depicts the distribution of this variable. The height of the first

bar shows that more than three-quarters of opioid takers have less than ten MED over the six month period, most likely consistent with acute use over a limited period. 2.5% of opioid takers obtain enough opioids to have a daily dosage of more than 120 MED. Guidelines provided to providers, e.g. Washington State Agency Medical Directors’ Group (2010), suggest that chronic usage above this threshold is associated with an escalation in risk. A final quantity-based measure is the prevalence of overlapping claims – i.e., having multiple prescriptions for the same drug at a point in time. We create a binary flag for whether an individual fills a second claim for the same ingredient more than a week before the first prescription should have been finished, based on its days supply.6 If the two claims are for different ingredients, we attribute their overlap to the patient experimenting with other opioids and do not code the individual as having overlapping claims. This measure is similar to the concurrently supplied measure in previous research (Jena et al., 2014). More than 91% of individual-half years have no overlapping claims. Most of those with overlapping claims have only one, but individuals in the top percentile have six claims overlapping by more than a week in a six month period. The data collected by a state PDMP is generally limited to prescriptions that have been written and filled in that state. Similarly, access to the data is generally limited to physicians, pharmacists and law enforcement officials in that state. Therefore, to the extent that a PDMP does curtail doctor-shopping it may push some drug abusers to obtain or fill prescriptions in other states. Indeed state and Federal policymakers have complained for more than a decade that uncoordinated state efforts are vulnerable to this behavior (GAO, 2002). Since the Medicare claims data provides information on the geographic location of enrollees, prescribers and pharmacies, we can test for this type of avoidance behavior. One challenge, however, is that it is not uncommon for Medicare beneficiaries to fill prescriptions out-of-state, either because they live near a state border, are traveling, or live part of the year in another state (e.g., “snow-birds”). To distinguish border-crossing for the purpose of obtaining opioids from other reasons that a patient would fill prescriptions in more than one state, we calculate an individual-level measure of “excess” opioid claims from out-of-state prescribers as 4 We collected the conversion of each opioid ingredient×strength×form from the following three sources: Palliative.org (2016); CMS (2015); Ohio Bureau of Workers’ Compensation (2016). The exact conversion is available upon request. 5 We chose to normalize total MEDs obtained by 180 days because we want to identify large quantities of opioids obtained for the purposes of diversion. An alternative measure would test the intensity of opioid treatment by dividing the total MED by the number of actual days supply rather than 180. In results available upon request we find no impact of a “must access” PDMP on this alternative measure. 6 Because we follow a person longitudinally across years, we can properly account for prescriptions filled in, say, late December and early January.

9

Excess Opioid Claims from OOS Prescriber = P

Opioid Prescriptions from OOS Prescribers P Opioid Prescriptions

−

P

Non-opioid P Prescriptions from OOS Prescribers Non-opioid Prescriptions

We construct an analogous measure of excess claims from out-of-state pharmacies; on average, individuals are less likely to fill opioids at an out-of-state pharmacy than to fill non-opioids at such pharmacies, making the mean of this measure in Table 3 negative. As with our other proxies, this type of behavior is rare; only a few percent of opioid takers are more likely to obtain opioids from out-of-state than from in-state providers. Lacking guidance from the literature to suggest a cutoff for these variables, we analyze them as continuous outcomes. Although we would expect that individuals who misuse prescription painkillers will obtain large amounts of those drugs, perhaps from multiple prescribers, other factors, such as poorly coordinated care, may lead to such patterns of utilization. Therefore, in order to interpret our opioid-based measures it is useful to construct similar measures of “misuse” for other classes of drugs that are not thought to be subject to abuse or diversion. Table 13 reports the mean of our outcome measures for two other classes of drugs commonly used by Medicare enrollees: statins and antidepressants. It is apparent from looking at this table that it is vanishingly rare to “doctor shop” or “pharmacy shop” for antidepressants and statins. This suggests that it is unlikely that there are benign explanations for such measures when it comes to opioids. In contrast, high days supply and overlapping claims are quite common among those two classes. One interpretation is that individuals commonly experiment with drugs in these categories, and that prescribers are comfortable allowing overlapping claims and large quantities among drugs with low potential for overdose and abuse. Similarly, our quantity-based measures could be consistent with experimentation among opioid types. Finally, we see that individuals obtain relatively more opioids from out-of-state prescribers than the other drug types, but the reverse holds for out-of-state pharmacies. Table 4 reports the correlation of our seven misuse measures with each other and opioid poisonings. If our misuse measures were uncorrelated, we might worry that we are mistaking appropriate use for misuse. We are therefore reassured that, of those who obtain opioids from five or more pharmacies, half also obtain them from five or more prescribers. On the other hand, 35% of those who obtain more than seven months’ supply see only one prescriber, suggesting that “doctor shopping” is not the only pathway to potential misuse. This suggests the utility of analyzing a number of proxy measures which capture various aspects of opioid misuse. The relationship between our measures of misuse and opioid poisonings also informative for understanding whether they truly reflect the type of medication use that PDMPs are intended to reduce. In contrast to studies which relate aggregate utilization to aggregate poisonings, our microdata allow us to observe the joint distribution at the individual level. An opioid poisoning incident is indicated by an ICD9 code for an 10

“opioid related overdose” in any position on an inpatient, outpatient, or physician claim (Coben et al., 2010; Bourgeois et al., 2010).7 The rate of opioid poisoning incidents in a half-year is nearly one per hundred. The bottom row of Table 4 indicate that the prescription-based measures are, in fact, positively correlated with opioid poisoning, as measured in the medical claims. The relationship can also be seen by running probit regressions of our poisoning outcome on the misuse proxies. Estimated marginal effects from such regressions are reported in Table 5. Consistent with the work of (Jena et al., 2014), nearly all of our misuse measures bilaterally predict opioid poisoning. Some measures raise the rate of opioid poisoning incidents fivefold. The exception to our findings is our measure of excess out-of-state pharmacies; perhaps this measures is more indicative of diversion than misuse by the beneficiary himself.

3.2

Estimation Strategy

Our principal analysis tests the impact of a state policy on state-halfyear aggregates of opioid use, misuse, and poisoning incidents. Our regression model is

Yst = δs + δt + β1(policyst ) + εst ,

(1)

where Yst is an opioid outcome averaged over a state×halfyear and 1(policyst ) is 1 if the state has a particular type of PDMP in the halfyear. One outcome, the percentage of Part D enrollees using any opioids, is calculated using data from all enrollees in a given state and half-year. For all outcomes related to misuse or abuse, we use the number of enrollees with at least one opioid prescription as the denominator. Each observation is weighted by the number of enrollees represented in the denominator. Each regression contains fixed effects for states δs and years δt and clusters the standard errors at the state; we also report wild cluster-bootstrap percentile-t confidence intervals (Bertrand et al., 2004; Cameron et al., 2008). We use this basic setup to evaluate the impact of two types of PDMPs: those that do and do not require providers to access the data collected by the program. Seventeen states implemented a PDMP without a “must access” provision between 2007 and 2012; six others implemented a “must access” PDMP. Our main hypothesis is that PDMPs will have a stronger impact when providers are required to access the data. To test this, we first estimate the effect of PDMPs that do not have a “must access” provision, dropping the six “must access” states. Then using the full sample of states we estimate two models to evaluate the effect of the “must access” laws. The first model includes the indicator for non-“must access” laws as an independent variable. In the second specification, we combine the states with the PDMPs without a “must 7 We exclude heroin poisoning from this definition, although several authors have noted the substitutability of heroin and opioids (Dart et al., 2015; Unick et al., 2013). In results available upon request, we find no effect of a “must access” PDMP on heroin poisonings. We also exclude other ICD9 codes considered suggestive of an opioid poisoning incident in the medical literature (Jena et al., 2014; Hartung et al., 2007), such as Respiratory Failure (51881, 51882); Alteration of Consciousness (7800∗ ); and Malaise, Fatigue, or Lethargy (7808∗ ).

