 The US opioid medication epidemic has resulted in serious health consequences for patients. Formulary management tools adopted by payers, specifically prior authorization (PA) policies, may lower the rates of opioid medication abuse and overdose. We compared rates of opioid abuse and overdose among enrollees in plans that varied in their use of PA from “High PA” (ie, required PA for 17 to 74 opioids), with “Low PA” (ie, required PA for 1 opioid), and “No PA” policies for opioid medications.

 Retrospective cohort study of patients initiating opioid treatment in Pennsylvania Medicaid from 2010 to 2012.

Generalized linear models with generalized estimating equations were employed to assess the relationships between the presence of PA policies and opioid medication abuse and overdose, as measured in Medicaid claims data, adjusting for demographics, comorbid health conditions, benzodiazepine/muscle relaxant use, and emergency department use.

The study cohort included 297,634 enrollees with a total of 382,828 opioid treatment episodes. Compared with plans with No PA, enrollees in High PA (adjusted rate ratio [ARR], 0.89; 95% confidence interval [CI], 0.85-0.93; 

<.001) and Low PA plans (ARR, 0.93; 95% CI, 0.87-1.00; 

= .04) had lower rates of abuse. Enrollees in the Low PA plan had a lower rate of overdose than those within plans with No PA (ARR, 0.75; 95% CI, 0.59-0.95; 

= .02). High PA plan enrollees were also less likely than No PA enrollees to experience an overdose, but this association was not statistically significant (ARR, 0.88; 95% CI, 0.76-1.02; 

 Enrollees within Medicaid plans that utilize PA policies appear to have lower rates of abuse and overdose following initiation of opioid medication treatment.

Health insurance payers can implement policies to help curb the opioid epidemic. This retrospective cohort study of Pennsylvania Medicaid data examined the associations between Medicaid plans that utilized prior authorization (PA) policies for opioid medications and enrollees developing opioid medication abuse or experiencing overdose. 

Enrollees within plans that subjected opioid medications to PA policies had lower rates of opioid medication abuse and overdose after initiating opioid medication treatment. 

Future research should work to extend these findings in order to support systematic and large-scale implementation of PA policies for opioid medications.

The US opioid medication epidemic has had serious effects on public health, including opioid-related overdoses and mortality.

 Perhaps the largest system-level investment in the United States to address abuse and prevent overdose has been prescription drug monitoring programs,

 which have shown mixed results for protecting patient health.

 Another system-level intervention is a lock-in program, wherein patients who exceed filling pattern thresholds are limited to specific providers/pharmacies to receive future opioid medications

; this intervention has shown some promise for improving medication monitoring and reducing diversion.

 Formulary management tools may represent a valuable set of interventions that payers can employ to control opioid medication consumption, deter shopping behaviors, and improve quality and safety.

One formulary management tool that may be used to address the opioid crisis is prior authorization (PA), a requirement placed on some medications by payers to verify that the medication is necessary and/or patients meet the medical criteria for use.

 An extensive body of literature has shown cost-saving benefits of PA policies for a variety of medications, often expensive name brand drugs

; however, limited research has been conducted on their impact on patient-related opioid and quality-of-care outcomes.

 PA policies applied in public or commercial health insurance plans frequently result in reductions in medication use,

 but this result is often accomplished by placing administrative burdens on clinicians.

Medicaid programs serving low-income populations have federally allowable co-payments mandating that minimal out-of-pocket costs can be charged to enrollees.

 Therefore, these programs are particularly reliant on PA policies, as opposed to other formulary management tools that use cost sharing to influence demand. Research has shown that approximately one-fourth of Medicaid patients who regularly use opioid medications (>90 days) are engaged in problematic opioid consumption behaviors.

 On average, Medicaid enrollees receive more than double the total annual opioid dose compared with the privately insured.

 To date, 19 states’ Medicaid programs have required PA for long-acting opioids, and study results show these policies can reduce long-acting opioid fills.

 The extent to which PA policies can help reduce the problematic opioid-related outcomes of abuse and drug overdose is unknown. We hypothesized that enrollees within Medicaid fee-for service (FFS) programs and managed care plans employing PA policies for opioid medications would have lower rates of abuse and opioid medication overdose compared with patients enrolled in Medicaid plans without PA. Understanding the potential associations between PA and abuse and overdose may provide health systems and payers with an additional tool to address problematic opioid-related outcomes.

This investigation was a retrospective cohort study that utilized Pennsylvania Medicaid data from 2010 to 2012. The Pennsylvania Medicaid program is among the largest in the United States in both expenditures and enrollment; the state's healthcare utilization, access,

 and statewide demographic profile (with the exception of lower rates of Hispanics)

 are similar to those seen across the nation. Pennsylvania has the eighth highest overdose rate in the country, and opioid prescribing rates are consistently above national averages.

 We obtained Pennsylvania Medicaid data directly from the Pennsylvania Department of Human Services (PADHS) for all FFS and managed care enrollees.

We used Medicaid enrollment data and pharmacy/medical claims to establish an analytic cohort of Medicaid enrollees who initiated a new opioid medication not used for addiction treatment (eAppendix [eAppendices available at 

]). We included patients in the study cohort who were aged 18 to 64 years, not dually eligible for Medicare (given that we could not capture medication use for those >64 years and dually eligible), without previous cancer treatment, not in long-term care for 90 or more days, and not receiving hospice services (as opioid use patterns would likely differ for these groups). We identified the index opioid exposure event as patients’ first oral, transdermal, or submucosal opioid medication fill.

To identify new episodes of opioid medication treatment, we excluded individuals from the cohort that possessed a record of filling any opioid medication, had an opioid use disorder, or experienced an opioid medication overdose in the 6 months prior to the index opioid fill. This step in the cohort construction allowed us to create a “clean” baseline period for patients before they were exposed to opioid medications and potentially developed abuse or experienced an overdose. Lapses in fills greater than 6 months following the index fill ended patients’ eligible treatment episodes. We selected a 6-month gap in fills to end the episode to be consistent with prior studies validating this approach in behavioral health populations.

 We examined numbers of patient episodes by plan PA status, and no major differences were detected (results not shown). This study was designated exempt by the University of Pittsburgh Institutional Review Board.

