MONITORING OPIOID ABUSE AND DOCTOR SHOPPING: AN EPIDEMIOLOGIC PERSPECTIVE IN PERSON, PLACE, AND TIME By CHRIS DELCHER A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT O F THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2014
Â© 2014 Chris Delcher
To my sons, Ian and Lucas, and my wife, Diana
4 ACKNOWLEDGMENTS I have heard people refer to me a s a "non traditional" student. This is taken to mean a student who doesn't transition directly from an undergra duate program to a PhD program -someone who did "something else" between the two. It also means that 1) I am older than the average student fin ishing a dissertation and 2) I have had more time to encounter people in my life that have helped me get to this point. Unfortunately, it may also mean that I will forget to acknowledge many of them. I will do my best. I am writing the acknowledgements o n Mother's Day 2014 so I'll start with Mom and Dad. Thank you for believing in me and teaching me that I could do anything if I just tried hard enough. M y sister , Missi , for simply being the best little sister a big brother can have. M y best friends , To m and Farra Noel , whom I met at Florida State University. I didn't realize it at the time, but the many hours spent studying with you trained me well for this dissertation. More importantly, you taught me how to laugh at myself, another skill I needed in great quantities to finish it. Tom, you will always be my best man. M y Peace Corps friends and families from El Salvador that kept me safe and inspired me to pursue a career in public health. T he Wrangham family who brought me along on a journey to Uga nda that continues to inspire me to this day. M y friend Dan Koski who convinced me to move to Chapel Hill , North Carolina . My mentors and friends from the University of North Carolina at Chapel Hill. Especially, Yoshi Ampo, Cheolwoo Park and Jeongyoun A hn who helped and mentored me there, but, more importantly, became life long friends. M y friends and colleagues at the Virginia Department of Health. Jeff Stover gave me the flexibility to pursue research ideas in that environment. He hired me as an epi demiologist with little to no experience in epidemiology. At s ome point, the light flipped on , I realized that I needed more training,
5 and the PhD seed started to grow. My colleagues at Virginia Health Information and Michael Lundberg who really taught m e how to "know your big data" by example. M y friends and colleagues at NASTAD and Haiti, especially Gen Meredith, Mark Griswold, and Barbara Roussel who have helped me stay grounded in public health reality and entertained my research ideas. M y colleagues at the Prescription Drug Monitoring Program. Thank you for your insight during this My peers at the University of Florida, especially Taj Azarian, for the wonderful classroom interactions. Linda Cottler for getting me jump started with prescription drug surveillance. Betsy Shenkman for providing me with a fellowship , facilitating access to critical data, and ope ning the right doors for me. Alex Wagenaar, Bob Cook, and Bruce Goldberger, my committee members. As a "non traditional" student, your criticism has sometimes been a blow to my ego and a difficult "pill for me to swallow" (pun intended). I'm sure you've all seen that at some point during this process but I really appreciate your academic investment in me. Thank you for the many of the funny, non dissertation related conversations. M y mentor and chair Mildred Maldonado Molina. You waited two years for me to get to Gainesville. Oddly, you are not associated with this dissertation. Now wait, before you get worried about that last statement, let me explain that I mean that in the highest and most honorable epidemiologic sense. You are not a mere associ ation in this effort, but, rather, a direct cause in the completion of my journey. Without your
6 guidance, support, encouragement, and training, this dissertation would have never made it beyond the title page. You will always be welcomed to pizza night. M y beautiful boys, Ian and Lucas. (I put the beautiful part in to embarrass you when you get old enough to read this.) I can't tell you how much I have drawn inspiration from the little things. I still have your bookmark that says "SAS worker, may the F orce be With You", your ceramic coffee mug that says "SAS worker", your construction paper picture in orange and blue that says "PAPI"; your note that says that I'm the greatest graduate student in the world. You blindly believed in me because I'm your Da d. That's true love and that's what I needed the most. M y beautiful wife, Diana. (I put the beautiful part in because it's true.) This is just one of the outcomes (for two more see ) of a path that you and I started 16 years ago in El S alvador. In the long hours of the night in North Carolina, I can still remember the glow of your face in front of the computer screen helping me enter data for my m aster s degree. Next, I took you to live in the city with the highest rates of sexually tra nsmitted diseases in the country so that I could learn something about epidemiology . In our little blue house there, we started our family. Then, you moved to Gainesville by yourself, without a job and with two kids to help get us settled in for this jou rney. Hopefully, you knew what you were getting into from day one with this "Peace Corps guy . " Even so, I know this has been difficult in ways that we never anticipated. Through it all, you have been by my side, truly dedicated to me, Ian and Lucas, and our shared dreams as a family. I wish I could promise that great riches will come from this dissertation. The only thing I can really promise is to not get another PhD. I love you and I'm still your "Peace Corps guy . "
7 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ .......... 11 LIST OF FIGURES ................................ ................................ ................................ ........ 12 ABSTRACT ................................ ................................ ................................ ................... 14 CHAPTER 1 BACKGROUND AND SIGNIFICANCE ................................ ................................ ... 16 Epidemiology ................................ ................................ ................................ .......... 16 Prescription Drug Abuse ................................ ................................ ................... 16 Prescription Pain Reliever Abuse ................................ ................................ ..... 17 A Conceptual Model: Prescription Opioid Seeking Episodes and Outcomes ......... 19 Entry into the Healthcare System ................................ ................................ ..... 19 A Drug Seeking Episode ................................ ................................ .................. 21 The Use Spectrum ................................ ................................ ........................... 24 Defining Doctor Shopping from the Epidemiologic Perspective .............................. 25 Public Health Impact of Doctor Shopping ................................ ......................... 28 Risk Fact ors Associated with Doctor Shopping ................................ ................ 29 Prescription Drug Monitoring Programs (PDMPs) ................................ .................. 31 Study Goal, Aims, and Research Questions ................................ ........................... 33 2 PROGRAM ON OXYCODONE CAUSED MORTALITY: A MONTHLY TIME SERIES ANALYSIS, 2003 2012 ................................ ................................ ............. 40 Overview ................................ ................................ ................................ ................. 40 Florida ................................ ................................ ................................ .............. 41 PDMP and Opioid Mortality ................................ ................................ .............. 43 C urrent Study ................................ ................................ ................................ ... 44 Methods ................................ ................................ ................................ .................. 44 Research Design ................................ ................................ .............................. 44 Data Sources ................................ ................................ ................................ .... 45 Primary Effect ................................ ................................ ................................ ... 46 Independent Variables ................................ ................................ ..................... 47 Covariates ................................ ................................ ................................ ........ 49 Statistical Analyses ................................ ................................ .......................... 51 Parameter i dentification, e stimation, and m odel d esign ............................. 52 Secondary and s ensitivity a nalyses ................................ ........................... 54 Results ................................ ................................ ................................ .................... 55 Descriptive ................................ ................................ ................................ ........ 55
8 Uni variate Time Series Analysis ................................ ................................ ....... 56 Multivariate Time Series Analysis ................................ ................................ ..... 57 Discussion ................................ ................................ ................................ .............. 60 3 RISK FACTORS FOR DOCTOR SHOPPING FOR SCHEDULE II PRESCRIPTION OPIOIDS IN THE TEXAS MEDICAID POPULATION ................. 80 Overview ................................ ................................ ................................ ................. 80 Public Health Impact ................................ ................................ ........................ 81 Risk Factors Associated with Doctor Shopping ................................ ................ 83 Current Study ................................ ................................ ................................ ... 84 Methods ................................ ................................ ................................ .................. 85 Research Design ................................ ................................ .............................. 85 Dependent Variable (Doctor Shopping) ................................ ............................ 85 Independent Variables: Risk Factors ................................ ................................ 86 Overview of c ontextual d omains ................................ ................................ 86 Statistical Analyses ................................ ................................ .......................... 91 Results ................................ ................................ ................................ .................... 92 Study Population ................................ ................................ .............................. 92 Doctor Shoppers V ersus N on Doctor Sh oppers ................................ ............... 92 Bivariate Associations ................................ ................................ ...................... 94 Associations within Contextual Domain ................................ ............................ 94 Full Model ................................ ................................ ................................ ......... 95 Discussion ................................ ................................ ................................ .............. 96 4 RISK FACTORS FOR PRESCRIPTION OPIOID ABUSE AND DEPENDENCE IN THE TEXAS MEDICAID POPULATIO N, 2011 ................................ ................. 109 Overview ................................ ................................ ................................ ............... 109 Medicaid ................................ ................................ ................................ ......... 110 Risk Factors ................................ ................................ ................................ ... 111 Screening Tools ................................ ................................ ............................. 112 The Role of Administrative Claims B ased Studies ................................ ......... 112 Methods ................................ ................................ ................................ ................ 113 Research Design ................................ ................................ ............................ 113 Dependent Variable: Clinically R ecognized Opioid Abuse and Dependence . 115 Independent Variables: Risk Factors ................................ .............................. 115 Statistical Analyses ................................ ................................ ........................ 119 Results ................................ ................................ ................................ .................. 120 Prevalence of Opioid Abuse or Dependence ................................ .................. 120 Opioid Abusers V ersus Non Abusers ................................ ............................. 121 Bivar iate and Multivariate Associations ................................ .......................... 123 Domain Adjusted ................................ ................................ ..................... 123 Fully Adjusted ................................ ................................ .......................... 123 Receiver Operator Curve Characteristics of Two Models ........................ 124 Discussion ................................ ................................ ................................ ............ 125
9 5 CONCLUSIONS ................................ ................................ ................................ ... 138 Accomplishments of the Dissertation ................................ ............................. 138 Future Directions and Policy Recommendations ................................ ............ 140 Natio nal level Prescription Drug Monitoring Program Study .................... 140 Negative Health Outcomes Proximal to Mortality ................................ ..... 141 Improving Drug related Surveillance in Florida ................................ ........ 142 Prospective use of PDMP Data to Identify Prescription Opioid Abuse ..... 144 Network Analysis of Doctor Sho ppers ................................ ...................... 146 Provider level Effects on Doctor Shopping ................................ .............. 147 APPENDIX A DRUG DEATH REPORTING SYSTEM ................................ ................................ 150 B OPIOID AND BENZODIAZEPINE PRESCRIPTION DRUGS REPORTED TO .......... 151 C DE IDENTIFIED PATIENT ACTIVITY REPORT (PAR) REQUESTED FROM THE FLORIDA PRESCRIPTION DRUG MONITORING PROGRAM ................... 152 D NUMBER OF INACTIVE PAIN CLINIC LICENSES AND % CHANGE BY MONTH AND YEAR, FLORIDA (AS OF NOVEMBER 2013) ............................... 153 E PERCENTAGE CHANGE IN PURCHASES OF OXYCODONE BY FLORIDA PHARMACIES FROM 2010 TO 2011 ................................ ................................ ... 154 F GEOGRAPHIC DISTRIBUTIONS ................................ ................................ ......... 155 G THE VINCENTY EQUATION ................................ ................................ ................ 156 H PRESCRIPTION DRUG MONITORING PROGRAM OPERATIONAL STA TUS, 2011 ................................ ................................ ................................ ...................... 157 I ICD 9 CM CODES USED TO CLASSIFY SUBSTANCE USE DISORDERS ....... 158 J SCHEDULE II COMPARED TO OTHER PRESCRIPTION OPI OIDS ................... 159 K DISTINCT MULTIPLE PROVIDER EPISODE MATRIX ................................ ........ 161 L SPEARMAN CORRELATION COEFFICIENTS FOR CATEGORICAL VARIABLES IN THE HEALTH STATUS DOMAIN ................................ ................ 162 M ESTIMATED ODDS RATIOS FOR THE RELATIONSHIP BETWEEN MENTAL HEALTH OUTCOMES AND PAINFUL CONDITIONS VARIABLES, TEXAS MEDICAID 2011 ................................ ................................ ................................ ... 163
10 N ESTIMATED ODDS RATIOS AND 95% CIs FOR THE RELATIONSHIP BETWEEN POTENTIAL MEDIATING PATHWAYS AND OPIOID ABUSE/DEPENDENCE, TEXAS MEDICAID 2011 ................................ ............... 164 P THE CASCA DING EFFECTS OF PRESCRIPTION OPIOID ABUSE IN THE UNITED STATES ................................ ................................ ................................ . 166 Q EMERGENCY DEPARTMENT VISIT RATE FOR NONMEDICAL USE OF PHARMACEUTICALS, MIAMI FORT LAUDERDALE, 2008 2011 ....................... 167 R SCREEN SHOT OF THE INTERACTIVE DATA QUERY SYSTEM CREATED RELATED DEATH SURVEILLANCE SYSTEM ................................ ................................ ................... 168 S THE NUMBER OF FLORIDA PATIENTS IDENTIFIED AS DOCTOR SHOPPERS BY QUARTER, 2012 2013 ................................ ............................... 169 T RESPONSES FROM FLORIDA PRESCRIPTION DRUG MONITORING PROGRAM SURVEY ................................ ................................ ............................ 170 REFERENCES ................................ ................................ ................................ ............ 171 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 189
11 LIST OF TABLES Table page 2 1 Characteristics of select drug ................................ ................. 73 2 2 Characteristics of oxycodone caused deaths in Florida, 2003 2012 .................. 74 2 3 Evaluation and selection of candidate univariate models for oxycodone cause mortality in Florida, 2003 2012 ................................ ................................ . 75 2 4 Bivariate associations of oxycodone caused mortality with covariates ............... 76 2 5 Model building steps for estimating the effect of the Prescription Drug Monitoring Program (PDMP) on oxycodone caused mortality ............................ 77 2 6 Univariate ARIMA modeling and bivariate model results for other licit and illicit drugs ................................ ................................ ................................ ........... 78 2 7 Estimates for potential reductions in proximal outcomes related to the PDMP using CDC based estimates for all opioids ................................ ......................... 79 3 1 Descriptive statistics of Texas Medicaid enrollees exposed to a Schedule II prescription opioid by doctor shopping (4 or more prescribers and 4 or more dispensers) status, 2011 ................................ ................................ .................. 104 3 2 Bivariate associations with doctor shopping by contextual domain, Tex as Medicaid, 2011 ................................ ................................ ................................ . 106 3 3 Adjusted associations between variables within contextual domains, Texas Medicaid, 2011 ................................ ................................ ................................ . 107 3 4 Risk factor model for doctor shopping for Schedule II prescription opioids, Texas Medicaid, 2011 ................................ ................................ ...................... 108 4 1 Descriptive statistics for Texas Medicaid enrollees exposed to at least one Schedule II pr escription opioid with clinically recognized opioid dependence/abuse, 2011 ................................ ................................ ................. 130 4 2 Bivariate associations between risk factors and having clinically recognized opioid abuse and dependence, Te xas Medicaid, 2011 ................................ ..... 133 4 3 Contextual domain modeling results. Odds ratios are only adjusted for other variables with the contextual domain block ................................ ...................... 134 4 4 2011 and 2012 modeling results ................................ ................................ ....... 135
12 LIST OF FIGURES Figure page 1 1 The conceptual framework for the current study. ................................ ................ 36 1 2 Population health surveillance process components. ................................ ......... 37 1 3 The prescription drug monitoring surveillance system with key proces ses and stakeholders. ................................ ................................ ................................ ...... 38 1 4 Literature review summary of doctor shopping threshold parameters, prevalence and other study characteristics, 2004 2013. ................................ .... 39 2 1 Monthly counts of oxycodone caused mortality in Florida, January 2003 ( t=1 ) to December 2012 ( t=120 Commission. ................................ ................................ ................................ ....... 65 2 2 Exp ected monthly count of oxycodone caused mortality (squares) overlayed on observed counts within the year (circles). ................................ ...................... 66 2 3 The natural log transformed and differenced (1 st order) oxycodone ca used mortality series, 2003 2012. ................................ ................................ ............... 67 2 4 Natural log transformed counts of oxycodone caused mortality (line) overlayed on observed counts (points). ................................ .............................. 68 2 5 Expected value of oxycodone caused mortality by month of the year. Observed values for each month in each year (10 years) is shown (circles). Note: In sensitivity analyses, we specified this seasonal pattern as an ARIMA(0,1, 1)x(0,1,1) 6 parameter. ................................ ................................ ...... 69 2 6 Autocorrelation function (ACF, top) and partial autocorrelation (PCF, bottom) plots for oxycodone caused mortality with 95% confidence interval bands. Note: The slow linear decay in the ACF indicates trend. The PCF indicates that monthly values of oxycodone caused mortality are significantly correlated out to the 3 rd time lag. ................................ ................................ ........ 70 2 7 Autocorrelation function (ACF) and partial autocorrelation function (PCF) plots of the oxycodone caused (transformed, differenced) mortality series. Note: In the ACF plot, the significant correlation at lag(1) suggests a 1 st order moving average model [ ARIMA(0,1,1) ]. I n the PCF plot, significant correlation at lag(1) lag(3) suggests an 2 nd or 3 rd order autoregressive model [ ARIMA(2,1,0) or ARIMA (3,1,0) ]. ................................ ................................ ....... 70 2 8 Univariate ARIMA(0,1,1) model (black) of obs erved oxycodone caused mortality (circles) with 95% confidence interval bands (grey). ............................ 71
13 3 1 Flow chart for variable sources, inclusion criteria, and sample size in this study, Texas Medicaid , 2011. *Linked to NPI Standard Database from Centers for Medicare and Medicaid Services to obtain practice location and geocoded **STAR, STAR+PLUS, NorthSTAR, Fee for service, and PCCM.***ICD 9 CM 140.0 208.9****Place of service codes (31, 32, 33, and 54) . Enrollees with a CPT code indicating methadone maintenance therapy were excluded (n=59). ................................ ................................ ...................... 102 3 2 Random sample of 12 doctor shoppers and their Schedule II opioid prescribers (4 or more disti nct prescribers, dispensers not shown) in the Texas Medicaid population, 2011. Blue=enrollee, red=prescriber. In this illustration, six doctor shoppers are connected to 1 prescriber (in box) and three doctor shoppers (on perimeter) are not connected to o ther doctor shoppers via any prescriber. ................................ ................................ ............. 103 4 1 Forest plot of adjusted odds ratios for 2011. ................................ .................... 136
14 Abstract of Dissertation Presented to the Graduate Schoo l of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy MONITORING OPIOID ABUSE AND DOCTOR SHOPPING: AN EPIDEMIOLOGIC PERSPECTIVE IN PERSON, PLACE, AND TIME By Chris Delcher August 2014 Chair: Mildred Maldonado Molina Major: Epidemiology Prescription drug abuse is a national epidemic in the United States. P rescription pain relievers , in particular, are highly abused despite regulatory oversight to control their distribution. Ele ctronic databases, known as prescription drug monitoring programs (PDMPs) , are designed to monitor the prescribing and dispensing of controlled subs tances , including prescription pain relievers , and identify patients who inappropriately seek drugs from mul itple providers doctor shopping . Despite national calls for PDMP implementation, little is known regarding their effect on reducing deaths from prescription pain relievers like oxycodone. The prevalence and risk factors for doctor shopping and presc ription opioid abuse or , is not well studied either. To address these gaps in the literature, w e conducted three studies to address the role of PDMPs on prescription opioid mo rtality , the risk factors associated with doctor shopping and , likewise, for diagnosed cases of opioid abuse and dependence in a large Medicaid population . First, u sing a quasi experimental , interrupted time series
15 analysis , we found a n immediate 25 perce nt decline in oxycodone caused mortality after the implementation of the Florida PDMP. Secondly, u sing enrollment, encounter, and pharmacy data from Medicaid, we examined a population of doctor shoppers using an ecologic framework with risk factors char acterized into multiple domains . We identified several unique risk factors including greater risk of being identified as a doctor shopper when providers have recent encounters with other doctor shoppers. Further research on the nature of doctor shopping networks is warranted . The prevalence of opioid abuse or dependence among adult enrollees in the Texas Medicaid program exposed to a Schedule II prescription opioid in 2011 was 6.3 percent . Adopting an epidemiologic approach, w e identified several unique r isk factors including doctor shopping behavior , having a major mental health disorder, and a co morbid non opioid substance use disorder. Our models highlight the importance of treating mental health conditions in addition to p ain. Using a PDMP like mode l, we found evidence that the data available to PDMPs may be a helpful screening tool for the challenging task of identifying prescription opioid abuse or dependence at the clinical level.
16 CHAPTER 1 BACKGROUND AND SIGNIFICANCE Epidemiology Prescription Drug Abuse Therapeutic use of prescription drugs is highly prevalent in the United States. According to the US Centers for Disease Control and Prevention (CDC), 48.5% of the US population used at least one prescription drug in the past month during the pe riod 2007 to 2010. 1 The percentage of the US population using 1 or more, 2 or more, and 5 or more prescription drugs has increased linearly in each category over the past decade. 2 Not surprisingly, increasing trends in prescription drug use have given rise to an increasing number of negative population health outcomes. Prescription drug related overdoses a nd death have reached epidemic proportions in the United States. 3 5 In 2010, the leading cause of injury related death for 25 to 65 year olds was unintentional poisoning from the ingestion of prescription drugs. 6 In 2008, for the first time, t he poisoning mortality rate per 100,000 population surpassed that of fatal car crashes in the United States. 7 In Florida, one of the US states we examine in this dissertation, it has been estimated that 8 people die per day of a prescription drug overdose with national estimate s at 105 deaths per day. 8,9 From 1997 to 2008, inpatient hospitalizations among 18 to 24 year olds for overdoses rel ated to alcohol, drugs, or a combination of the two increased by 25%, 56%, and 76%, respectively. 10 The c lass of prescription drugs , known as psycho th erapeutic s , that target the central nervous system (i.e., sedatives, stimulants, tranquilizers, and pain relievers) are of particular concern because t hese drugs are high ly addictive and frequently abused. 11 According to the 2012 National Survey on Dr ug Use and Health ( NSDUH ), the
17 percentage of the US population that had ever used sedatives, stimulants , tranquilizers , and pain relievers was 2.9%, 7.4%, 9.1%, and 14.3%, respectively . 12 This dissertation is specifically focused on mortality outcomes and the addictive behaviors associated with pain relievers. Prescription Pain Reliever Abuse Pain relievers (also known as analgesics , pain killers, narco tics) are indicated for the treatment of acute and chronic pain . The active ingredients are opioids which function by binding to opioid receptors in the brain to reduce the perception of pain. It is important to note that the most powerful pain relievers and the illegal drug heroin have essentially the same chemistry and bind to the same receptors in the brain. 13 Pain relievers have become the most frequently prescribed class of all medications regardless of whether patients present to physician offices, outpatient settings, or emergency departments. 1 The relatively high prevalence of pain reliever use reported by NSDUH is consistent with, but not entirely explained by, the high prevalence of chronic pain in the US population which has been estimated at approximately one third of the adult population. 14 Two studies by Mazer Amirshahi et al. in adults and adolescents demonstrate that opioids have been prescribed at higher rates than the prevalence of pain in emergency departments would warrant. 15,16 National estimates for specific chronic painful conditions requiring prescription opioids are lacking but chronic lower back pain increased from 3.9% to 10.2% f rom 1996 to 2006 in a representative household study in North Carolina. 17 Oxycodone ( e.g. brand names OxycontinÂ®, PercocetÂ®) is one of the most p owerful active ingredients in the prescription opioid family which also includes prescription hydrocodone ( e.g. brand name VicodinÂ®) . 18 R etail sales of oxycodone
18 increased by 866% between 1997 and 2007. 19 In 2005, h y drocodone surpassed cholesterol lowering drugs to become the most prescribed medication in the United States . 20 A s the prevalence of chronic painful conditions and medical availability of powerful pain relievers has increased over the past decade so too has the ir misuse and abuse . According to the National Institute on Drug Abuse (NIDA) , in 2010 , 2.7% of the US population was taking a psychotherapeutic drug nonmedically , with the pain relievers most frequently abuse d . 11 According to Healthy People 2020, the past year preval ence of non medical use of pain relievers in the US population 12 and over was 4.3% in 2011. 21 According to the CDC, 55% of those that abused prescription opioids obtained them from a friend or relative and 17.3% obtained them from one prescri ber. 22 We n ote that the prevalence of pain reliever abuse will differ according to the definition used. NIDA defines prescription drug abuse without a prescription; in a way other than as prescribed; or for the experience or feeling elici nonmedical use. 23 In this dissertation, we operationalize d a definition of opioid abuse and dependence based on the International Classification of Dise ases, Ninth Revision, Clinical Modification (ICD 9 CM). 24 We discuss the strengths and limitations of this definition in Chapter 4. The national epidemic persists d espite the fact that pain relievers with the highest abuse potential (e.g. Oxyc odone) are strictly controlled and monitored by the U nited potential. 25 The Controlled Substance Act Schedule ranges from 1 (illegal drugs) to 5 (controlled drugs with least abuse potential) with Oxycodone being a Schedule II
19 (high est abuse potential for a legal drugs ). T here is a chain of population level health effects associated with the misuse and abuse of opioids. T he CDC estimates that for every prescription opioid related death, there are 10 treatment admissions for abuse, 32 emergency department visits for mis use or abuse, 130 people who abuse or are dependent, and 825 non medical users of prescription opioids. 22 This dissertation examines three negative health outcomes associated with Schedule II prescription opioids , including o xycodone caused mortality, aberrant drug seeking behavior known as doctor shopping , and clinically recognized opioid abuse or depende nce. Further dis cussion of the epidemiology of prescription opioid misuse and abuse is found the introductory material of each chapter . One of the major concepts presented in this dissertation involves compulsive drug (opioid) seeking behavior and measuring opioid abuse and dependence using healthcare encounter data. The next section presents a conceptual model to help understand this process . A Conceptual Model : Prescription Opioid S eeking Episodes and Out comes I t is helpful to begin with a conceptual model for in the US health care system (see Figure 1 1 ) . The complete model is adapted from components developed by the Institute of Medicine , relevant literature , and knowledge of how controlled substances are monitored in the population . 26,27 Entry into the Healthc are System T he model shows a patient ( P i ) that ente rs the health care system seeking a prescription opioid for legitimate chronic pain . T he model used in this dissertation implicitly assumes that m ost patients seek pain medication for a bona fide medical condition. 28 Keller et al. found that 81.3% of physicians believed that "legitimate pain"
20 explained why most op ioid dependent patients began using opioids. 29 Even so , the mod el represents an open system whereby il legitimate patien ts (e.g. street users) may enter and legitimate patients can exit to obtain opioids from non medical sources. 30 We focus on c hronic pain here because the underlying painful conditions (e.g. arthritis, back pa in, headaches) can require chronic opioid therapy increasing the risk of negative health outcomes from long term exposure. 14,31 The American Society of Interventional Pain Physicians provides a complete working definition of chronic pain . 32 For the dissertation , we examine three painful conditions associated with chronic pain and apply the distinction made in the literature between cancer related and non cancer chronic pai n (CNCP) . 33 Thus, for this dissertation we exclude patients with cancer related pain. Understanding the etiology and risk factors associated with CNCP is beyond the scope of this study but we note the complex interactions of pathological, historic, social, cognitive, and emotional risk factors associated with pain for a given individual as conceptualized by Turk et al . 33 W e represent a diagnosis of pain only ( Dx (PAIN) ) or pain in the presen ce of a co morbid condition ( Dx (PAIN) + Dx (N) ) where N represents any number of underlying health problems such as a major mental health disorder . Mental health problems (e.g. depression, anxiety) have been associated with both CNCP and opioid use. 34,35 M ediating and moderating factors are illustrated at nearly all stages of the model . For this dissertation, we operationalize evidence ba sed risk factors into domains using variables available in Medicaid administrative claims data. Further details and complete definitions for all variables are provided in each methodology section.