11

access” provision with states having no PDMP at all. In addition to our main analysis, which is based on state-year averages for all opioid users, we also analyze the same outcomes calculated for key subsamples of Medicare enrollees. As noted, it is generally believed that the problem of opioid abuse is most serious among lower-income Medicare beneficiaries and those who qualify on the basis of a disability. Therefore we conduct separate analyses by Medicare eligibility category. We also estimate separate regressions for state-level measures calculated for enrollees with and without a diagnosis of cancer, as indicated by their medical claims, and for enrollees who die or do not die in the half-year. The rationale for cutting the data this way is that an effective PDMP should reduce questionable prescribing without adversely affecting the treatment of patients with a legitimate need for pain relief. A finding that PDMP laws affected the prescription drug utilization of cancer patients or those near death would suggest either an unintended consequence of the policy or, perhaps, spurious correlation. Finally, we also estimate a set of “placebo” regressions that test for an effect of PDMP laws on the use or “abuse” of statins and antidepressants. A finding that “must access” PDMP affects these outcomes would suggest a problem with our research design.

3.3

Methodological Limitations

The use of Medicare claims data for this analysis brings several advantages: individual-level data allows us to measure extremes of the distribution, the large sample size provides many observations of rare outcomes, and the medical claims allow us to directly observe the linkage between opioid utilization and poisonings. However, the Part D program only began in 2006, several years into the escalation of opioid utilization. We therefore have only a short time period on which to assess state-specific trends in opioid misuse prior to PDMP implementation. We don’t observe outcomes involving prescribers or pharmacies in 2007, shortening the period even further for those outcomes. Other limitations of our dataset are inherent to administrative claims. Part D claims could undercount an individual’s utilization if they purchase opioids on the street. In addition, a Medicare beneficiary can fill opioid prescriptions using cash instead of Part D benefits, although over our sample period there was no reason to do so. This limitation is less likely to apply to opioid-related medical encounters, although there we rely on the accuracy of medical coding for fee-for-service beneficiaries. Finally, we only observe “successful” doctor- or pharmacy-shopping; if a PDMP makes any visit to a physician or pharmacy less likely to yield opioids, but Medicare beneficiaries counter by increasing their visits, we would observe no change in provider-shopping.8 8 In principle, we could look in the medical claims for an increase in medical encounters that do not yield opioid prescriptions, but there are several practical obstacles to this. Nothing in a medical claim denotes whether the encounter resulted in a prescription, or if so, which prescription. Instead, we must look for matches in the national provider identifiers (NPI) of a beneficiary’s prescribers and medical service providers. However, medical services are sometimes billed under group NPIs, while

12

4

Results

4.1

Effect of a PDMP Without a “Must Access” Provision on Opioid Misuse and Poisonings

In Table 6, we report the results of our difference-in-difference strategy for the PDMPs that do not require providers to access the database, dropping the six states that implement a “must access”’ PDMP from the analysis. Our point estimates indicate that a PDMP without a “must access” provision is associated with both increases and decreases in measures of misuse and poisonings; none of our point estimates can be distinguished from no effect. In order for a PDMP to provide useful information to a prescriber or pharmacist who checks it, the PDMP needs to first populate its database of utilization.9 Therefore, in the third panel of this table, we again test for the impact of a non-must access PDMP lagging the implementation date by six months. That is to say, we assume a newly-implemented PDMP is only effective six months after its actual implementation date. We again find that a PDMP without a “must access” provision is largely ineffective. Overall, the results in Table 6 are consistent with the opinion expressed by numerous experts that PDMP that rely entirely on the voluntary participation of providers are unlikely to be effective.

4.2

The Effect of “Must Access” PDMPs on Opioid Misuse and Poisonings

In Table 7, we report the results of estimating Equation 1 using a “must access” PDMP alone as our policy variable. The models reported in the upper panel include indicators for the presence of a non-“must access” PDMP. In effect, this analysis compares states with a “must access” provision against a two-part control group of all states with no changes in PDMPs and the seventeen states who implement a non-“must access” PDMP. Because we find no significant differences between these seventeen states and other states that did not implement a new PDMP during this period, our preferred model combines these two categories of states into a single control group. Results from this more parsimonious specification are reported in the lower panel of Table 7. For every estimate, we report in brackets the wild cluster-bootstrap percentile-t confidence interval, with CIs that exclude zero in bold (Cameron et al., 2008). Brewer et al. (2013) show that this bootstrap provides correct inference – hypothesis testing of correct size – in the presence of autocorrelated prescriptions are sometimes issued under a prescriber’s Drug Enforcement Agency (DEA) number instead of their NPI. As a result, many Part D prescriptions can’t be matched to the medical encounter that generated them. 9 Among our six states implementing a “must access” PDMP, only Delaware did not first have a non-“must access” PDMP which had already begun collecting data. Our estimates of the policy impact of a “must access” PDMP are robust to lagging, although lagging changes the states used to estimate the policy impact, since Louisiana and Nevada can now be used for all outcomes whereas Kentucky (who implements in the last half-year) is now excluded.

13

errors when the number of treated groups is as low as five. Our inference is unaffected by this procedure.10 The two models produce similar point estimates for the effect of a “must access” PDMP. In the model that controls for other types of PDMPs, a “must access” PDMP is associated with a 13% reduction in the number of beneficiaries obtaining opioids from five or more prescribers (-0.00212/0.0158 = -0.134). In the model with only an indicator for “must access” PDMPs, the percent effect is 12%. The implementation of a “must access” PDMP also appears to reduce the number of Medicare beneficiaries who fill prescriptions at five or more pharmacies in a six month period. In the upper panel, the estimated coefficient is -0.00117, with a p-value of 0.10. In the lower panel, the estimated coefficient is -0.00175 (p =0.004). Relative to mean of the dependent variable, which is 0.005, this is a 35% effect. In contrast to these results for doctor and pharmacy “shopping”, we find insignificant effects on our quantity-based measures. A “must access” PDMP does not statistically significantly affect obtaining more than 210 days of opioids, obtaining more than 120 MEDs/day, or the share of opioid claims that overlap. The passage of a “must access” PDMP may prompt individuals to cross state lines in search of less-regulated prescribers and pharmacies. In columns (6) and (7), we test whether the introduction of a “must access” PDMP in a state raises the rate at which the state’s residents obtain opioids from out-of-state providers. We find evidence consistent with that hypothesis for the case of out-of-state pharmacies: a “must access” PDMP raises the rate at which individuals fill excess claims at out-of-state pharmacies by about a third of a standard deviation. The point estimate for out-of-state prescribers is also positive and large relative to the sample mean, but the standard errors are also large, resulting in a statistically insignificant estimate. The final column tests whether a “must access” PDMP improves an important health outcome: opioid poisonings. We find a negative but statistically insignificant effect on poisonings. Table 8 repeats our baseline analysis for opioid use outcomes, rather than misuse. A “must access” PDMP does not affect the share of those taking opioids, nor does it affect the mean number of prescribers, mean number of pharmacies, or mean days supply among those taking opioids. These results suggest that a “must access” PDMP has few effects at the mean, which is consistent with a well-targeted policy that is not impacting appropriate access to opioid utilization. Researchers testing the effect of opioid policies on mean outcomes may mistakenly conclude that a policy has no effect when, instead, the effects are concentrated among misuse in the long right tail of utilization. 10 We also computed the pairs cluster-bootstrap percentile-t CIs. The pairs cluster assigns a given state’s outcomes to an alternative state’s policy history, obtaining the distribution of estimates under a null hypothesis of no effect. We find similar CIs using this procedure; results available upon request.