 We identified opioid medication abuse following previously published approaches

International Classification of Diseases, 9th Edition

) coding classifications and pharmacy claims. After the index fill, enrollees who had any code for an opioid use disorder (304.0, 304.00, 304.01, 304.02, 304.03, 304.7, 304.70, 304.71, 304.72, 304.73, 305.5, 305.50, 305.51, 305.52, 305.53) or opioid medication poisoning (965.00 [opium poisoning], 965.02 [methadone poisoning], 965.09 [opiate poisoning—not elsewhere classified], E.850.1 [accidental methadone poisoning], and E.850.2 [accidental opioid poisoning&mdash;not elsewhere classified])

 and had any overlapping fill for a opioid pain medication were categorized as having abuse (no abuse = 0, abuse = 1).

 Patients meeting this definition of abuse have been observed to have both a heightened overdose risk

 and serious behavioral, mental, and/or physical health problems.

 We recognize this definition of abuse does not match the 

Diagnostic and Statistical Manual for Mental Disorders 

definition, but we chose to employ this term given its previous use in the literature.

The opioid overdose indicator used in this analysis followed previously established methods for identifying prescription opioid overdose using 

 The overdose indicator occurred after the index opioid fill, comprised opioid medication poisoning codes (965.00, 965.02, 965.09, E.850.1, E.850.2), and was dichotomized (no overdose = 0, overdose = 1). These codes capture nonfatal and fatal overdose events resulting in hospitalization, emergency department (ED) visits, and/or other medical care. We did not capture overdose events outside of the healthcare system, which may have largely been untreated within Pennsylvania given the limited and variable availability of naloxone to public safety and prehospital healthcare professionals during the study years. We acknowledge that abuse and overdose are both constructed using poisoning claims and that there is some overlap between these measures; however, we chose this approach (ie, not removing the poisoning codes from the abuse indicator) to remain consistent with the previous literature.

 Specification of plans in Pennsylvania Medicaid with PA took place in partnership with the Bureau of Managed Care within PADHS. Officials from PADHS provided FFS PA information and contacted all managed care plans (N = 8) via e-mail requesting historical formulary medication management policy information between January 1, 2010, and December 31, 2012. Qualitative responses were transferred into a data-tracking template. Given the variation in use of PA across plans in our study data, we followed an ordinal classification approach for categorization of policies similar to those previously employed in the literature.

 One insurance plan was labeled “Low PA” (ie, required PA for 1 opioid); 2 plans were labeled “High PA” (ie, required PA for 17-74 opioids); and 6 plans were labeled “No PA” based on the number of generic, brand name, and combination product medications subjected to PA (Table 1). PA policies were active before or on the first day of our study observation period (January 1, 2010), thus limiting our ability to compare the differences among plans across time. We therefore conducted a cross-sectional comparison of enrollees across plan types.

 Covariates were measured in the enrollees’ baseline periods. Demographic covariates included in the model were age (18-29, 30-39, 40-49, 50-64 years), sex, race/ethnicity (white, black, Hispanic, other), Medicaid eligibility category (General Assistance, Supplemental Security Income, Temporary Assistance for Needy Families), Medicaid plan type (FFS, managed care organization), and urban/rural county of residence (coded using Rural-Urban Continuum Codes

We likewise included measures of comorbidity in the models, which were also measured at baseline. Specific comorbidities included: alcohol use disorders (abuse/dependence), nonopioid drug use disorders (abuse/dependence [eg, cocaine, marijuana] not including Not Elsewhere Classified codes, which clinicians may have used in lieu of opioid use disorder codes), several indicators for mental health disorders (adjustment, anxiety, mood, personality, miscellaneous), separate indicators for pain diagnoses (back, neck, arthritis/joint, headache/migraine), and HIV/AIDS.

 We included in the model a modified Elixhauser Comorbidity Index, an indicator that used 

 codes to measure patient comorbidity within administrative claims data from hospitals and physician services. This indicator was modified by removing comorbidities described above that we included as individual covariates. ED use was also included in the model (≥1 visit = 1, <1 visit = 0).

We included morphine milligram equivalents (MME) following the index fill but before the occurrence of abuse or overdose. MME was constructed by converting the total within-episode opioid supply into morphine equivalents, dividing by the days' supply, and coding into 4 levels: ≥100 MME/day, 50 to <100 MME/day, 20 to <50 MME/day, <20 MME/day.

 Indicators of medication use that are known correlates with abuse/overdose were also added as covariates in the model, which included any use of benzodiazepines and muscle relaxants in the baseline period. All covariates were categorical with the exception of the Elixhauser Comorbidity Index, which was a count indicator, and age, which was ordinal.

With the exception of descriptive demographic characteristics that were calculated at the person level for patients enrolled in the 3 PA plan types, analyses for this study were conducted at the episode level. Our modeling strategy needed to account for 2 features of our data: heterogeneity in the duration of opioid use across episodes and some enrollees having multiple episodes. The importance of accounting for episode-level events for individual enrollees is based on the dynamic nature of patient behaviors and health status across time, which can alter an individual’s risk. We therefore employed generalized linear models with generalized estimating equations (GEE) using log link function and Poisson distribution where follow-up length (day) was treated as offset in the model, and the exchangeable covariance structure was employed to account for standard error correlation.

These models were able to account for greater exposure to opioids and PA policies within an episode and greater numbers of episodes, and they were applied to examine the association between the outcome variables of opioid medication: 1) abuse and 2) overdose and the predictor variable of PA adjusted for all covariates described above. We also report abuse and overdose rates with 95% confidence intervals (CIs) by PA type adjusted for all covariates and offset log length of episode. All analyses were conducted with SAS version 9.4 (SAS Institute; Cary, North Carolina).

In an alternative model specification, we estimated both the abuse and overdose outcome analyses using a propensity score matching approach wherein we matched individuals in the High and Low PA plans to those in No PA plans. Results showed no substantive differences; therefore, we chose to present the adjusted GEE results instead of the matched sample for the purpose of simplicity and to maximize the sample size included in our analyses.

The analytic cohort included 297,634 individual plan enrollees with a total of 382,828 opioid treatment episodes. Many enrollees within the cohort had multiple opioid treatment episodes, with patients having an average of 2 episodes (median = 1; results not shown). Table 2 presents descriptive patient-level demographic and episode-level health and medication use characteristics. The largest proportions of patients were aged 18 to 29 years (n = 140,876; 47.3%) and were female (n = 212,209; 71.3%). The largest proportional differences among PA plans were in race and rural/urban living location. White enrollees were most prominent in High PA plans (77.2%; n = 79,965), and Black (47.9%; n = 15,950) and Hispanic (31.2%; n = 10,386) enrollees were most prominent in the Low PA plan. Most Low (99.9%; n = 33,226) and No PA (95.1%; n = 152,913) enrollees lived in urban locations compared with 63.7% (n = 65,948) of High PA plan enrollees.