21 A Drug Seeking Episode When a patient enters the system seeking a prescription opioid , an episode ( E i ) with a provider ( Prov j ) begins. T he ter often refer to any healthcare professional licensed to prescribe or dispense a controlled substance. This dissertation differentiates between prescriber s (e.g., emergency medicine physician, family practitioner, advanced nurse practitioner , dentist) and dispenser s (i.e., pharmacist). We do so because the concept of s , which is used as an indicator of aberrant drug see king behavior known as doctor shopping , requires prescribers and dispensers to be counted independently (see b elow for complete definition). Also, p rescribers and dispensers generally operate in different practice locations in geographic space . The model specifically includes this parameter because understanding how drug seekers operate over geographic distances is addressed in this dissertation. During a normal healthcare encounter, t he prescriber pe r forms a med ical evaluation to decide whether to pres cribe an opioid or not. This decision process is represented as a function of a multiple variables ( f[Z] ) which could includ pain levels, urine screens, behavior or state of agitation, socio demographics , history of substance abuse and menta l health disorders , and prescription history ( f[ Rx h ] ). 36,37 The may be incomplete if the patient has visited multiple providers in different healthcare system or states, and/ or withholds prescription related information (intentionally or not). In theory, the provider can obtain the most complete controlled substance history by
22 One of the central questions of this dissertation is whether PDMPs are effective tools for reducing mortality associated with prescription opioids. Briefly, PDMPs are large databases designed to provide complete controlled substance information to providers. The d ecision to check the PDMP is itself a complex function of factors such as legal mandates, provider awareness of the database , registration status, perceptions of usefulness, and user friendliness. Each of these factors have been cited in the literature as barriers to PDMP usage. 38 When the barriers are minimized , the provider initiates an electronic PDMP q uery and obtains the controlled substance history . Providers that use PDMPs tend to 1) find them useful , 2 ) change their prescribing patterns, and 3 ) make clinical decisions as opposed to legal decisions . 39 41 This process of obtaining information from the PDMP can be described as reactive or proactive (a.k.a. ) . The p rocess is proactive if providers proactively detected. In many states, an example multiple provider episodes for the same controlled substance within a certain window of time. At the close of the provider patient episode , a medical billin g claim is filed to an insurance system to report the relevant information associated with the encounter including patient socio demographics, diagnostic conclusions , and information to identify providers. The term IC repr esents any of several major insuran ce claims system s in the Compensation, and private insurance companies . If the prescription is not written , then
23 the episode ( E i ) can end or the patient can seek another p rovider or therapy ( E i+1 ), or the patient may otherwise attempt to circumvent the traditional healthcare system ( E ix ). The focus of this dissertation is on Medicaid but our methods are applicable to nearly any of the administrative systems mentioned above. If the prescription is written, the patient must fill it at a dispensing pharmacy ( PHARM k ). Dispensers decide whether or not to dispense the controlled substance; representing the final control point in obtaining the medication . Like the prescriber, pharmacists in many states have access to a PDMP . In Florida, examined in this dissertation, e d users are pharmacists. 42 A li ne of communication be tween prescribers and di spenser s is explicitly shown in the model because they interact with each other around dispensing decisions. If the prescription is not filled, the patient may re engag e the health care system as described above. If the prescription opioid is filled, then the pharmacy episode is complete and prescription related parameters (e.g., date of fill, dosage , days supply) are transmitted to multiple data systems including a prescription information exch ange for large retail pharmacies ( e.g., Walmart Â® , personal communication), 3) the insurer, and 4) the state PDMP . Through time, multiple provider episodes (MPE) can be measured from the pharmacy transactions in any or all of these systems . Clarifying th e definition of multiple provider episodes (MPE) is central to this dissertation because , from the epidemiologic perspective, this concept is treated as a proxy for the type of compulsive drug seeking associated with addictive behavior ( see next section ) w hen MPEs are high . We will discuss the risk factors associated doctor
24 shopping behavior in more detail after completing our discussion of the conceptual model. The Use Spectrum Once the patient has obtained the prescription opioid , the complex system of therapeutic use, pain relief, tolerance, dependence, misuse , abuse, addiction, and diversion begins . We note that d efining each of the terms remains an on going challenge for researchers in this field. NIDA distinguishes the physiological phenomena of physical dependence from addiction because the latter induces compulsive drug seeking behavior. 13 Tolerance is defined as the need to increase dosage to achieve the desired effect. Prescription opioid abuse w as described by Compton and Volk ow bona fide this definition differs from the Diagnostic and Statistical Manual, Fourth Edition (DSM IV). 43 The optimal outcome of therapeutic opioid use is effective treatment of pain and , if accomplished, the patient disengages from the healthcare system for this episode. If the pain is not treated or under treated, the patient may 1) seek alternative treatment, 2) live with the pain, 3) develop complications, or 4) experience worsening pain. 26 Medical misuse rs of prescription opioids can include patients that intentionally 1) self escalate their d ose, 2) change the route of administration, 3 ) give the opioid to a relative/friend who needs it, 4 ) obtain additional prescription opioids without a prescription , 5 ) use for a non bona fide medical reason and /or 6 ) sell pills on the illicit market for eco nomic gain (i.e., diversion). 13,44 Preventing medical misuse is complex and this dissertation will show that independent actions taken by stakeholders from
25 public health ( e.g. symptom recognition), industry (e.g. tamper resistant formulations), and l aw enforc ement (e.g. diversion control) play a role in chang ing population level outcomes. This dissertation does not directly address the complex interactions between these entities but the model illustrates this concept. For example, healthcare providers modify their prescribing patterns if they feel they are going to be singled out by law enforcement as prescribers. 29,45 Healthcare p ro viders can interact directly with law enforcement by report ing their patients directly to the DEA. A study in West Virginia, found that 37% of physicians had ever reported a doctor shopper to law enforcement. 46 As mentioned ab ove, prescription opioid use can lead to multiple health outcomes including a lethal overdose . Opioid use can result in a fatal respiratory depression where breathing slows, oxygen levels drop, and the brain and other vital organs stop functioning. 47 In Florida, one of the states examined here , a medical examiner investigates fatal overdoses and reports results to a drug related surveillance system m anaged by the Florida from this system in Chapter 2 to examine oxycodone caused deaths in Florida. The following section provides more details on two key concepts in this dissertation : doctor shopping and prescription drug monitoring programs . Additional details are provided in Chapters 2 4. Defining Doctor Shopping from the Epidemiologic Perspective A single provider episode occurs when a patient receives or fills a prescription opioid from a si ngle provider . M ultiple provider episodes can either be a series of prescriptions obtained from the same provider or different providers. Patients that engage in a large number of distinct multiple provider episodes (dMPEs) , with the
26 exception of cancer p atients, 37 The on of doctor shopping also implies tha t healthcare practitioners are 48 Thus, another important dimension of doctor shopping could be the presumption of fradulence. 48 While the operational definitions of doctor shopping vary by study, typical definitions include four parameters from longitudinal pharmacy transactions. These parameters include: 1) an active ingredient (e.g., opioid, benzodiazepine); 2) the Controlled Substance Act Schedule (e.g., Schedule II or I II); 3) the timeframe (e.g., 90 days); and 4) the number of distinct providers , either prescri bers, dispensers, or both. From the legal perspective, doctor shopping is usually treated as an activity motivated by the desire to divert medication. 49 National Drug Control Policy (ONDCP) and the 2013 National Drug Control Strategy lth issue, not just a criminal justice 50 This quote summarizes one of the guiding principles of this dissertation. From the epidemiologic perspective, doctor shopping is viewed as a dimension of compulsive drug seeking behavior driven by addiction. We acknowledge t hat s ome research shows that some patients are driven to doctor shop for prescription opioids both medical and non medical reasons . Davis and Johnson found that one type of street drug users in New York City sought prescription opioids from doctors for eu phoric purposes only while other groups of street drug users sought opioids for legitimate medical purposes but sold them as well. 30 Here, we explicitly take the epidemiologic perspective by examining patient health related and other healthcare structural risk fact ors associated
27 with doctor shopping in Chapter 3. We believe that the epidemiologic approach to this problem provides a more nuanced perspective on doctor shoppers that will improve support for public health prevention efforts in the long term . The epide miologic perspective highlights several key concepts : 1) some patients engage in high multiple provider episodes for legitimate purposes (e.g., cancer care, palliative care), 2 ) doctor shopping can occur in the context of legitimate medical need, and 3 ) c o morbid conditions are expected to be associated with doctor shopping particularly substance abuse and mental health problems (see substance abuse loop in the model ). 49 In Chapter 4, we specifically examine the association between clinically recognized opioid abuse and dependence and doctor shopping. The doctor sho pping literature frequently uses administrative claims data to study this behavior. In Chapter 4 we provide additional discussion of the role of claims data in this field. In our review of twelve studies using claims data to identify and study doctor sho pping from either the peer reviewed literature or government sources, five came from PDMP populations, five from privately insured populations, one from Medicare, and one from the Medicaid population sampled from five large US states. 51 53 The purpose of the latter report was to provide case studies of how doctor shoppers operate and quantify the excess payments associated with those enrollees. 51 The report did not attempt to identify any risk factors associated with doctor shopping in the Medicaid population. Given that Medicaid is the largest public insurer in the United States, covering more than 62 million Americans, we believe this is an important gap in the literature that this dissertation is d esigned to address. 54 Figure 1 4 shows the results of our review of the doctor shopping literature.
28 Public Health Impact of Doctor Shopping Doctor shopping for prescription opioids has been associated with both illicit drugs and pres cription drug overdose mortality . 55 58 The American Society of Interventional Pain Physicians identified doctor shopping as a major contributing factor to prescription drug fatalities. 32 Those that doctor shop for prescription opioids are exposed to high doses of daily morphine milligram equivalents which increases the ir risk of opioid overdose. 5 6,59 61 During the intensification of the current prescription opioid epidemic in the United States , doctor shopping has also increased in some states. 5 For example, in California, the prevalence of doctor shopping for Schedule II prescription opioids increased by 111% (18 44 year old males) to 213% (65+ females) during the period 1999 to 2007. 60 Doctor sho ppers for prescription opioids also frequently receive prescriptions for multiple classes of controlled substances which can increase the risk of adverse medication interactions and overdose. 62 In addition to the direct effects of opioid ex posure, doctor shoppers may also represent a subset of at risk patients that are being coerced into doctor shopping because they are poor, elderly, or otherwise vulnerable (e.g., HIV patients) to drug seekers who want their medications. 63 One study young men that eventually died of a heroin overdose. 58 Doctor shopping is a significant source of global pharmaceutical diversion , although to what extent is an area of active research. 64 66 While the prevalence of doctor shopping reported in the literature is relatively low (<1.0%), the economic and psychological effects of doctor shopping across the US healthcare system are widespread. 51 53,60,62,67,68 In 2007, the US Government Accountability Office reported $63 million in excess Medicaid claim payments associated with doctor shoppers. 51 In
29 West Virginia, 40% of all medical specialties surveyed and 90% of emergency departm ent (ED) physicians reported a suspected encounter with a doctor shopper on a weekly basis. 46 ED physicians in Massachusetts suspected that at least 30% of their patients were doctor shoppers. 69 Physician interaction with doctor shoppers may increase thei r fear of 1) being scrutinized by regulatory agencies, 2) contributing to increased abuse or addiction potential for patients with a history of substance abuse, and 3) increasing community level diversion. 45,70 Clinicians express frustration with the challenges of managing patients with chronic pain; doctor shoppers likely contribute to this frustration, in part, because they are difficult to identify with certa inty . 71 In West Virginia , a third of physicians did not their medical assessment. 46 One of the principle functions of state PDMPs, now operating in 48 US states, is to help healthcare providers and state authorities identify doctor shoppers. 72 However, the thresholds that many PDMPs use to define doctor shopping are designed to detect the most extreme cases and they are not designed to provide health related in formation . Earlier identification of the risk factors for doctor shopping at multiple levels remains a public health imperative and one of the objectives of this dissertation . Risk Factors Associated with Doctor Shopping Very little is known about the r isk factors associated with doctor shopping. Clinical level r isk factors include requesting medications by name, multiple visits for the same complaint, a suspicious history, and having symptoms out of proportion to examination results. 69 A t the population level , research has largel y focused on informatio n contained in administrative claims data in the public and private sectors
30 (e.g., Medicaid, Medicare, private insurance). 51 53 Wilsey et al . found younger age was asso ciated with multiple provider episodes but not gender. 62 Likewise, Han et al . found 18 44 year olds had the highest prevalence of doctor shopping for Schedule II opioids in California with no difference by gender. 60 Interestingly, a recent unpublished study from the New York City Department of Health and Mental Hygiene did not find any demographic risk factors associated with doct or shoppers , although the variables analyzed were not specified in the report . 73 One of the key concepts of this dissertation is geographic risk factors , therefore we provide an extended discussion below. Geographic Risk Factors . Understanding the spatial characteristics and geographic risk factors of doctor shopping is an emerging area of research and empirical evidence in this area is limited. Cepeda et al. shoppers traveled a median of 200 miles to fill opioid prescriptions and one fifth visited more than one state. 74 Wilsey et al. found that residents of non metropolitan areas were less likely to have multiple p rescribers than metropolitan residents. 62 Weiner et al. noted that drug seeking behavior for prescription drugs varied widely even between two urban academic health centers that were only 1.5 miles apart. 69 Cicero and Ellis found that 25% of patients seeking Tramadol , an opioid analgesic, from online pharmaciesdid so because they could not find a prescriber who would give it to them. 75 D octor shoppers seem to be aware of their own reputation among their geographically local healthcare providers and are known to cross state lines to obtain prescription pain medications. 74,76 well known destinations for drug seekers from many states. 77 Healthcare providers appear to be suspicious of possible doctor shopping when patients
31 visit their clinics from distant counties seeking controlled substances (R Cook, personal communication) or claim to be fr om out of town . 78 Several organizations are experimenting with the use of distance based crit eria to make dispensing decisions. T he American Academy of Pain Management recently pharmacists for their use in identifying possible doctor and the pharmacy (e.g. a patient from San Antonio, seeing a doctor in Houston, and filling the 79 At least one major retail pharmacy, Walgreens, appears to be using distance based criteria for a controversial checklist for deciding whether to dispense pain medication s . 80 The importance of research that quantif ies spatial risk factors is underscored by the fact that the geographic mobility of doctor shoppers is one of the primary issues driving the efforts to create regional and national level PDMPs that share prescription data across state lines. 81,82 Furthermore, understanding spatial variation in doctor shopping may contribute to improved understanding of community level patterns in prescription opioid overdoses or signal other trends, such as communities experiencing high rates of drug diversion. 83 85 Prescription Drug Monitoring Programs (PDMPs) S tate PDMPs are complex data systems designed to gather, centralize and disseminate information on prescription medications with high potential for abuse and diversion. 86 nt and public safety, providing patient level reports only to agencies operating with these mandates. 87 explicitly trying to implement goals with a public health component. 88
32 Unde rstanding PDMP effectiveness in reduc ing negative health outcomes, such as prescr iption drug mortality, remains a national priority. 28 Figu re 1 2 shows the theory of the public health surveillance process. 8 9 Surveillance is defined as the systematic, ongoing, collection, management, analysis, and interpretation of data followed by the dissemination of these data to stakeholders to stimulate action. 90 Surveillance is different than monitoring because, in theory, surveillance is an epidemiologically intelligent p rocess leading to actionable information (i.e., patients engaged in harmful activity) whereas monitoring may be a crude data collection process that does not provide useful information. PDMPs systematically collect controlled prescription drug informatio n, often provide a web based interface for disseminating these data to prescribers and dispensers (i.e., healthcare providers), law enforcement, and public health officials. In theory and practice , health care providers analyze and interpret these pharmac y transactions to determine if their patients are 1) using particular controlled substances, 2 ) visiting multiple health care providers for similar medications (i.e. doctor shopping), 3 ) at risk for adverse drug drug interactions, and/or 4 ) compliant with clinical protocols limiting patient use of particular drug classes . Then, providers take positive action to reduce the risk of negative health outcomes , such as overdose, by modifying prescribing decisions that would have otherwise increased medication ex posure . Law enforcement and agency level public health officials generally collaborate to identify and reduce doctor shopping behavior with no intervention in medication management. F igure 1 3 shows the PDMP surveillance system and interactions among key stakeholders .
33 The ONDCP recommends implementing PDMPs as a mechanism for reducing prescription drug abuse. 91 Other major public health organizations that support the use of PDMPs as a strategy for addressing the epidemic including the Centers for Medic are and Medicaid Services, the American Medical Association, and the Robert Wood Johnson Foundation. 82,92,93 There is limited research on whether PDMPs are effective for improving population level outcomes and even less attention has been paid to their screening characteristics an d clinical performance . 81 Additional details on PDMPs can be found in Chapter 2. Study Goal, Aims, and Research Question s The specific aims of this dissertation are: ( Study 1 ) To m easure the effect of a state PDMP on oxycodone caused mortality . W e use a quasi experimental design using the monthly time series of oxycodone caused deaths in Florida from 2003 to 2012. We develop autoregressive moving average (ARIMA) models of oxycodone caused deaths and measure the change in an indicator and continuous variable representing th r possible explanatory effects. We hypothesized that the PDMP would reduce oxycodone caused mortality. Specific pathways might include 1) identif ication and prevent ion of doctor shoppers from di verting deadly prescription opioids to street markets, 2) help ing healthcare providers detect aberrant drug use patterns, and 3) help ing healthcare providers detect drug drug interactions that would otherwise prove fatal. ( Study 2 ) To e xamine the risk fact ors associat ed with doctor shopping for Schedule II prescription opioids . W e use a cross sectional, observational study design
34 using enrollment, encounter, and pharmacy claims data from the Texas Medicaid program in 2011. We categorized risk factors int o five contextual domains based on the epidemiologic perspective: 1) sociodemographic, 2) health status including psychiatric and painful comorbid conditions, 3) provider episodes , 4 ) medication profiles and 5) spatial relationships. Briefly, s ociodemogr aphic variables included enrollee age, sex, race/ethnicity, and metropolitan status. Enrollee health status variables included diagnosed opioid abuse or dependence, non opioid abuse or dependence, Clinical Risk Groups (CRGs), mental health disorders, and p ainful conditions. Provider episode variables include Medicaid program type, total number of healthcare encounters, and prescriber encounters with other doctor shoppers. Medication profiles include overlapping prescriptions of benzodiazepines and daily d ose of morphine milligram equivalents. Spatial variables include distances between enrollees, prescribers, and dispensers and using out of state prescribers. Detailed definitions of each variable are found in the Methodology for Chapter 3 . Study 3 . To e x amine the risk factors associated with diagnosed opioid abuse or dependence among those exposed to Schedule II prescription opioids . First, w e use a cross sectional, observational study design using enrollment, encounter, and pharmacy claims data from the Texas Medicaid program in 2011. In addition to examining the risk factors modeled in Study 2, we examine the association of opioid abuse and dependence with doctor shopping. Second, we compared the relative performance of two different risk factor models: 1) the complete Medicaid model and 2) a PDMP like model (i.e., the full
35 Medicaid model minus the health status domain because diagnostic information is not available to many prescription drug monitoring programs.) We also performed a post hoc pat h analysis to examine pain and non opioid substance abuse as potential mediators between mental health disorders and opioid abuse or dependence.
36 Figure 1 1. The c onceptual f ramework for the current study .
37 Figure 1 2. Population health surveillance pro cess components. Adapted from: El Allaki F, Bigras Poulin M, Michel P, Ravel A. A population health surveillance theory. Epidemiol Health 2012; 34: e2012007. doi:10.4178/epih/e2012007 ( P age 4, Figure 2 ) Step 1: a. Dissatisfaction (Rx abuse epidemic ) b. Information need ( What is the nature of this epidemic? ) c. Motivation to act (National crisis) Step 2: Problem Formulation (need for Rx drug surveillance system for highly abused substances) Step 3 : Surveillance Planning (How to establish prescription drug monitoring programs (PDMPs) Step 4 : Surveillance Implementation (state PDMPs) Step 5 : Information communication The problem Knowledge Production Information Sharing
38 Figure 1 3. The prescription drug monitoring surveillance system with key processes and stakeholders.
39 F igur e 1 4 . Literature r eview summary of doctor shopping threshold parameters , prevalence and other study characteristics, 2004 2013.
40 CHAPTER 2 THE EFFECTS OF ON OXYCODONE CAUSED MORTALITY : A MONTHLY TIME SERIES ANALYSIS, 2003 2012 Overview Prescription drug abuse has reached epidemic proportions in the United States as have rates of prescription drug related overdose and deaths. 3,5,94,95 In 2010, the leading cause of injury related mortality for 25 to 65 year olds was unintentional drug poisoning of which most deaths occur from prescription drugs. 6 In 2008, for the first time, the poisoning mortality rate per 100,000 population surpassed that of fatal car crashes in the United States. 7 are the presc ription medications at the center of this pharmacoepidemic. 3,5 Prescription opioids have become the most prescribed class of medication , even surpassing cholestero l lowering drugs. 53 Oxycodone (OxycontinÂ®, PercocetÂ®) is a s emi synthetic active ingredient one of the most powerful (and preferred) types of medications in the opioid family, which also includes prescription hydrocodone (VicodinÂ®) and heroin. These trends are of national concern. One of the objectives of Healthy People 2020 is the prevention of s uch increases in the rate of poisoning death (90% of which are drug related) in the United States. 7,21 The set of proximal negative health outcomes related to prescription opioid abuse are notable . The US Centers for Disease Control and Prevention (CDC) estimates t hat for every prescription opioid related death, there are 10 treatment admissions for abuse, 32 emergency department visits for misuse or abuse, 130 people who abuse or are dependent, and 825 non medical users of prescription opioids. 22 State to state variation in d rug overdose deaths is pronounced. 96 From 2003 to 2009, the oxycodone caused mortality rate
41 increased by 265% in Florida 8 and the state was ranked 11 th for highest overdose mortality in the nation. 96 The current study focuses on understanding the trends in oxycodone caused mortality in Florida over the past decade. Florida In Florida, 8 people die every day of a prescription drug overdose. 97 The state top 100 dispensing U.S. physicians of oxycodone pills resided in Florida. 98 Recently, the oxycodone landscape changed in Florida. In 2011, the Medical oxycodone mortality in ne arly a decade. 99 , 100 A number of legal, pharmaceutical industry, policy, and public health actions may have contributed to this substantial downward trend in oxycodone mortality. 93 In this study, we hypothesize a priori that multiple actions plausibly contributed to this decline (presented in historic sequence below). First, in January 2010, the state of Florida required pain management clinics to register throug h the Division of Medical Quality Assurance (MQA) at the Florida or privately owned facility that advertises in any medium for any type of pain management services or where in any month a majority of patients are prescribed opioids, benzodiazepines, barbiturates, or carisoprodol for the treatment of chronic 101 Second, in August 2010, a reformulated version of extended release OxycontinÂ®, a popular brand of extended release oxycodone, was introduced to the nat ional market.
42 According to the Federal Drug Administration (FDA), the re formulation makes it more 102 While we are unaware of any studies that have examined the impact of the re formulation on fatal opioid overdoses, Butler et al. and Cicero et al. showed that abuse rates declined immediately after release and continued to fall well into 2012. 103 , 104 Severtson et al. showed that reports of diversion of extended release oxycodone in a national surveillance system decreased by 53%. 105 Third, in February 2011, the U.S. Drug Enforcement Agency (DEA) took action . 106 This operation resulted in the closure of hun dreds of illegitimate pain management clinics 107 I n the first six months of 2011, police closed nearly 400 pain clinics (prima rily in Southeast Florida). 106,108 This region also contains 3 of the 5 counties with the highest density of pill mill clinics per 100,000 population. 109 Fourth, in June 2011, Florida took ag gressive legislative action by passing House Bill 7095 which established or strengthened state regulation of activities by controlled substance dispensing physicians, pain clinics, pharmacies and wholesale controlled substance distributors. 110 Specific actions of the law included a requirement that practitioners use counterfeit proof prescription blanks, immediate seizures of pill inventories from suspect cli nics, and a series of arrests and closures. 111 Finally, in October 2011, Florida became the 37th state in the nation to implement a prescription drug monitoring pr ogram (PDMP). 97,112 PDMPs are data systems that centralize information on controlled, prescription medications with high
43 potential for abuse and diversion such as oxycodone (DEA Schedule II). Licensed dispensers (e.g. pha rmacists) are required to electronically notify the PDMP when oxycodone prescriptions are filled and these transactions form the basis of a surveillance system used by law enforcement, state agency personnel , healthcare prescribers and dispensers. Taken as a whole and given declines in oxycodone caused mortality, it is clear that oxycodone became less available in Florida in recent years. Oxycodone prescribing in Florida has been reduced by 97% and purchases of oxycodone by Florida pharmacies declined d ramatically. 77,98 Today, according to the DEA, none of the top 100 oxycodone prescribers reside in Florida. 113 Given the confluence of multiple explanatory events that arguably reduced the availability of oxycodone at the population level, it is not clear if the PDMP contributed significantly to the recent decline in oxycodone caused mortality in Florida. This is an import ant empirical question for health policy in Florida and there is growing national interest in understanding PDMP effectiveness as a tool for improving public health outcomes such as preventing drug overdoses. 114 implementing PDMPs a major strategic initiative for reducing prescription drug abuse. 91 Other major public health organizations that support the use of PDMPs as a strategy for addressing the epidemic , including the Centers for Medicare and Medicaid Services, the American Medical Ass ociation, and the Robert Wood Johnson Foundation. 82,92,93 PDMP and Opioid Mortality contributing factor to reducing oxycodone caused deaths in the state. Physicians and pharmacists accessed the Florida PDMP 2.6 million times from October 2011 to
44 September 2012. 99 As of September 2013, 11,653 pha rmacists (i.e., dispensers) and 7,305 medical doctors (i.e., prescribers) had registered for the PDMP. 115 This represents 43% and 11% of the pharmac ists and medical doctors licensed to practice in the state of Florida, respectively. mortality. 116 One ecologic analysis of 19 states showed no differences in opioid overdose mortality in states with operational PDMPs. 4 Green et al. responded by pointing out that the study failed to stratify the analysis by states with structural barriers that limit health care provider access. 4,117 Green et al. argued that states allowing healthcare provider access to the PDMP (e.g., Florida) might significantly modif y the risk of overdose in a community via a pathway that modifies prescribing patterns of prescription opioids. 117 Indeed, PDMPs have to been shown to influence prescribing patterns in an emergency room setting. 40 Current Study on reducing oxycodone caused mortality in the state. We conceptualized the study as a state level, natural experiment that tests the change in t he rate of oxycodone caused mortality before and after implementation of the PDMP in October 2011. Methods Research Design We used a time series quasi experimental research design. The objective was to s PDMP on oxycodone caused mortality while controlling for possible alternative explanations for any observed effect using comparison data series. The use of comparison time series controls for many internal
45 (e.g., Florida specific) and external factors ( e.g., nationwide trends) that might reasonably contribute to the observed effect but for which reliable measures are not available. The time series included 120 repeated monthly observations from January 2003 to December 2012, the most recent data availabl e. The study protocol was approved by the University of Florida Gainesville Health Science Center Institutional Review Board (#IRB201300144). Data Sources Oxycodone Caused Mortality. We obtained drug ssion (MEC), a division of the Florida Department of Law Enforcement (FDLE). 118 The MEC instructs district medical examiners (n=24) to determine whether a particular drug was a cause of death (i.e., detected in lethal concentrations) or merely present (i.e., detected in non le thal concentrations) at the time of death after examining the available evidence. After reviewing the autopsy and identified in the decedent. The vast majority of drug related deaths have multiple drugs identified. The MEC makes n o distinction f or primary versus secondary cause of death when multiple drugs are found in lethal concentrations. An MEC data quality committee reviews the reports prior to public release (ME C, personal communication). Appendix A diagrams the MEC data submission process. 119 W e note that the reporting system was fully automated in 2010. An MEC intern al analysis showed that automation likely led to an increase in reporting. 119 In this study, we included a binary indicator variable (0=pre automation, 1=post auto mation) to account for any change associated with reporting system automation.