14

4.3

Model Assumptions: Analysis of Pretrends and Results in Each Implementing State

An assumption underlying our difference-in-difference strategy is that states that implement a “must access” PDMP were similar prior to implementation to those that do not. A standard way to test this assumption is to estimate Equation 1 using leads of the policy variable. If the states that go on to implement a PDMP are different prior to implementation, the leads of the policy variable will have statistically-significant policy impacts. Table 9 reports the coefficients on our “must access” indicator as well as its half-year and year leads. Our short-T panel presents a challenge to this analysis because Nevada and Louisiana implement their “must access” PDMPs at the end of 2007 and the beginning of 2008 respectively. Since our data only begin in 2007, we cannot use Nevada in estimating the half-year lead’s impact, and we cannot use Nevada or Louisiana in estimating the year lead’s impact. Therefore, for the quantity and poisoning outcomes (where N=612), the treatment states change in each row of the table. Nevertheless, we generally find the policy leads have insignificant impacts on quantity and poisoning outcomes. The only exception is for a very large days supply, on which we find no effect in our baseline analysis. For the prescriber and pharmacy outcomes observed 2008 to 2012, the same four states (Delaware, Kentucky, Ohio, and Oklahoma) identify all three coefficients in Table 9. Here, we are most concerned by the finding that those obtaining opioids from five or more pharmacies was falling a year prior to “must access” PDMP implementation. Figure 8 shows this variable over our time period for 45 non-implementing states (thick black line) and four implementing states (colored lines, dashed in the “post” period). It is clear from Figure 8 that this finding is driven by Delaware and is not present for our other implementing states. Figures 7 and Figure 9 show the other outcomes for which we find significant effects in our baseline analysis. We also test the impact of a “must access” PDMP singly in each state. If our analysis is truly isolating the impact of a “must access” PDMP, all six states should have a similar experience. Table 10 repeats our baseline analysis using, in each panel, 45 non-implementing states and one of the six implementing states. Standard errors clustered at the state level are in parentheses; confidence intervals adjusted for the impact of a single treatment are in brackets (Conley and Taber, 2011), with CIs that exclude zero in boldface. We cannot estimate a difference-in-difference coefficient for the prescriber and pharmacy outcomes for Louisiana and Nevada because we first observe those outcomes in 2008, when those states already had a “must access” PDMP. The results in this table suggest that, in spite of variations in exact implementation, a “must access” PDMP has similar effects in each implementing state. In Figure 10, we show coefficients for the three outcomes on which we find significant effects overall: five or more prescribers, five or more pharmacies, 15

and excess out-of-state pharmacies. Each figure shows the overall effect (blue dot) and the effects in each implementing state, with bars representing the confidence intervals from clustered standard errors. In most cases the effects in implementing states cannot be distinguished from the overall effect. However, these charts also highlight the narrowness of the CIs in the single-state analyses, which could be driven by the bias towards over-rejection documented by (Conley and Taber, 2011). Reexamining our findings using their adjusted CIs, we find mostly null results; we can infer only that Delaware’s “must access” PDMP raised the rate at which Delaware residents obtained opioids from out-of-state prescribers and pharmacies, while Kentucky’s had a similar effect on out-of-state prescribers.

4.4

Results by Subsample

Table 11 tests the impact of a “must access” PDMP on aggregate outcomes within four Medicare subsamples. We find that the effects of “must access” PDMPs are concentrated among the disabled, and particularly the low-income disabled. Consistent with the literature reviewed in Section 2, the disabled are more likely to have utilization patterns that suggest “misuse” and have a much higher rate of opioid poisoning incidents. It’s possible that a “must access” PDMP has larger effects in the low-income disabled because the policy is well-targeted towards the type of “misuse” that is more common among this subpopulation. Alternatively, a “must access” PDMP may simply have a larger impact among the low-income disabled population. If physicians are more likely to suspect “misuse” among the low-income disabled, then a “must access” provision may result in more PDMP checks for this group than for elderly opioid users. Figure 11 illustrates both the difference in means and policy impact among the four eligibility subsamples for the three outcomes where we find effects in the general population. In each figure, the blue bar represents the mean of the variable in the subsample, and the red bar shows our “post” prediction. Table 12 describes the impact of a “must access” PDMP on subsamples defined by health status. An unintended consequence of a “must access” PDMP might be a “chilling effect” on appropriate opioid utilization. The top two panels of the table report the effect of a “must access” PDMP on state-halfyear aggregates of cancer patients and those without cancer. A first note is that our measures of opioid “misuse” are not markedly higher among cancer patients, suggesting we are not mislabeling high but appropriate utilization too often. The table shows that a “must access” PDMP reduces measures of misuse among those without cancer, without impeding the utilization of cancer patients. In fact, a “must access” PDMP is associated with high quantities (measured either via days supply or MED) among cancer patients, perhaps because physicians are more comfortable prescribing higher quantities when they are certain of the absence of misuse. The second two panels consider two different subsamples: those who die in the half-year versus those

16

who do not. Again, misuse measures are surprisingly similar among the subsamples.11 Our results demonstrate that the effect of a “must access” PDMP is driven by those without cancer and not near the end of life. However, we do find that the provision somewhat reduces the rate of dying individuals obtaining opioids from multiple prescribers or pharmacies. One possible explanation is that some of those individuals die as a result of opioid misuse (including misuse with suicidal intent), and that the provision appropriately impedes their access. In results available upon request, we find no association between a “must access” PDMP and all-cause mortality.

4.5

Placebo Test: “Must Access” PDMPs’ Effect on Non-opioids

If a “must access” PDMP is correlated for some reason with general trends in prescription drug utilization or prescribing patterns, we will incorrectly attribute the change in opioid misuse measures to the “must access” PDMP. As a placebo test, Table 13 looks at the impact of a “must access” PDMP on outcomes for statins and antidepressants. We find no significant effects of a “must access” PDMP on outcomes for drug types besides opioids.

5

Conclusion

In this paper, we provide the first empirical evaluation of state laws that require prescribers to access data collected by Prescription Drug Monitoring Programs. Several key conclusions emerge from our analysis. First, our results suggest that PDMPs that do not require provider participation are not effective in reducing questionable or inappropriate use of prescription opioids. In contrast, we do find evidence that “must access” PDMPs do have the desired effect of curbing certain types of extreme utilization. Specifically, such policies appear to reduce the prevalence of “doctor shopping” and “pharmacy shopping”. Several additional results suggest that these estimates represent causal effects. First, the point estimates are broadly similar for all states that implemented “must access” laws during our study period. Second, the effects are concentrated among low-income disabled enrollees, who have the highest rates of misuse measures. In contrast, we find no unintended consequences for utilization among cancer patients, among whom opioid abuse is less likely to be a problem. Although these results point to the effectiveness of “must access” PDMPs, others suggest their limits. We find no statistically significant effect on a key medical outcome: opioid poisoning incidents. There are several interpretations of these findings. Perhaps Medicare beneficiaries who are misusing opioids are able to maintain consumption and therefore do not have fewer poisonings. We find evidence that beneficiaries are turning to out-of-state sources (perhaps locating criminal or negligent out-of-state sources), but they may 11 We make no adjustment for the amount of time in the half-year the patient was alive. If utilization were identical, means of measures would still be lower among those who die in the half-year due to censoring.