The most common level of opioid consumption within episodes (60.9%-68.6%) was 20 to 49.9 MME per day following the index opioid fill. The unadjusted rate of abuse within episodes was 3.46 for High PA plans, 2.36 for the Low PA plan, and 3.39 for the No PA plans after the index opioid medication fill. The unadjusted rate of overdose in episodes was 0.26 in High PA plans, 0.19 for the Low PA plan, and 0.29 in No PA plans after the index fill (Table 3).

Table 4 shows the results of the GEE analyses adjusted for demographic and health status differences across plan types, which demonstrated that individuals in High PA plans were 11% less likely to develop opioid medication abuse after their index opioid medication fill compared with those within plans with No PA (95% CI = 0.85-0.93; 

<.001). Enrollment in the Low PA plan was also associated with a 7% lower rate of developing opioid medication abuse after the opioid medication index fill (95% CI, 0.87-1.00; 

In terms of the relationship between PA and overdose, enrollment in the Low PA plan was associated with a 25% lower rate of experiencing an overdose following the index opioid medication fill (95% CI = 0.59-0.95; 

 = .02). There was a nonsignificant 12% reduction (95% CI = 0.76-1.02; 

= .08) in overdose for enrollees in High PA plans. We recognize that the High PA plans had the highest unadjusted rate of abuse, but after adjustment, the No PA plan had the highest. To identify which set of covariates influenced this change, we re-estimated our model by adding blocks of variables to the model in a stepwise fashion (eg, block 1 = abuse, overdose, and MME; block 2 = demographics; block 3 = mental/behavioral health and co-occurring health conditions [ie, pain, Elixhauser Index]). Results showed that adding the demographic block resulted in the change. We also re-estimated the GEE analyses without the MME covariate to examine its influence on model outcomes. The magnitude and direction of all effects were unchanged after removing MME.

Table 5 reports adjusted rates based on the GEE analyses for abuse and overdose per person-days (where 452.1 [standard deviation = 299.2] was the average number of per-person follow-up days for subjects in the cohort). The adjusted rates of abuse were 2.49% for High PA plans, 2.58% for the Low PA plan, and 2.76% for No PA plans per average person-days. The adjusted rates of overdose were 0.21% for High PA plans, 0.17% for the Low PA plan, and 0.23% for No PA plans per average person-days.

The opioid medication epidemic has brought to the forefront of healthcare practice and policy the need to identify and intervene with patients engaged in abuse and at risk for overdose. Policy-level efforts that limit access to opioid medications and influence patient and prescriber behaviors have the potential to make an important impact on reducing negative patient outcomes. We analyzed data from Medicaid enrollees who developed opioid pain medication abuse or experienced an overdose after initiating opioid treatment who were within plans that utilized PA policies compared with plans that did not. Our findings showed a minority of plans implemented PA policies (3 of 9 plans) and there was substantial variation in the number of medications within plans subjected to PA policies (range = 1-74).

Enrollment in High and Low PA plans was associated with modestly lower adjusted rates of opioid medication abuse, and enrollment in the Low PA plan was associated with lower adjusted rates of overdose. These results are consistent with those of previously published studies that have examined the effects of PA on opioid medication fills. Specifically, our findings that PA was associated with 7% to 11% (

= .02) lower rates of overdose are consistent with studies that have reported 8% to 19% reductions in long-acting opioid medication fills among enrollees in plans that utilized PA policies.

A central point of importance for our findings is that they advance previous studies reporting reductions in long-acting opioid medication fills, which is an outcome metric with limited ability to differentiate between patients with problematic use and those who may benefit from opioid medications. Assessing only fills as an outcome also cannot disentangle patient and prescriber behaviors. Accordingly, evaluating the benefits of PA policies by changes in abuse and overdose more effectively discriminates reductions in potential repercussions of opioid use—perhaps demonstrating an outcome especially relevant to combatting the national opioid epidemic. Further, reductions in abuse, for example, could have valuable ramifications for health systems, payers, and prescriber stakeholders as patients with opioid use disorders have higher healthcare needs, utilization, and costs.

Future research should seek to extend our findings by examining the effect of PA within an analytical framework capable of examining both within- and between-group differences over time. Such studies should seek to account for legitimate pain management needs of patients. In particular, although abuse and overdose are both important outcomes for patient safety and health, future studies should examine potentially unintended consequences of PA plans on patients, such as undertreatment of painful conditions

 If future research continues to provide support for PA, broader implementation of these policies may necessitate streamlined and automated approaches to minimize the disruption to the medical/pharmacy workflow.

These findings should be viewed in light of certain limitations. First, while we recognize that the strengths of our study include possessing actual PA information from Medicaid plans, having complete FFS and managed care data, and Pennsylvania being similar to other programs in the nation with respect to healthcare utilization, access,

 it nonetheless represents 1 state in the United States.

Second, the last year for our data was 2012, and some analyses have shown reductions in opioid abuse and diversion in more recent years.

 Furthermore, the larger Medicaid landscape has evolved since this date, including the expansion of Medicaid through the Affordable Care Act in Pennsylvania and many other states.  Studies conducted with more recent data may yield different estimates as a result of these changes. PADHS also recently implemented additional restrictions on opioids

 that may yield even greater benefit. However, these policies went into effect after our study period ended so we were unable to evaluate them.

Third, we used a simple and straightforward approach to categorizing PA schemes (High, Low, and No based on the number of products subject to PA). It is possible our data collection method did not capture other aspects of these policies, such as the ease of use that may influence prescribing behavior. It is also possible that we did not capture information on all plan features or policies that may influence opioid prescribing and use. In light of this, we recognize our characterization of the policies may not capture the full range of interventions plans may have had in place. We note, however, that enrollees with evidence of opioid medication misuse have an equal possibility of enrollment in the state Medicaid agency-operated lock-in program.