46 This is a study of oxycodone caused deaths but we briefly describe each major type of opioid and benzodiazepine prescription drug monitored by the MEC in Appendix B . The o ther prescription drugs include Carisoprodol/meprobamate (SomaÂ®) and Zolpidem (AmbienÂ®). The complete drug list includes both licit (e.g., ethanol) and illicit drug (e.g., a ccording to its classification by Florida statute is provided in Appendix B . 120 Prescriptions filled for drugs that the US DEA schedules as 2, 3, or 4 are reported to and monitored by the Florida PDMP. The dependent variable in our study consists of 120 repeated monthly counts of oxycodone caused mortality in Florida from January 2003 to December 2012, the most recent data available. Primary Effect In October 2011, Florida implemented the PDMP. We note that the legal framework was codified in 2009 but the PDMP was not operationalized until 2011. 112 We modeled the PDMP effect using two approaches. First, we modeled the effect of the PDMP implementation as a b inary indicator variable (0=pre implementation, 1=post implementation). Second, we modeled the effect of the PDMP as a continuous variable. To do so , we obtained the number of PDMP queries from healthcare practitioners and the number of registered practitioners by mon th as of December 2012 from the Florida PDMP, the most recent data available. 42 A PDMP query is a single instance of an electronic request for a summary report of controlled prescription drug s filled by an individual
47 patient. W e calculated the query rate as the [# of queries]/[# of registered users] by month from Oct 2011 to Dec 2012. 117 We classified both dispensers and prescribers as healthcare practitioners for this purpose. A query o f the PDMP represents a single instance of an electronic request for a summary report containing each controlled prescription drug filled by an individual C provides a d e identified example of a PAR that a healthcare practitioner in Florida sees when querying the PDMP. To construct the monthly population base in Florida, we used proportional linear interpolation from annual population estimates from 2010 to 2012. Indepe ndent Variables As recommended for quasi experimental time series designs, we included an internal and external control series. 121 We decided a priori that controls would be retained in all models. Internal control series. We included an internal control series consisting of monthly counts of ethanol caused mortality (excl uding oxycodone deaths) for five reasons: 1) to control for overall population change and mortality trends in Florida (independent of the trend in oxycodone caused deaths), 2) to align the control series with the monthly temporal resolution of our outcome, 3) to control for unforeseen trends abusing population around the time of our intervention, 4) to account for any errors (e.g., budget, administrative) intrinsic to the MEC reporting system 119 and 5) to account for seasonal patterns around the time of PDMP intervention. Office of Demographic and Economic
48 Research. 122 We found th at yearly population growth in Florida was highly correlated with ethanol caused mortality ( R 2 =89%) in linear regression. 2011 monthly counts of mortality from all causes from the US Centers for Disease Control and Preve WONDER system. 123 To avoid missing data, we obtained 2012 estimates from the Florida Department of Health 124 and proportioned the yearly value to monthly counts. All cause mortality was correlated with ethanol caused mortality in time series regression at the 10% level ( ARIMA(0,1,1) , p=0.066 ). Third, seasonality analyses indicated that ethanol caused mortality tends to be slightly lower in November and that oxycodone caused mortality followed a similar trend. Given that the PDMP was first operational in October, we included the series to reduce the possibility that a seasonal November decline might explain any observed effect. External Control Series. As a second control series, we obtained uninte ntional, monthly poisoning deaths involving oxycodone for New York City (NYC) from 2003 to 2012, the most recent data available. We used NYC as an external control series for several reasons. First, the rate of opioid analgesic mortality increased steadi ly until 2011 and then declined; a trend similar to oxycodone caused deaths in Florida. 125 The function of this control is to capture the variatio n associated with unmeasured effects occurring in the natural history of the opioid epidemic that might explain declines in Florida. New York has had an operational PDMP since 1973 , so we do not think that recent declines in oxycodone caused deaths in NYC are associated with PDMP. 112 Second, the population base of New York City is large and diverse similar to Florida. Third, we attempted to obtain drug identified deaths from other states (e.g.,
49 Georgia), but administrative contacts indicated that staffing limitations and technical barriers were too high. We considered pooling other state level opioid poisoning deaths cause mortality dataset as the external con trol series. 126 However, this approach is not fe asible because the most recent data available is December 2011, which does not provide complete coverage for our study. Covariates The justification for adding covariates in our study is to ensure that observed effects due to the introduction of re formul ated OxycontinÂ®, Operation Pill Nation, and House Bill 7095 are sufficiently and parsimoniously integrated into the oxycodone caused mortality time series prior to estimating the effect of the PDMP. In preliminary analyses, we conceptualized the pattern of effect and developed a set of transfer functions representing each of the three variables. To reduce collinearity and model complexity, we examined the effects of a continuous variable pain clinic closures to characterize the variation from the three variables. We examined the correlation coefficient matrix of each variable with pain clinic closures. In sensitivity analyses, we also estimated the PDMP effect using both approaches and found no significant differences. Thus, our primary covariate in t his study is pain clinic closures. Below we describe this variable and how it operationalizes each of the three variables. Pain clinic closures. As of January 1, 2010, pain clinics were required to Quality Assurance (MQA) Services. We https://ww2.doh.state.fl.us/downloadnet/Main.aspx as of November 2013, the most recent data available. We calculated a monthly count of pain clinics reporting an D shows the
50 monthly count, running total, and percent change in the number of inactive pain clinics from 2010 to 2012 in Florida. Operation Pill Nation. The DEA reporte d that Operation Pill Nation which resulted in the closure of a large number of pain clinics reduced oxycodone sales significantly. We obtained the change in year to year percentage of oxycodone units sold to Florida pharmacies in 2010 and 2011 from the U S DEA (see Appendix E ). 98 These data points constituted the transfer function representing Operation Pill Nation. Re formulated OxycontinÂ®. We conceptualized the introduction of re formulated OxycontinÂ® onto the market as an ind icator effect initiating in August 2010, increasing linearly, and reaching maximum impact 6 months later. The 6 month time period was derived from Butler et al. results showing that the prevalence of the 30 day abuse rates from original oxycodone declined linearly and immediately, for approximately 6 to 9 months, before achieving a new, lower mean level after the re formulation. 104 Thus, the effect is hypothesized to decline rapidly as abuse deterrent oxycodone quickly replaced the highly abused forms in the drug abusing population. We hypothesized that pain clinics would have conceivably closed during this time due to limited supplies of the most preferred form of pain relief. House Bill 7095. We conceptualized House Bill 7095 as a short term, indicat or effect occurring over a 3 month period that resulted in more pain clinic closures. We did so on the basis of HB 7095 analysis provided by the MQA: Strike Forces entered the business p remises of identified dispensing practitioners and quarantined any Schedule II or III controlled substance inventory found on site. A total of 23 physicians in 24 locations statewide were inspected, with the bulk of visits occurring in south Florida, and o f the 105,579.5 pills found, 6,172 were transported to UPS by FDLE for a
51 reverse distributor and 99,407.5 were quarantined off site by FDLE. (p. 16). 111 Examining t he change in inactive pain clinics during this period indicates that, prior to HB 7095, the number of monthly inactive pain clinics increased by 9.2%, nearly doubled (17.7%) in the months after HB 7095, and then declined to 7.6% by September 2011. Thus, we argue that pain clinic license data adequately captures the rapid effect of HB7095 on pain management clinics. Appendix F shows the effect forms of each variable compared to the inactive pain clinics variable. Statistical Analyses Descriptive. For com pleteness, we report an annual demographic profile of oxycodone caused deaths in Florida and by region. To contrast a period of consistent mortality increase with recent declines, w e collapsed years 2003 through 2009 into a single time period. 97 Univariate Time Series Analysis. Our study uses the Box and Jenkins analytic approach to univariate time series analysis. 127,128 The approach consists of three distinct and iterative analytic phases to model a time series: 1) parameter identification, 2) model estimation, and 3) model diagnostics. In this study, the outcome consists of m onthly counts of oxycodone caused mortality ( Y t ) from January 2003 to December 2012, the most recent data available. The Florida PDMP was implemented in October 2011. The time series has 120 observations, 94 pre and 14 post PDMP implementation.
52 The go al is to find the best fitting and parsimonious Autoregressive Integrated Moving Average (ARIMA) model to statistically characterize the structural parameters of the time series. In ARIMA notation, we identified the best ARIMA (p,d,q) model where p is the autoregressive (AR) parameter, d is the differencing parameter, and q is the moving average (MA) parameter. We used the entire oxycodone cause mortality series in the modeling process as no abrupt changes in the series warranted modeling subset of the ser ies. 128 Parameter Identification, Estim ation, and Model Design In this phase, we examined a plot of Y(t) to assess t rend in mean and variance parameters, seasonality parameters, series outliers, and autocorrelation (i.e, serial dependence). We assessed mean, variance, and seasonality using two types of plots suggested by Williamson (1999). 129 To detrend, we used differencing Y(t) because it is a standard and parsimonious approach. To stabilize variance, we transformed Y(t) using the natural logarithmic fu nction (i.e., ln(Y(t) ). To assess outliers, we examined studentized residuals after linear regression of Y(t) on time. To assess autocorrelation, we interpreted the correlation coefficients plotted on the autocorrelation function (ACF) graph and partial autocorrelation function (PCF) graph of Y(t). 130 The number of significant lags suggests th e autoregressive (AR) or moving average (MA) parameters that might best fit Y(t) . To select AR or MA parameters, we added the parameters and tested for 1) statistical significance using maximum likelihood estimation and 2) residual autocorrelation using th e Q remains after parameterization). A Q statistic indicating significance at the 95% level indicated the possibility that univariate parameters did not sufficiently account for the
53 autocorrel ation in the series. We chose the best fitting model on the basis of lowest Akaike Information Criterion (AIC) and most efficient removal of serial dependence. As we did with oxycodone caused deaths, we characterized each regional and separate drug series using the Box and Jenkins approach. Overall, we found that one of four possible classes of ARIMA models were sufficient ARIMA (0,1,1), ARIMA (1,1,0), ARIMA (2,1,0) or ARIMA (1,1,1) so we limited our model selection process to these options. These speci fications are common for many social science studies. 128 Bivariate Analyses. The goal of bivariate analyses was to obtain preliminary effect estimates to use as baseline comparisons with the adjusted effect estimates. Using the best fitting ARIMA (p,d,q) model from the univariate modeling p hase, we tested the bivariate associations of oxycodone caused mortality. We did not eliminate variables from our models on the basis of p values. Multivariate Analyses. For model build ing, we specified the maximum model and added the PDMP primary effect last. 131 We examined two models; 1) a frequency model for the PDMP as an indicator variable (0/1) and 2) a frequency model for the PDMP as a continuous variable (i.e., the PDMP utilization rate ) . The general form of the frequency model is written: (1 B 1 )Y t 1 ) Z t 1 (1 B 1 )X * 1t 2 (1 B 1 )IC t + 3 (1 B 1 )EC t + 4 (1 B 1 )X 3t + B 5 (1 B 1 )X 5t ( eq. 2 1 ) where t is a monthly observation from t 1 (January 2003) to t 120 (December 2012); is a constant; is the first order moving average parameter operating on Z t which is a random error (white noise) component; 1 B 1 is the first order backshift operator (i.e., first order difference parameter). The asterisk on the PDMP term means that PDMP implementation is either a) modeled where X 1 =0 prior to Oct. 2011 and X 1 =1 thereafte r
54 ( i.e., a change in monthly count ) or b) where X 1 is a continuous variable equal to the [# of queries by PDMP health care providers ]/[# of registered PDMP healthcare provider user s] . 2 (1 B 1 )IC t is the differenced count of the internal control series; 3 (1 B 1 )EC t is the differenced count of the external control series. The term 4 (1 B 1 )X 3t represents the differenced count of inactive pain clinics. The term B 5 (1 B 1 )X 5t is an indica tor for MEC automation where X 5 =0 prior to Jan 2010 and X 5 =1 thereafter . In the case of binary indicator variables (coded 0/1) we interpret the model coefficients ( n ) in terms of percentage impact where 132 : Percent Impact=100*[Exp( n ) 1] ( eq. 2 2 ) Secondary and Sensitivity Analyses Regional Effects . We stratified oxycodone caused deaths by region. We classified the Miami (e.g., Miami Dade County), West Palm Beach (e.g., Palm Beach County), and Ft. Lauderdale (e.g., Broward County) medical examiner districts as the G ). Additional Drug Series. We exa mined several addi tional drug series monitored by the PDMP. Heroin was included because recent studies suggest an increase in heroin caused deaths as a result of shrinking oxycodone availability. These additional drug series included deaths caused by 1) any benzodiazepine , 2) alprazolam, 3) oxycodone (with alprazolam removed), 4) any opioid (with oxycodone removed), and 5) heroin. Pain clinic closure compared to 3 variables. Using the final model, we tested the effect of the PDMP using the 1) pain clin ic closure variable and 2) three separate variables representing re formulated Oxycontin, Operation Pill Nation, and HB 7095. We used the ARIMA procedure in SAS 9.3 for all analysis.
55 Results Descriptive Table 2 1 shows data available on four major dru g classes reported to the MEC. The MEC reports on 12 types of prescription opioids, 17 types of benzodiazepines, alcohol, and several illicit drugs including heroin and cocaine. From 2003 to 2012, 7,804 oxycodone caused deaths were reported (mean=65 death s per month). Of those, 2,784 deaths (36%) also had alprazolam identified in lethal concentrations. In the 14 month period pre and post PDMP implementation, the mean number of monthly deaths was 11 5 and 6 4 , respectively. Table 2 2 describes oxycodone caused decedents in Florida. Over the study period, 42% percent of oxycodone caused deaths were among 35 50 year olds and 66% were male. Age and sex proportions appear to have shifted in recent years. In 2012, for the first time, those aged 50 or greater died in higher proportions than those 35 50 years old. In 2012, the proportion (38%) of female deaths was higher than any other period. In Florida, oxycodone caused deaths occur almost exclusively (95%) among whites with much smaller proportions of Black s (3%) and Hispanics (1.1%). There was a small increase in the proportion of the deaths represented by Blacks (5.3%) Approximately 87% of oxycodone caused deaths are identified as accidental with suicides representing approximately 10% of the total. The re is a substantial increase in oxycodone caused deaths classified as suicides between 2011 (8.7%) and 2012 (13.3%). Alprazolam is present in lethal concentrations in 36% of oxycodone cause deaths with another 16% having Alprazolam present in non lethal co ncentrations. In
56 contrast, ethanol is only present in 23% of oxycodone caused deaths. The same demographic tables for the South Florida region are provided in Appendix H . Univariate Time Series Analysis Trend, Seasonality, and Outliers . Figure 2 1 plo ts the monthly time series of oxycodone caused mortality in Florida, 2003 2012. The series is an increasing function until 2010 when observed counts begin to decrease. Oxycodone caused deaths reached a maximum of 146 deaths in March 2010. Figure 2 2 plot s the intrayear mean overlayed on the observed monthly counts of oxycodone caused mortality. The intrayear mean ranges from 25 monthly deaths (2003) to 126 monthly deaths (2011), evidence that the mean is non stationary. This plot also suggests that the variance increases (i.e., heteroskedasticity) through time as the observed monthly counts show dispersion in the latter part of the series. Figure 2 3 shows the 1 st order difference applied to the series to make the mean stationary. Figure 2 4 shows the natural log transformed series to adjust for possible heteroskedasticity. Figure 2 5 shows the seasonal plot of oxycodone caused deaths averaged by month of the year. On average, the number of oxycodone caused deaths are highest in the month of March ( n=72 deaths), lowest in the month of June (n=59 deaths), and decrease from October to November. In sensitivity analyses, we tested a 6 month seasonal ARIMA model of the form ARIMA(0,1,1)x(0,1,1) 6 using the final set of covariates. The seasonal term was s ignificant but we did not include it in the final models because 1) none of the estimates changed substantially 2) AIC increased and 3) the internal control series already served as a seasonal adjustment parameter.
57 In outlier analysis (see Appendix I ), we identified June 2006 (n=15 deaths) as an outlier as it was nearly 3 standardized units below expected values from the other residuals in the series. The MEC indicated that this observation was likely the result of reporting error. In sensitivity analy ses, we replaced this value with the mean May 2006 and July 2006 (31 deaths). Estimates did not change but model fits improved substantially. Estimation of Structural Parameters . Figure 2 6 shows the ACF and the PCF for the observed oxycodone caused mor tality series with the 95% confidence interval. For both the observed and transformed series, the ACF is indicative of a strong linear trend while the PCF shows that monthly counts are significantly correlated to lag(3) of the series. The latter results suggests a second or third order autoregressive and/or a moving average parameters as possible candidates for the final model. Figure 2 7 shows that after differencing there is a sharp decay in the ACF after lag(1) with most remaining correlations coeffic ients within the 95% confidence interval. On the basis of this result, we added a first order moving average model to our selection process. ARIMA (p,d,q) Candidate Models . Table 2 3 shows the three candidate models ARIMA(0,1,1) , ARIMA(1,1,0) , ARIMA(2,1, 0) identified in the estimation phase. We added an ARIMA(1,1,1) model option because it is commonly used in health policy analysis. For oxycodone caused mortality, we selected the ARIMA (0,1,1) model for this study. Figure 2 8 displays the ARIMA(0,1,1) mo del overlayed on the observed values of oxycodone caused mortality with 95% confidence intervals. Multivariate Time Series Analysis Bivariate Associations . Appendix J shows the correlation matrix for the transfer functions (3 variables) and the inactive pain clinic continuous variable. The
58 pain clinic license variable was highly correlated with the re formulated OxycontinÂ® indicator (r= 0.77), oxycodone sales representing Operation Pill Nation (r=0.42), and the HB 7095 indicator (r= 0.49). Based on thes e results, we used the inactive pain clinic variable to model each effect. Table 2 4 shows the bivariate association of oxycodone caused mortality with each covariates and the PDMP. After controlling for the patterns accounted for in the ARIMA noise mode l, n either the internal control series ( 0.1438 , p= 0.4233 ) nor the external control series ( 0.1149 , p= 0.7569 ) were significantly caused mortality. Likewise, reporting system automation was not significantly associat ed with the outcome ( 15.1196 , p= 0.1861 ), but the sign was in the expected direction. Inactive pain clinics were significantly associated with oxycodone caused mortality ( 0. 0838 , p= <.0001). The negative sign indicates that as the number of inactive p ain clinics increased, oxycodone caused mortality decreased. The PDMP indicator ( 32.747 , p=0.000 7 ) and the query rate ( 0. 3424 , p= <0.0001 ) were both highly significant . The negative sign indicates that PDMP implementation and an increasing average n umber of quer ies per practitioner was associated with declines in oxycodone caused mortality. Effect of the PDMP on Oxycodone Caused Mortality . Table 2 5 shows the series and the report ing system automation indicator never reached significance in any model but we include for completeness.
59 In Step 3, we added the inactive pain clinic variable, which was highly significant ( 0.0841 , p <0.0001 ). After adding the PMDP indicator in Step 4, the pain clinic estimate attenuated slightly ( 0. 0626 , p=0.000 5 ). In Step 4 a ( model 1) , the PDMP indicator variable was highly significant ( 24.7610 , p=0.0 079 ). In other words, t he number of oxycodone caused deaths declined by 24. 7 [CI: 42. 9 , 6 .4 ] immediately after PDMP implementation . In model 2, u sing the PDMP query rate per practitioner instead of the indicator variable , the effect remained significant ( = 0. 229 , p= 0.0 02 1 ) with similar model performance. The coefficient indicates that for a one unit change in the average number of PDMP queries per practitioner , oxycodone caused mortality decline d by 0.229 persons . Given the close timing of House Bill 7095 (July 2011) and the start of the PDMP (October 2011), we tested a 1 month, 2 month, and 3 month indicator variable for HB 7095 in our final model as a sensitivity analysis. PDMP estimates varied from 22.45 to 27.43 (data not shown) and remained significant at the 9 5% level. Figure 2 9 shows the p redicted values of oxycodone caused mortality before and after implementation of the prescription drug monitoring program (PDMP). Squares (with linear trend line) are monthly observations 14 months prior to the PDMP and tri angles (with linear trend line) are 14 months after the PDMP. The immediate change in level was approximately 25 deaths (102 to 77 per month) or approximately 25%. Additional Drug Series. Table 2 6 summarizes the univariate model selection process and shows the estimate of the PDMP indicator that was bivariately associated with other drug series available from the MEC . At the state level, the PDMP was associated with a significant decline in oxycodone caused mortality after excluding
60 deaths due to alprazolam ( 18.95 , p=0.0 037 ). However, the PDMP was not associated with a change in overall opioid mortality when oxycodone caused deaths were excluded ( 3.12 , p= 0. 6290 ) . The PDMP was associated with significant declines in alprazolam caused deaths ( = 23.12 , p=0 .0 002 ) but not deaths from all benzodiazepines ( 13.75 , p=0. 1867 ). Our bivariate results indicate that the PDMP was not associated with a significant increase in heroin caused deaths 2.47 , p= 0.3499 ) . Discussion g program was immediately associated with a 25% decline in oxycodone caused mortality . We observed this association even in the context other legal, policy, and pharmacy related events, operating partially via pain clinic closures, that likely contributed to a dramatic decline in oxycodone availability in the state. We found that the PDMP significantly contributed to reducing oxycodone cause mortality in the state, even after controlling for 1) the introduction of re formulated OxycontinÂ® to the market whi ch made it more difficult to abuse, 2) law enforcement regulations around controlled substance prescribing introduced by House Bill 7095. Trends in oxycodone caused deaths did not appear to be driven solely by trends in alprazolam caused deaths because our results were consistent after excluding alprazolam deaths from the analysis. The PDMP reduced oxycodone caused deaths even after additional sensitivity analyses to accou nt for the close timing of HB 7095. We also found an association between oxycodone caused mortality and the average number of times that healthcare practitioners queried the PD M P database. T his result extends our findings to support the hypothesis that he althcare practitioners
61 using the PDMP may directly impact drug related mortality at the population level. 117 This hypothesis is plausible in Florida for several reasons. First, healthcare practitioners have statutory authority to access the PDMP. In some states, t he PDMP is operated by an agency (e.g., law enforcement) that bars direct access to healthcare practitioners. 117,133 Second, healthcare practitioners are accessing the Florida PDMP. As of September 2013, prescribers and dispensers requested approximately 6.5 million patient activity reports. Moreover, the healthcare practitioner use rate rapidly increased af national model, indicates that it took several years to reach the current rate in Florida. 39 Third, healthcare practitioners are known to change their prescribing hab it s when they have access to PDMP information . One study in Ohio showed that emergency department physicians changed their con trolled substance prescribing habits when a PDMP was available. 40 Our study, however, does not rule out reasonable alternatives to the healthcare causal pathway. For example, law enforcement freq uently uses the PDMP to conduct investigations of individuals suspected of visiting multiple prescribers and dispensers for the purpose of diverting prescription drugs (i.e., doctor shoppers). Therefore, the PDMP may serve to reduce the doctor shopping po pulation through successful case 2011. 99 It is plausible that deaths among doctor shoppers themselves may have been averted as prosecution acted as prevention. Doctor shopping behavior was found to be prevalent among drug related decedents in one study. 61 Alternatively, deaths among sub populations supplied by doctor shoppers via diversion o f oxycodone may have
62 been averted as supply dwindled. More research is needed on these mechanisms and sub populations. Our bivariate models of other drug series indicate that the PDMP may have primarily reduced deaths from oxycodone and alprazolam rather than the general classes of prescription opioids and benzodiazepines . Given the high number of oxycodone deaths with the presence of alprazolam in lethal concentrations and the fact that oxycodone roduced by alone. 77 However, the effects of the PDMP were consistent when we examined an outcome specific to just oxycodone. Our demographic analyses were not comprehensive but su ggestive. They may indicate shifting demographics in oxycodone caused deaths that are epidemiologically important. These include a larger proportion of older individuals (> 50 years old) and women as well as a larger proportion of deaths classified as su icides in 2012. Our results seem consistent with recent data from the CDC showing 1) larger increase in the proportions of women dying from prescription pain killer overdoses, 2) that women 45 54 years old have the highest risk of dying from pain killers, and 3) that 10% of the deaths from pain killers among women are reported as suicides. 134 We note that our The current study should be interpre ted in light of several limitations. First, while the full data series contains sufficient time points for stable time series modeling, the post intervention period includes only 14 monthly data points for analysis. These are the most recent data availab le. Second, findings apply only to Florida and may not apply
63 to PDMPs and oxycodone caused mortality in other states. Third, we cannot completely rule out alternative explanations that might explain oxycodone caused declines around this time. Longer foll ow ups in Florida and replications of the study in other states are warranted to firmly establish effects of PDMPs. Study findings are important for several reasons. First, this is the first study to show that a state level PDMP is effective in reducing o xycodone caused mortality , with similar effect s for deaths in other drug classes monitored by the PDMP. While the reductions in oxycodone caused mortality may be modest (and potentially offset by substitution effects), the PDMPs impact on more proximal ou tcomes may be much greater. When we applied the annual number of deaths averted as a weighting factor to prescription opioid abuse, the PDMP immediately reduc ed : trea tment admissions for abuse by 250 people, emergency department visits for misuse or abuse by 800 people, people who abuse or are dependent by 3,250 people, and people who take prescription painkillers for nonmedical use by 20,630 people (see Table 2 7). T his is a crude estimate but o ur quasi experimental time series analytic framework provides a method for testing these hypothesized reductions in future analyses. The current study has several strengths. First, we used a quasi experimental research desig n which is very strong design in health policy analysis when randomization is not possible. 121 Second, we conceptualized the PDMP as an indicator variable, as is common for this research design, but strengthened our inference by testing an actual metric of PDMP utilization. Third, the specificity of the medical
64 ata allowed us to construct drug specific time series , minimiz ing misclassification. In conclusion, this is the first study to show that the PDMP had a significant effect on reducing oxycodone caused mortality in Florida. Future research can contribute to our understanding of how the PDMP operates to avert deaths from prescription drugs monitored to optimize the most effective components of this strategic tool for reducing prescription drug abuse.
65 Figure 2 1. Monthly counts of oxycodone caused mor tality in Florida, January 2003 ( t=1 ) to December 2012 ( t=120 Commission .
66 Figure 2 2. Expected monthly count of oxycodone caused mortality (squares) overlayed on observed counts within the year (circles).
67 Fi gure 2 3. The natural log transformed and differenced (1 st order) oxycodone caused mortality series, 2003 2012 .
68 Figure 2 4. Natural log transformed counts of oxycodone caused mortality (line) overlayed on observed counts (points).
69 Figure 2 5. Expected value of oxycodone caused mortality by month of the year. Observed values for each month in each year (10 years) is shown (circles). Note: In sensitivity analyses, we specified this seasonal pattern as an ARIMA(0,1,1)x(0,1,1) 6 parameter .
70 Fi gure 2 6. Autocorrelation function (ACF, top) and partial autocorrelation (PCF, bottom) plots for oxycodone caused mortality with 95% confidence interval bands. Note: The slow linear decay in the ACF indicates trend. The PCF indicates that monthly values of oxycodone caused mortality are significantly correlated out to the 3 rd time lag . Figure 2 7. Autocorrelation function (ACF) and partial autocorrelation function (PCF) plots of the oxycodone caused (transformed, differenced) mortality series. Note: In the ACF plot, the significant correlation at lag(1) suggests a 1 st order moving average model [ ARIMA(0,1,1) ]. In the PCF plot, significant correlation at lag(1) lag(3) suggests an 2 nd or 3 rd order autoregressive model [ ARIMA(2,1,0) or ARIMA (3,1,0) ].
71 Figure 2 8. Univariate ARIMA(0,1,1) model (black) of observed oxycodone caused mortality (circles) with 95% confidence interval bands (grey) .
72 Figure 2 9. P redicted values of oxycodone caused mortality (line) before and after implementation of the prescription drug monitoring program (PDMP). Squares (with linear trend line) are monthly observations 14 months prior to the PDMP and triangles (with linear trend line) are 14 months after the PDMP. X=raw counts. The immediate change in level was appr oximately 25 deaths (102 to 77 per month) or approximately 25%.
73 Table 2 1. Characteristics of select drug 2003 to 2012 Type Drug* Months Available** Total Deaths*** Monthly Mean Opioids (n=12) (1) All Prescription Opioids 120 18,457 154 (1a.) Oxycodone 120 7,804 65 (1b.) Other opioids 120 10,653 89 Buprenorphine 23 31 1 Codeine 35 135 4 Fentanyl 120 1,171 10 Hydrocodone 120 2,530 21 Hydromorphone 99 551 6 Me peridine 40 51 1 Methadone 120 6,354 53 Morphine 120 2,788 23 Oxymorphone 48 420 9 Propoxyphene 100 663 7 Tramadol 108 507 5 Sedative Hypnotic: Benzodiazepines (n=17) (2) All Benzodiazepines 120 8,268 69 (2a.) Alprazolam 120 6,072 51 (2b.) O ther Benzodiazepines 108 2,196 18 Licit (3) Alcohol 120 4,542 38 Illicit (4) Cocaine 120 6,427 54 (5) Heroin 120 1,087 9 *Includes counts where a drug was determined to be the cause of death and not merely present. **Values less than 120 months indic ates missing data due to varying start dates for reporting or zero counts for the month.***Counts are not mutually exclusive. For example, of the 7,804 oxycodone caused deaths, 2,784 deaths (36%) also had alprazolam identified in lethal concentrations .
74 Table 2 2. Characteristics of oxycodone caused deaths in Florida, 2003 2012 2003 2009 (N=4,306) 2010 (N=1,516) 2011 (N=1,247) 2012 (N=735) All (N=7,804) Variable N % N % N % N % N % Age Group <18 53 1.2 12 0.8 7 0.6 5 0.7 77 0.99 18 2 5 634 14.7 177 11.7 108 8.7 60 8.2 979 12.54 26 34 831 19.3 361 23.8 283 22.7 116 15.8 1,591 20.39 35 50 1,874 43.5 618 40.8 511 41.0 272 37.0 3,275 41.97 >50 913 21.2 348 23.0 337 27.0 282 38.4 1,880 24.09 Sex Female 1,391 32.3 487 32.1 441 35.4 283 38.5 2,602 33.34 Male 2,913 67.6 1,029 67.9 806 64.6 452 61.5 5,200 66.63 Race/Ethnicity White 4,167 96.8 1,446 95.4 1,172 94.0 678 92.2 7,463 95.63 Black 108 2.5 42 2.8 43 3.4 39 5.3 232 2.97 Hispanic 20 0.5 24 1.6 26 2.1 15 2.0 8 5 1.09 Other 11 0.3 4 0.3 6 0.5 3 0.4 24 0.31 Manner of Death Accident 3,693 85.8 1,347 88.9 1,116 89.5 592 80.5 6,748 86.47 Homicide 2 0.0 1 0.1 1 0.1 2 0.3 6 0.08 Natural 14 0.3 5 0.3 0 0.0 17 2.3 36 0.46 Undetermined 123 2.9 39 2.6 21 1. 7 26 3.5 209 2.68 Suicide 474 11 124 8.2 109 8.7 98 13.3 805 10.32 Alprazolam Status Absent 2,133 49.5 642 42.3 594 47.6 402 54.7 3,771 48.32 Co cause 1,400 32.5 627 41.4 513 41.1 244 33.2 2,784 35.67 Present 773 18.0 247 16.3 140 11.2 89 12 .1 1,249 16.00 Ethanol Status Absent 3,296 76.5 1,188 78.4 972 77.9 558 75.9 6,014 77.06 Co cause 532 12.4 203 13.4 157 12.6 92 12.5 984 12.61 Present 478 11.1 125 8.2 118 9.5 85 11.6 806 10.33 Notes: Age group and sex had 2 unknown values over the study period .