17

also substitute towards street sources of opioids. In addition, Medicare beneficiaries may be engaging in less opioid diversion but not actually less misuse. Another possibility is that a must access PDMP reduces the rate at which individuals become opioid users, delaying the effects on poisonings beyond our sample period. To our knowledge, no evaluation yet exists of the changes CMS made in 2013 to opioid access policies in Part D; however, our research suggests that CMS should consider implementing some sort of “must access” policy among Part D prescribers (pursuing legislation if the mandate to combat “waste, fraud, and abuse” cannot enable this step.) A collaboration between CMS and state PDMPs could allow states access to CMS’s data resources, allowing improved identification of suspicious patterns of prescribing and utilization and smoother sharing of information across state lines.

18

References Arditi, Lynn, “New law: Health-care providers must register in prescription database,” Providence Journal, 2014. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, “How Much Should We Trust Differences-In-Differences Estimates,” The Quarterly Journal of Economics, 2004, 119 (1). Blum, Jonathan, “Curbing Prescription Drug Abuse in Medicare, Testimony before U.S. Senate Committee on Homeland Security and Government Affairs,” Technical Report, Director, Center for Medicare Management, Washington, DC June 2013. Available at http://www.hhs.gov/asl/testify/2013/06/4483.html. Bourgeois, Florence T., Michael W. Shannon, Clarissa Valim, and Kenneth D. Mandl, “Adverse Drug Events in the Outpatient Setting: An 11-year National Analysis,” Pharmacoepidemiology and Drug Safety, September 2010, 19 (9). Brady, Joanne E., Hannah Wunsch, Charles J. DiMaggio, Barbara H. Lang, James Giglio, and Guohua Li, “Prescription drug monitoring and dispensing of prescription opioids,” Public Health Rep, 2014, 129. Brewer, Mike, Thomas F. Crossley, and Robert Joyce, “Inference with Difference-in-Differences Revisited,” November 2013. IZA Discussion Paper No. 7742. Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller, “Bootstrap-Based Improvements for Inference with Clustered Errors,” The Review of Economics and Statistics, 2008, 90 (3). Chen, Li Hui, Holly Hedegaard, and Margaret Warner, “Drug-poisoning Deaths Involving Opioid Analgesics: United States, 1999–2011,” Technical Report, National Center for Health Statistics, Washington, DC September 2014. NCHS Data Brief No. 166. CMS,

“Opioid

Morphine

Equivalent

Conversion

Factors,”

https://www.cms.gov/

Medicare/Prescription-Drug-Coverage/PrescriptionDrugCovContra/Downloads/ Opioid-Morphine-EQ-Conversion-Factors-March-2015.pdf 2015. Accessed: February 2, 2016. Coben, Jeffrey H., Stephen M. Davis, Paul M. Furbee, Rosanna D. Sikora, Roger D. Tillotson, and Robert M. Bossarte, “Hospitalizations for Poisoning by Prescription Opioids, Sedatives, and Tranquilizers,” American Journal of Preventive Medicine, May 2010, 38 (5). Conley, Timothy G. and Christopher R. Taber, “Inference with “Differences in Differences” with a Small Number of Policy Changes,” The Review of Economics and Statistics, 2011, 93 (1), 113–125. 19

Dart, Richard C., Hilary L. Surratt, Theodore J. Cicero, Mark W. Parrino, Geoff Severtson, Becki Bucher-Bartelson, and Jody L. Green, “Trends in Opioid Analgesic Abuse and Mortality in the United States,” New England Journal of Medicine, 2015, 372 (3). Davis, Corey S., Matthew Pierce, and Nabarun Dasgupta, “Evolution and Convergence of State Laws Governing Controlled Substance Prescription Monitoring Programs, 1998-2011,” American Journal of Public Health, August 2014, 104. GAO, “State Monitoring Programs Provide Useful Tool to Reduce Diversion,” Technical Report, Government Accountability Office, Washington, DC May 2002. GAO-02-634. , “Instances of Questionable Access to Prescription Drugs,” Technical Report, Government Accountability Office, Washington, DC September 2011. GAO-11-699. Gourlay, Kristin, “Fighting Prescription Drug Abuse, One Log In At a Time,” Rhode Island Public Radio, 2013. Haegerich, Tamara M., Leonard J. Paulozzi, Brian J. Manns, and Christopher M. Jones, “What we know, and dont know, about the impact of state policy and systems-level interventions on prescription drug overdose,” Drug and Alcohol Dependence, 2014, 145, 34 – 47. Haffajee, Rebecca L., Anupam B. Jena, and Scott G. Weiner, “Mandatory Use of Prescription Drug Monitoring Programs,” JAMA, 2015, 313 (9), 891–892. Hartung, Daniel M., Luke Middleton, Dean G. Haxby, Michele Koder, Kathy L. Ketchum, and Roger Chou, “Rates of Adverse Events of Long-Acting Opioids in a State Medicaid Program,” Annals of Pharmacotherapy, June 2007, 41. Islam, M. Mofizul and Ian S McRae, “An Inevitable Wave of Prescription Drug Monitoring Programs in the Context of Prescription Opioids: Pros, Cons and Tensions,” BMC Pharmacology & Toxicology, 2014, 15 (46). Jena, Anupam B., Dana Goldman, Lesley Weaver, and Pinar Karaca-Mandic, “Opioid Prescribing By Multiple Providers in Medicare: Retrospective Observational Study of Insurance Claims,” British Medical Journal, 2014, 348. Kunins, Hillary V., Thomas A. Farley, and Deborah Dowell, “Guidelines for opioid prescription: why emergency physicians need support,” Annals of Internal Medicine, 2013, 158.

20

Levy, Benjamin, Leonard Paulozzi, Karin A. Mack, and Christopher M. Jones, “Trends in Opioid AnalgesicPrescribing Rates by Specialty, U.S., 20072012,” American Journal of Preventive Medicine, 2015, 49 (3), 409 – 413. Li, Guohua, Joanne E. Brady, Barbara H. Lang, James Giglio, Hannah Wunsch, and Charles DiMaggio, “Prescription drug monitoring and drug overdose mortality,” Injury Epidemiology, 2014, 1 (1), 1–8. Manchikanti, Laxmaiah, Bert Fellows, Hary Ailinani, and Vidyasagar Pampati, “Therapeutic use, abuse, and nonmedical use of opioids: a ten-year perspective,” Pain Physician, February 2010, 13. Morden, Nancy E., Jeffrey C. Munson, Carrie H. Colla, Jonathan Skinner, Julie P.W. Bynum, Weiping Zhou, and Ellen Meara, “Prescription Opioid Use Among Disabled Medicare Beneficiaries: Intesity, Trends, and Regional Variation,” Medical Care, September 2014, 52 (9). NSDUH, “National Survey on Drug Use and Health: Summary of National Findings,” September 2014. Available at http://www.samhsa.gov/data/sites/default/files/NSDUHresultsPDFWHTML2013/ Web/NSDUHresults2013.htm; some results from authors’ tabulation at http://www.icpsr.umich.edu/ icpsrweb/SAMHDA/studies/35509/datasets/1/sdaxml. Ohio Bureau of Workers’ Compensation, “MED Table,” https://www.bwc.ohio.gov/downloads/ blankpdf/MEDTable.pdf 2016. Accessed: February 2, 2016. Palliative.org, “INSTRUCTIONS FOR MORPHINE EQUIVALENTDAILY DOSE (MEDD),” http:// palliative.org/NewPC/_pdfs/tools/INSTRUCTIONsMEDD.pdf 2016. Accessed: February 2, 2016. Pasero, Christine L. and Margo McCaffery, “Pain Ratings: The Fifth Vital Sign,” American Journal of Nursing, February 1997, 97. Paulozzi, Leonard J., Edwin M. Kilbourne, and Hema A. Desai, “Prescription Drug Monitoring Programs and Death Rates from Drug Overdose,” Pain Medicine, May 2011, 12 (5), 747–54. PDMP Center of Excellence, “Mandating PDMP participation by medical providers: current status and experience in selected states,” Technical Report, Brandeis University, Waltham, MA October 2014. Available at http://www.pdmpexcellence.org/sites/all/pdfs/COE_briefing_mandates_2nd_rev.pdf. Poston, Rebecca, “E-FORCSE 2011–2012 Prescription Drug Monitoring Program Annual Report,” December 2012.