Fourth, whereas the abuse measure is one of the more common and valid indicators in the field,

 it has the potential to misclassify individuals engaged in legitimate use of opioids. Moreover, while we have been able to adjust for a number of patient-level characteristics in our analyses that could have introduced bias into our findings, other individual-level factors and regional variations in our outcomes could have influenced study outcomes. Future research should seek to employ quasi-experimental designs with comparison groups, such as difference-in-differences analyses, to help better understand the impact of PA.

Last, we recognize opioid use disorders are likely undercoded within claims data

 and claims data do not account for cash payments to prescribers/pharmacies, which could influence observed associations were these data available.

This study examined associations between PA requirements and developing abuse or experiencing an opioid medication overdose following an index opioid medication fill for Medicaid enrollees. These findings extend previous research by demonstrating improved outcomes among patients within PA plans in terms of lower rates of abuse and overdose. Future research should seek to extend these findings within more rigorous causal designs/analyses and among other Medicaid and commercial payer data to continue building evidence for PA policies reducing problematic opioid behaviors and consequences. 

 School of Social Work (GC), and School of Medicine (C-CHC), Department of Psychiatry (GC), and School of Medicine, Division of General Internal Medicine (AJG, WFG, WF), and Center for Pharmaceutical Policy and Prescribing (GC, AJG, WFG, JMD), and Graduate School of Public Health (C-CHC, CL, EC, PZ, JMD), University of Pittsburgh, Pittsburgh, PA; VA Pittsburgh Healthcare System (AJG, WFG), Pittsburgh, PA; University of Arizona, College of Pharmacy (WL-C), Tucson, AZ; Pennsylvania Department of Human Services, Medical Assistance Programs (DK), PA.

 Centers for Disease Control and Prevention (U01CE002496).

 Dr Gordon has received grants from NIH, VA, and other public grants; the contents do not represent views of the US department of Veterans Affairs or the US government. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (GC, AJG, WFG, WF, DK, JMD); acquisition of data (GC, PZ, JMD); analysis and interpretation of data (GC, AJG, WFG, WL-C, DK, CL, EC, WF, PZ, JMD); drafting of the manuscript (GC, AJG, C-CHC, WL-C, CL, JMD); critical revision of the manuscript for important intellectual content (GC, AJG, WFG, WL-C, CL, EC, WF, DK, JMD); statistical analysis (GC, C-CHC, WL-C, PZ); obtaining funding (GC, JMD); administrative, technical, or logistic support (EC, JMD); and supervision (JMD).

 Gerald Cochran, PhD, University of Pittsburgh, 4200 Forbes Ave, 2117 CL, Pittsburgh, PA 15260. E-mail: gcochran@pitt.edu.&ensp;

1. Rudd RA, Aleshire N, Zibbell JE, Gladden RM. Increases in drug and opioid overdose deaths—United States, 2000-2014. 

. 2016;64(50-51):1378-1382. doi: 10.15585/mmwr.mm6450a3.

2. Compton WM, Jones CM, Baldwin GT. Relationship between nonmedical prescription-opioid use and heroin use. 

. 2016;374(2):154-163. doi: 10.1056/NEJMra1508490.

3. Clark T, Eadie J, Knue P, Kreiner P, Strickler G; The Prescription Drug Monitoring Program Center of Excellence; Heller School for Social Policy and Management, Brandeis University. Prescription drug monitoring programs: an assessment of the evidence for best practices. The PEW Charitable Trusts website. http://www.pewtrusts.org/~/media/assets/0001/pdmp_update_1312013.pdf. Published September 20, 2012. Accessed May 2017.

4. As the prescribing practices of emergency providers come under enhanced scrutiny, watch for red flags of drug-seeking behavior. 

5. Paulozzi LJ, Kilbourne EM, Desai HA. Prescription drug monitoring programs and death rates from drug overdose. 

. 2011;12(5):747-754. doi: 10.1111/j.1526-4637.2011.01062.x.

Articles

WEB EXCLUSIVE

Enrollees in Medicaid plans employing prior authorization policies for opioid medications may have lower rates of opioid medication abuse and overdose.

Medicaid Prior Authorization and Opioid Medication Abuse and Overdose

 To examine the impact of prescription drug monitoring programs (PDMPs) on drug overdose deaths.

We used variation in the timing of state PDMP legislation and implementation to estimate the impact of these programs on drug overdose mortality rates across all drug categories from 1999 to 2014 and separately for each category from 1999 to 2010. Data used include US all-jurisdiction mortality data, estimated population data, and sociodemographic data from the CDC and the US Census Bureau.

 Multivariate regression models were applied to state panel data, including state and year fixed effects and state-specific linear time trends. Preprogram tests were used to assess the common trends assumption underlying our empirical approach. 

The implementation of PDMPs was not associated with reductions in overall drug overdose or prescription opioid overdose mortality rates relative to expected rates in the absence of PDMPs. For most categories, PDMPs were associated with increased mortality rates, but the associations were statistically insignificant. In a subsample analysis of states with PDMPs in operation for 5 or more years, the programs were found to be associated with significantly higher mortality rates in legal narcotics, illicit drugs, and other and unspecified drugs.

 PDMPs were not associated with reductions in drug overdose mortality rates and may be related to increased mortality from illicit drugs and other, unspecified drugs. More comprehensive and prevention-oriented approaches may be needed to effectively reduce drug overdose deaths and avoid fatal overdoses from other drugs used as substitutes for prescription opioids.

Prior studies of state prescription drug monitoring programs (PDMPs) and fatal drug overdoses have either not examined specific drug categories or examined only selected specific drug categories, focusing on prescription opioids. This study examined the impact of PDMPs on drug overdose mortality rates across all drug categories from 1999 to 2014 and separately for each category from 1999 to 2010, including illicit drugs and other and unspecified drugs, and found that: 

		PDMPs, on average, have not been effective in reducing drug overdose mortality rates. 

		PDMPs may be associated with increased overdose mortality rates in drug categories other than prescription opioids.

Prescription drug overdoses have become one of the fastest growing and most serious public health concerns in the United States. The number of deaths has increased more than 7-fold: from about 6100 in 1980 to 47,055 in 2014,

 or approximately 129 deaths every day. In 2014, the overall US mortality rate was 823.7 per 100,000, and the drug overdose mortality rate was 14.7 per 100,000.

 In addition, recent reports indicate that fatal drug overdoses significantly contributed to the unexpected increase in mortality among midlife non-Hispanic whites.

 More than half of drug overdose deaths are caused by prescription drugs, and more than 70% of prescription drug overdose deaths are caused by opioid pain relievers.