75 Table 2 3. Evaluation and selection of candidate univariate models for oxycodone cause mortality in Florida, 2003 2012 Candidate Autoregressive Integrated Moving Average Models [ARIM A(p,d,q)]* Model Parameters ARIMA(0,1,1)** ARIMA(1,1,0) ARIMA(2,1,0) ARIMA(1,1,1) MA (q) 0.67983 t =9.86 (p<0.0001) No MA term No MA term 0.64026 t =6.21 (p<0.0001) AR 1 (p) No AR 1 term 0.48676 t = 6.04 (p<0.0001) 0.64229 t = 7.26 (p<0.0001) 0 .10399 t = 0.78 (p=0.4332) AR 2 (p) No AR 1 term No AR 2 term 0.32545 t = 3.69 (p=0.0002) No AR 2 term Model Notation ln[Y(t)] = Z(t) 0.68Z(t 1) ln[Y(t)] = 0.49Y(t 1) + Z(t) ln[Y(t)] = 0.64Y(t 1) + 0.32Y(t 2) + Z(t) ln[Y(t)] = 0.10Y(t 1)+ Z(t) 0 .64Z(t 1) Model Selection Index (AIC) 12.90 6.66 4.09 11.68 *p =autoregressive (AR) parameter, d =differencing (d) parameter, and q =moving average (MA) parameter. All models are natural log transformed, first order differenced (d=1) and mean centered. **ARIMA(0,1,1) was the best model on the basis of lowest AIC and no residual autocorrelation. ln[Y(t)]=natural log transformed oxycodone caused mortality where t=1 to t=120 months
76 Table 2 4. Bivariate associations of oxycodone caused mortality with covariates *All MA terms remained significant at p<0.0001 ** Models the release of re formulated OxycontinÂ®, oxycodone sales, and HB 7095 in this study. ***Separate models Model Parameter Description SE t p value MA term* AIC Controls Ethanol caused mortality (Florida) 0.1438 0.1796 0.80 0.4233 0.57 949 Oxycodone caused mortality (New York City) 0.1149 0.3714 0.31 0.7569 0.57 950 Covariates Reporting system automation 15.1196 11.4348 1.32 0.18 61 0.57 948 Pain clinic inactive licenses** 0.0838 0.0178 4.72 <0.0001 0.73 938 Primary Effect*** PDMP (indicator) 32.7470 9.6295 3.40 0.0007 0.66 941 PDMP (query rate) 0.3424 0.0806 4.25 <0.0001 0.70 938
77 Table 2 5. Model buildin g steps for estimating the effect of the Prescription Drug Monitoring Program (PDMP) on oxycodone caused mortality IC=Internal Control, EC=External Control, SYS=reporting automation, LIC=inactive pain clinic licenses, PDMP=Prescription Drug Mo nitoring Program, I=indicator form, Q=query rate form. *all MA terms were significant at p<0.0001 Step Covariate se t p value MA term* AIC 1 IC 0.1460 0.1804 0.81 0.4182 0.57 951 EC 0.1262 0.3722 0.34 0.7347 2 IC 0.1362 0.1801 0. 76 0.4496 0.57 951 EC 0.1011 0.3727 0.27 0.7862 SYS 14.56 11.5024 1.27 0.2055 3* IC 0.1339 0.1828 0.73 0.4638 0.73 941 EC 0.0851 0.3725 0.23 0.8192 SYS 15.761 11.4778 1.37 0.1697 LIC 0.0841 0.0179 4.70 <0.0001 4a** IC 0.1234 0.1 804 0.68 0.4940 0.76 936 EC 0.2061 0.37229 0.55 0.5799 SYS 16.3035 11.2837 1.44 0.1485 LIC 0.0626 0.0178 3.50 0.0005 PDMP(I) 24.7610 9.3146 2.66 0.0079 4b IC 0.1216 0.1812 0.67 0.5021 0.78 935 EC 0.1181 0.3668 0.32 0.7476 SYS 17.0937 11.291 1.51 0.1301 LIC 0.0554 0.0173 3.19 0.0014 PDMP(Q) 0.229 0.07449 3.08 0.0021
78 Table 2 6. Univariate ARIMA modeling and bivariate model results for other licit and illicit drugs *p = autoregressive (AR) parameter, d =differencing (d) parameter, and q =moving average (MA) parameter. The numbers refer to the number of terms included. None=no significant autocorrelation remaining after mo deling. All=all lags significant after modeling (numbers indicate which lags were significant). All candidate models were assessed using log transformed, first order differenced ( d =1) and mean centered. t ARIMA(p,d,q) that assesses the association of the PDMP with each of the drug series. Drug caused Deaths Candidate Autoregressive Integra ted Moving Average Models [ARIMA(p,d,q)]* PDMP indicator* ARIMA (0,1,1) ARIMA (1,1,0) ARIMA (2,1,0) ARIMA (1,1,1) se t p Oxycodone (alprazolam excluded) AIC: 11.24 None (best) AIC: 33.23 All AIC: 26.76 All AIC: 13.11 Lag**: None 18.9481 6.5338 2.90 0.0 037 Opioids (oxycodone excluded) AIC: 151.36 Lag: None AIC: 121.45 Lag: All AIC: 124.96 Lag: All AIC: 149.47 Lag: None (best) 3.1160 6.4490 0.48 0.6290 Alprazolam AIC: 57.36 None (best) AIC: 50.67 Lag: 6,12,18 AIC: 59.97 Lag: None AIC: 60.12 Lag: None 23.1222 6.2705 3.69 0.0002 All Benzodiazepines AIC: 67.65 Lag: 12 AIC: 67.51 Lag: 12 (best) AIC: 68.84 Lag: 12,18 AIC: 68.82 Lag: 12,18 13.7530 10.4154 1.32 0.1867 Heroin AIC:169.94 Lag: None (best) AIC:190.32 Lag: All AIC: 177.57 Lag: None A IC: 171.91 Lag: None 2.4761 2.6491 0.93 0.3499
79 Table 2 7. Estimates for potential reductions in proximal outcomes related to the PDMP using CDC based estimates for all opioids *Source: http://www.cdc.gov/injury/about/focus rx.html **Estimate from this study. Outcome (prescription pain kill ers) CDC Ratio* Immediate Reduction Deaths 1 2 5** Treatment admissions for abuse 10 25 0 Emergency department visits for misuse or abuse 32 80 0 People who abuse or are dependent 130 3 , 25 0 People with nonmedical use 825 2 0 , 6 25
80 CHAPTER 3 RISK FACTORS FOR DOCTOR SHOPPING FOR SCHEDULE II PRESCRIPTION OPIOIDS IN THE TEXAS MEDICAID POPULATION O verview seeking patients who episodically obtain controlled substance prescriptions from multiple health care practitioners when the practitioners are unaware of the use of the other prescriptions. 48 Doctor shopping is a public health concern because it usually involves highly abused controlled substances such as prescription opioids (e.g., oxycodo ne, hydrocodone). 37 However, doctor shopping is not well studied from the pub lic health perspective as resources have tended to focus on law enforcement model (i.e., punitive model) has been accepted nationwide as nearly all US states have laws against this practice. In a few states doctor shopping is a felony. 48,135 Furthermore, pu blic and private insurance agencies and companies devote considerable effort to identifying fraudulent payments and punishing doctor shoppers (and even their providers). For example, in 2007, the US Government Accountability Office reported $63 million in fraudulent Medicaid drug claim payments associated with doctor shoppers. 51 CVS with doctor shoppers. 136 The purpose of this study is to examine the risk and protective factors associated with doctor shopping using an alternative to the punitive model: the risk factor model approac h of epidemiology and public health.
81 Public Health Impact Doctor shopping has been associated with both illicit and prescription drug overdose mortality. 55 58 The American Society of Interventional Pain Physicians recognizes doctor shopping as a major contributing factor to prescription drug fatalities. 32 At the individual level, the high risk of overdose and other adverse health ingredient, 2) overlapping prescriptions o f other powerful classes of controlled substances (e.g., anti anxiety medications like benzodiazepines) and 3) illicit drugs. 56,59 61 At the population level, doctor shopping populations have been associated with the prescription drug epidemic of the last decade in the United States. In California, the prevalence of doctor shopping for S chedule II prescription opioids increased by 111% (younger males) to 213% (older females) from 1999 2007. 5,60 In addition to the direct deleterious effects of toxic medication exposure, doctor shopping behavior is concerning for public health for other reasons. First , the behavior may be associated with vulnerable groups requiring additional monitoring . For example , one study found that doctor shopping behavior may have been a desperate from young men that eventually died of a heroin overdose. 58 In a South Florida study, some subjects had been coerced into doctor shopping by drug seekers who wanted their medications because they were poor, elderly, or HIV patients. 63 Second, doctor shoppers may have a negative impact on the health system via negative provider interactions. Some providers report frequent encounters with doctor shoppers despite the fact the prevalence of do ctor shopping reported in the literature is relatively low (<1.0%). 51 53,60,62,67,68 In West Virginia, 40% of all medical specialties
82 surveyed and 90% of emergency department (ED) physicians reported a suspected encounter with a doctor shopper on a weekly basis. 46 ED physicians in Massachusetts suspected that at least 30% of their patients were doctor shoppers. 69 We have limited information on the effects of these frequent interactions but they may increas e the anxiety that providers experience about 1) being scrutinized by regulatory agencies, 2) contributing to addiction especially for patients with a history of substance abuse and 3) increasing the risk of community level pharmaceutical diversion. 45,70 Despite relatively common encounters, providers have difficulty identifying and characterizing doctor shoppers. 71 In one s tudy, one third of physicians did not report 46 While negative urine tests is a clinical tool used to detect possible diversion (i.e., a patient is not taki ng their medication, suggesting that they might be selling it) which is associated with doctor shopping, providers generally have limited set of options to identify other risk factors. 36 Many providers have access to Prescription Drug Monitoring Programs (PDMPs) in 48 US states to help identify doctor shoppers but usage rates remain low. PDMPs and other administrative data systems typically identify doctor shoppers by a high number of distinct multiple provider episodes (dMPEs). Indeed, PDMPs are actively encouraged to send alerts to providers and law enforcement agencies using their pre defined dMPE thresholds. 72,87 Once identified, PDMPs offer limited sociodemographic data (e.g., age, race, sex) and no health status information regarding these patients. Thus, identifying and improving risk factor models remains an important challenge for public health.
83 Ri sk Factors Associated with Doctor Shopping At the clinician patient interaction level, some risk factors for doctor shopping have been identified. These include requesting medications by name, multiple visits for the same complaint, a suspicious history , and having symptoms out of proportion to examination results. 69 At the population level, research has largely focused on just private sectors (e.g., Medicaid, Medicare, private insurance) but som e risk factors have been ascertained. 51 53 Younger age is the only risk factor identified by two such studies. Wilsey et al. found younger age was associated with multiple provider episodes but not gender. 62 Likewise, Han et al. found 18 44 year olds had the highest prevalence of doctor shopping for Schedule II opioids in California but found no difference by gender. 60 A recent unpublished study from the New York City Department of Health and Mental Hygiene did not find any demographic risk factors associated with doctor shoppers although the variables analyzed were not spe cified in the report. 73 One emerging area of ris k factor research is understanding the spatial relationships between doctor shoppers and their providers. Cepeda et al. found that one fifth visited more than one stat e. 74 Doctor shoppers seem to be aware o f their own lines to obtain prescription pain medications. 74,76 Wilsey et al. found that residents of non metropolitan areas were less likely to have multiple prescribers than metropolitan residents. 62 Weiner et al. noted that drug seeking behavior for prescription drugs varied widely even between two urban academic health centers that were only 1.5 miles apart. 69 Some PDMPs even offer geocoded maps of prescriber and dispenser locations
84 to providers on the patient rep ort, but it is unclear how this information is being used in practice. (Florida PDMP, personal communication). doctor shopping. For example, the index increases when patien ts visit their clinics from distant counties seeking controlled substances or claim to be from out of town. 78 The doctor, the patient, and the pharmacy (e.g. a patient from San A ntonio, seeing a doctor 79 At least one major retail pharmacy, to identify doctor shoppers for pain medications. 80 Current Study The purpose of this study is to examine the risk factors associa ted with doctor shopping for Schedule II prescription opioids in five contextual domains 1) sociodemograhic, 2) health status including painful comorbid conditions, 3) provider episode s 4) spatial relationships and 5 ) medication profiles. We hypothesized these contextual domains using risk factors identified from our review of the literature and our conceptual model (see Figure 1 1). Texas Medicaid claims data were appropriate for this study for two reasons. First, Texas Medicaid data include detailed en rollment, encounter, and pharmacy claims providing a rich source of risk factor variables not available from other administrative data sources such as PDMPs. Second, regression based doctor shopping studies using administrative data require large populati ons. In 2010, Texas had the 3rd largest Medicaid enrollment population (n= 4,844,337) in the United States. 137 Inciardi et al. found that Medicaid recipients were involved in
85 . 138 To our knowledge, this is the first study to identify risk factors for doctor shopping using a n adult Medicaid population. Methods Research Design This is a ob servational study of the risk factors associated with doctor shopping. We in c l uded Texas Medicaid enrollees exposed to at least one Schedule II prescription opioids during 2011 that 1) were age 18 or older 2) had non malignant conditions (i.e., cancer) and 3) did not resid e in a n assisted living facility (e.g., nursing home) . 139 W e selected enrollees with defined by the American Society of Health System PharmacistsÂ® . 140 W database to determin e if the opioid is current ly classified as Schedule II . 141 T he eligible sample consisted of 11, 379 enrollees. Figure 3 1 shows an overview of the data and variable sources, inclusion criteria, and sample sizes for this study. Dependent Variable (Doctor Shopping) Based on previous studies with claims data , we a pri ori defined doctor shopping as having 4 or more prescribers (j) and 4 or more dispensers (k) j k) . 51 53,60,62,67,68 While some agencies (e.g., PDMPs) define doctor shopping on a 90 day timeframe we kept the time frame to 1 year. Rice et al. found that 6 month and 12 month time frames produced similar results in models to identify risk factors for prescription opioid abuse. 67
86 Independent Variables: Risk Factors Overview of Contextual Domains We categorized risk factors into five contextual do mains: 1) sociodemographic, 2) health status including painful comorbid conditions, 3) provider episodes , 4) spatial relationships, and 5 ) medication profiles. Sociodemographic variables included enrollee age, sex, race/ethnicity, and metropolitan status. Enrollee health status variables included opioid abuse or dependence, non opioid abuse or dependence , Clinical Risk Groups (CRGs), mental health disorders, and painful conditions. Provider episode variables include Medicaid program type, total number of healthcare encounters, and prescriber encounters with other doctor shoppers. Medication profiles include overlapping prescriptions of benzodiazepines and daily dose of morphine milligram equivalents. Detailed descriptions of each variable are found below . Sociodemographics . enrollment , sex, race/ethnicity (White non Hispanic, Black non Hispanic, Hispanic, American Indian or Alaskan, Asian/Pacific Islander, Unknown, Other), and United States Department of Codes (RUCC). We categorized age into three groups (18 24, 25 44, and >44 years old). (n= 21 ) and Asian, Pacific (n= 61 ) 142 Health Status . Enrollees were pre classified by the state of Texas into CRGs to determine health status. 143 CRGs use more than 2,000 ICD 9 CM code s and some Current Procedural Terminology codes assigned by providers at the time of health care
87 bers who had no medical encounters during the measurement period or were seen only for routine care; , , llnesses that can usually be , in severity and progression, can be complicated, and require extensive care, such as , sses that often result in progressive deterioration, debilitation, and death, excluding active malignancies. Enrollees classified CRG classification. Opioid abuse or depe ndence was defined as ICD 9 CM 304, 304.01, 304.02, 304.7, 304.71, 304.72 305.5, 305.51, or 305.52. Non opioid substance use disorders (e.g., alcohol abuse/dependence) were defined as ICD 9 CM 305.00 305.93 and 303.00 304.93 exclusive of the opioid related codes. Mental health diagnoses included major disorders (293.0 302.9, excluding any of the substance use disorders described above) and personal history of mental disorder (V11.0, V11.1, V11.2, V11.8, and V11.9). Painful co morbid conditions included b ack pain (720.0 through 724.9), arthritis (710.0 through 739.9, excluding the back pain codes) and headaches ( inclusive of migraines (346.0 through 346.9), tension headaches (307.81), and headache symptom (784.0)). Provider Episodes . Texas Medicaid prov ides health care services to most enrollees (>75%) through a managed care model using health plans organized into four
88 programs: STAR, STAR Health, STAR +PLUS, and NorthSTAR. A Fee for S ervice (FFS) program is also available. We included Medicaid program type as a potential confounder for several reasons . First, managed care programs primarily operate in metropolitan service delivery areas (SDAs). NorthSTAR provides services to eligible residents of Dallas, Ellis, Collin, Hunt, Navarro, Rockwall and Kau fman counties. Second, enrollees in the STAR+PLUS program include Supplemental Security Income ( SSI ) /SSI related clients with a disability or who are age 65 and older and have a disability. Third, outpatient prescriptions are capped at three per month fo r adults in FFS. 144 This administratively const rains the possible number of doctor shoppers in FFS. Other enrollees receive unlimited prescriptions if they are deemed medically necessary. In 2011, 34 counties transitioned to Medicaid managed care . We created a nty of residence to account for any effect that the transition would have on the likelihood of doctor shopping. We defined a binary variable that represented prescribers with 2 or more doctor shoppers in their Schedule II prescribing population. We include d this variable because we hypothesized that provider experience with other doctor shoppers would affect prescribing patterns. 41,46 Spatial Relationships . To examine spatial relationships between enrollees and their providers, we included two spatial var practice location was a state other than Texas and another that measured the geographic distance between enrollees and their providers. Using the state variable, we also determined if providers practiced in a n on PDMP state.
89 To calculate geographic distance, we obtained the latitude and longitude F shows step by step instructions and the full software parameters that were used for geocoding . We geocoded 87% of the residential addresses to the street level which did not vary by doctor shopping status and 98% of enrollees reported their most recent address in Texas. Using the National Provider Identifier (NPI) reported with each pharmacy claim, we obtained the prescriber and dispenser primary practice locations from the NPI Standard database provided by the Centers for Medicare and Medicaid Services . 145 We chose the practice location address rather than the mailing address for this analysis and included practices located out of state because active Texas Medicaid enrollees maintaining residence in Texas are permitted to obtain services outside of Texas . We geocoded 95% of all providers to the street level. Appendix M shows the geographic distribution of enrollees, prescribers and dispensers with claims for at least one Schedule II prescription opioid in this study. Dispensers appear to be more geographically constrained to Texas state boundaries than either enrollees or prescribers. For example, 6.4% of Schedule II opioid prescribers are located out of state compared to <1% of dispensers . Calculating Geographic Distance . We calculated and summed the straight line SAS GEODIST function. 146 Appendix N provides the equation for the SAS GEODIST function.
90 We chose geographic distance for several reasons. First, calculating geographic distance is a) straightforward g iven latitude and longitude, b) accomplished without additional geographic information system softwar e, and c) calculated without the aid of street network file which can be computationally burdensome. Thus, we believe that geographic distances are more l ikely to be implemented by other agencies with access to administrative claims data (e.g., PDMPs). Second, geographic distance was found to be comparable to network street measures for health studies that do not require high precision measurement (e.g., e mergency vehicle routing). In sensitivity analyses, we tested the effect using the distances between the in state providers only due to our concerns that a few enrollees traveling very long distances out of state might skew the data. The results remained the same. Prior to regression modeling, we log transformed distances. We determined if providers were located in states with or without operational prescription drug monitoring programs (PDMPs) as of 2011 (n=14) , Appendix O shows a map of US states by PDM P operational status during this study period (2011). Medication Profiles . Benzodiazepines are often associated with the use of prescription opioids. We identified benzodiazepine prescription s and defined a binary variable representing whether an enrol lee had one or more benzodiazepine prescription during the study period. We calculated the daily morphine milligram equivalents (MME) for each prescription opioid using strength and conversion factors provided by the CDC and the Prescription Drug Monitor ing Program Technical and Training Center. 147 Using quantity
91 (in milligrams) and supply (in days) in the pharmacy claims, we calculated daily dose MME as: Daily Dose=(Quantity*Strength*MME Conversion Factor)/Supply High average daily doses have been associa ted with increased risk of overdose (>40mg MME, odds ratio: 12.2) . 148 The risk appears to increase proportionally by MME dose . 59 as an adverse indicator for PDMPs. 147 Prior to regression modeling, we log transformed daily dose . We obtained all data from the state of Texas and the Texas External Quality Review Health Center IRB 01 (#IRB201300683) . Statistical Analyses First, we calculated descript ive statistics (with Chi categorical variables and t te sts for continuous variables) . Second, we conducted bivariate logistic regression to obtain unadjusted odds ratios and 95% confidence intervals (CIs) . We decided a priori to retain race/ethnicity, sex and age variables in this analysis regardless of statis tical significance. Third, we examine d predictor variables within five contextual domain s . We a) examine d the Spearman correlation coefficients for variables that might indicate collinearity, b ) observe d any ab rupt changes to the size of the estimates, c ) eliminate d weakly associated variables (p>0.05) , and d ) obtain ed the c statistic for comparing the relative predictive performance of each contextual domains. The c statistic quantifies the ability of a logistic regression model over its continuous rang e of predicted probabi lities to discriminate between doctor shoppers and non doctor shoppers . 149 It is equivalent to the area under the curve (AUC) in receiver
92 operator curve (ROC) analysis. Fourth, we selected our final variable set and calculated fully adjusted O Rs and 95%CI. We conducted two sensitivity analys es: 1) we compared our selection with backward stepwise elimination and 2) we estimated the final model using alternative definitions of doctor shopping (i.e., outcome variable). M odel fits and predictive performance were ass essed throughout using the AIC statistic and c statistic , respectively . We used SAS version 9.2 for all statistical analyses. 130 Results Study Population In 2011, there were 638,462 Texas Medicaid enrollees exposed to at least one prescription opioid , 454,671 of which were non institutionalized adults 18) with non mali gnant conditions . Of those, 11, 379 enrollees had at least one Schedule II prescription opioid . These enrollees utilized 23,523 (46%) prescribers and 4,087 (90%) dispensers . The Schedule II opioid exposed population (n=11,379) is the focus of our analysi s . D escriptive statistics for enrollees exposed to Schedule II compared to other schedules are included in Appendix P for reference . There were 180 enrollees (1.6% ) classified as doctor shoppers in our analysis . The number of distinct prescribers used ran ged from 1 to 12 and the number of distinct dispensers used ranged from 1 to 10. Appendix Q ( see Katz et al. ) shows the full range of variation in distinct prescriber and dispensers for doctor shoppers in this study. 88 Doctor Shoppers versus non Doctor Shoppers Table 3 1 shows the characteristics of enrollees exposed to a Schedule II prescription opioid by their doctor shopping status . Sociodemographics . The majority (52.2%) of doctor shopper s and non doctor shoppers were non Hispanic White s . V ariation by doctor shopping status was found for
93 non Hispanic Black s and Hispanic enrollees . For example, Hispanics were less frequently classified as doctor shoppers (1 1. 7% versus 21.5%). The majority of enrollees prescribed a Sch edule II opioid were women (65. 7%) , although se x did not vary significantly by doctor shopping status ( p=0.5062 ) . Doctor shoppers were younger on average than non doctor shoppers by approximately 2 years ( 45.7 v ersus 47.8 year s old, p=0.0330 ) . Doctor shoppers more frequently resided in metropolitan areas than non doctor shoppers (92. 7% versus 84.1 %, p=0.0015 ). Health Status . Doctor shoppers were more often diagnose d with a major chronic health condition, opioid abuse or depe ndence, a non opioid substance abuse problem , a major mental health disorder, and a diagnosis of backpain, arthritis, or headache. While b ackpain and arthritis were highly prevalent in the overall study population, 73.8% and 88. 5 %, respectively , doctor sh oppers still had these diagnoses reported more frequently on their encounter claims . Provider Episodes . D octor shoppers had a median of 158 health care encounters compared to 90 for non doctor shoppers and they were more frequently enrolled in the STAR + PLUS Medicaid program . A pproximately one quarter of doctor shoppers lived in a county that transitioned to managed care in 2011 but no significant variation was seen by doctor shopping status ( we excluded this variable from further analysis ) . A very larg e proportion (98%) of doctor shoppers had received a Schedule II opioid prescription from a provider that had also prescribed to two or more patients that engaged in doctor shopp ing behavior during the study period. Spatial Relationships . Nearly half (48. 3%) of the doctor shoppers used an out of state prescriber compared to one quarter (25. 1 %) of non doctor shoppers. Doctor
94 shoppers were more likely to use a prescriber in a non PDMP state than non doctor shoppers (15.6 % versus 7.3 %, p<0.0001 ) . The median distance between doctor shoppers and their providers was 60.02 miles compared to 36.29 miles for non doctor shoppers. Medication Profile . A larger proportion of doctor shoppers had one or more prescriptions for a benzodiazepine medication compared to n on doc tor shoppers (62.2% versus 50.0 %, p=0.0018 ). The median daily dose morphine milligram equivalent for doctor shoppers was 123 MME/day versus 90 MME/day for non doctor shoppers. Bivariate Associations Table 3 2 shows the biva riate associations of va riables with doctor shopping by contextual domain . We examined the correlation coefficient for variables in the health status domain and found relatively high correlation between arthritis and back pain ( = 0.61, p<0.0001 ). On this basis, we tested backp ain and arthritis in the same model and eliminated arthritis because it was no longer statistically significant. Our state level variables were also moderately correlated ( = 0.48, p<0.0001 ) and after adding them into the same model, we eliminated the sta te PDMP variable for the same reason . The STAR+PLUS Medicaid program was significantly different from fee for service but none of the other program types were so we collapsed this variable into STAR+PLUS versus ther . Associations within Contextual Domain Table 3 3 shows the adjusted associations between variables within contextual domains and fit statistics for the five contextual models. In this stage, non opioid substance abuse or dependence , major mental health disorder, headache, and Medicaid p rogram type were excluded from further analysis on the basis of statistical
95 non significance. Distance between enrollees and their providers was not significant (OR: 1.11 CI:0.97, 1.27) when entered into a model with the out of state variable. The contex tual domain with the least predictive power was the sociodemographic domain (c statistic=0.614) while the provider episode domain had the highest (c statistic=0.750). This excess predictive power was primarily the due to the influence of having a provider that prescribed to 2 or more doctor shoppers ( OR=46.01 CI: 11.38, 186.10) . Full Model Table 3 4 shows the results of our final model. Metropolitan status, having a chronic health care condition, and having one or more benzodiazepine prescription were no longer si gnificant after full adjustment using all contextual domains. In summary, Hispanics were less likely to be doctor shoppers when compared to non Hispanic Whites (OR: 0.58, CI:0.67,1.53). Younger age was significantly associated with increas ed likelihood of doctor shopping . Having a clinically recognized diagnosis of opioid abuse or dependence or backpain were both significantly associated with doctor shopping. As expected, h aving a provider that prescribed to 2 or more doctor shoppers rema ined highly significant with little attenuation of the odds ratio . Figure 3 2 visualizes several doctor shopping enrollees and the connectivity in their network of prescribers. In this figure , six doctor shoppers are connected to one prescriber ( inside t he box). In other results, e nrollees with one or more out of state prescribers and e scalating daily dose MME were as sociated with doctor shopping. The full model had AIC = 1583 and a c statistic =0.833 . We performed two sensitivity analyses. First, we ad ded all variables into one model and applied backward stepwise elimination to examine any differences between
96 automated selection and our analytic process . T he same set of variables were selected with the exception of sex which did not enter the final mod el . Second, we used the final set of covariates to model different operational definitions of doctor shopping. To do so, we re classified doctor shopping as either the 2j x 2k level or the 3j x 3k level. The same set of covariates were significant but A IC increased substantially and the c statistic decreased for the 3j x 3k level and 2j x 2k level, respectively. This indicated poorer model fit with less predictive discrimination for the alternative definition s of doctor shopping. Discussion In this stu dy , we identified unique risk factor s f or doctor shopping among Texas Medicaid enrollees exposed to Schedule II prescription opioids. To our knowledge, this is t he first study to identify risk factor s for doctor shopping in a state Medicaid population. This is the first doctor shopping study to examine the variation associated with race/ethnicity . It is unclear why Hispanic enrollees were less likely to be classified as doctor shop pers than White non Hispanic s . In Kentucky , where use of the prescriptio n drug monitoring program is relatively high, Hispanics in the Medicaid program were more likely to discuss their prescription drug monitoring report with their physician. 150 One hypothesis is that providers monitored the prescription drug pattern s, perhaps with the state PDMP, of their Hispanic patients more than other race/ethniciti es. This idea seems unlikely given that the Texas PDMP is operated by a law enforcement agency , an administrative structure that tends to limit healthcare provider access to information . I n 2011 , Texas PDMP u sers accessed patient history reports at a rat e of 183 per 100,000 population which was an order of magnitude lower than in Florida (1,761 per
97 100,000) which allows healthcare provider access. 151 Data on PDMP utilization by Medicaid providers in Texas would help address this question. Since doctor shopping , by definition , requires multiple visits to healthcare providers, the differences that we found may be partially explained by disparities in access to healthcare . Younger age was associated with doctor shopping but gender did not emerge as a risk factor . Although men more frequently receive d a Schedule II opioid prescription , women were as frequently identified as doctor shoppers. Our results are consistent with age and gender findings from two doctor shopping studies using PDMP data of subjects expose d to Schedule I I opioids. 60,62 Having cli nically recognized opioid abuse or dependence was associated with doctor shopping even after adjusting for higher probability of being diagnosed due to high utilization of the healthcare system . Rice et al. found a similar association with increasing multiple provider episodes (i.e., a proxy for doctor shopping) using a risk factor model for prescription opioid abuse in a privately insured population . P rescription opioid abusers used 3 to 4 times as many dispensers and prescribers, respectively, as non abusers. 67 Th e association between provider utilization and substan ce use disorder is consistent with the underlying compulsive drug seeking behavior associated with opioid addiction. 23 It is important to identify subpopulations of doctor shoppers because some gro ups are likely to have medical needs with co occurring addiction, that if addressed, might reduce the likelihood of engaging in multiple provider episodes . Having one or more out of state prescribers of Schedule II prescription opioids was significant ly associated with doctor shopping. Cepeda et al. reported similar finding using a private insurance claims database. 74 W e found that including geographic
98 distances between Texas Medicaid enrollees and their providers did not provide any explanatory power in our final model. In a post hoc a nalysis , w e compared the within the Schedule II and non Schedule II opioid using populations . W e found that doctor shoppers for Schedule II opioids utilized providers that were nearly 4 times as distant (60 miles ve rsus 15 miles) . Within each population, the doctor shopper to non doctor shopper distance ratio (doctor shoppers traveled two times as far) was nearly identical. T he median daily morphine milligram equivalent (MME) dose for doctor shoppers is cause for c oncern . The median MME was 123 mg/day which is higher than the minimum high risk threshold (>=100 mg/day) recommended by the CDC. 147 McDonald and Carlson (2013) estimated MME for doctor shoppers in a large retail pharmacy at 109 MME/day. 152 This result indicates the ongoing n eed for monitoring the dosage associated with prescription opioids and doctor shoppers using available tools, including PDMPs , and in formats easily and quickly accessible to providers. Unexpectedly, we found that a prescription patterns (i. e., whether the re was claims based evidence of prescribing to two or more enrollees that were doctor shopping for Schedule II opioid s ) was very strongly associated with doctor shopping enrollees . Small cell size s and the unusually large odds ratio prompte d us to perform a sensitivity analysis by changing our definition of doctor shopper . When we defined doctor shopping as having prescribers and 3 dispensers, the odds ratio remained quite large and significant . Thus, while the point estimate varied wi dely by the number of prescribers and dispensers used to define doctor shopping , the positive association remains robust and important .