Available at http://www.floridahealth.gov/statistics-and-data/e-forcse/

news-reports/_documents/2011-2012pdmp-annual-report.pdf. 21

Reifler, Liza M., Danna Droz, J. Elise Bailey, Sidney H. Schnoll, Reginald Fant, Richard C. Dart, and Becki Bucher-Bartelson, “Do Prescription Monitoring Programs Impact State Trends in Opioid Abuse/Misuse?,” Pain Medicine, 2012, 13. Unick, George Jay, Daniel Rosenblum, Sarah Mars, and Daniel Ciccarone, “Intertwined Epidemics: National Demographic Trends in Hospitalizations for Heroin- and Opioid-Related Overdoses, 1993-2009,” PLOS One, February 2013, 8 (2). Volkow,

Nora D., “Testimony before the Senate Caucus on International Narcotics Control

on Americas Addiction to Opioids:

Heroin and Prescription Drug Abuse,”

2014.

Avail-

able at www.drugabuse.gov/about-nida/legislative-activities/testimony-to-congress/2015/ americas-addiction-to-opioids-heroin-prescription-drug-abuse. Washington State Agency Medical Directors’ Group, “Interagency guideline on opioid dosing for chronic non-cancer pain: an educational aid to improve care and safety with opioid treatment,” http: //www.agencymeddirectors.wa.gov/files/opioidgdline.pdf 2010. Accessed: February 5, 2016. West, Nancy A., Stevan G. Severtson, Jody L. Green, and Richard C. Dart, “Trends in Abuse and Misuse of Prescription Opioids Among Older Adults,” Drug and Alcohol Dependence, 2015, 149. White tor 2015.

House, Efforts

“Obama

to

Administration

Address

Available

Prescription at

Announces

Drug

Abuse

Public and

and

Heroin

Private Use,”

SecOctober

https://www.whitehouse.gov/the-press-office/2015/10/21/

fact-sheet-obama-administration-announces-public-and-private-sector. Wisconsin Prescription Drug Monitoring Program, “Training Guide for Wisconsin Practitioners and Pharmacists,” http://dsps.wi.gov/Documents/PDMP/Practitioners_Pharmacists_Training_Guide. pdf October 2013. Accessed: February 2, 2016.

22

9 From this window, you may perform the following functions:

pharmacists. (Wisconsin Prescription Drug Monitoring Program, 2013)

Accessing PDMP Data

Note: Your search criteria and the recipient names you selected are located above Figure 1: Wisconsin PDMP Training Screenshot your report. You may click the down arrow in the Recipients field to view the The Wisconsin Department Safety and ProfessionalinServices patients you of chose to include your provided report.this screenshot in an October 2013 training manual for prescribers and

Your report results are displayed similar to the following:

Wisconsin Department of Safety and Professional Services - Pharmacy Examining Board Training Guide for Wisconsin Practitioners and Pharmacists

23

2008

2009

2010

2011

2012

23

Opioid Prescriptions Poisoning Incidents 2007

2008

2009

2010

2011

40

Opioid Prescriptions (millions) 24 26 25

50 45 Poisoning Incidents (thousands)

27

55

2007

300

350 450 400 Poisoning Incidents (thousands)

500

220 Opioid Prescriptions (millions) 190 200 210 180

Opioid Prescriptions Poisoning Incidents

2012

Figure 2: Opioid Utilization and Poisoning, Overall (top) and in Medicare Part D (bottom) Overall opioid prescriptions collected from Volkow (2014). Overall poisoning incidents, which reflect emergency department visits alone, are collected from the Drug Abuse Warning Network Emergency Department Data for 2007 to 2011 and extrapolated to 2012 (the dataset was discontinued in 2011). Medicare opioid prescriptions and poisoning incidents, based off of inpatient, outpatient, and physician claims, are from a random sample of enrollees in free-standing Part D and fee-for-service Medicare in 2007-2012.

24

State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming

Table 1: State Laws First PDMP Law Implemented “Must Access” Law Implemented August, 2004 – September, 2008 – September, 2007 – July, 2011 – January, 1998 – June, 2005 – October, 2006 – January, 2011 January, 2011 – – July, 2009 – July, 2011 – January, 1998 – April, 2000 – April, 2000 – January, 1998 – May, 2006 – July, 2008 – July, 1998 July, 2012 April, 2007 January, 2008 September, 2003 – October, 2011 – January, 1998 – January, 2002 – July, 2007 – June, 2006 – – – July, 2011 – August, 2011 – November, 2002 October, 2007 June, 2012 – August, 2009 – July, 2004 – May, 2001 – January, 2006 – April, 2007 – August, 2008 November, 2011 June, 2004 November, 2010 July, 2009 – January, 1998 – January, 1998 – June, 2006 – July, 2010 – January, 2003 – January, 1998 – January, 1998 – July, 2006 – April, 2003 – July, 2007 – January, 1998 – June, 2010 – July, 2003 –

This table reports on laws that take effect between 1998 and 2012. Laws in effect in January 1998 are listed as beginning that month.

Table 2: Summary Statistics: Beneficiaries Share taking opioids Share dually eligible for Medicaid Share entitled to Medicare due to disability Share with cancer Share who die this half-year Number of person X half-years Mean number of half-years observed

Full Sample 0.175 0.410 0.245 0.095 0.053 11,256,940 9.639

Opioid Takers 1.000 0.522 0.358 0.125 0.062 1,965,499 5.661

Full sample are individuals enrolled in Part D and fee-for-service Medicare (not Medicare Advantage) in the period 2007–2012. Opioid takers are those filling at least one opioid prescription.

Table 3: Summary Statistics: Opioid Outcomes Number of Prescribers 5+ Prescribers Number of Pharmacies 5+ Pharmacies Total Days Supply 211+ Days Supply Mean Daily MED 120+ Daily MED Share with Claims Overlapping Excess Share of Claims from OOS Prescriber Excess Share of Claims from OOS Pharmacy Rate of Opioid Poisoning Incidents

Mean 1.475 0.016 1.222 0.005 81.556 0.078 12.246 0.025 0.088 0.009 -0.027 0.007

Median 1.0

99th percentile 5.0

1.0

4.0

35.0

420.0

1.3

193.3

0.0 0.0

0.9 0.5

Outcomes reported for opioid takers sample described in Table 2. “OOS” signifies out-of-state and “MED” signifies morphine equivalent dosage. See text for definition of morphine-equivalent dosage measures, share of claims overlapping, and excess OOS claims.

26

1.5e+04

.8

Frequency 1.0e+04

.6

5000

Fraction .4

0

.2 0 0

10

20

30

40

50

0

10

20

30

40

50

0

0

1000

.2

Fraction .4

Frequency 2000 3000

.6

4000

.8

5000

Figure 3: Number of prescribers of opioids, each bar=1,vertical line at 5 prescribers. Right figure depicts the frequency for only 5+ prescribers.

0

5

10

15

20

25

5

10

15

20

25

0

0

.1

1.0e+04

Fraction .2

Frequency 2.0e+04

.3

.4

3.0e+04

Figure 4: Number of pharmacies for opioids, each bar=1, vertical line at 5 pharmacies. Right figure depicts the frequency for only 5+ pharmacies.