Inappropriate prescription drug use not only affects health outcomes, but is also correlated with increasing fraud, waste, and additional costs for taxpayers, employers, and insurers.

 For example, a study that examined medical and pharmacy claims data from 16 self-insured employer health plans reported that enrollees identified as having drug abuse or dependence had hospitalization rates 12 times higher than those without these conditions and annual healthcare and drug costs 8 and 5 times higher, respectively.

 The total societal costs of prescription opioid overdose, abuse, and dependence in the United States in 2013 was estimated at $78.5 billion, including healthcare costs, lost productivity, and criminal justice costs.

To address prescription drug misuse and abuse, an increasing number of states (currently more than 40) have implemented a prescription drug monitoring program (PDMP). These programs maintain statewide databases, collecting data on the prescribing, dispensing, and purchasing of controlled substances.

 These data can be used to identify suspected illegal activities, such as prescription diversions, doctor shopping, and pill mills; to inform public health initiatives; and to facilitate the treatment of drug addiction, among others. The Office of National Drug Control Policy (ONDCP) considers PDMPs an important tool to combat prescription drug overdose deaths. Implementing PDMPs was one of the 4 major areas in ONDCP’s 2011 Prescription Drug Abuse Prevention Plan, with the goals of having legislation in all 50 states establishing PDMPs within 36 months and decreasing by 15% the number of unintended opioid-related overdose deaths within 60 months.

Despite the expansion of PDMPs, evidence on their effectiveness in reducing prescription drug abuse or misuse is inconclusive.

 For example, 1 study analyzed opioid abuse treatment admission data and found that PDMPs mitigated the increasing trend of opioid abuse and misuse.

 However, another study found no demonstrable decrease in prescription opioid abuse associated with PDMPs.

 Reviewing patients’ prescription history in a university medical center emergency department changed clinicians’ opioid prescription plans for 74 of 179 patients, with 61% receiving fewer opioids and 39% receiving more opioids than originally planned.

 A more recent study (2016) used the 2001 to 2010 data from the National Ambulatory Medical Care Survey and found that PDMPs were associated with a significant decrease in Schedule II opioids prescribing rates.

The existing literature on PDMPs and drug overdose deaths is limited. The results of a small number of earlier studies on this topic suggest that PDMPs might not be effective in reducing fatal drug overdoses, but the findings were not consistent.

 A more recent study using an interrupted time series design found that implementing a PDMP was associated with a decrease in prescription opioid-related overdose mortality rates.

 These studies focused on prescription opioid-related overdose deaths and/or overall drug overdose deaths. To our knowledge, no study has broken down drug overdose deaths across different classes of drugs, an important gap in the literature. Although PDMPs monitor only prescription controlled substances, they might also affect the use of other drugs, as individuals may switch to nonprescription drugs or find alternative ways of obtaining prescription medications.

Our study aimed to investigate whether PDMPs were effective in reducing fatal drug overdoses across all drug categories and separately for each category. Given the magnitude of the prescription drug epidemic and the expansion of PDMPs at the national level, evaluating PDMPs’ overall impact may be helpful for assessing the relative effectiveness of public policies designed to reduce drug overdoses. 

We conducted our analysis in 2 parts. First, using publicly available mortality data from the Centers for Disease Control and Prevention (CDC),

 we examined overall drug overdose mortality rates in PDMP and non-PDMP states in the United States between 1999 and 2014. This database suppresses mortality data when the number of deaths is fewer than 10. Next, we applied the same analysis to the unsuppressed CDC mortality data obtained from the National Center for Health Statistics (NCHS) from 1999 to 2010—the maximal range available from the multiple-cause-of-death data at the time of our analysis&mdash;identifying drug overdose deaths by subcategory. 

In our study sample, 19 states began operating PDMPs sometime between 2002 and 2010: Alabama, Arizona, Colorado, Connecticut, Iowa, Louisiana, Maine, Minnesota, Mississippi, North Carolina, North Dakota, New Mexico, Ohio, South Carolina, Tennessee, Virginia, Vermont, West Virginia, and Wyoming. From 2011 to 2014, 15 more states implemented PDMPs: Alaska, Arkansas, Delaware, Florida, Georgia, Kansas, Maryland, Montana, Nebraska, New Hampshire, New Jersey, Oregon, South Dakota, Washington, and Wisconsin. Two states (“states” hereafter includes the District of Columbia) did not begin to operate PDMPs by the end of 2014: Missouri and the District of Columbia.

 The remaining 15 states were excluded as they implemented PDMPs before 2000 and did not have sufficient pre-implementation years available to support the empirical approach used in this paper. 

International Classification of Diseases Codes 10th Revision

 codes): X40-X44, X60-X64, X85, and Y10-Y14 for the underlying-cause-of-death data and T36.0-T50.9 for the multiple-cause-of-death data. A complete list of these 

The data used to measure drug overdose mortality rates include: 1) CDC Wide-ranging Online Data for Epidemiologic Research (WONDER) database of mortality

; 2) de-identified individual-level unsuppressed mortality data for all jurisdictions of the United States, obtained from the NCHS; and 3) estimated population data produced by the US Census Bureau and NCHS, which were also extracted from the CDC WONDER website.

 The mortality data are based on the information from all death certificates filed in all jurisdictions in the United States. Each death certificate contains a single underlying cause of death and up to 20 multiple causes. Data used for covariates include population estimates from the US Census Bureau and NCHS

and sociodemographic data from the Current Population Survey, produced by the US Census Bureau.

To estimate the effect of PDMPs on fatal drug overdoses, we used multivariate regression models with state and year fixed effects and state-specific linear time trends. The inclusion of state fixed effects allows each state to serve as its own control group, eliminating all time-invariant unobserved differences across states. National and state trends in PDMP operation and drug overdose mortality rates, which might otherwise confound our estimates, were accounted for with year fixed effects and state-specific linear time trends.

 This approach contrasts with the analysis in Patrick et al, which relied on state fixed effects and a single national linear time trend.

 The exposure was defined as the state- and year-specific PDMP operation status (operated = 1, not operated = 0). The outcome variable was state-level, year-specific drug overdose mortality rates measured by the number of deaths per 100,000 individuals. To control for other differences across states, we included state-level, time-varying covariates that might be associated with drug overdose mortality rates and PDMP operation status; in particular, the percentages of a state’s population that is male, white, high school educated or better (age 25 or older), uninsured, and enrolled in the Medicaid program. We also controlled for median household income (in 2015 US$ for 1999-2014 and 2011 US$ for 1999-2010). Mortality rates were crude rates because the covariates we obtained were not age-adjusted.