99 This finding might suggest that doctor shoppers are clustered in to smaller provider networks than non doctor shoppers. It is plausible that intense law enforcement efforts have had the combined effect of removing the highest risk individuals from the doctor shopping pool while simultaneously concentrating (i.e., clustering) them into specific provider networks through tim e. Green et al. reported that Texas had the highest rate of PDMP utilization by law enforcement among 19 states they examined. 117 Also, i n late 2010, Texas passed one of the only pain clinic laws in the country. Our finding suggests that public heal th agencies might focus limited resources on provider education and improving access to the PDMP for identifying especially in networks. This concept recognizes the ecologic framework for addressing risks at the p atient and structural levels. A risk based approach may be necessary in states, such as Texas, whose PDMPs are administered by a law enforcement agency rather than a public health agency. 117,153 Our study has several important limitations. First, validation is needed to confirm that our claims based definition of doctor shopping is consistent with other dimensions of doctor shopping such a s a motivation to deceive for the purpose of obtaining medication , or both. 135 We are u naware of any such validation study and the existing studies of identified doctor shoppers are qualitative and primarily designed to reveal the logistics of how they operate. 66,76,154 In our analysis, t o minimize the possibility of including patients that would legitimately see multiple providers , we excluded cancer patients. We found an
100 association with cli nically recognized opioid abuse or dependence , but we do not know if provider knowledge of doctor shopping behavior prompted such a diagnosis. Second, Medicaid claims data do not capture cash transactions, transactions from other payer types (e.g., Medicare) or I nternet based purchases . Doctor shoppers that have ever paid in cash have been found to use double the number of prescribers compared to doctor shopper s who only use insurance such as Medicaid. 152 Thus, we are likely underestimating the prevalence of doctor shoppers . Third, we do not know the extent to which our results are generalizable to other managed Medicai d program s or insurance systems . Our study has important implications for a number of on going efforts to identify, understand, and prevent doctor shopping at multiple levels. At the clinic level, we found that may help identify doctor shoppers. This is a metric that agencies with access to administrative claims data, including PDMPs, will want to explore. For example, p atient activity reports (PARs) from PDMPs can be modified to include a clear indication when the patient belongs to a high risk doctor shopping network . While there is little research on PARs from the perspective of information design, the reports tend to take the form of a line list of prescriptions which may make trend , pattern , and risk factor an alysis difficult in the limited amount of time available during a clinical encounter. Our geographic findings support the value of continued national efforts of insurance systems and PDMPs to share data across state borders. We showed that doctor shoppers appear to have serious chronic health care con ditions including substance use, abuse, or dependence . This finding may help distinguish between doctor shopping behavior that is driven by an underlying addiction
101 and behavior motivated for the purpose of di version. The mean daily morphine milligram equivalent observed in our study indicates that some doctor shoppers may be at an immediate risk of a fatal overdose. Earlier identification of doctor shopping and the associated risk factors that we identified may reduce the health risk s including mortality of prescription opioid abuse.
102 Figure 3 1. Flow chart for variable sources, inclusion criteria, and sample size in this study, Texas Medicaid, 2011. *Linked to NPI Standard Database from Centers for Medicare and Medicaid Services to obtain practice location and geocoded **STAR, STAR+PLUS , NorthSTAR, Fee for service, and PCCM .***ICD 9 CM 140.0 208.9 ****Place of service codes (31, 32, 33, and 54) . Enrollees with a CPT code indicating methadone maintenance therapy were excluded (n= 59). 1+ prescription opioid (n=638,462) S chedule II Study Population (n=11,379, 2.5%) Other Schedule (n= 443,292 ) Age>=18 with Non malignant diagnosis (n=461,243) Enrollee age <18 (n=161,372) Malignant diagnosis*** (n=15,804) Inclusion Exclusions Institutional (n= 6,5 72) Age>=18, non cancer, non institutional (n=454,671, 71.2%)
103 Figure 3 2. Random sample of 12 doctor shoppers and their Schedule II opioid prescribers (4 or more distinct prescribers, dispensers not shown) in the Texas Medicaid population, 2011. Blue=enrollee, red=prescriber. In this illustration , six doctor shoppers are connected to 1 prescriber (in box) and three doctor shoppers (on perimeter) are not connected to other doctor shoppers via any prescriber .
104 Table 3 1. Descriptive statistics of Texas Medicaid enrollees exposed to a Schedule II prescription opioid by doctor shopping (4 or more prescribers and 4 or more dispensers) status, 2011 Domain Variable Doctor Shopper All (N=11,379) p value Yes (N=180) No (N=11,199) N % N % N % Sociodemogra phic (1) Race/Ethnicity 94 52.22 5,885 52.55 5,979 52.54 White 0.0019 Black 34 18.89 1,599 14.28 1,633 14.35 Hispanic 21 11.67 2,395 21.39 2,416 21.23 Other 31 17.22 1,320 11.79 1,351 11.87 Age Group 18 24 17 9.44 707 6.31 724 6.36 0.0092 25 44 61 33.89 3,411 30.46 3,472 30.51 >44 102 56.67 7,081 63.23 7,183 63.13 Sex 114 63.33 7,332 65.47 7,446 65.44 Female 0.5498 Male 66 36.67 3,867 34.53 3,933 34.56 Age (SD) 45.78 (12.87) 47.81 (12.70) 47.78 (12.70) 0. 0330 Metropolitan 167 92.78 9,412 84.04 9,579 84.18 Yes 0.0014 No 13 7.22 1,787 15.96 1,800 15.82 Health Status (2) Clinical Risk Group 0 0 399 3.56 399 3.51 Healthy <0.0001 Significant Acute 2 1.11 150 1.34 152 1.34 Minor 2 1. 11 414 3.70 416 3.66 Moderate 11 6.11 909 8.12 920 8.09 Major Chronic Health Condition 132 73.33 5,916 52.83 6,048 53.15 Unassigned 33 18.33 3,411 30.46 3,444 30.27 Opioid abuse/dependence 28 15.56 742 6.63 770 6.77 <0.0001 Yes No 15 2 84.44 10,457 93.37 10,609 93.23 Non opioid abuse/dependence 46 25.56 1,944 17.36 1,990 17.49 Yes 0.0041 No 134 74.44 9,255 82.64 9,389 82.51 Mental health disorder 114 63.33 5,642 50.38 5,756 50.58 Yes 0.0006 No 66 36.67 5,557 49.62 5,623 49.42 Backpain 161 89.44 8,014 71.56 8,175 71.84 Yes <0.0001 No 19 10.56 3,185 28.44 3,204 28.16 Arthritis 174 96.67 9,619 85.89 9,793 86.06 Yes <0.0001 No 6 3.33 1,580 14.11 1,586 13.94 Headache 63 35. 00 3,073 27.44 3,136 27.56 Yes 0.0043 No 117 65.00 8,126 72.56 8,243 72.44
105 Table 3 1. Continued Domain Variable Doctor Shopper All (N=11,379) p value Yes (N=180) No (N=11,199) N % N % N % Provider Episodes (3) Median En counters 158 90 92 Medicaid Program Type 28 15.56 2,487 22.21 2,515 22.10 Fee for Service 0.0001 NORTHSTAR 8 4.44 594 5.30 602 5.29 PCCM 33 18.33 2,720 24.29 2,753 24.19 STAR 13 7.22 1,248 11.14 1,261 11.08 STAR+PLUS 98 54.44 4,15 0 37.06 4,248 37.33 Transitioned to managed care 44 24.44 2,535 22.64 2,579 22.66 Yes 0.5653 No 136 75.56 8,664 77.36 8,800 77.34 Prescriber w/ 2+ doctor shoppers Yes 178 98.89 6,806 60.77 6,984 61.38 <0.0001 No 2 1.11 4,393 39. 23 4,395 38.62 Spatial (4) 1+ Out of state prescriber 93 51.67 8,390 74.92 8,483 74.55 Yes <0.0001 No 87 48.33 2,809 25.08 2,896 25.45 1+ Prescriber in non PDMP state Yes 28 15.56 806 7.20 834 7.33 <0.0001 No 152 84.44 10,393 92. 80 10,545 92.67 Median Distance (miles) 60.02 36.29 36.68 Medication Profile (5) 1+ Benzodiazepine in past year 112 62.22 5,603 50.03 5,715 50.22 Yes 0.0012 No 68 37.78 5,596 49.97 5,664 49.78 Median Daily Dose (MME) 123 90 90 No tes: Medians are presented when quantitative variables are highly skewed; no means testing was performed; these variables were transformed prior to regression modeling. p values from Chi square testing. White and Black race are non Hispanic ethnicity. MM E=Mean morphine equivalent
106 Table 3 2. Bivariate associations with doctor shopping by contextual domain, Texas Medicaid, 2011 Domain Variable Odds Ratio Estimate Lower 95% CI Upper 95% CI Socio demographic (1) Black 1.33 0.90 1.98 Hispanic 0.55 0.34 0.88 Other 1.47 0.98 2.22 Male 1.10 0.81 1.49 Age 0.99 0.98 1.00 Metropolitan 2.44 1.38 4.30 Health Status (2) Major Chronic Health Condition 2.46 1.76 3.42 Opioid abuse/dependence 2.60 1.72 3.91 Non opioid abuse/dependence 1.63 1.17 2.29 M ental health disorder 1.70 1.25 2.31 Backpain 3.37 2.09 5.43 Arthritis* 4.76 2.11 10.77 Headache 1.42 1.05 1.94 Encounters (log 10) 2.90 2.05 4.10 Provider Episodes (3) NORTHSTAR 1.20 0.54 2.64 PCCM 1.08 0.65 1.79 STAR 0.93 0.48 1.79 STARPLU S 2.10 1.37 3.20 Prescriber w/ 2+ doctor shoppers 57.41 14.25 231.39 Spatial (4) Average distance (log10) 1.33 1.21 1.48 1+ Out of State Prescriber 2.79 2.08 3.76 1+ Prescriber in non PDMP State** 2.38 1.58 3.58 Medication Profile (5) 1+ benzodiaze pine in past year 1.65 1.21 2.23 Daily dose MME (log 10) 4.73 3.02 7.40 Notes: Reference groups: Race/ethnicity=non Hispanic White, Medicaid program type=Fee for Service. *Eliminated from further analysis because variable was no longer significant at t he 95% level when added to a model with backpain. **Eliminated from further analysis because variable was no longer significant at the 95% level when added to a model with out of state prescriber.
107 Table 3 3. Adjusted associations between variables with in contextual domains, Texas Medicaid, 2011 Domain Effect Odds Ratio Estimate Lower 95% CI Upper 95% CI AIC c statistic Socio demographic (1) Black 1.21 0.81 1.80 1830 0.614 Hispanic 0.50 0.31 0.80 Other 1.38 0.92 2.09 Male 1.11 0.82 1.51 Age 0.99 0.98 1.00 Metropolitan 2.43 1.37 4.29 Health Status (2) Major Chronic Health Condition 1.94 1.37 2.73 1795 0.667 Opioid abuse or dependence 2.09 1.35 3.22 Non opioid abuse or dependence 1.10 0.77 1.59 Mental health disorder* 1. 13 0.82 1.57 Back Pain 2.62 1.60 4.28 Headache* 1.11 0.81 1.52 Provider Episodes (3) STARPLUS* 1.20 0.89 1.62 1675 0.750 Health care encounters 2.00 1.41 2.85 Prescriber with 2+ doctor shoppers 46.01 11.38 186.10 Spatial (4) 1+ Out of State Prescriber 2.26 1.51 3.39 1807 0.643 Distance 1.11 0.97 1.27 Medication Profile (5) 1+ benzodiazepine in past year 1.50 1.11 2.04 1801 0.659 Daily dose MME 4.51 2.87 7.08 Reference for Race/ethnicity is White, non Hispanic. *Candidates fo r exclusion on the basis of statistical significance.
108 Table 3 4. Risk factor model for doctor shopping for Schedule II prescription opioids, Texas Medicaid, 2011 Domain Effect Estimate SE Odds Ratio Estimate Lower 95% CI Upper 95% CI Socio demograp hic (1) Black 0.0849 0.8160 1.01 0.67 1.53 Hispanic 0.4700 0.1554 0.58 0.36 0.95 Other 0.3110 0.1809 1.27 0.83 1.94 Male 0.0887 0.1587 1.19 0.87 1.63 Age 0.0236 0.0800 0.98 0.96 0.99 Metro 0.1024 .00652 1.23 0.68 2.23 Health Status (2) Major chronic health condition 0.0402 0.1527 1.08 0.73 1.61 Opioid abuse/dependence 0.2254 0.1003 1.57 1.02 2.41 Backpain 0.4103 0.1094 2.27 1.38 3.75 Provider Episodes (3) Health care encounters 0.4628 0.1279 1.59 1.07 2.36 Prescriber with 2+ doctor shop pers 1.8569 0.2021 41.01 10.09 166.69 Spatial (4) 1+ Out of state provider 0.2838 0.3577 1.76 1.30 2.40 Medication Profile (5) 1+ benzodiazepine 0.0863 0.0787 1.19 0.87 1.63 Daily dose MME (log 10) 1.9508 0.0809 7.03 4.29 11.54
109 CHAPTER 4 RISK FAC TORS FOR PRESCRIPTION OPIOID ABUSE AND DEPENDENCE IN THE TEXAS MEDICAID POPULATION , 2011 Overview The abuse of p rescription opioid s, or pain relievers , is a national epidemic. 5,28,155 In 2010, the U . S . incidence of prescription pain reliever misuse was approximately 5,500 people per day according the U . S . Centers for Disease Control and Prevention (CDC). 22,23 Misuse wa s . In 2010, the prevalence of pain killer misuse was approximately 5.1 million persons, or 1.7% of the US population . 23 In the US, opioid use and misuse increased 1,448% and 4,680%, respectively, from 1996 to 2011. 156 Opioid abuse is higher in North America than in other regions, but the problem is a growing global concern as well. 157 In 2010, an estimated 15.5 million people globally were opioid dependent. 158 Opioid misuse appe ars to initiate a cascade of negative health events with system wide implications . The CDC estimates that for every 825 people that use opioids non medically , there are 130 people who abuse or are dependent and 32 people who visit the emergency department for misuse or abuse . 22 National d ata from the Drug Abuse Warning Network indicate d that , in 2011, narcotic pain relievers represent ed approximately one third of 1.2 million emergency department visits attributable to pharmaceuticals. 159 Boscarino et al. found that approximately one third of patients on chronic opioid therapy for chronic non ca ncer pain had a lifetime opioid use disorder using both the Diagnostic and Statistical Manual, Version 5 ( DSM 5 ) and DSM 4 cr iteria. 160
110 Medicaid Research suggests that opioid abuse disorder may be more problematic in Medicaid compared to the general population . Usin g the National Survey on Drug Use and Health (NSDUH), Becker et al. found that respondents with opioid use disorder were more likely to be on Medicaid than those without substance use disorders (SUDs) . 161 Edlund et al. found th at approximately 3% of an Arkansas Medicaid population exposed to chronic opioid therapy had a diagnosis of opioid abuse or dependence. 162 Prescription opioid abuse is estimated to be 10 times higher fo r those insured by Medicaid compared to private plans, resulting in excess medical and drug costs of approximately $7 billion/year. 61,67,163,164 Furthermore, the rate of co morbid substance use disorders are high in Medicaid popu lations , treatment rates are low , and access to treatment varies from state to state. 165 According to estimates from the National Comorbidity Survey and NSDUH, the prevalence of SUDs range from 9.2% to 13.8% for those insured by Medicaid compared to 6% of adults in the general population. 166,167 O nly 15% of Medicaid enrollees had access to treatment for opioid abuse in the past year . 161 Approximately, 20% of SUD treatment funding in the United States comes from Medicaid. Using the National Ambulatory Medical Care Survey, Olsen et al. found that Medicaid recipients were twice as likely to be prescribed opioids. 168 In a study by Cummings et al. , Texas had the third highest proportion of counties that do not offer access to an outpatient SUD facility that accepts Medicaid. 165 Birbaum et al. estimated that the total US so cietal costs of prescription opioid abuse at $55.7 billion in 2007 (USD in 2009). 164
111 Risk Factors Identifying the risk factors associated with prescription opioids abuse and dependence remains a significant challenge. 169 Sociodemographic risk factors that have been associated with prescription opioid abuse include being male , younger age, and White but vary by study and data source. 170 For example, in a study using the 1991 National Household Survey on Drug Abuse (NHSDA), Simoni Wastila et al. found that women were more likely to be non medi cal users of opioid analgesics but a recent study using private insurance claims found that men were more likely to be clinically recognized abusers. 67,170 Health related risk factors include chronic pain, patient history of substance abuse and/ or underlying mental health disorders . 17 1 Cheatle summarized by citing a most reliable, and evidence 33,172 Pasquale et al. found that the increased costs associated with prescription opioid abuse are driven by pain and psychiatric comorbiditie s in Medicare. 173 The healt h care system and prescribing patterns have played a role in this epidemic . Approximately 20% of non medical users of prescription opioids rec eive a prescription from one or more physicians. In fact, Jones et al. found that as the number of days of medical misuse increased (i.e., heavier misuse) , the likelihood of obtaining a prescription from a physician also increased (27.3% of non medical us ers). 174 Two studies by Mazer Amirshahi et al. in adults and adolescents show opioids have been prescribed at higher rates than the prevalence of pain in emergency departments would warrant . 15,16 Rice et al. found that the number of dispensers was a significant predictor
112 of opioid abuse. 67 Similarly, a study of Medica id administrative data in Wisconsin concluded that the number of dispensers was associated with substance abuse. 175 Screening Tools M ultiple clinical screening tools including physician administered questionnaires, patient self reports, physician assessment of aberrant behavior, urine tests, and claims based ap proaches are available . Sehgal et al. and Frankel et al. provide systematic review s of the opioid risk assessment literat ure . 176 Frankel et al. concluded that , from el even risk assessment tools currently available to physicians , the best stand alon e clinical tool for predicting opioid misuse was the Revised Screener and Opioid Assessment for Patients with Pain (SOAPP R) . This conclusion was based on a single validation study (N= 302 pain management clinic participants from five centers in the US ); the SOAPP R predicted future opioid misuse . 177 In their guidelines fo r responsible opioid prescribing, Manchikanti et al. concluded that opioid abuse screening tests still lack sufficient evidence of reliability and accuracy . T ools such as the early identification of abusive patterns using pharmaceutical transaction data a v ailable to prescription drug monitoring programs (PDMPs) are promising. 32 We are unaware of any published studies that utilize PDMP derived data, such as evidence of escalating dosages , as a possible screening tool for opioid abuse and dependence . The Role of Administrative Claims based S tudies U sing administrative claims data provides researchers an opportunity to identify risk factors in multiple domains (e.g., demographic, diagnostic, health service utilization, and medications) required to improve knowledge of complex conditions such as prescription opioid abuse. 178 Claims based studies have been used to identify risk factors for opioid abuse and dependence in privately and publicly insured populations
113 such as Medicare and veterans but for Medicaid they are scarce . 163,179 181 To our knowledge, Sullivan et al. and Edlund et al. provide the only studies to utilize Medicaid claims data to identify the risk factors associated with prescription opioid abuse. 162,171 Both of these studies used the same data from the Trends and Risks of Opioid Use for Pain (TROUP) study from Arkansas . We do not know how representative of other Medicaid populations these studies are , but clearly additional studies from other states are warranted. The purpose of this study is to 1) estimate the preval ence of opioid abuse and dependence , 2) identify the risk and protective factors associated with opioid abuse and dependence using an ecologic framework, and 3) com pare the diagnostic characteristics of a Medicaid and presc ription drug monitoring program like model of opioid abuse and dependence in a large Medicaid population . Methods Research Design This is a observational study of the risk factors associated wit h clinically recognized opioid abuse and dependence from administrative data . The analytic dataset consists of Texas Medicaid enrollees age 18 or older who were exposed to at least one Schedule II prescription opioid in 2011 . We defined prescription opi oid as a ny from the guidelines of the American Society of Health System PharmacistsÂ® ( AHFS ) . 140 We used the US Food and Drug as Schedule II. 141
114 We e xcluded three types of Medicaid enrollees : those 1) who were diagnosed with malignancies (International Classification of Diseases, Clinical Modification, Ninth Revision (ICD 9 CM) codes between 140.0 and 208.9; n = 15,804 ) , 2) residing in a facility ( Place of Service codes=31, 32, 33, or 54 , n= 6,572 ) 139 , and 3) receiving metha done maintenance therapy (Curr ent Procedural Terminology code H0020 , n= 59 ) . Texas Medicaid administrative claims data were appropriate for this study for several reasons. First, Texas had the 3rd largest Med icaid enrollment population (n= 4,844,337) in th e United States in 2010. 137 Second, Texas is the state with the thir d highest proportion of counties without access to Medicaid funded treatment facilities for substance use disorders in the United States. 165 Third , Texas Medicaid data includes detailed enrollment, encounter, and pha rmacy claims providing a rich source of demographic and clinical variables (i.e., possible risk factors) not available from other administrative data sources such as PDMPs. Enrollment data was captured at a single point in time (i.e., at enrollment) so , f or enrollees that may have changed their Medicaid program within the calendar year, we used the enrollment variables from the program of longest enrollment. Encounter and pharmacy claims were transactional , meaning that these data w ere submitted on a cont inuous basis during our study period. The eligible sample consisted of 11,326 enrollees. Figure 3 1 shows an overview of the data and variable sources, inclusion criteria, and sample sizes for this study (the same sample as Chapter 3) . We obtained all d ata from the state of Texas and the Texas External Quality Review Organization (EQRO) . This study was approved 01 (#IRB201300683).
115 Dependent Variable : Clinically recognized Opioid Abuse and Dependence Cli nically recognized o pioid abuse or dependence was defined as at least one encounter claim indicating at least one of the following ICD 9 CM codes 305.50 305.53 (opioid abuse), 304 .00 304.03 (opioid dependence) or 304.70 304.73 (opi oid with other drug depen dence) . We included any occurrence of opioid abuse and dependence from the diagnoistic fields. Independent Variables: Risk Factors Overview of Contextual Domains . We categorized risk factors into five contextual domains : sociodemographic, health stat us including psychiatric and painful comorbid conditions , Medicaid program type, provider episodes and network characteristics, and prescription medications . Sociod emographic variables included enrollee age, sex, race/ethnicity, and rurality. H ealth statu s variables included non opioid substance use disorders , Clinical Risk Groups (CRGs) , mental health disorders , and painful conditions. Medicaid program types are known as STAR, STAR Health, STAR+PLUS, NorthSTAR, and fee for service (FFS). Provider episode variables include total number of enrollee encounters, multiple provider episodes, US state location of provider practice , and prescriber experience with doctor shoppers. Prescription medication profiles include whether the enrollee had a benzodiazepine prescription and the average daily dose of morphine milligram equivalents calculated from all opioids . Below we describe in detail each of the variables mentioned above. Sociodemographic . W e obtained age at enroll ment , sex, race/ethnicity (White non Hispanic, Black non Hispanic, Hispanic, American Indian or Alaskan, Asian/Pacific Islander, Unknown, Other), and United States Department of
116 Codes (RUCC) for counties . Due to small cell sizes , we included (n=21) (n=61) located in a RUCC 1 3 county 9 county metropolitan 142 Health Status . Enrollees were pre classified by the state of Texas into CRGs to determine chroni c health status. 143 CRGs use more than 2,000 ICD 9 CM codes and some Current Procedural Terminology codes assigned by providers at the time of their health care encounter and are likely to accurately re flect the enr who had no medical encounters during the measurement period or were seen only for , luding illnesses or injuries, placing an , , illnesses that vary in severity and progr ession, can be complicated, and require , often result in progressive deterioration, debilitation, and death, such as active U means that the enro llee was not in a Medicaid program long enough to receive a CRG classification. Non opioid substance use disorders were defined as ICD 9 CM 305.00 305.93 and 303.00 304.93 exclusive of the opioid related codes. Appendix I shows the ICD 9 CM code s related to substance use in this study . Mental health diagnoses included the major disorders (293.0 302.9, excluding any of the substance disorders described above) and a personal history of mental
117 disorder (V11.0, V11.1, V11.2, V11.8, and V11.9). Painful condit ions included b ack pain (720.0 through 724.9), arthritis (710.0 through 739.9, excluding the back pain codes ) and h eadache s [ migraines (346.0 through 346.9), tension headaches (307.81 ), and headache symptom (784.0) ] . Medicaid Program Types . Texas Medic aid provides health care services to most enrollees (>75%) through a managed care model using health plans organized into four programs: STAR, STAR Health, STAR+PLUS, and NorthSTAR. A fee for service (FFS) program is also available. For this analysis, t here are several structural characteristics of these programs worth noting. First, managed care programs primarily operate in metropolitan service delivery areas (SDAs). NorthSTAR provides services to eligible residents of Dallas, Ellis, Collin, Hunt, Na varro, Rockwall and Kaufman counties. Second, enrollees in the STAR+PLUS program include Supplemental Security Income ( SSI ) /SSI related clients with a disability or who are age 65 and older and have a disability. Third , outpatient prescriptions are cappe d at three per month for adults in FFS. 144 This restriction constrains the possible number of distinct prescribers and dispensers (used for classifying doctor shoppers) in FFS. Other enrollees receive unlimited prescriptions if they are deemed medically necessary. I n 2011, 34 counties transitioned to Medicaid managed care in February (n=13) residence to test for this effect on diagnosing opioid abuse and dependence . Provider episodes . Based on previous studies with claims data 51 53,60,62,67,68 , we a priori defined five, mutually exclusive categories of distinct multiple provider episodes (dMPEs) according to claims submitted from distinct prescribers ( j ) and
1 18 disp ensers ( k ) for Schedule II prescription opioids . Multiple provider episodes are widely considered to be indicative of the compulsive behavior associated with addiction. The five ordinal levels j k), ( j k), j k), ( j k), j k). A small percentage of enrollees (<1%, n=4,237) had prescriptions with missing prescriber/dispenser ID s and were included the j k category. j k j k providers for their Schedule II prescription opioids within the calendar year. The timeframe used to define doctor shopping in the literature varies from 3 to 12 months. Rice et al. f ound that 6 month and 12 month timeframes produced comparable results in models to identify risk factors for prescription opioid abuse. 67 Using our definition of doctor shopping, we determined if enrollees received a Schedule II prescription opioid from a prescriber that prescribed a Schedule II prescription opioid to one or more doctor shoppers during the study period. Using the National Provider Identifier (NPI) reported with each pharmacy claim, we obtained the prescriber and dispenser primary practice state location by linking th e NPI with the NPI Standard database provided by the Centers for Medicare and Medicaid Services. 145 Less than 1% of p rescriber NPIs were missing from pharmacy claims in 2011 . We included practices located out of state because active Texas Medicaid enrollees maintaining residence in Texas are permitted to obtain services outside of Texas. 182 We determined if providers were located in states with or without operational prescription drug monitoring programs (PDMPs). Appendix N shows a map of US states with operational PDMP s in 2011 .