0

500

1000

1500

2000

0

500

1000

1500

2000

Figure 5: Days supply of opioids in a half-year, each bar=30 days, vertical line at 211 days. Right figure depicts the frequency for only 211+ days. 27

5000 4000

.8

Frequency 2000 3000

.6 Fraction .4

0

1000

.2 0 0

200

400

600

800

1000

200

400

600

800

1000

Figure 6: Daily Morphine-Equivalent Dosage obtained in a half-year, each bar=10 daily MED, vertical line at 120 MED each day in the half-year. Right figure depicts the frequency for only 120+ daily MED.

Table 4: Correlations Among Measures of Misuse (1) (2) (3) (4) (5) (6) (7) (8)

5+ Prescribers 5+ Pharmacies 211+ Days Supply 120+ Daily MED Overlapping Claims Excess OOS Prescribers Excess OOS Pharmacies Opioid Poisonings

(1) 1 0.2670 0.1625 0.0559 0.1597 0.0077 0.0127 0.0313

(2)

(3)

(4)

(5)

(6)

(7)

1 0.1402 0.0619 0.1172 0.0043 0.0127 0.0234

1 0.3652 0.4473 0.0006 0.0288 0.0267

1 0.2168 -0.0011 0.0142 0.0174

1 -0.0030 0.0362 0.029

1 0.1784 0.0005

1 0.0044

Pairwise correlations for measures of misuse and opioid poisoning incidents. Means for each variable available in Tables 2 and 3. Bold text signifies that the correlation differs from zero at the 0.001 level.

28

29 1,955,031

0.00791*** (0.000483)

1,955,031

0.00755*** (0.000252)

1,636,876

0.00933*** (0.000636) 0.00455*** (0.000687) 0.00231*** (0.000223) 0.00171*** (0.000295) 0.00344*** (0.000230) -0.00032 (0.000295) 0.00162*** (0.000288)

(8) Poisonings

This table reports the marginal effects of probits predicting an opioid poisoning incident in a half-year using contemporaneous measures of misuse among opioid takers enrolled in Part D and fee-for-service Medicare (not Medicare Advantage) in the period 2007–2012. The number of observations varies because in 2007 we do not observe prescribers and pharmacies, and in any year a few prescribers and pharmacies cannot be reliably assigned to states.

1,636,925

0.00012 (0.000306)

1,636,909

1,955,031

0.00752*** (0.000268)

Observations

1,643,982

0.02665*** (0.001905)

0.00221*** (0.000299) 1,643,982

0.02021*** (0.000947)

(7) Poisonings

Excess OOS Pharmacies

Excess OOS Prescribers

Overlapping Claims

120+ Daily MED

211+ Days Supply

5+ Pharmacies

5+ Prescribers

VARIABLES

Misuse measures and opioid poisonings, Sample mean = 0.0075 (1) (2) (3) (4) (5) (6) Poisonings Poisonings Poisonings Poisonings Poisonings Poisonings

Table 5: Probits on Opioid Poisoning Incidents

30

0.1800 540

-0.000551 (0.0007)

0.1800 540

1.4770 450

0.0007 (0.0004)

1.4770 450

5+ Pharmacies 0.000501 (0.0004)

1.2200 450

0.00243 (0.0030)

1.2200 450

211+ Days Supply 0.00303 (0.0028)

81.1800 540

-0.000545 (0.0007)

81.1800 540

120+ Daily MED -0.000191 (0.0009)

0.0161 450

0.00126 (0.0012)

0.0161 450

Overlapping Claims 0.000828 (0.0014)

0.0049 450

-0.000166 (0.0008)

0.0049 450

Excess OOS Prescribers 0.000149 (0.0007)

0.0777 540

-0.00127 (0.0043)

0.0777 540

Excess OOS Pharmacies -0.000149 (0.0036)

0.0255 540

0.000524 (0.0009)

0.0255 540

0.00125 (0.0007)

Poisonings

This table reports the impact of a PDMP without a “must access” provision. The top panel estimates Equation 1 using a binary measure of whether a PDMP without a “must access” provision is in place, excluding from the analysis the six states with a “must access” PDMP. The bottom panel repeats the top with the implementation of a non-must access PDMP lagged by six months, e.g. a PDMP implemented in January 2008 is said to come into effect in July 2008. Each observation is weighted by the number of opioid takers represented. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Analyses with 540 observations represent 45 states between 2007h1 and 2012h2. Outcomes involving prescribers and pharmacies are unavailable for 2007.

Mean of LHS N

Lagged Non-“Must Access” (excl “Must”)

Mean of LHS N

Non-“Must Access” (excl “Must”)

5+ Prescribers -0.000187 (0.0008)

Table 6: Effects of Non-“Must Access” PDMPs

0.0050 510

0.0777 612

[ -0.0065 , 0.0039 ]

-0.00118 (0.0019)

0.0777 612

0.00253 (0.0027)

0.00157 (0.0035)

211+ Days Supply

0.0253 612

[ -0.0010 , 0.0026 ]

0.000706 (0.0007)

0.0253 612

-0.000223 (0.0008)

0.000462 (0.0012)

120+ Daily MED

0.0880 612

[ -0.0208 , 0.0136 ]

-0.00389 (0.0037)

0.0880 612

0.000626 (0.0014)

-0.00321 (0.0041)

Overlapping Claims

0.0090 510

[ -0.0163 , 0.0226 ]

0.0027 (0.0029)

0.0090 510

0.000154 (0.0007)

Excess OOS Prescribers 0.00286 (0.0030)

-0.0268 510

[ 0.0022 , 0.0095 ]

0.00595** (0.0019)

-0.0268 510

-0.000245 (0.0035)

0.00569 (0.0033)

Excess OOS Pharmacies

0.0075 612

[ -0.0029 , 0.0011 ]

-0.000937 (0.0009)

0.0075 612

0.00126 (0.0006)

0.000434 (0.0011)

Poisonings

This table estimates the impact of a “must access” PDMP on opioid misuse and poisonings in Medicare Part D. In the top panel, we estimate Equation 1 with two independent variables: whether a PDMP with a “must access” provision is in effect and whether a PDMP without a “must access” provision is in effect. The bottom panel, our baseline specification, uses “must access” alone. Each column is the result of OLS regression on state-halfyear aggregates of opioid takers; each observation is weighted by the number of opioid takers represented. Fixed effects for states and halfyears are always included, and standard errors (in parentheses) are clustered at the state. In brackets, we report the wild cluster-bootstrap percentile-t CI, with CIs that exclude zero in bold. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Analyses with 612 observations represent 51 states between 2007h1 and 2012h2. Outcomes involving prescribers and pharmacies are unavailable for 2007.

0.0158 510

[ -0.0029 , -0.0005 ]

[ -0.0034 , -0.0004 ]

Mean N

-0.00175** (0.0006)

-0.00191** (0.0006)

“Must Access” BS CI

0.0050 510

0.000554 (0.0004)

0.0158 510

-0.000205 (0.0007)

Non “Must Access”

-0.00117 (0.0007)

5+ Pharmacies

Mean N

-0.00212* (0.0010)

“Must Access”

5+ Prescribers

Table 7: Effect of “Must Access” PDMP Law on Opioid Misuse and Poisonings

32

Mean prescribers -0.0077 (0.0073) 1.475 510

Mean pharmacies -0.0046 (0.0052) 1.222 510

Mean days supply 0.0805 (1.1980) 81.56 612

This table reports our baseline results on the impact of a “must access” PDMP on opioid use and other outcomes in Medicare Part D. Each column is the result of OLS regression on state-halfyear aggregates of opioid takers. The first regression is weighted by the number of enrollees, while all others are weighted by the number of opioid takers represented. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Analyses with 612 observations represent 510 states between 2007h1 and 2012h2. Outcomes involving prescribers and pharmacies are unavailable for 2007.