Clustered standard errors were used to correct for arbitrary patterns of serial correlation within states.

A key threat to establishing a cause-and-effect relationship between PDMPs and fatal drug overdoses is the possibility that states adopt PDMPs in response to changes in overdoses that depart from the state-specific linear time trends included in our models, or that adoption coincides with changes in these trends for other reasons. We tested for this possibility in an extended model presented in 

by adding indicator variables for the year prior to enactment of the PDMP law and for the 2 years prior to enactment, labeled PRE1 and PRE2, respectively. If these preprogram indicators are small and statistically insignificant, it suggests that after adjusting for national and state trends, states adopting PDMPs would have experienced similar changes in fatal drug overdoses as the nonadopting states in the absence of a PDMP.

We also examined potential PDMP enactment effects for the period following the enactment of a PDMP law, but prior to the PDMP becoming operational, by including an indicator labeled PEPO (post enactment and pre-operation) in the extended model for these periods (years of post enactment and pre-operation = 1; the other years = 0). In addition, to examine potentially important differences in the PDMP effect based on program duration, we also conducted a subsample analysis of PDMPs operating for 5 or more years.

Statistical analysis was conducted using SAS version 9.4 (SAS Institute Inc; Cary, North Carolina) and Stata/IC version 14 (StataCorp LP; College Station, Texas). Institutional review board approval was not needed because no human participants were involved in this study. 

Overall, mortality rates from prescription drug overdoses increased from 1999 to 2014, in both PDMP and non-PDMP states. The PDMP coefficients from our regression models (

) capture the difference between the mortality rates expected to arise in the absence of a PDMP, as predicted by all other covariates in the model, and the rates that occur when a PDMP is present. We organized the results to show the PDMP effect in: 1) overall drug overdose mortality rates, from the underlying-cause-of-death data and the multiple-cause-of-death data, for 1999 to 2014 and 1999 to 2010, separately (Table 1); and 2) subcategories of multiple-cause-of-death data in 2 ways: a) subcategories, which are not mutually exclusive, but collectively comprehensive and b) subcategories with high mortality rates (mortality rates >1 per 100,000 for both the PDMP and non-PDMP states) for 1999 to 2010 (Table 2). 

 illustrate the results in Tables 1 and 2, respectively.

For the overall overdose mortality rates, our estimates from the publicly available data (1999-2014) were 0.08 (95% confidence interval [CI], —0.89 to 1.04) and 0.02 (95% CI, –0.97 to 1.00) for underlying cause of deaths and multiple cause of deaths, respectively. The estimates from the unsuppressed data (1999-2010) were 0.02 (95% CI, –1.05 to 1.08) and 0.02 (95% CI, –1.07 to 1.11), respectively. In the extended model, the PDMP coefficients slightly increased, but were statistically insignificant (eAppendix Table 2 and eAppendix Figure 1). These results suggest that PDMP implementation had little impact on overall overdose mortality rates. Throughout the primary and extended models, all PDMP coefficients pertaining to overall overdose death rates were positive; however, their magnitude and statistical significance varied across model specifications.

In the subcategory analysis, PDMP coefficients were 0.02 (95% CI, —0.81 to 0.84) for legal narcotics and 0.85 (95% CI, –0.08 to 1.78) for other and unspecified drugs (Table 2). In the extended model, PDMPs were associated with significantly increased mortality rates for illicit drugs (0.92; 95% CI, 0.15-1.69) and cocaine (0.71; 95% CI, 0.11-1.31) (eAppendix Table 3 and eAppendix Figure 2).

Based on the subsample analysis for states with a PDMP in place for 5 or more years, all PDMP coefficients were positive for overall mortality rates and were significant for 1999 to 2010 (Table 1 and eAppendix Table 1). In the subcategory analysis, longer-standing PDMPs were associated with significantly increased mortality rates in several categories, including legal narcotics, illicit drugs, cocaine, other and unspecified drugs (Table 2), and illicit drugs and cocaine (eAppendix Table 2).

As shown in the extended model results, none of the preprogram indicators (PRE1 and PRE2) were significant, lending support to the model specification employed in our initial analysis. The PEPO indicator was positive and significant for some categories, which suggests an increase in mortality rates for those categories in the post enactment and preoperation periods. 

This study investigated PDMP effects on fatal drug overdoses in the United States from 1999 to 2014. We found that PDMPs were not associated with a reduction in either overall or prescription opioid drug overdose mortality rates. Moreover, during the period from 1999 to 2010, for which we conducted the subcategory analysis, PDMPs were often associated with increased mortality rates in drug categories other than prescription opioids, such as illicit drugs or other and unspecified drugs, particularly among the states with longer-standing PDMPs. Although our study period was shorter when examining mortality rates for different drug categories, our results may reflect an unintended consequence of PDMPs, at least up through 2010, whereby reduced access to prescription drugs may have led some individuals with addictive disorders to seek out substitute drugs.

Our findings have several policy and clinical implications. First, PDMPs do not seem to have been successful in reducing drug overdose mortality rates, even in the target categories of prescription opioids (T40.2) and legal narcotics (T40.2-T40.4). This is consistent with some previous studies.

 There are many possible reasons for this outcome. For example, PDMPs may not be able to fully address prescription diversions, doctor shopping, or other abusive behaviors, and under these programs, potential drug-related illegal activities are only detectable through prescription fillings. The rapidly growing online pharmaceutical sale space may have also increased the opportunities for individuals to evade state or federal regulations. (More than 90% of internet pharmacies are estimated to be illegal.

) Further, PDMPs may drive patients away from doctors who could help them address drug abuse or dependence. 

Second, our results imply that PDMPs might be related to increases in drug overdose mortality rates attributable to illicit drugs or other and unspecified drugs. The existing literature has raised the possibility of these unintended consequences of PDMPs, although there has been little empirical evidence to date.

 By analyzing overdose deaths for different drug categories, including illicit drugs and other and unspecified drugs, our study provides some evidence for such a possibility. Future research is needed to further explore the unintended consequences of PDMPs and the potential mechanisms contributing to them (eg, PDMPs’ influence on clinical practices and individual behaviors). If the underlying problems of drug addiction or drug abuse are not effectively addressed, PDMPs might trigger some people to obtain illicit drugs as potential substitutes. 