119 Medication Profiles . We identified a b enzodiazepine prescription in the pharmacy claims data using the appropriate AHFS terminology. We calculated the daily morphine milligram equivalents (MME) for each prescription opioid using strength and conversion factors prov ided by the CDC and the Prescription Drug Monitoring Program Technical and Training Center. 147 Using quantity (in milligrams) and supply (in days) in the pharmacy claims, we calculated daily dose MME as: Daily Dose=(Quantity*Strength*MME Conversion Factor)/ Supply High average daily doses have been associated with increased risk of overdose (>40mg MME, odds ratio: 12.2) . 148 The risk appears to increase proportionally by MME dose . 59 as an adverse indicator for PDMPs. 147 Statistical Analyses We used a pseudo split sample analytic design to model the risk factors for opioid abuse and dependence using 2011 data (i. e. training set) , and then we also applied the final model parameters to 2012 data . We refer to this pseudo some of the same enrollees. Firs t, we calculated descriptive statistics ( Chi s quare d tests for categorical variables and t tests for continuous variables) . Second, we conducted exact bivariate logistic regression . Third, w e used exact logistic regression to examine the contextual domai n adjusted odds ratios ( da ORs) and 95% confidence intervals (CIs). We also examined the Spearman rank correlation coefficients for variables in each contextual domain. W e used the domain based modeling process to : a) observe any abrupt changes to the es timates , b) remove variables that were highly correlated to avoid
120 problems with collinearity, c ) eliminate variables not contribut ing significantly to model s , and d ) obtain the c statistic for comparing the relative predictive performance of each domain. The c statistic quantifies the ability of a logistic regression model , over its continuous range of predicted probabilities , to discriminate between enrollees with opioid abuse and dependence and those without. 149 The c statistic is equivalent to the area under the c urve (AUC) in receiver operator curve (ROC) analysis. Fourth, we used backward stepwise elimination to confirm variable selection for 2011 data and to obtain the fully adjusted odds ratios (AOR) and 95% CIs. Fifth, we performed a sensitivity analysis by applying the parameters to 2012 data. After modeling was complete, we performed two additional analyses. W e compared the relative performance of two different models: 1) the complete Medicaid model and 2) a PDMP like model (i.e., the full Medicaid model minus the health status domain because diagnoses are not available to many prescription drug monitoring programs.) And, we performed a post hoc path analysis to examine potential mediators between mental he alth disorders and opioid abuse and dependence. All analyses were conducted in SAS 9.3. Results Prevalence of Opioid Abuse or Dependence In 2011, there were 638,462 Texas Medicaid enrollees exposed to at least one prescription opioid, 454,618 of which were non institutionalized adults (age >=18) with non malignant conditions or any indication of methadone medication therapy. Of those, 11,326 enrollees had at least one Schedule II prescription opioid (i.e., our analytic sample). The prevalence of clinically recognized opioid abuse or dependence in our sample was 6.3% (n=713) which was significantly higher than for enrollees exposed to
121 less addictive opioids (0.86%). Descriptive statistics for the risk factors included in this study for enrollees exposed to Schedule II compared to other schedules are i ncluded in Appendix P for reference. Opioid Abusers versus Non Abusers Table 4 1 shows the characteristics of enrollees exposed to a Schedule II prescription opioid by their opioid abuse or dependence (OA/D) status . Sociodemographics . E xamining enroll ees diagnosed with OA/D, the sociodemographic profile is characterized as White non Hispanic (5 3 . 7 %) , female (63.3%), older than 44 years old (60.5%) , and residing in a metropolitan area ( 87.4 %) . C ompared to enrollees without OA/D , enrollees with OA/D were mo re frequently Hispanic ( 2 4 . 1 % vs 20. 9 % , p=0.0294 ) , metropolitan residents ( 87 . 3 % vs 83.9 % , p=0.0136 ) , and younger (25 to 44 years old, 37.2% vs 29.8%, p<0.0001 ). Gender was not associated with diagnosed OA/D . Health Status . The health status profile of enrollees diagnosed with OA/D is characterized as having a major chronic health condition ( 66.5%), a history of mental health disorders (77%), a diagnosis of backpain (80.2%) and arthritis (89.5%). Approximately, half of the OA/D diagnosed enrollees had a t least one claim during the study period with a non opioid substance use disorder (SUD) indicated. This percentage was significantly higher than the distribution non opioid SUDs among enrollees without OA/D (49.1% vs 15.1%, p<0.0001 ). Likewise, all of t he painful conditions included in our study were significantly more prevalent among OA/D enrollees . Provider Episodes . Approximately, one third of the analytic sample was enrolled in the STAR+PLUS Medicaid program . However, this program had a larger p roportion
122 of enrollees with opioid abuse or dependence (45.9% vs 36.7%, p<0.0001) . Twenty three percent of the sample resided in a county that trans itioned to Medicaid managed care in 2011 . The transition was not associated with OA/D status and therefore we excluded this variable from further analysis . We found that 1.58% of our sample met criteria for doctor shopping (i.e., using 4 distinct prescribers and 4 distinct dispensers for their Schedule II prescription opioids during a one year period ) . While the absolute number was small (n=179), doctor shopping by enrollees with OA/D occurred at approximately twice the rate (3.78% v s 1.44%) than non OA/D enrollees . Other provider episode characteristics associated with OA/D enrollees included higher use of a) out of state prescribers (31% vs 25%, p=0.0004 ), b) prescribers in non PDMP states (10.5% vs 7.2%, p<0.0001 ), and c) provider s that have prescribed to one or more doctor shoppers ( 72.9% vs 60.5%, p<0.0001 ). The O A /D enrollees had a significantly higher number of encounter s ( 197 vs 130 ) than the non O A /D enrollees ( t = 8.42 , p<0.0001 ). Medication Profile . More than two thirds of the OA/D diagnosed enrollees had at least one prescription for a benzodiazepine during the study period, which wa s significantly higher than non OA/D enrollees (66.4% vs 49.1%, p<0.0001 ). T he number of dispensed opioids per enrollee compared to the num ber of benzodiazepines (8.6 vs 9. 5 prescriptions , respectively ) suggests that these prescriptions were often concurrent during the study period . The absolute number of prescriptions for either an opioid or a benzodiazepine did not differ by O A /D status. The daily dose of morphine milligram equivalents was significantly higher for O A /D enrollees (134. 5 mg/day vs 119. 8 mg/day, t=3.78 , p =0.0002 ).
123 Bivariate and Multivariate Associations Domain Adjusted Table 4 2 shows the bivariate associations of variables w ith opioid abuse or dependence by contextual domain. Table 4 3 shows the results of domain based modeling of opioid abuse or dependence . The v ariables that were eliminated on the basis of statistical non significance and/or backward stepwise elimination (with entry and retention significance lev els set to 10%) are indicated. The health status domain provide d the best model fit ( AIC =4803) and highest predictive discrimination ( c statistic =0.73 6 ) , followed by the provider episode domain (AIC= 5148 , c statis tic =0.6 34 ) , the medicat ion profile domain (AIC=5221, c s tatistic =0.614), and the sociodemographic domain (AIC=5283, c statistic =0.579) . Fully Adjusted Table 4 4 shows the result of the full model for opioid abuse or dependence for 2011 and 2012. Figure 4 1 shows a forest plot of adjusted odds ratios and 95% confidence intervals for the variables in the final model (2011). Blacks were less likely (OR=0.72, CI: 0.56 0. 94 ) and Hispanics were more likely (OR=1.3, CI: 1.0 1.5) to have been diagnosed with O A /D compared to White non Hispanic enrollees . Men were m ore likely than women to have O A /D (OR=1.2, CI: 1.0 1.4). Enrollees 25 to 44 years old were more likely than those 18 to 24 years old to have O A /D (OR=1.8, CI: 1.1 3.0). Not surprisingly, co occurr ing non opioid SUDs were a strong risk factor in our model. Enrollees with these conditions were approximately 4 times as likely to have O A / D (OR=3.9, CI: 3. 3 4. 6 ). O A / D was highly associated with a having a major mental health disorder other than an SUD . E nrollees with a major mental health disorder were approximately 2 times as likely to be diagnosed with O A / D
124 (OR: 1. 9 , CI: 1. 6 2. 4 ). Enrollees classified as doctor shoppers were approximately 2 times as likely to be diagnosed with OA/D (OR=1.8, CI: 1.1 2.7 ) . E nrollees were more likely to be diagnosed with OA/D when they filled a Schedule II opioid prescription from a prescriber with doctor shopping enrollees in their practice (OR=1.3, CI: 1.1 1.5). E nrollees that filled at least one benzodiazepine pre scription during the study period (OR=1.3, CI:1.1 1.5) were more likely have an OA/D diagnosis . The full model (2011), with an AIC =4699 and c statistic =0.761 , performed better than any single domain (see Table 4 4 ). The risk and protective factor s were l argely consistent when we performed the sensitivity analysis using 2012 data with two exceptions : the s tatistical d ifferences between H ispanics and White non Hispanics and age categories were no longer significant at the 5% level . Receiver Operator Curve C haracteristics of Two Models Figure 4 2 shows the receiver operator curves (ROC) for the full model, the PDMP like model, and a hypothetical model with no discriminatory ability (i.e., no better than random) . For reference, a horizontal line is shown at 0 .80 (sensitivity) and a vertical line at 0.52 (specificity) representing the characteristics of the R SOAPP screener for opioid misuse. 177 The full m odel has a c statistic or an area under the curve (AUC) of 0.76 ; interpreted as the probability that when two enrollees are selected at random, that the enrollee with the higher predicted probability from the model will be diagnosed with OA/D and the other will not . While we expected the full model to perform better, t he PDMP like model is better than random (AUC= 0.65) . Given the role of mental health disorders in our study, w e examined two independent mediation pathways between mental he alth disorder s and opioid abuse and dependence. Path 1 was potentially mediated by painful conditions (e.g. arthritis,
125 headache, or back pain) and Path 2 by non opioid SUDs . To be considered a mediator, path variables needed to be associated with 1) opioid abuse and dependence (outcome), 2) the mental health disorder category (causal exposure), and 3) modify the direct relationship between the outcome and the causal exposure ( i.e., a significant change in the odds ratio ). Both pathways met the first two criteria. W hen we modeled Path 1 , the odds ratio for mental health disorder did not change significantly , suggesting no effect modification. By contrast, the odds ratio along Path 2 decreased from 3.53 (CI: 2.95, 4.22) to 2.54 (CI: 2.11, 3.06) , indicating that non o pioid SUDs partially mediated the relationship between opioid abuse or d ependence and mental health. See Appendix M and N for the results of the mediation analysis. Discussion To our knowledge, this is one of only two studies that identify risk factors as sociated with clinically recognized opioid abuse or dependence for Medicaid enrollees exposed to Sc hedule II prescription opioids. We estimated the prevalence of opioid abuse or dependence in the Texas Medicaid population exposed to at least one Schedule II prescription opioid at 6.3%. According to Edlund et al. the prevalence of opioid abuse or dependence was 2.9% in the Arkansas Medicaid population , but that sample was restricted to just the population receiving 90 days of continuous opioid therapy of any Schedule type . 162 However, amo ng subjects receiving Schedule II prescription opioids in the Edlund study , the prevalence ranged from 3.8% to 7.8% , which is consistent with our finding . We identified several unique risk factors in this study . This is the first study of its kind to show that Black non Hispanics are less likely to have clinically recognized opioid abuse or dependence than White non Hispanics. Two other claims based studies of
126 opioid abuse, t he TROUPE study using Arkansas Medicaid data and a nationally representative study of the commercially insured population , did not evaluate the role of race and ethnicity in their models . 67,162 Edlund et al. reported similar findings for Blacks and Whites but from a population of v eterans . 180 Explanations for this finding are unclear but Blacks in the Texas Medicaid program appear to have less exposure to the most addictive opioids because they were significantly underrepresented in the Schedule II population ( see Appendix J ). T wo studies conducted in emergency department settings also show ed that B lacks were less li kely to receive an opioid analgesic. 183,184 Becker et al. found that Blacks were more likely to receive 2 of the 3 recommended guidelines for opioid monitoring in a primary practice. 185 As expected, o ur findings indicate that patients diagnosed with opioid abuse or dependence are more likely to have co occurring substance abuse disorders and other major mental health disorders. Having a painful physical condition (e.g., arthritis, headaches, or back pain ) was not associat ed with opioid abuse or dependence when there was a documented co occurring mental health disorder . For thes e patients, our finding s suggest that interventions to address opioid abuse or dependence should consider 1) mental health needs and 2) existing substance use problems . Our preliminary mediation path anaylsis has limited interpretability given the cross s ectional nature of our analysis. This is one of only a handful of studies to examine the association of doctor shopping and the provider patient network characteristics associated with opioid abuse or dependence. Cepeda et al. found that doctor shoppi ng behavior and oxycodone abuse were closely correlated in a private claims database. 179 We found that doctor
127 shoppers 4 distinct prescribers and distinct dispensers of a Schedule II opioid during a one year period ) were more likely to have clinically recognized opioid abuse or dependence. Also, w e found that doctor shoppers tended fill prescrip tions from prescribers that had e ncounters with other doctor shoppers. Our finding s have several important implications . First, at least one prescriber recogniz ed an abusive opioid behavior . Given the cr oss sectional nature of this study, w e do not know if the knowledge of the doctor shopping behavior gave rise to the diagnosis. We believe this is unlikely in Texas because the PDMP, one of the primary tools available for identifying doctor s hoppers, is administered by a law enforcement agency and is not used frequently by prescribers. PDMP utilization data available from state PDMPs and data linkage to administrative claims can help address this research gap. Our findings highlight the fact that there is a larger group of prescribers that encountered a n enrollee that was doctor shopping for a Schedule II prescription opioid and did not diagnose opioid abuse or dependence . Second, provider encounters with multiple doctor shoppers in their p ractice might modify a Schedule II opioids. Baldacchino et al. described physician attitudes towards chronic non cancer pain patients with a history of substance abuse that were doctor shopping as a mentality that resulted in a reluctance to prescribe. 186 The median daily morphine milligram equivalent (MME) dose for Te xas Medicaid enrollees is cause for concern. The median MME for enrollees exposed to Schedule II prescription opioids regardless of their opioid abuse or dependence status was higher
128 than the minimum high risk threshold (>=100 mg/day) recommended by the C DC. 147 Our study is likely an underestimate of the MME given that cash payments and non Medicaid prescriptions are not included . One study f ound that in a six month period , 478 (0.02%) clients in Was h both Medicaid and cash prescriptions for the same c ontrolled substance with 25% of those clients filling on the same day. 187 Our risk factor model based on administrative data had reasonable preliminary receiver operator curve characteristics. When we restricted the regression model to variables captured by PDMPs , the model was less sensitive and specific for opioid abuse and dependence but better than random chance . Given the clinical challenges associated with identifying opioid abuse, t his finding suggests that data systems with pharmacy transactions (e.g. Medicaid, PDMPs ) might offer some ability to augment and improve risk based screening for abusive behaviors when u sed in combination with an instrument like R SOAPP . The National Associat ion of Boards of Pharmacy offers a proto type screening tool based on this principle known as NAR x CHECK but the diagnostic characteristics have not been reported. 188 Our study is subject to several important limitations. We are likely underestimating the prevalence of opioid abuse or dependence because we 1) utilize d physician recognized cases of opioid abuse or dependence which may represent only the most severe cases in the population and 2) examined diagnoses recorded within a single year. We do not know the relative proportion of enrollees abusing or dependent o n prescription opioids versus heroin . However, the growing availability of prescription opioid s during this period in Texas is similar to the overall US trends towards higher
129 prevalence of prescription opioids abuse than heroin. 189 Our use of administrative claims data was consistent with other studies, but we are unaware o f research that has cross validated ICD 9 CM codes with accepted clinical instruments (e.g., R SOAPP ) for identifying opioid abuse or dependence . T he cross sectional nature of our study did not permit us to examine the risk factors for opioid use or depen dence prospectively . Our PDMP like administrative model offers some preliminary insight on detecting opioid abuse or dependence however our data does not capture pharmacy transactions made in cash or by non Medicaid payers. Finally, we do not know the ex tent to which our results are generalizable to other managed Medicaid programs. This is the largest study of the risk and protective factors for opioid abuse or dependence using Medicaid administrative claims to date. Our findings suggest that addressin g the prescription opioid abuse epidemic in this population requires a focus on underlying mental health, including other substance use disorders, and deserves particular attention in Texas where access to Medicaid funded outpatient substance use disorder treatment is among the lowest in the country.
130 Table 4 1. Descriptive statistics for Texas Medicaid enrollees exposed to at least one Schedule II prescription opioid with clinically recognized opioid dependence/abuse, 2011 Opioid Abuse/Dependence All (N= 11,326) p value Yes (n=713) No (n=10,613) N % N % N % Race/Ethnicity 383 53.72 5,562 52.41 5,945 52.49 White non Hispanic 0.0294 Black non Hispanic 81 11.36 1,550 14.60 1,631 14.40 Hispanic 172 24.12 2,228 20.99 2,400 21.19 Other 77 10 .80 1,273 11.99 1,350 11.92 Sex 451 63.25 6,967 65.65 7,418 65.50 Female 0.1934 Male 262 36.75 3,646 34.35 3,908 34.50 Resides in Metropolitan Area 623 87.38 8,903 83.89 9,526 84.11 Yes 0.0136 No 90 12.62 1,710 16.11 1,800 15.89 Age group 17 2.38 710 6.69 727 6.42 18 to 25 years old <0.0001 25 to 44 years old 265 37.17 3,164 29.81 3,429 30.28 > 44 years old 431 60.45 6,739 63.50 7,170 63.31 Clinical Risk Group 7 0.98 393 3.70 400 3.53 Healthy <0.0001 Significant Acute 4 0.56 147 1.39 151 1.33 Minor 3 0.42 414 3.90 417 3.68 Moderate 49 6.87 862 8.12 911 8.04 Major 474 66.48 5,537 52.17 6,011 53.07 Unassigned 176 24.68 3,260 30.72 3,436 30.34
131 Table 4 1 Continued Opioid Abuse/Dependence All (N=1 1,326) Yes (n=713) No (n=10,613 ) N % N % N % p value Non opioid related substance use disorder 350 49.09 1,604 15.11 1,954 17.25 Yes <0.0001 No 363 50.91 9,009 84.89 9,372 82.75 Mental health disorder 549 77.00 5,168 48.69 5,717 50.48 Yes <0.0001 No 164 23.00 5,445 51.31 5,609 49.52 Backpain 572 80.22 7,559 71.22 8,131 71.79 Yes <0.0001 No 141 19.78 3,054 28.78 3,195 28.21 Arthritis 638 89.48 9,107 85.81 9,745 86.04 Yes 0.0062 No 75 10.52 1,506 14.19 1,5 81 13.96 Headache 263 36.89 2,861 26.96 3,124 27.58 Yes <0.0001 No 450 63.11 7,752 73.04 8,202 72.42 Medicaid Program 119 16.69 2,387 22.49 2,506 22.13 FFS 0.2743 NORTHSTAR 49 6.87 549 5.17 598 5.28 PCCM 153 21.46 2,594 24.44 2,74 7 24.25 STAR 65 9.12 1,189 11.20 1,254 11.07 STARPLUS 327 45.86 3,894 36.69 4,221 37.27 Transitioned to managed care 164 23.00 2,411 22.72 2,575 22.74 Yes 0.8092 No 549 77.00 8,202 77.28 8,751 77.26 Multiple Provider Episode Level (j=presc riber, k=dispenser) 482 67.60 8,490 80.00 8,972 79.22 1j x 1k <0.0001 2j x 2k 156 21.88 1,556 14.66 1,712 15.12 3j x 3k 48 6.73 415 3.91 463 4.09 4j x 4k 17 2.38 109 1.03 126 1.11 5j x 5k 10 1.40 43 0.41 53 0.47
132 Table 4 1. Continued Opioid Abuse/Dependen ce All (N=11,326) Yes (n=713) No (n=10,613 ) N % N % N % p value Has Out of State prescriber in Rx Network 492 69.00 7,952 74.93 8,444 74.55 No 0.0004 Yes 221 31.00 2,661 25.07 2,882 25.45 Has prescriber in non PDMP state 638 89.48 9,852 92.83 10,490 92.62 No 0.0009 Yes 75 10.52 761 7.17 836 7.38 Provider has prescribed to 2+ doctor shoppers 520 72.93 6,425 60.54 6,945 61.32 Yes <0.0001 No 193 27.07 4,188 39.46 4,381 38.68 >=1 benzodiazepine prescription in past year 473 66.34 5,211 49.10 5,684 50.19 Yes <0.0001 No 240 33.66 5,402 50.90 5,642 49.81
133 Table 4 2. Bivariate associations between risk factors and having clinically recognized opioid abuse and dependence , Texas Medicaid, 2011 Domain Variable Esti mate Standard Error Wald Chi Square Pr > Chi Square OR LL UL Sociodemographic (1) Black 0.2756 0.1256 4.8134 0.0282 0.76 0.59 0.97 Hispanic 0.1143 0.0951 1.4442 0.2295 1.12 0.93 1.35 Other 0.1296 0.1287 1.0147 0.3138 0.88 0.68 1.13 Male 0.1044 0. 0803 1.6904 0.1935 1.11 0.95 1.30 25 to 44 y.o. 1.2522 0.2536 24.3771 <.0001 3.50 2.13 5.75 > 44 y.o. 0.9825 0.2504 15.3953 <.0001 2.67 1.64 4.36 Metropolitan 0.2846 0.1158 6.0387 0.014 1.33 1.06 1.67 Health Status (2) Major chronic health conditi on 0.5978 0.0817 53.5691 <.0001 1.82 1.55 2.13 Non opioid substance abuse 1.6893 0.0797 449.6421 <.0001 5.42 4.63 6.33 Mental health disorder 1.2602 0.0911 191.46 <.0001 3.53 2.95 4.22 Backpain 0.4941 0.0964 26.2497 <.0001 1.64 1.36 1.98 Arthritis 0.3407 0.1252 7.4078 0.0065 1.41 1.10 1.80 Headache 0.4597 0.0806 32.4943 <.0001 1.58 1.35 1.86 Provider Episodes (3) NORTHSTAR 0.5824 0.1762 10.923 0.0009 1.79 1.27 2.53 PCCM 0.1682 0.1255 1.7959 0.1802 1.18 0.93 1.51 STAR 0.0922 0.1583 0.3393 0.5602 1.10 0.80 1.50 STAR + PLUS 0.5214 0.1102 22.4019 <.0001 1.68 1.36 2.09 Doctor shopper 0.9965 0.2125 21.9838 <.0001 2.71 1.79 4.11 Prescriber w/ 2+ doctor shoppers 0.5632 0.0866 42.2938 <.0001 1.76 1.48 2.08 Spatial (4) Out of state prescriber 0.2946 0.084 12.2947 0.0005 1.34 1.14 1.58 N on PDMP out of state prescriber 0.4199 0.1277 10.8086 0.001 1.52 1.19 1.96 Medication Profile (5) 1+ benzodiazepine in past year 0.7145 0.0816 76.6678 <.0001 2.04 1.74 2.40 Daily dose MME (log 10) 0.6909 0 .1156 35.6992 <.0001 2.00 1.59 2.50 Significant associations in bold. Reference categories: White non Hispanic, 18 25 years old.
134 Table 4 3. Contextual domain modeling results. Odds ratios are only adjusted for other variables with the contextual dom ain block Domain Effect Odds Ratio Estimate Lower 95% Confidence Upper 95% Confidence AIC c statistic Socio demographic (1) Black 0.749 0.584 0.959 5283 0.579 Hispanic 1.118 0.926 1.350 Other 0.961 0.745 1.240 Male 1.175 1.002 1.378 25 to 4 4 years old 3.589 2.175 5.923 > 44 years old 2.728 1.666 4.469 Metropolitan residence* 1.352 1.075 1.700 Health Status (2) Major chronic health condition* 1.240 1.046 1.470 4803 0.736 Non opioid SUD 4.164 3.538 4.900 Mental health disorder 2.383 1.966 2.889 Back Pain* 0.973 0.795 1.192 Headache* 1.161 0.983 1.372 Provider Episodes (3) NORTHSTAR* 1.496 1.053 2.124 5148 0.634 PCCM 1.194 0.931 1.531 STAR 1.031 0.753 1.411 STARPLUS 1.278 1.024 1.597 Health care encounte rs (log (10)) 2.244 1.869 2.693 Doctor Shopper 1.971 1.289 3.013 1+ Out of State Prescriber* 1.019 0.859 1.210 Provider with 2+ doctor shoppers 1.351 1.124 1.623 Medication Profile (4) 1+ benzodiazepine in past year 1.964 1.673 2.307 5221 0. 614 Daily dose MME (log(10)) 1.829 1.455 2.301 Reference for race/ethnicity is White, non Hispanic., Substance Use Disorder (SUD) *candidates for exclusion on the basis of statistical significance. **We re ran this domain (minus the Medicaid program type) to mimic the limited information available to most PDMPs. The AIC was 5770.6 and c statistic 0.651. Summary of backward elimination: out of state , headache, backpain, CRG, and metro are removed.
135 Table 4 4. 2011 and 2012 modeling results 2011 ( AIC=4699, c statistic =0.761) 2012 (AIC=5905, c statistic=0.768) Domain Effect Odds Ratio Estimate Lower 95% Confidence Upper 95% Confidence Odds Ratio Estimate Lower 95% Confidence Upper 95% Confidence Sociodemographic Black 0.721 0.556 0.935 0.499 0.393 0.632 Hispanic 1.260 1.035 1.535 1.007 0.844 1.202 Other 0.940 0.720 1.225 0.929 0.754 1.143 Male 1.186 1.003 1.402 1.262 1.091 1.460 25 to 44 years old 1.785 1.059 3.007 1.237 0.859 1.781 > 44 years old 1.307 0.782 2.187 0.904 0.632 1.294 Heal th Status Non opioid substance use disorder 3.862 3.266 4.566 3.623 3.136 4.186 Mental health disorder 1.989 1.629 2.429 1.909 1.598 2.282 Provider Episodes Doctor Shopper 1.767 1.128 2.768 1.483 1.038 2.118 Provider with 2+ doctor shoppers 1.265 1.05 4 1.518 1.235 1.057 1.442 Medication Profile 1+ benzodiazepine in past year 1.279 1.074 1.524 1.180 1.014 1.373 Daily dose MME (log(10)) 2.288 1.777 2.947 1.960 1.563 2.458 Notes: The 2012 sample contained 975 patients diagnosed with opioid abuse or dependence and 11,177 controls. *Adjusted for the number of health care encounters
136 Figure 4 1. Forest plot of adjusted odds ratios for 2011.
137 Figure 4 2. Receiver operator curves (ROC) for the full and PDMP like opioid abuse or dependence models, Texas Medicaid, 2011. The area under the curve for the full and PDMP like models are 0. 76 and 0.65, respectively. Note: The full model uses all the variables shown in Table 4 4 whereas the PDMP model is restricted to just sociodemographi c s, provider epi sodes, and medication profiles ( excluding the health status domain).
138 . CHAPTER 5 CONCLUSIONS Accomplishments of the Dissertation Prescription opioid abuse is a national epidemic. 5,155 The objective s of this dissertation were to examine 1) the effect of prescription drug monitoring on oxycodone caused mortality , 2) risk factors associated with an aberran t drug seeking behavior associated with opioid addiction risk factors associated with opioid abuse and dependence including doctor shopping . We expected the first objective to provide new knowledge of the effectiveness o f existing prescription drug monitoring programs for reducing deaths from oxycodone. Results have national implications because the call for expanded use of PDMPs is widely supported. 28,82,116 We expected our second objective to provide novel insight on the risk factors associated with doctor shopping for Schedule II prescriptio n opioids. Doctor shoppers o btain large quantities of prescription opioids from medical sources. Doctor shoppers are generally treated as criminals who divert medication to the illicit market and a problem for law enforcement. This dissertation sought to bring an epidemiologic pers pective to this problem. We used an ecologic modeling approach to gain insight on the risk factors related to socioeconomics, mental health conditions, the healthcare system, overlapping medications, and geographic location. In doing so, we provide insig ht to change the current perspective from preventing criminal behavior towards preventing the compulsive drug seeking associated with addictive behavior. By doing so, we can pursue public health interventions to address the underlying causes of opioid addi ction. Likewise, our third objective was to use the same modeling framework to examine the risk factors associated with prescription opioid abuse and dependence.
139 We also wanted to understand how prescription drug monitoring programs might be made more us eful to providers on the front lines of the prescription drug abuse epidemic. We summarize our key findings below. We showed that prescription drug monitoring helped reduce oxycodone caused deaths in Florida by approximately 4% , or 30 people per year . We hypothesize d that one possible effect mechanism is related to reduc ing the population of doctor shopp ers. In the two year period from 2012 to 2013, the number of patients identified as doctor shoppers in the Florida PDMP decreased by approximately half . However, questions remains regarding the mechanism of the effect, for example was the reduction driven by law enforcement agents using the PDMP database to identify and eliminate doctor shoppers ? or did medical providers use the PDMP database to make cl inical decisions that improved patient health outcomes including better opioid management? We showed that doctor shoppers were more likely to be in provider networks with other doctor shoppers a large Medicaid population. The extent to which this structu ral association is a function of PDMP use by some providers and not others is an area of future research. We found a strong positive association between patients diagnosed with opioid abuse and dependence and doctor shopping behaviors . We also found tha t information contained in PDMP s provides better than random diagnostic capability for these patients . To do this, we created a model from Medicaid claims data that was similar to the sociodemographic, prescription, and geographic data elements available from most PDMPs.