“Must Access” PDMP

P(taking opioids) -0.0033 (0.0026) 0.183 612

Table 8: Effects of “Must Access” PDMP Law on Average Opioid Use

33 510 0.0158

-0.00169* (0.0008) 0.00021 (0.0011) -0.00056 (0.0009) 510 0.0050

5+ Pharmacies -0.00006 (0.0005) -0.00038 (0.0006) -0.00198* (0.0009) 612 0.0777

211+ Days Supply -0.00196 (0.0013) -0.00255** (0.0008) 0.00456 (0.0030) 612 0.0253

120+ Daily MED -0.000113 (0.0003) -0.00020 (0.0005) 0.00160 (0.0016) 612 0.0880

Overlapping Claims -0.00552** (0.0019) 0.00174 (0.0041) 0.00042 (0.0032) 510 0.0090

Excess OOS Prescribers 0.00429 (0.0031) -0.00519* (0.0021) 0.00401 (0.0022)

510 -0.0268

Excess OOS Pharmacies 0.00597*** (0.0017) -0.00156 (0.0024) 0.00187 (0.0028)

612 0.0075

-0.00081 (0.0006) -0.00055 (0.0005) 0.00050 (0.0014)

Poisonings

This table analyzes whether states that implement a “must access” PDMP differed from those that did in not prior to implementation. Each column is the result of OLS regression on state-halfyear aggregates of opioid takers, weighted by the number of opioid takers represented. The first row reports the estimate of contemporaneous effects; the second and third rows report the estimate of a half-year and year lead of the treatment variable. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Analyses with 612 observations represent 510 states between 2007h1 and 2012h2. Outcomes involving prescribers and pharmacies are unavailable for 2007.

N mean of LHS

Year Lead

Half-Year Lead

“Must Access” PDMP

5+ Prescribers

Table 9: Pretrends Analysis: Effects of “Must Access” PDMP Prior to its Implementation

.04

Ohio Oklahoma

0

.01

5+ Prescribers .02

.03

Kentucky Delaware

2008h1

2009h1

2010h1

2011h1

2012h1

.02

Figure 7: 5+ Prescribers in four “Must Access” states, 2008-2012. Thick black line is 45 states without a “must access” PDMP; LA & NV (who implement prior to 2008) are excluded. Dashes represent “post”.

Ohio Oklahoma

0

.005

5+ Pharmacies .01

.015

Kentucky Delaware

2008h1

2009h1

2010h1

2011h1

2012h1

Figure 8: 5+ Pharmacies in four “Must Access” states, 2008-2012. Thick black line is 45 states without a “must access” PDMP; LA & NV (who implement prior to 2008) are excluded. Dashes represent “post”. 34

-.02 Excess OOS Pharmacy -.06 -.04 -.08 -.1

Kentucky Delaware

2008h1

2009h1

2010h1

2011h1

Ohio Oklahoma 2012h1

Figure 9: Excess out-of-state pharmacies in four “Must Access” states, 2008-2012. Thick black line is 45 states without a “must access” PDMP; LA & NV (who implement prior to 2008) are excluded. Dashes represent “post”.

35

-0.00174*** (0.0002) [ -0.0050 , 0.0029 ]

-0.00215*** (0.0005)

[ -0.0095 , 0.0068 ]

[ -0.0052 , 0.0011 ]

[ -0.0013 , 0.0065 ]

[ -0.0146 , 0.0045 ]

[ -0.0125 , 0.0101 ]

0.00110*** (0.0002)

-0.00304*** (0.0005)

-0.00239*** (0.0002)

[ -0.0067 , 0.0012 ]

[ -0.0106 , 0.0057 ]

-0.00108** (0.0004)

-0.00345*** (0.0002)

-0.00325*** (0.0005)

[ -0.0143 , 0.0204 ]

0.00263* (0.0013)

[ -0.0208 , 0.0227 ]

-0.00344* (0.0014)

[ -0.0150 , 0.0324 ]

0.00270* (0.0012)

[ -0.0161 , 0.0251 ]

-0.00431*** (0.0012)

[ -0.0146 , 0.0292 ]

0.00357* (0.0014)

[ -0.0179 , 0.0168 ]

-0.000943 (0.0013)

211+ Days Supply

Overlapping Claims

[ -0.0377 , 0.0136 ]

[ -0.0094 , 0.0282 ]

[ -0.0125 , 0.0266 ]

[ -0.0297 , 0.0145 ]

[ -0.0295 , 0.0085 ]

[ -0.0067 , 0.0123 ]

[ -0.0225 , 0.0288 ]

Oklahoma Only 0.00195*** 0.00501*** (0.0005) (0.0008)

[ -0.0085 , 0.0112 ]

Ohio Only -0.000214 -0.0103*** (0.0005) (0.0009)

[ -0.0097 , 0.0157 ]

Nevada Only -0.00113* -0.0160*** (0.0005) (0.0010)

[ -0.0082 , 0.0079 ]

Louisiana Only -0.000391 -0.00156 (0.0005) (0.0008)

[ -0.0095 , 0.0179 ]

Kentucky Only 0.00168** 0.00135 (0.0005) (0.0009)

[ -0.0054 , 0.0136 ]

Delaware Only 0.00331*** -0.0101*** (0.0005) (0.0008)

120+ Daily MED

[ -0.0031 , 0.0168 ]

0.00591*** (0.0004)

[ -0.0115 , 0.0138 ]

-0.00206*** (0.0004)

[ 0.0009 , 0.0289 ]

0.00780*** (0.0005)

[ 0.0022 , 0.0220 ]

0.0111*** (0.0004)

Excess OOS Prescribers

[ -0.0022 , 0.0301 ]

0.00640*** (0.0015)

[ -0.0067 , 0.0406 ]

0.00486* (0.0019)

[ -0.0048 , 0.0422 ]

0.00723** (0.0022)

[ 0.0027 , 0.0350 ]

0.0112*** (0.0015)

Excess OOS Pharmacies

[ -0.0095 , 0.0092 ]

-0.00333*** (0.0004)

[ -0.0106 , 0.0085 ]

-0.000137 (0.0006)

[ -0.011 , 0.0093 ]

-0.00141** (0.0004)

[ -0.0081 , 0.0092 ]

-0.00107*** (0.0003)

[ -0.0142 , 0.0132 ]

0.000347 (0.0007)

[ -0.0029 , 0.0157 ]

0.00325*** (0.0004)

Poisonings

This table repeats our baseline analysis using, in each panel, only one of the six states that implement a “must access” PDMP between 2007 and 2012 (excluding the five other states). Each observation is weighted by the number of individuals represented. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. Brackets report CIs adjusted for a single treatment in the manner of Conley and Taber (2011), and CIs that exclude zero are in boldface. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Outcomes involving prescribers and pharmacies are unavailable for 2007, which means we cannot estimate a difference-in-difference coefficient for those outcomes in Lousiana and Nevada (which implement the policy in 2008h1.)