Third, our findings suggest that PDMPs may need to be combined with more comprehensive and prevention-oriented approaches to address drug overdose deaths. Examples of prevention-oriented approaches include: 1) improving patient education on the appropriate use of drugs, 2) ensuring proper access to prescription drugs for those with medical needs, 3) expanding treatment programs for those with drug abuse problems, and 4) improving provider education and clinical practices for pain management. Such approaches are consistent with the recent program, "Prescription Drug Overdose: Prevention for States," funded by the CDC in 16 states.

Fourth, our findings suggest that policy makers, insurers, and managed care organizations might need to consider the effects of PDMPs when designing health plans, including such features as reimbursement, overall benefit design, and coverage criteria for specialty treatment of drug abuse. Currently, CMS provides the Medicare Part D Opioid Prescriber Summary File on its website, which includes the individual provider’s National Provider Identifier, last name, zip code, and number/percentage of prescription claims for opioid drugs.

 Although this information is based on Medicare claims data, physicians might be concerned about the potential effects the PDMP data could have on reimbursement when combined with administrative data.

 Reimbursement policies based on the number of opioid prescriptions, without consideration of medical need or value, might negatively affect quality of care and, ultimately, increase costs for payers if physicians limit opioid prescriptions in a way that runs counter to optimal patient care. In addition, there may also need to be better coordination between primary and specialty care providers for patients with drug problems. 

Finally, this study suggests that restricting attention to overdose deaths caused by opioids or prescription drugs might not fully capture the impact of PDMPs. Researchers and policy makers may need to be cautious about the heterogeneous effects across different drug categories, partly due to drug substitutability. The results also call attention to the incomplete information on the cause of death in mortality data. The “other and unspecified drugs” category (T50.9) has a larger number of fatal overdoses than any other category (eAppendix Figure 3) and warrants further investigation. 

First, our analysis of overdose mortality rates by drug category only extended to 2010. Second, although our empirical approach mechanically eliminated many potential confounders and our preprogram tests provided support for the common trends assumption necessary for its use, we cannot rule out all sources of bias, such as those created by the time-varying factors we were unable to control for. Finally, our analysis uses binary indicators for PDMP implementation status, thus only estimating an average PDMP effect. The existing literature has documented heterogeneity in the design and implementation of PDMPs across states.

 Hence, the effects of PDMPs on drug overdose deaths may also differ across states. Previous studies examined the effects of some PDMP characteristics on overdose mortality. These studies focused on factors such as the type of governing agency, statutory authority to monitor noncontrolled substances, the requirement for committee oversight, exempting practitioners from the obligation to access PDMP data, and the provision of unsolicited reports to healthcare practitioners and law enforcement agencies, and did not find significant protective effects of any of these features.

 Another study reported that monitoring 4 or more drug schedules and more frequent updating of PDMP data were associated with greater reductions in overdose deaths.

 Future studies are warranted to evaluate the effectiveness of other PDMP characteristics. 

PDMPs were not associated with a decrease in drug overdose mortality rates, even in the target category of prescription opioids. They may be associated with increased mortality rates in categories other than prescription opioids, especially in states where PDMPs have been operating for longer periods of time. Further research is needed to better understand the heterogeneous impacts of PDMPs. More comprehensive, prevention-oriented approaches, including improvement in patient education and clinical practices for pain management, may be needed to effectively reduce mortality caused by drug overdoses.

The authors are grateful to the National Center for Health Statistics, the CDC, and the various vital statistics jurisdictions for providing data used in this study.

 University of Pennsylvania (YN), Philadelphia, PA; The Pennsylvania State University (DGS, YS, JRM), University Park, PA.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (YN, DGS, YS, JRM); acquisition of data (YN, DGS); analysis and interpretation of data (YN, DGS, YS, JRM); drafting of the manuscript (YN, DGS, YS, JRM); critical revision of the manuscript for important intellectual content (YN, DGS, YS, JRM); statistical analysis (YN, JRM). 

 Young Hee Nam, PhD, Center for Pharmacoepidemiology Research and Training, Center for Clinical Epidemiology and Biostatistics, Department of Biostatistics, Epidemiology and Informatics, Perelman School of Medicine, University of Pennsylvania, 826 Blockley Hall, 423 Guardian Dr, Philadelphia, PA 19104. E-mail: ynam@mail.med.upenn.edu. 

POLICY

This study investigates the impact of state prescription drug monitoring programs on drug overdose mortality rates for all drug categories.

State Prescription Drug Monitoring Programs and Fatal Drug Overdoses

To replicate and extend a recently published analysis of the drivers of excess costs of opioid abuse.

Retrospective data analysis using de-identified claims data from the Truven MarketScan Commercial Claims and Encounter database.

 Medical and prescription drug claims from beneficiaries covered by large self-insured US companies were used to select patients with incident diagnoses of opioid abuse between 2012 and 2015. Two cohorts, abusers and nonabusers, were matched using propensity score methods. Excess healthcare costs were estimated over a 6-month baseline period and 12-month follow-up period. Cost drivers were assessed by diagnosis (3-digit 

International Classification of Diseases, Ninth Revision, Clinical Modification

 The analysis included 73,714 matched pairs of abusers and nonabusers. Relative to nonabusers, abusers had considerably higher annual healthcare costs of $10,989 per patient, or $1.98 per member per month. Excess costs were similar, yet lower, than the previous analysis using another commercial claims database. In both analyses, a ramp-up in excess costs was observed prior to the incident abuse diagnosis, followed by a decline post diagnosis, although not to baseline levels. Key drivers of excess costs in the 2 studies included opioid use disorders, nonopioid substance misuse, and painful and mental health conditions. From 2010 to 2014, the prevalence of diagnosed opioid abuse doubled, with incidence rates exhibiting an increasing, though flatter, trend than earlier in the period.

 Opioid abuse imposes a considerable economic burden on payers. Many abusers have complex healthcare needs and may require care beyond that which is required to treat opioid abuse. These results are robust and consistent across different data sources.

		Commercially insured patients who receive a diagnosis of prescription opioid abuse, dependence, or overdose/poisoning incur annual excess healthcare costs of $10,989 (or $1.98 per member per month) relative to matched nonabusers, confirming a large burden to commercial payers as shown by the results of a recently published study.