140 Future Directions and Policy Recommendations Below we propose 6 new studies based on knowledge gained from the current dissertation findings: 1) replication of the PDMP study in multiple states , 2) examining the effect of PDMPs on addition al population level health outcomes, 3) improving drug related surveillance, 4) prospectively evaluating PDMP data as a screening tool for prescription opioid abuse, 5) network analysis of providers, dispensers, and enrollees receiving Schedule II prescrip tion opioids in Medicaid , and 6) examining the effects the provider level characteristics on doctor shopping in Medicaid. National level P rescription D rug Monitoring P rogram S tudy The first study of the dissertation focused on evaluating the presc ription drug monitoring program implemented in 2011 in Florida. To extend this work, w e propose a national evaluation of prescription drug monitoring programs using an interrupted time series quasi experimental design to further test and validate the effe ct of implementing PDMPs on prescription drug related deaths . In contrast to our study in Florida, most states implemented PDMPs several years prior to 2011 . 133 Appendix O is a visual guide showing that approximately 26 states operationalized PDMPs prior to 2008. Data from these states should provide a s ufficient number of pre post monthly observatio ns required for statistical balance and power in the time series. Monthly prescription drug related mortality can be obtained from t he all cause mortality data and is publicly available through the end of 2011. 123 Analytically, w e propose to 1) develop state specific auto regressive integrated moving average ( ARIMA ) models , 2) use related mortal ity as external controls, and 3) use inverse weighting methods to pool results across states for a single effect size estimate . 190 Variation in PDMP implementation and operations
141 such as mandatory versus voluntary prescriber registration , proactive versus reactive patient prescription history repor ting, and administration by a public health agency versus a law enforcement agency provides an opportunit y to understand PDMP mechanisms moving forward. During this dissertation, we built a database to monitor state by state differences in PDMPs using nat ional resources from organizations such as the Prescription Drug Monitoring Program Training and Technical Assistance Center and the Public Health Law Research project of the Robert Wood Johnson Foundation. 151,191 Negative Health Outcomes P roxim al to M ortality The second proposed study focuses on other negative health outcomes associated with prescription opioid abuse. Increases in substance abuse treatment admissions, emergency department visits, and, most disturbingly, overdose deaths attributable to prescription drug abuse place enormous burdens upon communities across the country. R.Gil Kerlikowske, Director of the Office of Nat ional Drug Control Policy 93 The quote above from the Director of the Office of National Drug Control Policy references a chain of outcomes related to prescription drug abuse. The CDC quantified of outcomes for prescription opioids in a Morbidity and Mortality Weekly Report in 2008 . Appendix P shows a graphic representation of the prescription opioid 22 To extend this work, w e propose an empirical test t heory using the time series analytic framework we presented in Chapter 2 . 96 To review, we used the total deaths averted from our results and stepped backwards along the cascade to estimate that the PDMP reduced (per 100 population at risk): treatment admissions for abuse by 41 people, emergency
142 department visits for misuse or abuse by 132 people, people who abuse or are dependent by 533 people, and people who take prescription painkillers for nonmedical use by 3,383 people (see Table 2 7). Several of these suggested outcomes are estimable using currently available data. For example, the Drug Abuse Warning N etwork (DAWN) collects detailed data from emergency department visits from several metropolitan areas in the US, including the Miami Fort Lauderdale area. Appendix Q shows the annual rate per 100,000 population of emergency department visits for non medic al use of all opiates, oxycod one combinations, and oxycodone only in the Miami Fort Lauderdale metropolitan area from 2008 to 2011. The downward trend in misuse starting in 2010 is consistent with the decline in oxycodone caused deaths around the same tim e. By obtaining monthly data from the DAWN system and confirming its validity for this purpose , we can test the hypothesis that the PDMP contributed to a decrease in emergency department visits for misuse or abuse. This would be an important contribution to the PDMP effectiveness literature and help bolster support fro m key stakeholders at the state level. Improving Drug related S urveillance in Florida The third proposed study is a policy recommendation on improving state based surveillance systems on d rug related mortality. In Florida, drug related deaths are investigated by medical examiners (ME) in 24 districts and reported vertically to a central agency known as the Medical Examiners Commission (MEC). The public use version of these data are advant ageous because it includes active ingredient specific causes of death and the reports are more timely than the national equivalent for monitoring drug mortality maintained by the CDC . For example, t he MEC reports oxycodone caused rather than just opioid c aused ( which is a drug class that include s
143 hydrocodone ) and the dataset in Florida is available more than one year in advance of national data. Despite these advantages , the MEC data are still delayed by more than a year. This time lag can limit the epi demiologic impact of these data for detecting new mortality trends. For example, in late 2011, oxycodone caused deaths significantly decreased in Florida, but heroin caused deaths subsequently increased (unpublished data) . This trend was detected late and only recently ha ve we begun s ys tematic research on possible reasons for the increase (e.g., substitution effects, changes in pricing ). Increasing the timeliness of the MEC reporting system will increase the opportunities for faster detecti on of similar t rends in the future. We believe that streamlin ing the data collection process by creating a single web based, secure data entry portal will be a cost effective approach for improving timeliness . Currently, the MEC emails 24 separate database s containing the data collection form s to each of the ME districts at the beginning of the year. Use of the database is problematic for some districts due to limited local technical capacity and some programming glitches. As a result, some districts modify the data structure , send data in different formats , and/or face technical barriers resulting in delay . Furthermore, the data are only reported at two times during the year June (interim) and December (final). We propose the creation of a secure, web based data collection system to resolve these technical issues, standardize data collection, and provide an on going data repository. This new database can be queried at any time during the year by medical examiners , the MEC data quality committee and others with t he appropriate
144 credentials. T he MEC commission could establish a new policy for incentivizing on going data entry to the system such as real time data quality feedback reporting that would minimize errors throughout the year and reduce end of year workloa d. Improving this process would allow staff to respond to critical drug surveillance questions faster and detect local, district, and state emerging threats earlier. This format would also encourage and facilitate higher utilization among public health r esearchers. Appendix R shows a screen shot of an online interactive data visualization of MEC data created during this dissertation work and presented to the MEC in February 2014. Prospective use of PDMP Data to Identify Prescription Opioid Abuse The four th proposed study is prospective analysis of prescription drug monitoring programs as a clinical tool for detecting prescription opioid abuse and dependence. Systematic reviews of the literature on identifying opioid misuse and abuse in clinical populatio ns have concluded that the Revised Screener and Opioid Assessment for Patients with Pain ( R SOAPP is the most reliable prospective screening tool for opioid abuse to date . Most stakeholders recognize that the complex nature of opioid abuse requires mult iple and simultaneous strategies that might reasonably be implemented by providers . 32,176,192,193 One such strategy is the use of p rescription drug monitoring programs (PDMPs) which were designed to i dentify u nu sual prescription patterns such multiple overlapping prescriptions and large numbers of multiple providers. A wide range of health care stakeholders recommend that PDMPs be integrated into routine clinical practice. 28,82,176 For example, Sehgal et al. provide an algorithmic approach for long term opioid therapy in chronic pain in which checking a state PDMP is integrated into routine medication adherence monitoring . 176 The current
145 evidence suggests that when PDMP information is integrated into practice, physici ans change medical practice including their prescription choices. 40,41 In the current dissertation work, we identif ied risk factors for prescription opioid abuse and dependence and we examined the predictive characteristics of PDMP like information ab ility to discriminate bet ween patients with opioid abuse and dependence and those without . T he PDMP like information we modeled from a Medicaid administrative dataset performed better than random with an area under the curve suggesting that it would be an informative screening tool (AUC=0.65) . We propose a follow up validation study to compare the R SOAPP alone t o a combined, clinical assessment tool consisting of the R SOAPP and the predicted probabilities established for identifying opioid abuse from PD MP based models like the one presented in our study . R esults of the current study can be applied in a clinical setting . For instance, u sing the receiver operator curve ( ROC ) generated from our binary regression model (see Figure 4 2 ) we can determine t he optimal cut point (i.e., the predicted probability produce d by the model where the ratio of the true positive rate to the false positive rate is maximized ) . The combined clinical assessment tool will consist of an R SOAPP score and a yes/no score for PDMP identified opioid abuse based on the cut point after accounting for patient specific effects (e.g., age, race/ethnicity, gender) entered into the model. The field appears to moving in this direction. For example, t he National Association of Boards of Pharmacy offers a proto type screening tool known as NAR x CHECK but the diagnostic characteristics have not been reported. 188
146 Network Analysis of Doctor Shoppers A fifth study could focus on improving our understanding of prescription opioid prescribing networks . Using network analysis, w e propose a follow up study of doctor shopping for Schedule II prescription opioids using our current Medicaid dataset . We performed a brief literature review in PubMed and either [prescription drug, prescription medication, opioid, prescrib er, doctor shopper, Schedule II (2), Medicaid ] and no time frame restriction. We only identified three relevant studies. 194 196 The following is a brief review of those studies within the conte xt of this dissertation. Cavallo et al. concluded that prescription drug networks in Italy showed scale free behavior . 194 A discussion of scale free behavior is beyond the scope of this dissertation but Schneeberger et al. provides a nice summary of the policy and intervention implications of scale free network structures from the perspective of sexually transmitted disease partnerships. 197 They argue that observing a scale free structure supports the notion of targeting the highest risk individuals in the network and that transmission can only be eliminated by reducing the number of new networ k contacts. This concept is consistent with providers associated with doctor shopping networks. The next steps would involve formal test s of the hypothesis that doctor shopping networks for Schedule II presc ription opioids are scale free in the Medicaid population. Jonas et al. analyzed a network of Oxycontin Â® abusers in rural Appalachia and found that all drug users than marijuana only abusers. 195 The Medicaid administrative data does not support a comprehensive understanding of social relationships . F urther refinement of our geographic analyses
147 might help us infe r a certain degree of socioeconomic connectivity in the absence of true observed relationships. Fraser et al. found no demographic distinctions in the street drug networks of opioid and non opioid abusers but found multiple statistical differences betwee n their network characteristics . Those differences explained approximately one third of the variation between drug abuser groups . Our findings reported here are consistent with . We observed relatively modest sociodemographic differences ; some n etwork characteristics appeared to be more explanatory for doctor shoppers . Using network analysis, we can test the hypothesis that networks of both doctor shoppers and those diagnosed with opioid abuse and dependence are quantitatively different than the ir respective controls . Provider level Effects on Doctor Shopping Finally, we propose a follow up study to examine the outcomes used in this dissertation and their associations with provider level characteristics. Using an ecologic framew ork that included prescriber and dispenser level variables, we found that prescriber characteristics, such as prior experience with prescribing Schedule II opioids to other doctor shoppers, are factors highly associated with a propen sity for d octor shopping . Additional factors might include physician practice setting (i.e., emergency room, family practice, pain clinic), medical specialty (e.g., psychiatry, emergency medicine) , and frequency of utilization of state prescription drug m onitoring programs (PDMPs). States such as Kentucky and Florida have report ed significant reductions in the prevalence of doctor shopping once the ir PDMP s beg an operating . Appendix S shows the quarterly reduction in the reported number of doctor shoppers from 2012 to 2013 in Florida. Based on this limited evidence, PDMPs
148 appear to be effective at reducing doctor shopping. 114 However, the mechanism by which doctor shoppers are reduced and/or prevented from entering the he alth care system is not clear. This dissertation hypo thesize s the existence of two primary mechanisms for reducing doctor shopping : 1) law enforcement prosecution and 2) health provider intervention . We propose a secondary data analysis of survey responses collected by the Florida PDMP from its registered user s that would begin to address this research question . In preliminary analyses of these survey responses (n=1,046) from health care professionals who use the PDMP in Florida, we found that the vast majority (>90%) of both prescribers and dispensers find the PDMP useful in cont rolling doctor shoppin (see Appendix T). When asked you taken as a result of using the PDMP system to monitor controlled substance approximately 15% o f providers reported that they confirmed that a patient was doctor shopping. Taken together, these responses raise three important questions for future research . First, how do providers process the information provided by the PDMP to reach the decision t hat a patient is doctor shopping? The Florida PDMP did not assist providers in directly identifying doctor shoppers . (Florida PDMP, personal communication) In the absence of a direct report busy providers must analyze a activity report of prescr ibe d medications to reach this decision (see Appendix C for an example of the report in Florida ). Second, how do providers respond when a doctor shopper has been identified? Green et al. found that when the Connecticut PDMP was used in practice, a clini cal response (e.g. referral , drug abuse screen, revisit pain/treatment agreement, urine screen) was
149 more likely to occur than a legal intervention. 41 The Florida PDMP survey is consistent with that study. The vast majority of providers do not appear to report doctor shopping patie nts to law enforcement (i. e., less than 0.17% report). Third , to what degree do prescribers, dispensers, and law enforcement communicate with each other to identify doctor sh oppers? In terms of the health care providers, t he PDMP does appear to increase c ommunication . This program will likely increase communication between providers tely 72% of PDMP users agreed. However, t he nature and effect of this communication is unclear . Currently, the American Medical Association is concerne d that dispensers intrud on prescriber decisions. 198 It is not clear whether the PD MP is germane to this debate but further research is warranted. In conclusion, we found that prescription drug monitoring in Florida reduced oxycodone cause mortality in that state by 4%. We identified several unique risk factors for doctor shoppin g for Schedule II prescription opioids as well as opioid abuse or dependence among those exposed to Schedule II prescription opioids in a large Medicaid population. The individual level, health related risk factors (e.g., major chronic health conditions , opioid abuse or dependence , mental health disorders) that we identified highlight the need for an epidemiologic approach to preventing doctor shopping . Finally, w e found that doctor shopping is associated with structural level characteristics such as a pr . These findings may provide the opportunity to intervene at multiple levels moving forward.
150 APPENDIX A DRUG DEATH REPORTING SYSTEM The MEC is a division of the Florida Department of Law Enforcement. The green side of the graph shows data flow after the system was automated in 2010. Source : Adapted from John Thogmartin, MD Pinellas County Chief Medical Examiner .
151 APPENDIX B OPIOID AND BENZODIAZ EPINE PRESCRIPTION DRUGS REPORTED TO Type Family Drugs Reported* FL CSA** Narcotic Opioids Oxycodone, buprenorphine, codeine, fentanyl, heroin, hydrocodone, hydromorphone, meperidine, methadone, mo rphine, oxymorphone, propoxyphene, Tramadol 2 Buprenorphine 3 Sedative hypnotic Benzodiazepines Alprazolam, alpha hydroxyalprazolam, clonazepam, 7 aminoclonazepam, chlordiazepoxide, desalkyflurazepam, diazepam, estazolam, flunitrazepam, flurazepam, lor azepam, midazolam, nordiazepam, oxazepam, temazepam, triazolam, alpha hydroxytriazolam 4 *Other prescription drugs reported include carisoprodol, meprobamate, and Zolpidem. Ethanol (i.e., alcohol) and illicit drugs such as cocaine and heroin are also rep orted. **CSA=Controlled Substance Act. The Florida Prescription Drug Monitoring Program (PDMP) monitors prescriptions for controlled substances scheduled 2,3, and 4.
152 APPENDIX C DE IDENTIFIED PATIENT ACTIVITY REPORT (PAR) REQUESTED FROM THE FLORIDA PRE SCRIPTION DRUG MONITORING PROGRAM
153 APPENDIX D NUMBER OF INACTIVE PAIN CLINIC LICENSES AND % CHANGE BY MONTH AND YEAR, FLORIDA (AS OF NOVEMBER 2013) Date Running Total % Change (monthly) % Change (yearly) Jan 10 0 0.0% Feb 10 0 0.0% Mar 10 0 0.0% A pr 10 0 0.0% May 10 2 0.0% Jun 10 2 0.0% Jul 10 2 0.0% Aug 10 2 0.0% Sep 10 97 4750.0% Oct 10 121 24.7% Nov 10 145 19.8% Dec 10 196 35.2% Jan 11 288 46.9% Feb 11 324 12.5% Mar 11 354 9.3% Apr 11 390 10.2% May 11 444 13.8% 22100% Jun 11 485 9.2% 24150% Jul 11 571 17.7% 28450% Aug 11 673 17.9% 33550% Sep 11 724 7.6% 646% Oct 11 752 3.9% 521% Nov 11 777 3.3% 435% Dec 11 811 4.4% 313% Jan 12 842 3.8% 192% Feb 12 893 6.1% 175% Mar 12 937 4.9% 164% Apr 12 980 4.6% 151% May 12 1010 3.1% 127% Jun 12 1026 1.6% 111% Jul 12 1050 2.3% 83% Aug 12 1056 0.6% 56% Sep 12 1068 1.1% 47% Oct 12 1072 0.4% 42% Nov 12 1079 0.7% 38% Dec 12 1093 1.3% 34%
154 APPENDIX E PERCENTAGE CHANGE IN PURCHASES OF OXYCODONE BY FLORIDA PHARMACIES FROM 2010 TO 2011 Month % Change Jan +1.5% Feb +2.7% Mar +2.8% Apr 16% May 12% Jun 10% Jul 14% Aug 13% Sep 22% Oct 25% Nov 29% Dec 29% Source: US Drug Enforcement Agency .
155 A PPENDIX F GEOGRAPHIC DISTRIBUTIONS Notes: Prescribers (n=23,523, triangle), dispensers (n=4,087, circles), and enrollees (n=16,212, squares) with claims for at least one Schedule II prescription opioid, Texas Medicaid, 2011 .
156 APPENDIX G THE VINCENTY EQUATION SAS PROC GEODIST uses the Vincenty equation for calculating the geodetic distance between two points on a spherical surface We refer to geodetic distance as geographic distance in this study. The equation is given by: (eq. 3 1) 1 1 ential address, and 2 2 ) are the coordinates for a provider practice location.
157 APPENDIX H PRESCRIPTION DRUG MONITORING PROGRAM OPERATIONAL STATUS, 2011 Notes: Gray=states with no operational PDMPs, white=states with operati onal PDMPs. 10 states lacked operational PDMPs: Georgia, Maryland, Arkansas, Wisconsin, Montana, Delaware, South Dakota, New Jersey, Alaska, and Washington. Four more began operating in 2011: Florida (October), Oregon (September), Nebraska (May), and Kans as (April).
158 APPENDIX I ICD 9 CM CODES USED TO CLASSIFY SUBSTANCE USE DISORDERS OPIOID RELATED Opioid Abuse (305.50, 305.51, 305.52, 305.53) Opioid Dependence (304.00, 304.01, 304.02, 304.03) Opioid with Other Drug Dependence (304.70, 304.71, 304.72, 30 4.73) NON OPIOID RELATED Alcohol Dependence and Abuse (303.90, 303.91, 303.92, 305.00, 305.01, 305.02) Amphetamine Dependence and Abuse (304.40, 304.41, 304.42, 305.70, 305.71, 305.72) Anti depressant Abuse (305.80, 305.81, 305.82) Barbiturate Dependence and Abuse (304.10, 304.11, 304.12, 305.40, 305.41, 305.42) Cannabis Dependence and Abuse (304.30, 304.31, 304.32, 305.20, 305.21, 305.22) Cocaine Dependence and Abuse (304.20, 304.21, 304.22, 305.6, 305.61, 305.62) Hallucinogen Dependence and Abuse (304.50 , 304.51, 304.52, 305.30, 305.31, 305.32) Drug Dependence and Abuse (304.60, 304.61, 304.62, 304.90, 304.91, 304.92, 305.90, 305.91, 305.92), Combination Drug Dependence (304.80, 304.81, 304.82).
159 APPENDIX J SCHEDULE II COMPARED TO OTHER PRESCRIPTION O PIOIDS Domain 2011 Schedule II Other Opioids All N % N % N % Sociodemographic (1) Race/Ethnicity 5,979 52.54 133,541 30.12 139,520 30.69 White Black 1,633 14.35 97,836 22.07 99,469 21.88 Hispanic 2,416 21.23 179,329 40.45 181,745 39.97 O ther 1,351 11.87 32,586 7.35 33,937 7.46 Age Group 18 24 724 6.36 177,091 39.95 177,815 39.11 25 44 3,472 30.51 177,190 39.97 180,662 39.73 >44 7,183 63.13 89,011 20.08 96,194 21.16 Sex 7,446 65.44 335,496 75.68 342,942 75.43 Female Male 3,933 34.56 107,796 24.32 111,729 24.57 Metropolitan 9,579 84.18 386,308 87.15 395,887 87.07 Yes No 1,800 15.82 56,984 12.85 58,784 12.93 Health Status (2) Clinical Risk Group 399 3.51 98,019 22.11 98,418 21.65 Healthy Sign ificant Acute 152 1.34 27,136 6.12 27,288 6.00 Minor 416 3.66 21,859 4.93 22,275 4.90 Moderate 920 8.09 44,797 10.11 45,717 10.05 Major 6,048 53.15 67,897 15.32 73,945 16.26 Unassigned 3,444 30.27 183,584 41.41 187,028 41.13 Opioid abuse and depe ndence 770 6.77 3,824 0.86 4,594 1.01 Yes No 10,609 93.23 439,468 99.14 450,077 98.99 Non opioid abuse/dependence 1,990 17.49 27,415 6.18 29,405 6.47 Yes No 9,389 82.51 415,877 93.82 425,266 93.53 Mental health disorder 5,756 50.58 91,092 20.55 96,848 21.30 Yes No 5,623 49.42 352,200 79.45 357,823 78.70 Backpain 8,175 71.84 111,588 25.17 119,763 26.34 Yes No 3,204 28.16 331,704 74.83 334,908 73.66 Arthritis 9,793 86.06 186,875 42.16 196,668 43.26 Yes No 1,586 13.94 256,417 57.84 258,003 56.74 Headache 3,136 27.56 68,971 15.56 72,107 15.86 Yes No 8,243 72.44 374,321 84.44 382,564 84.14 Provider Episodes (3) Medicaid Program Type 2,515 22.10 79,639 17.97 82,154 18.07 FFS NORTHSTAR 602 5.29 45,908 10.36 46,510 10.23 PCCM 2,753 24.19 126,152 28.46 128,905 28.35 STAR 1,261 11.08 144,181 32.53 145,442 31.99 STAR+PLUS 4,248 37.33 47,412 10.70 51,660 11.36 Transitioned to managed care 2,579 22.66 118,406 26.71 120,985 26.61 Yes No 8,800 77.34 324,886 73.29 333,686 73.39 Multiple Provider Episode Level 9,012 79.20 354,988 80.08 364,000 80.06 1j x 1k 2j x 2k 1,723 15.14 62,065 14.00 63,788 14.03 3j x 3k 464 4.08 15,968 3.60 16,432 3.61 4j x 4k 126 1.11 5,64 7 1.27 5,773 1.27 5j x 5k 54 0.47 4,624 1.04 4,678 1.03 Prescriber with 2+ doctor shoppers
160 Domain 2011 Schedule II Other Opioids All N % N % N % Yes 6,984 61.38 154,449 34.84 161,433 35.51 No 4,395 38.62 288,843 65.16 293,238 64.49 Spatial (4) 1+ Out of state prescriber 8,483 74.55 381 ,046 85.96 389,529 85.67 No Yes 2,896 25.45 62,246 14.04 65,142 14.33 1+ prescriber in non PDMP state 10,545 92.67 427,350 96.40 437,895 96.31 No Yes 834 7.33 15,942 3.60 16,776 3.69 Medication Profile (5) 1+ benzodiazepine in past y ear 5,715 50.22 56,240 12.69 61,955 13.63 Yes No 5,664 49.78 387,052 87.31 392,716 86.37 All 11,379 100.00 443,292 100.00 454,671 100.00
161 APPENDIX K DISTINCT MULTIPLE PROVIDER EPISODE MATRIX Distinct Dispensers 4 5 6 7 8 9 n Disti nct Prescribers 4 49 19 8 1 3 2 126 5 25 9 9 3 1 2 37 6 8 6 3 2 0 0 11 7 6 3 4 3 0 0 4 8 4 1 1 1 0 0 2 9 0 1 1 0 0 0 10 1 1 0 0 2 0 0 12 0 1 0 0 0 0 0 4 x 4 5 x 5 6 x 6 7 x 7 8 x 8 other N=180
162 APPENDIX L SPEARMAN CORRELATION COEFFIC IENTS FOR CATEGORICAL VARIABLES IN THE HEALTH STATUS DOMAIN Health Status Variables Clinical Risk Group* Mental Health Disorder Backpain Arthritis Headache** Non opioid SUD Clinical Risk Group 1 0.24999 0.19489 0.28464 0.12269 0.13122 Mental Health Disor der 1 0.22369 0.22701 0.18235 0.22725 Backpain 1 0.58107 0.17298 0.13277 Arthritis 1 0.15276 0.09712 Headache 1 0.09637 Non opioid SUD 1 Disorder, **i ncludes mi graines and tension headaches.
163 APPENDIX M ESTIMATED ODDS RATIOS FOR THE RELATIONSHIP BETWEEN MENTAL HEALTH OUTCOMES AND PAINFUL CONDITIONS VARIABLES, TEXAS MEDICAID 2011 Mental Health Outcomes (OR, 95% CI) Model Painful Condition Mental Health Disor der Non opioid substance abuse* Bivariate Arthritis 4.44 (3.91,5.03) 2.69 (2.23,3.25) Back pain 3.17 (2.91,3.46) 2.45 (2.15,2.79) Headaches 2.39 (2.20,2.61) 1.66 (1.49,1.84) Added to same model Arthritis 2.24 (1.92,2.60) 1.42 (1.13,1.79) Back pain 2.05 (1.84,2.28) 1.97 (1.68,2.32) Headaches 2.02 (1.85,2.21) 1.46 (1.32,1.62) *Shown for reference only. The OR and 95% CI for mental health disorders and non opioid substance use disorders is 3.82 (3.41, 4.27).
164 APPENDIX N ESTIMATED ODDS RATIOS AND 95% CIs FOR THE RELATIONSHIP BETWEEN POTENTIAL MEDIATING PATHWAYS AND OPIOID ABUSE/DEPENDENCE, TEXAS MEDICAID 2011 Component Variable Bivariate Path 1 Path 2 Causal Exposure: Mental Health Disorder Mental health disorders 3.53 (2.95, 4.22) 3.37 (3.41,4 .27) 2.54 (2.11,3.06) Mediator 1: Painful Conditions Arthritis 4.44 (3.91,5.03) 0.67 (0.49,0.94) -Back pain 3.17 (2.91,3.46) 1.36 (1.06,1.75) -Headaches 2.39 (2.20,2.61) 1.27 (1.08,1.49) -Mediator 2: Other substance abuse Non opioid substance abuse 5.42 (4.63,6.33) -4.28 (3.64,5.03)
165 APPENDIX O OPERATIONAL STATUS OF PRESCRIPTION DRUG MONITORING PROGRAMS AS OF APRIL 2014
166 APPENDIX P THE CASCADING EFFECTS OF PRESCRIPTION OPIOID ABUSE IN THE UNITED STATES Source: U.S. Centers for Diseas e Control and Prevention.
167 APPENDIX Q EMERGENCY DEPARTMENT VISIT RATE FOR NONMEDICAL USE OF PHARMACEUTICALS , MIAMI FORT LAUDERDALE, 2008 2011 0 20 40 60 80 100 120 140 2008 2009 2010 2011 Rate per 100,000 population Opiates/opioids Oxycodone/combinations oxycodone
168 APPENDIX R SCREEN SHOT OF THE INTERACTIVE DATA QUERY SYSTEM CREATED RUG RELATED DEATH SURVEILLANCE SYSTEM
169 APPENDIX S THE NUMBER OF FLORIDA PATIENTS IDENTIFIED AS DOCTOR SHOPPERS BY QUARTER, 2012 2013 Source: E FORSCE Annual Report 2013 .
170 APPENDIX T RESPONSES FROM FLORIDA PRESCRIPTION DRUG MONITORING PROGRAM S URVEY Dispenser Chain Other Prescriber In your experience, how useful has the PDMP been so far in helping to control "doctor shopping" by patients seeking to access or abuse controlled substances? N % N % N % Useful 347 95 214 91 265 91 Not Useful 6 2 8 3 5 2 As a result of using the PDMP system, do you communicate more with patients? 1 Yes, Definitely 183 50 125 53 156 53 2 Yes, Somewhat 152 42 65 28 94 32 3 No 27 7 44 19 37 13
171 REFERENCES 1. Centers for Disease Control and Prevention (CDC). Fast Stats. Therapeutic Drug Use. 2013. Available at: http://www.cdc.gov/nchs/fastats/drugs.htm. Accessed February 2, 2014. 2. Gu Q, Dillon CF, Burt VL. Prescription drug use contin ues to increase: U.S. prescription drug data for 2007 2008. NCHS Data Brief 2010: 1 8. 3. Paulozzi LJ. Prescription drug overdoses: A review. J Safety Res 2012; 43 : 283 289. doi:10.1016/j.jsr.2012.08.009. 4. Paulozzi LJ, Kilbourne EM, Desai HA. Prescript ion Drug Monitoring Programs and Death Rates from Drug Overdose. Pain Med 2011; 12 : 747 754. doi:10.1111/j.1526 4637.2011.01062.x. 5. Centers for Disease Control and Prevention. CDC Grand Rounds: Prescription Drug Overdoses a U.S. Epidemic. Morb Mortal Wkly MMWR 2012; 61 : 10 13. 6. National Center for Health Statistics (NCHS), National Vital Statistics System. 10 Leading Causes of Injury Deaths by Age Group Highlighting Unintentional Injury Deaths, United States 2010. Available at: http://www.cdc.gov/ injury/wisqars/pdf/10LCID_Unintentional_Deaths_2010 a.pdf. Accessed April 20, 2013. 7. Warner M, Chen LH, Makuc DM, Anderson RN, MiniÃ±o AM. Drug poisoning deaths in the United States, 1980 2008. NCHS Data Brief 2011: 1 8. 8. Goldberger B, Thogmartin J, J ohnson H, Paulozzi L, Rudd R, Ibrahimova A. Drug Overdose Deaths Florida, 2003 2009. Morb Mortal Wkly MMWR 2011; 60 : 869 872. 9. US Centers for Disease Control and Prevention. Drug Overdose in the United States: Fact Sheet. Home Recreat Saftey 2014. Ava ilable at: http://www.cdc.gov/homeandrecreationalsafety/overdose/facts.html. Accessed February 14, 2014. 10. White AM, Hingson RW, Pan I J, Yi H Y. Hospitalizations for alcohol and drug overdoses in young adults ages 18 24 in the United States, 1999 2008: results from the Nationwide Inpatient Sample. J Stud Alcohol Drugs 2011; 72 : 774 786. 11. National Institute on Drug Abuse. Topics in Brief: Prescription Drug Abuse. Sci Drug Abuse Addict 2011. Available at: http://www.drugabuse.gov/publications/topics i n brief/prescription drug abuse. Accessed February 14, 2014.