“Must Access” PDMP

“Must Access” PDMP

“Must Access” PDMP

“Must Access” PDMP

“Must Access” PDMP

“Must Access” PDMP

5+ Pharmacies

5+ Prescribers

Table 10: Effect of “Must Access” PDMP Law: Each State Individually

0 -.001 -.002

-.0011

-.0019

-.003

-.0022

-.003 -.0032

-.004

5+ Prescribers Mean .0158 Must Access KY only OH only

DE only

OK only

.002

All

0

.0011

-.0017

-.002

-.0018 -.0024

-.0034

-.004

5+ Pharmacies Mean .005

Must Access KY only OH only

DE only

OK only

.015

All

.01

.011

.0072 .0064

.005

.0059 .0049

0

Excess OOS Pharmacies Mean -.0268 All

DE only

Must Access KY only OH only

OK only

Figure 10: Estimated effect sizes for 5+ Prescribers (top), 5+ Pharmacies (middle), and Excess OOS Pharmacies (bottom) overall and for each state implementing a “must access” PDMP. 37

38

-0.00125* (0.0005) 0.00509

“Must Access” PDMP mean of LHS

-0.0000509 (0.0002) 0.000597

-0.000365 (0.0005) 0.00129

-0.00258* (0.0011) 0.00847

-0.00480** (0.0017) 0.0133

5+ Pharmacies

0.00196 (0.0030) 0.00956

0.00229 (0.0020) 0.00465

0.00323 (0.0039) 0.0115

Elderly & Dual Eligible 0.00107 -0.000345 0.00381* (0.0022) (0.0015) (0.0017) 0.0499 0.0214 0.0738 Elderly & Not Dual Eligible 0.00535** 0.00163 -0.00386 (0.0018) (0.0011) (0.0023) 0.0338 0.0105 0.0600

Excess OOS Prescribers

Disabled & Not Dual Eligible 0.00198 0.00142 0.0038 (0.0114) (0.0047) (0.0084) 0.155 0.0446 0.1310

Overlapping Claims

0.00227 (0.0054) 0.00926

120+ Daily MED

Disabled & Dual Eligible -0.0171*** -0.000637 -0.0135* (0.0029) (0.0016) (0.0053) 0.141 0.0435 0.1270

211+ Days Supply

0.00753 (0.0038) -0.0529

-0.00145 (0.0027) -0.00804

0.00715 (0.0042) -0.0235

0.00558*** (0.0009) -0.00859

Excess OOS Pharmacies

0.000533 (0.0014) 0.00271

-0.00138 (0.0013) 0.00475

-0.00459 (0.0034) 0.0108

-0.00103 (0.0022) 0.0158

Poisonings

This table repeats our baseline analysis on subsamples defined by eligibility category in Medicare Part D between 2007 and 2012. The panels show the results for OLS regressions on state-halfyear aggregates of opioid takers in each of the specified eligibility subsamples. Each observation is weighted by the number of individuals represented. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Outcomes involving prescribers and pharmacies are unavailable for 2007.

0.00106 (0.0017) 0.00775

-0.00281* (0.0014) 0.0211

“Must Access” PDMP mean of LHS

“Must Access” 0 PDMP mean of LHS

-0.00481*** (0.0013) 0.0367

“Must Access” PDMP mean of LHS

5+ Prescribers

Table 11: Effect of “Must Access” PDMP Law: Disabled and Dual-Eligible

.04 0

.01

5+ Prescribers .02

.03

Sample Mean in Blue, Estimated Post in Red

.015

Dual

Non-Dual Disabled

Dual

Non-Dual Elderly

0

5+ Pharmacies .01 .005

Sample Mean in Blue, Estimated Post in Red

Non-Dual Disabled

Dual

Non-Dual Elderly

-.05

Excess OOS Pharmacies -.03 -.02 -.01 -.04

0

Dual

Sample Mean in Blue, Estimated Post in Red Dual

Non-Dual Disabled

Dual

Non-Dual Elderly

Figure 11: Sample means (blue bar) and estimated “post” effect for Medicare subsamples (red, bars represent 95% CI) for 5+ Prescribers (top), 5+ Pharmacies (middle), and Excess OOS Pharmacies (bottom) 39

40

-0.00157* (0.0008) 0.0159

“Must Access” PDMP mean of LHS

-0.00172** (0.0006) 0.0051

-0.00238 (0.0012) 0.0024

-0.00193*** (0.0005) 0.0050

-0.000432 (0.0014) 0.0044

5+ Pharmacies

0.00119 (0.0048) 0.0059

0.00279 (0.0028) 0.0092

-0.00489 (0.0068) 0.1230

Die in the Half-Year -0.00302 -0.00182 (0.0051) (0.0038) 0.0438 0.0206

Do Not Die in the Half-Year -0.00116 0.000827 -0.0037 (0.0020) (0.0007) (0.0035) 0.0799 0.0256 0.0856

0.00325 (0.0023) 0.0085

-0.00463 (0.0032) 0.0853

-0.00126 (0.0080) 0.0123

Excess OOS Prescribers

Not Cancer Patients -0.00286 0.000218 (0.0020) (0.0007) 0.0792 0.0254

Overlapping Claims

0.00283 (0.0087) 0.1070

120+ Daily MED

Cancer Patients 0.0110*** 0.00436** (0.0029) (0.0015) 0.0671 0.0240

211+ Days Supply

0.00629** (0.0020) -0.0274

-0.0000279 (0.0048) -0.0163

0.00585** (0.0017) -0.0259

0.00665 (0.0044) -0.0329

Excess OOS Pharmacies

-0.000862 (0.0010) 0.0071

-0.00256 (0.0032) 0.0131

-0.000934 (0.0008) 0.0074

-0.000934 (0.0018) 0.0078

Poisonings

This table repeats our baseline analysis on subsamples defined by health status in Medicare Part D between 2007 and 2012. The panels show the results for OLS regressions on state-halfyear aggregates of opioid takers with vs. without cancer (top two panels) as well as opioid takers near the end of life (those who die in the halfyear) vs. those who do not. Each observation is weighted by the number of individuals represented. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Outcomes involving prescribers and pharmacies are unavailable for 2007.

-0.00752* (0.0033) 0.0143

-0.00193* (0.0010) 0.0149

“Must Access” PDMP mean of LHS

“Must Access” PDMP mean of LHS

-0.00173 (0.0033) 0.0220

“Must Access” PDMP mean of LHS

5+ Prescribers

Table 12: Effect of “Must Access” PDMP Law on Health Status Subsamples

41

-0.00003 (0.0003) 0.0027

-0.00003 (0.0000) 0.0001

“Must Access” PDMP mean of LHS

“Must Access” PDMP mean of LHS

-0.00004 (0.0000) 0.00003

-0.00003 (0.0001) 0.00040

-0.00175** (0.0006) 0.00495

5+ Pharmacies

Overlapping Claims

-0.00417 (0.0038) 0.0880

-0.0046 (0.0034) 0.2070

-0.000938 (0.0045) 0.1400

211+ Days Supply Opioids -0.00118 (0.0019) 0.0777 Antidepressants -0.000615 (0.0021) 0.2220 Statins -0.00149 (0.0012) 0.0406

-0.0001 (0.0004) -0.0016

-0.0005 (0.0007) (0.0014)

0.0027 (0.0029) 0.0090

Excess OOS Prescribers

-0.00101 (0.0013) 0.0217

-0.000674 (0.0008) 0.0057

0.00595** (0.0019) -0.0268

Excess OOS Pharmacies

This table reports our baseline results on the impact of a “must access’ PDMP on utilization of the indicated drug type in Medicare Part D between 2007 and 2012. Each column is the result of OLS regression on state-halfyear aggregates of takers of the indicated drug; each observation is weighted by the number of takers represented. Fixed effects for states and halfyears are always included, and standard errors are clustered at the state. *, **, and *** represent significance at the 5, 1, and 0.1 percent levels, respectively. Outcomes involving prescribers and pharmacies are unavailable for 2007.

-0.00191** (0.0006) 0.0158

“Must Access” PDMP mean of LHS

5+ Prescribers

Table 13: Effect of “Must Access” PDMP Law on Placebo Drugs