		An assessment of the key cost drivers associated with opioid abuse highlights that abusers often face multiple mental health, substance use, and/or other health issues beyond opioid abuse, suggesting a need for more comprehensive care. 

		The prevalence of diagnosed opioid abuse doubled between 2010 and 2014, with incidence rates exhibiting a flatter, although also increasing, trend.

Prescription opioid pain relievers can be highly effective in providing relief for patients suffering from chronic pain.

 At the same time, prescription opioid abuse, dependence, overdose, and poisoning (hereinafter “abuse”) have collectively become a national public health concern. Policy interventions, such as states’ prescription drug monitoring programs and rescheduling certain opioids, as well as technological innovations, such as the introduction of abuse-deterrent formulations, have been implemented with the goal of curbing opioid misuse while trying to maintain appropriate access to care for patients. 

Opioid abuse is also costly; societal costs of opioid abuse in 2007 were estimated at $55.7 billion.

 Payers face a large proportion of these costs; previous estimates of the annual per-patient economic burden to payers of opioid abuse range from $10,000 to $20,000.

 Despite growing concern, the current literature has only begun to address the drivers of these excess costs of opioid abuse in terms of place of service and comorbid conditions. 

This paper first seeks to replicate the findings of a newly published study performed using a commercial claims database from OptumHealth Care Solutions, Inc (Optum). The original study presented recent refined estimates of the excess costs of opioid abuse from a payer perspective and investigated the trajectory and drivers of those excess costs. The goal of this study was to replicate the findings using a different, more expansive database: the Truven MarketScan Commercial Claims and Encounter database (Truven). In addition, we extend those findings by presenting updated rates of diagnosed prevalence and incidence of opioid abuse, as well as the incremental per-member-per-month (PMPM) cost of abuse to payers.

We analyzed de-identified administrative claims data from Truven, which cover more than one-third of the US population (ie, 117 million commercially insured beneficiaries, including employees, children, spouses, and retirees) of large, self-insured companies located throughout the United States and in a broad array of industries and job types. The databases aggregate data from a large number of plan types (eg, fee-for-service and health maintenance organizations [HMOs]).

 The data include medical claims (service date, diagnoses received, procedures performed, place of service, and payment amounts), pharmacy claims (fill dates, national drug codes, and payment amounts), and eligibility information (patient demographic characteristics and enrollment history) for the period of January 2010 to September 2015. The Optum data used in the prior study contained a similar commercially insured population from different companies, industries, and regions. 

To ensure replicability, the methodology applied for the replication analysis paralleled that in the original analysis to the extent possible. Full details of the sample selection procedure and baseline characteristics can be found in the original study.

Incident abuse diagnoses between January 2012 and September 2015 were identified in the medical claims using the following 

) codes for opioid abuse, dependence, or overdose/poisoning: 304.0x, 304.7x, 305.5x, 965.00, 965.02, and 965.09.

 The index date for abusers was defined as the date of the first abuse, dependence, or overdose/poisoning diagnosis. For nonabusers, the index date was the date of a random medical claim during this same period. 

The follow-up period over which each individual was observed was defined as the 6-month period prior to the index date and 0 to 6 months thereafter. Including the 6 months prior to the index date in the follow-up period was important because abuse may result in greater use of medical resources prior to a formal diagnosis. Claims and costs 7 to 12 months before the incident diagnosis (the baseline period) were used in propensity score matching. A graphical depiction of the study period can be found in the eAppendix (

Two nonoverlapping cohorts of incident opioid abusers and nonabusers were selected. A 10% random sample of nonabusers was used. Patients in both cohorts had to be aged 18 to 64 years throughout the study period. In addition, we required continuous eligibility in non-HMO plans throughout the baseline and preindex follow-up periods (ie, up to and including the index date). After that, we followed patients for an additional 6 months, regardless of eligibility, to avoid bias from dropping patients who either died or lost healthcare coverage due to their abuse.

Propensity score matching of abusers and nonabusers was used to account for observable confounding factors associated with healthcare costs. A complete list of baseline characteristics, which were used as covariates in the propensity score model, and clinical codes can be found in the eAppendix. Abusers were matched 1:1 to nonabusers based on propensity scores ± one-fourth of a standard deviation using a greedy-matching algorithm.

 Baseline period characteristics for the postmatch cohorts were then compared to assess the quality of the match. 

We compared abusers’ and nonabusers’ average healthcare costs to payers during 6-month increments of the study period: 1) the baseline period (7 to 12 months before the index event), the preindex period (1 to 6 months before the index event), and the postindex period (0 to 6 months following the index event). Medical costs were categorized by place of service: inpatient, emergency department (ED), outpatient/other (eg, skilled nursing facilities), and rehabilitation facilities (costs attributed to rehabilitation facilities include those for care received in inpatient and outpatient substance abuse facilities, including all treatment received for other, both abuse and non—substance abuse-related conditions). All costs were inflated to 2015 US dollars using the medical care component of the Consumer Price Index. To assess cost drivers, the top diagnoses contributing to excess medical costs were reported by 3-digit 

We also calculated annual prevalence and incidence rates of opioid abuse for 2010 to 2014. Prevalence and incidence rates for 2015 were not available due to the incomplete year of data available at the time of the study. Prevalence rates reflect all patients with at least 1 claim for opioid abuse within a given year, whereas incidence rates reflect patients with incident diagnoses within a given year. The denominator for both incidence and prevalence comprises patients with at least 1 month of coverage within a year. 

The healthcare burden of opioid abuse is substantial; abusers often have complex healthcare needs and may require care beyond that which is required to treat abuse.

May 2017

Medicaid Prior Authorization and Opioid Medication Abuse and Overdose

State Prescription Drug Monitoring Programs and Fatal Drug Overdoses

Drivers of Excess Costs of Opioid Abuse Among a Commercially Insured Population

The Association of Mental Health Program Characteristics and Patient Satisfaction

Assessing the Effect of the VHA PCMH Model on Utilization Patterns Among Veterans With PTSD

Disparities in Diabetes and Hypertension Care for Individuals With Serious Mental Illness

Opioid Prescribing for Chronic Pain in a Community-Based Healthcare System

Critical Incident Stress Debriefing After Adverse Patient Safety Events

The Cost of Adherence Mismeasurement in Serious Mental Illness: A Claims-Based Analysis

Prescription Opioid Registry Protocol in an Integrated Health System