172 12. United States Department of Health and Human Services. Substance Abuse and Mental Health Services Administration. Center for Behavioral Health Statistics and Quality. National Survey on Drug Use and Health, 2012. 2013. Available at: http://doi.org/10.3886/ICPSR34933.v1. 13. National Institute on Drug Abuse. How do opioids affect the brain and body? Prescr Drugs Abuse Addict 2011. Available at: http://www.drugabuse.gov/publications/research r eports/prescription drugs/opioids/how do opioids affect brain body. Accessed February 14, 2014. 14. Johannes CB, Le TK, Zhou X, Johnston JA, Dworkin RH. The prevalence of chronic pain in United States adults: results of an Internet based survey. J Pain Of f J Am Pain Soc 2010; 11 : 1230 1239. doi:10.1016/j.jpain.2010.07.002. 15. Mazer Amirshahi M, Mullins PM, Rasooly I, van den Anker J, Pines JM. Rising opioid prescribing in adult U.S. emergency department visits: 2001 2010. Acad Emerg Med Off J Soc Acad Em erg Med 2014; 21 : 236 243. doi:10.1111/acem.12328. 16. Mazer Amirshahi M, Mullins PM, Rasooly IR, van den Anker J, Pines JM. Trends in prescription opioid use in pediatric emergency department patients. Pediatr Emerg Care 2014; 30 : 230 235. doi:10.1097/PE C.0000000000000102. 17. Freburger JK, Holmes GM, Agans RP, et al. The Rising Prevalence of Chronic Low Back Pain. Arch Intern Med 2009; 169 : 251. doi:10.1001/archinternmed.2008.543. 18. The Center for Substance Abuse. Profile: Oxycodone. Available at: ht tp://www.cesar.umd.edu/cesar/pubs/oxy.pdf. Accessed February 15, 2014. 19. Manchikanti L, Fellows B, Ailinani H, Pampati V. Therapeutic use, abuse, and nonmedical use of opioids: a ten year perspective. Pain Physician 2010; 13 : 401 435. 20. Kuehn BM. Opi oid Prescriptions Soar. JAMA 2007; 297 : 249. doi:10.1001/jama.297.3.249. 21. U.S. Department of Health and Human Services. Healthy People 2020 summary of objectives: Substance abuse. 2012. Available at: http://www.healthypeople.gov/2020/topicsobjectives20 20/overview.aspx?to picid=40.
173 22. US Centers for Disease Prevention and Control. Policy Impact: Prescription Painkiller Overdoses. Inj Prev Control 2013. Available at: http://www.cdc.gov/HomeandRecreationalSafety/pdf/PolicyImpact PrescriptionPainkillerOD .pdf. 23. National Institute on Drug Abuse. Prescription Drugs: Abuse and Addiction . National Institute on Drug Abuse, 2011. Available at: http://www.drugabuse.gov/sites/default/files/rrprescription.pdf. Accessed April 8, 2013. 24. US Centers for Disease Control and Prevention. International Classification of Diseases, Ninth Revision, Clinical Modification (ICD 9 CM). Classif Dis Funct Disabil . Available at: http://www.cdc.gov/nchs/icd/icd9cm.htm. Accessed May 5, 2014. 25. Office of Diversion Control. Co ntrolled Substance Schedules. 2012. Available at: http://www.deadiversion.usdoj.gov/schedules/index.html#list. 26. Institute of Medicine (IOM). Relieving Pain in America: A Blueprint for Transforming Prevention, Care, Education, and Research. Washington, DC, 2011. 27. Delcher, C, Maldonado Program: Temporal and Spatial Trends in Provider Utilization. 2013. 28. White House. crisis ., 2011. Available at: http:// www.whitehousedrugpolicy.gov/publications/pdf/rx_abuse_plan.pdf. Accessed April 18, 2013. 29. Keller CE, Ashrafioun L, Neumann AM, Van Klein J, Fox CH, Blondell RD. Practices, perceptions, and concerns of primary care physicians about opioid d ependence associated with the treatment of chronic pain. Subst Abuse Off Publ Assoc Med Educ Res Subst Abuse 2012; 33 : 103 113. doi:10.1080/08897077.2011.630944. 30. Davis WR, Johnson BD. Prescription opioid use, misuse, and diversion among street drug us ers in New York City. Drug Alcohol Depend 2008; 92 : 267 276. doi:10.1016/j.drugalcdep.2007.08.008. 31. Boudreau D, Von Korff M, Rutter CM, et al. Trends in long term opioid therapy for chronic non cancer pain. Pharmacoepidemiol Drug Saf 2009; 18 : 1166 117 5. doi:10.1002/pds.1833. 32. Manchikanti L, Abdi S, Atluri S, et al. American Society of Interventional Pain Physicians (ASIPP) guidelines for responsible opioid prescribing in chronic non cancer pain: Part 2 -guidance. Pain Physician 2012; 15 : S67 116.
174 3 3. Turk DC, Wilson HD, Cahana A. Treatment of chronic non cancer pain. Lancet 2011; 377 : 2226 2235. doi:10.1016/S0140 6736(11)60402 9. 34. Sullivan MD, Edlund MJ, Steffick D, UnÃ¼tzer J. Regular use of prescribed opioids: association with common psychiatr ic disorders. Pain 2005; 119 : 95 103. doi:10.1016/j.pain.2005.09.020. 35. Pade PA, Cardon KE, Hoffman RM, Geppert CMA. Prescription opioid abuse, chronic pain, and primary care: a Co occurring Disorders Clinic in the chronic disease model. J Subst Abuse T reat 2012; 43 : 446 450. doi:10.1016/j.jsat.2012.08.010. 36. Christo PJ, Manchikanti L, Ruan X, et al. Urine drug testing in chronic pain. Pain Physician 2011; 14 : 123 143. 37. Longo LP, Parran T Jr, Johnson B, Kinsey W. Addiction: part II. Identification and management of the drug seeking patient. Am Fam Physician 2000; 61 : 2401 2408. 38. Barrett K, Watson A. Physician Perspectives on a Pilot Prescription Monitoring Program. J Pain Palliat Care Pharmacother 2005; 19 : 5 13. doi:10.1080/J354v19n03_03. 39. KASPER Evaluation Team, Blumenschein K, Fink J, et al. Independent Evaluation of the Impact and Effectiveness of the Kentucky All Schedule Prescription Electronic Reporting Program (KASPER) . Lexington, Kentucky: Institute for Pharmaceutical Outcomes and Policy, 2010. Available at: http://chfs.ky.gov/NR/rdonlyres/24493B2E B1A1 4399 89AD 1625953BAD43/0/KASPEREvaluationFinalReport10152010.pdf. 40. Baehren DF, Marco CA, Droz DE, Sinha S, Callan EM, Akpunonu P. A Statewide Prescription Monitoring Program Affects Emergency Department Prescribing Behaviors. Ann Emerg Med 2010; 56 : 19 23.e3. doi:10.1016/j.annemergmed.2009.12.011. 41. Green TC, Mann MR, Bowman SE, et al. How does use of a prescripti on monitoring program change medical practice? Pain Med Malden Mass 2012; 13 : 1314 1323. doi:10.1111/j.1526 4637.2012.01452.x. 42. FORCSE Prescription Drug Monitoring Program Monthly Registration & Utiliza tion. 2012. Available at: http://www.doh.state.fl.us/mqa/pdmp/Monthly_Registration_Utilization.html. 43. Compton WM, Volkow ND. Major increases in opioid analgesic abuse in the United States: Concerns and strategies. Drug Alcohol Depend 2006; 81 : 103 107. doi:10.1016/j.drugalcdep.2005.05.009.
175 44. Beaudoin FL, Straube S, Lopez J, Mello MJ, Baird J. Prescription opioid misuse among ED patients discharged with opioids. Am J Emerg Med 2014. doi:10.1016/j.ajem.2014.02.030. 45. Fishman SM. Risk of the View Thr ough the Keyhole: There Is Much More to Physician Reactions to the DEA Than the Number of Formal Actions. Pain Med 2006; 7 : 360 362. doi:10.1111/j.1526 4637.2006.00194.x. 46. Virginia . W V Med J 2010; 106 : 10 14. 47. United Nations Office on Drugs and Crime,. Opioid overdose: preventing and reducing opioid overdose mortality . Geneva, Switzerland: World Health Organization, 2013. Available at: http://www.unodc.org/docs/treatment/overdo se.pdf. Accessed May 9, 2014. 48. Centers for Disease Control and Prevention, National Center for Injury Prevention and Control, Division of Unintentional Injury Prevention. Law: Doctor Shopping. 2012. Available at: http://www.cdc.gov/homeandrecreationals afety/Poisoning/laws/dr_shoppin g.html#states. 49. Neutel CI, Skurtveit S, Berg C, Sakshaug S. Multiple prescribers in older frequent opioid users -does it mean abuse? J Popul Ther Clin Pharmacol J ThÃ©rapeutique Popul Pharamcologie Clin 2013; 20 : e397 405. 50. Office of National Drug Control Policy (ONDCP). 2013 National Drug Control Strategy. White House 2013. Available at: http://www.whitehouse.gov/ondcp/2013 national drug control strategy. Accessed February 16, 2014. 51. Kutz GD. Medicaid: Fraud and ab use related to controlled substances identified in selected states . DIANE Publishing, 2010. 52. United States Government Accountability Office. MEDICARE PART D Instances of Questionable Access to Prescription Drugs ., 2011. Available at: http://www.gao.gov /new.items/d11699.pdf. 53. Cepeda MS, Fife D, Chow W, Mastrogiovanni G, Henderson SC. Opioid Shopping Behavior: How Often, How Soon, Which Drugs, and What Payment Method. J Clin Pharmacol 2012. 54. The Henry J. Kaiser Family Foundation. Medicaid. Medicai d Primer Key Income People 2013. Available at: http://kff.org/medicaid/issue brief/medicaid a primer/. Accessed June 11, 2014.
176 55. Hall AJ, Logan JE, Toblin RL, et al. Patterns of abuse among unintentional pharmaceu tical overdose fatalities. JAMA J Am Med Assoc 2008; 300 : 2613 2620. doi:10.1001/jama.2008.802. 56. Peirce GL, Smith MJ, Abate MA, Halverson J. Doctor and pharmacy shopping for controlled substances. Med Care 2012; 50 : 494. 57. Dhalla IA, Mamdani MM, Siv ilotti MLA, Kopp A, Qureshi O, Juurlink DN. Prescribing of opioid analgesics and related mortality before and after the introduction of long acting oxycodone. Can Med Assoc J 2009; 181 : 891 896. doi:10.1503/cmaj.090784. 58. Martyres RF, Clode D, Burns JM. Seeking drugs or seeking help? Med J Aust 2004; 180 : 211 214. 59. Gomes T, Mamdani MM, Dhalla IA, Paterson JM, Juurlink DN. Opioid Dose and Drug Related Mortality in Patients With Nonmalignant Pain. Arch Intern Med 2011; 171 . doi:10.1001/archinternmed.2011.117. 60. Han H, Kass PH, Wilsey BL, Li C S. Increasing trends in Schedule II opioid use and doctor shopping during 1999 2007 in California. Pharmacoepidemiol Drug Saf 2013. doi:1 0.1002/pds.3496. 61. Coolen P, Best S, Lima A, Sabel J, Paulozzi L. Overdose deaths involving prescription opioids among Medicaid enrollees Washington, 2004 2007. MMWR Morb Mortal Wkly Rep 2009; 58 : 1171 5. 62. Wilsey B, Fishman S, Gilson A, et al. Profi ling multiple provider prescribing of opioids, benzodiazepines, stimulants, and anorectics. Drug Alcohol Depend 2010; 112 : 99 106. 63. Green TC, Bowman SE, Ray M, Zaller N, Heimer R, Case P. Collaboration or Coercion? Partnering to Divert Prescription Opi oid Medications. J Urban Health 2013; 90 : 758 767. doi:10.1007/s11524 012 9784 5. 64. Inside Diversion Control, US Department of Justice, Office of Diversion. 2011. Available at: http://www.deadiversion.usdoj.gov/prog_dscrpt/index.html. 65. Fischer B, Bi bby M, Bouchard M. The global diversion of pharmaceutical drugs non medical use and diversion of psychotropic prescription drugs in North America: a review of sourcing routes and control measures. Addict Abingdon Engl 2010; 105 : 2062 2070. doi:10.1111/j.13 60 0443.2010.03092.x. 66. Cicero TJ, Kurtz SP, Surratt HL, et al. Multiple Determinants of Specific Modes of Prescription Opioid Diversion. J Drug Issues 2011; 41 : 283 304.
177 67. Rice JB, White AG, Birnbaum HG, Schiller M, Brown DA, Roland CL. A Model to I dentify Patients at Risk for Prescription Opioid Abuse, Dependence, and Misuse. Pain Med 2012; 13 : 1162 1173. doi:10.1111/j.1526 4637.2012.01450.x. 68. Parente ST, Kim SS, Finch MD, et al. Identifying controlled substance patterns of utilization requiring evaluation using administrative claims data. Am J Manag Care 2004; 10 : 783 790. 69. Weiner SG, Griggs CA, Mitchell PM, et al. Clinician Impression Versus Prescription Drug Monitoring Program Criteria in the Assessment of Drug Seeking Behavior in the Emer gency Department. Ann Emerg Med 2013. doi:10.1016/j.annemergmed.2013.05.025. 70. Primm BJ, Perez L, Dennis GC, et al. Managing pain: The Challenge in Underserved Populations: Appropriate Use Versus Abuse and Diversion. J Natl Med Assoc 2004; 96 : 1152 1161 . 71. Spitz A, Moore AA, Papaleontiou M, Granieri E, Turner BJ, Reid MC. with chronic non cancer pain: a qualitative study. BMC Geriatr 2011; 11 : 35. doi:10.1186/1471 2318 11 35. 7 2. Prescription Drug Monitoring Programs Training and Technical Assistance Center. Status Prescr Drug Monit Programs Dec 2013 . Available at: http://www.pdmpassist.org/pdf/pmpprogramstatus2013_a.pdf. Accessed January 25, 2014. 73. Paone, Denise. Applicati on in Public Health Highlighting PMP. 2013. Available at: http://www.pdmpassist.org/pdf/PPTs/National2013/25 1 A_B_C%20Parker_Paone_HartmanOConnell.pdf. Accessed January 27, 2014. 74. Cepeda MS, Fife D, Yuan Y, Mastrogiovanni G. Distance traveled and freq uency of interstate opioid dispensing in opioid shoppers and nonshoppers. J Pain Off J Am Pain Soc 2013; 14 : 1158 1161. doi:10.1016/j.jpain.2013.04.014. 75. Cicero TJ, Ellis MS. Health outcomes in patients using no prescription online pharmacies to purcha se prescription drugs. J Med Internet Res 2012; 14 : e174. doi:10.2196/jmir.2236. 76. Rigg KK, March SJ, Inciardi JA. Prescription Drug Abuse & Diversion: Role of the Pain Clinic. J Drug Issues 2010; 40 : 681 702. 77. DEA Office of Diversion Control. Drug Trends March 2012 . Drug Enforcement Agency, 2012. Available at: http://www.deadiversion.usdoj.gov/mtgs/pharm_awareness/conf_2012/mar ch_2012/drug_trends_0329.pdf. Accessed April 4, 2013.
178 78. Gerhardt AM. Identifying the Drug Seeker: The Advanced Practice Role in Managing Prescription Drug Abuse. J Am Acad Nurse Pract 2004; 16 : 239 243. doi:10.1111/j.1745 7599.2004.tb00445.x. 79. Robert Twillman. Memo on Changes to Walgreens Policy on Filling Prescriptions for Controlled Substances. 2013. Availabl e at: http://fapmmed.net/WalgreenMemo.pdf. 80. pills.http://www.wthr.com/story/23469086/2013/09/18/walgreens secret checklist reveals controversial new policy on pain pills. Published September 18, 2013. Accessed February 5, 2014. 81. Deyo RA, Irvine JM, Millet LM, et al. Measures such as interstate cooperation would improve the efficacy of programs to track controlled drug prescriptions. Health Aff Proj Hope 2013; 32 : 603 613. doi:10. 1377/hlthaff.2012.0945. 82. American Medical Association. Drug diversion and prescription drug monitoring programs. 2012. Available at: http://www.ama assn.org/resources/doc/washington/prescription drug monitoring issue brief.pdf. Accessed August 26, 2013 . 83. Nordmann S, Pradel V, Lapeyre Mestre M, et al. Doctor shopping reveals geographical variations in opioid abuse. Pain Physician 2013; 16 : 89 100. 84. Peter Kreiner. Geospatial Analyses of Opioid Overdose Rates and Rates of Questionable Activity in M assachusetts. 2012. Available at: http://www.slideshare.net/OPUNITE/peter kreiner. Accessed March 31, 2013. 85. Rossen LM, Khan D, Warner M. Hot spots in mortality from drug poisoning in the United States, 2007 2009. Health Place 2013; 26C : 14 20. doi:10. 1016/j.healthplace.2013.11.005. 86. Controlled Substances Act (CSA) ., 1970. 87. Clark TE. Prescription Drug Monitoring Programs: An Assessment of the Evidence for Best Practices . Heller School for Social Policy and Management, Brandeis University, 2012. Available at: http://www.pdmpexcellence.org/sites/all/pdfs/Brandeis_PDMP_Report_final .pdf. 88. Katz N, Panas L, Kim M, et al. Usefulness of prescription monitoring programs for surveillance -analysis of Schedule II opioid prescription data in Massachusett s, 1996 2006. Pharmacoepidemiol Drug Saf 2010; 19 : 115 123. doi:10.1002/pds.1878.
179 89. El Allaki F, Bigras Poulin M, Michel P, Ravel A. A population health surveillance theory. Epidemiol Health 2012; 34 : e2012007. doi:10.4178/epih/e2012007. 90. Buehler JW public health surveillance in the 21st century. Introduction. Morb Mortal Wkly Rep Surveill Summ Wash DC 2002 2012; 61 Suppl : 1 2. 91. Crisis . Executive office of the President of the United States, 2011. Available at: http://www.whitehouse.gov/sites/default/files/ondcp/issues content/prescription drugs/rx_abuse_plan.pdf. Accessed April 8, 2013. 92. Centers for Medicare and Medicaid Ser vices. Drug Diversion in the Medicaid Program State Strategies for Reducing Prescription Drug Diversion in Medicaid. 2012. Available at: https://www.cms.gov/Medicare Medicaid Coordination/Fraud Prevention/MedicaidIntegrityProgram/downloads/drugdiversion.pd f. 93. Levy J, Segal L, Fuchs Miller A. Prescription Drug Abuse, Strategies to Stop the Epidemic Foundation, 2013. Available at: http://www.rwjf.org/content/dam/farm/reports/reports/2013/rwjf408045. 94. Jo nes CM. Pharmaceutical Overdose Deaths, United States, 2010. JAMA 2013; 309 : 657. doi:10.1001/jama.2013.272. 95. Bruehl S, Apkarian AV, Ballantyne JC, et al. Personalized Medicine and Opioid Analgesic Prescribing for Chronic Pain: Opportunities and Challe nges. J Pain 2013; 14 : 103 113. doi:10.1016/j.jpain.2012.10.016. 96. Centers for Disease Control and Prevention (CDC). Vital signs: overdoses of prescription opioid pain relievers --United States, 1999 -2008. MMWR Morb Mortal Wkly Rep 2011; 60 : 1487 1492. 97. Goldberger BT. Drug Overdose Deaths Florida, 2003 2009. Morb Mortal Wkly MMWR 2011; 60 : 869 870. 98. News from Drug Enforcement Agency (DEA), Domestic Field Divisions, Miami. Fla Law Enforc Prescr Drug Efforts Prod Posit Results Purch Oxycodone De cline 2012. Available at: http://www.justice.gov/dea/divisions/mia/2012/mia013012p.html. Accessed October 15, 2013. 99. Florida Department of Health. 2011 2012 Prescription Drug Monitoring Program Annual Report (E FORCSE). 2012. Available at: http://www. eforcse.com/docs/2012AnnualReport.pdf.
180 100. Drugs Identified in Deceased Persons . Tallahassee: Florida Department of 101. Pain Management Clinic Registration Information and Registration Relinquishmen t Form. Available at: http://www.floridahealth.gov/licensing and regulation/pain management clinics/_documents/relinquish clinic registration form.pdf. Accessed November 6, 2013. 102. FDA approves abuse deterrent labeling for reformulated OxyContin. 2013. Available at: http://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/ucm348 252.htm. Accessed October 28, 2013. 103. Cicero TJ, Ellis MS, Surratt HL. Effect of abuse deterrent formulation of OxyContin. N Engl J Med 2012; 367 : 187 189. doi:10.1056/NEJMc 1204141. 104. Butler SF, Cassidy TA, Chilcoat H, et al. Abuse Rates and Routes of Administration of Reformulated Extended Release Oxycodone: Initial Findings From a Sentinel Surveillance Sample of Individuals Assessed for Substance Abuse Treatment. J Pain Off J Am Pain Soc 2012. doi:10.1016/j.jpain.2012.08.008. 105. Severtson SG, Bartelson BB, Davis JM, et al. Reduced Abuse, Therapeutic Errors, and Diversion Following Reformulation of Extended Release Oxycodone in 2010. J Pain Off J Am Pain Soc 2013. doi: 10.1016/j.jpain.2013.04.011. 106. The New York Times . http://www.nytimes.com/2011/09/01/us/01drugs.html. Published August 31, 2011. Accessed February 11, 2013. 107. George M. Investigation: Florida has more Burger Kings. ABC Action News . http://www.abcactionnews.com/dpp/news/region_tampa/investigation%3A florida has more pain clinics than mcdonald%27s#ixzz2NQNHYlxJ. Published March 1, 2011. 108. ackdown Sets Off a Rush Into Georgia. Wall Street Journal . http://online.wsj.com/article/SB10001424127887324478304578173341194 754984.html#articleTabs%3Darticle. Published December 25, 2012. Accessed February 11, 2013. 109. Sauber Schatz EK, Mack KA, Diekm an ST, Paulozzi LJ. Associations between pain clinic density and distributions of opioid pain relievers, drug related deaths, hospitalizations, emergency department visits, and neonatal abstinence syndrome in Florida. Drug Alcohol Depend 2013. doi:10.1016/ j.drugalcdep.2013.05.017.
181 110. Prescription Drugs. HB 7095, Regular Session 2011 (Fla 2011). 111. Florida Department of Health, Division of Medical Quality Assurance. 2011 2012 ANNUAL REPORT and LONG RANGE PLAN . Tallahassee, FL, 2012. Available at: http: //www.floridahealth.gov/licensing and regulation/reports and publications/_documents/annual report 12a.pdf. Accessed November 4, 2013. 112. Alliance of States with Prescription Drug Monitoring Programs. PMP Legislation & Operational Dates. 2013. Available at: http://www.pmpalliance.org/content/pmp legislation operational dates?order=field_pmp_operational_value&sort=desc. Accessed August 26, 2013. 113. 114. Brandeis University PMP Center of Excellence. Briefing on PMP Effectiveness. 2011. Available at: http://www.pmpexcellence.org/sites/all/pdfs/pmp_effectiveness_brief_final. pdf. 115. Florida Department of Health. 2012 2013 Prescription Drug Monitoring Program Annual Report (E FORCSE). 201 3. 116. Gugelmann HM, Perrone J. Can prescription drug monitoring programs help limit opioid abuse? JAMA J Am Med Assoc 2011; 306 : 2258 2259. doi:10.1001/jama.2011.1712. 117. Green TC, Zaller N, Rich J, Bowman S, Friedmann P. Revisiting Paulozzi Pain Med 2011; 12 : 982 985. doi:10.1111/j.1526 4637.2011.01136.x. 118. Drugs Identified in Deceased Persons by Florida Medical Examiners . Florida Department of L, 2013. Available at: http://www.fdle.state.fl.us/Content/getdoc/9c4df7c2 4b07 4f7a acfb 6461719c0e44/2012 interim drug reportFINAL.aspx. Accessed April 8, 2013. 119. Analysis of 2005 2008. 2008. 120. T he Florida Statutes ., 2013. Available at: http://www.leg.state.fl.us/Statutes/index.cfm?App_mode=Display_Statute& Search_String=&URL=0800 0899/0893/Sections/0893.03.html#1. Accessed October 15, 2013. 121. Cook TH, Campbell DT. Quasi Experimentation . Boston , MA: Houghton Mifflin, 1979.
182 122. Population and Demographic Data. Fla Off Econ Demogr Res 2013. Available at: http://www.edr.state.fl.us/Content/population demographics/data/index.cfm. Accessed March 11, 2013. 123. Centers for Disease Control and Preve ntion, National Center for Health Statistics. Underlying Cause of Death. Available at: http://wonder.cdc.gov/ucd icd10.html. Accessed August 9, 2013. 124. Florida Department of Health, Division of Public Health Statistics & Performance Management. Fla CHA RTS Community Health Assess Resour Tool Set . Available at: http://www.floridacharts.com/charts/default.aspx. 125. Paone D. Opioid Analgesics in New York City. 2013. Available at: http://www.pdmpassist.org/pdf/PPTs/National2013/25 1 A_B_C%20Parker_Paone_Ha rtmanOConnell.pdf. Accessed October 25, 2013. 126. Delcher C, Maldonado Molina MM, Wagenaar AC. Effects of alcohol taxes on alcohol related disease mortality in New York State from 1969 to 2006. Addict Behav 2012; 37 : 783 789. doi:10.1016/j.addbeh.2012.02 .019. 127. Box G, Jenkins GM. Time Series Analysis: Forecasting & Control . 2nd ed. New Jersey: Prentice Hall, 1976. 128. McDowall D. Interrupted time series analysis . Newbury Park [u.a.: Sage, 1980. 129. Williamson GD, Weatherby Hudson G. A monitoring s ystem for detecting aberrations in public health surveillance reports. Stat Med 1999; 18 : 3283 3298. 130. SAS Institute. . Cary, NC: SAS Institute, Inc., 2008. 131. Cheng J, Edwards LJ, Maldonado Molina MM, Komro KA, Muller KE. R eal longitudinal data analysis for real people: Building a good enough mixed model. Stat Med 2010; 29 : 504 520. doi:10.1002/sim.3775. 132. Giles D. Interpreting dummy variables in semi logarithmic regression models: exact distributional results. 2011. 133 . Alliance of States with Prescription Drug Monitoring Programs. Status of ., 2012. Available at: http://www.pmpalliance.org/pdf/pmpstatustable2012.pdf.
183 134. Centers for Disease Control and Prevention (CDC). Vital Signs: Overdoses of Prescription Opioid Pain Relievers and Other Drugs Among Women -United States, 1999 2010. MMWR Morb Mortal Wkly Rep 2013; 62 : 537 542. 135. Worley J, Hall JM. Doctor shopping: a concept analysis. Res Theory Nurs Pract 2012; 26 : 262 278. 136. Betses M, Brennan T. Abus ive Prescribing of Controlled Substances A Pharmacy View. N Engl J Med 2013; 369 : 989 991. doi:10.1056/NEJMp1308222. 137. Henry J. Kaiser Family Foundation. State Health Facts. Total Medicaid Enroll FY2010 . Available at: http://kff.org/medicaid/state in dicator/total medicaid enrollment/. Accessed January 2, 2014. 138. Inciardi JA, Surratt HL, Kurtz SP, Cicero TJ. Mechanisms of prescription drug diversion among drug involved club and street based populations. Pain Med 2007; 8 : 171 183. 139. Texas Medica id and Healthcare Partnership. Texas Medicaid Provider Procedure Manual (2012 online edition). Place Serv Coding . Available at: http://www.tmhp.com/HTMLmanuals/TMPPM/2012/Vol1_06_Claims_Filing. 08.034.html. Accessed March 14, 2014. 140. American Society of Health System Pharmacists. Drug Information. AHFS Drug Inf 2014. Available at: http://www.ahfsdruginformation.com/index.aspx. Accessed March 14, 2014. 141. US Food and Drug Administration. Drugs. Natl Drug Code Dir 2014. Available at: http://www.fda.gov/ Drugs/InformationOnDrugs/ucm142438.htm. Accessed March 14, 2014. 142. United States Department of Agriculture, Economic Research Service. Rural Classifications. Available at: http://www.ers.usda.gov/briefing/Rurality/RuralUrbCon. 143. Neff JM, Valentine J, Park A, et al. Trends in pediatric hospitalizations of children in washington state by insurance and chronic condition status, 1991 1998. Arch Pediatr Adolesc Med 2002; 156 : 703 709. 144. Texas Health and Human Services Commission. Texas Medicaid and C HIP in Perspective Ninth Edition ., 2013. Available at: http://www.hhsc.state.tx.us/medicaid/reports/pb9/pinkbook.pdf. Accessed January 19, 2014.
184 145. Centers for Medicare and Medicaid Services. NPI Files. NPI Files . Available at: http://nppes.viva it.com/ NPI_Files.html. Accessed March 16, 2014. 146. GEODIST Function, SAS(R) 9.2 Language Reference: Dictionary, Fourth Edition. Available at: http://support.sas.com/documentation/cdl/en/lrdict/64316/HTML/default/vie wer.htm#a003113162.htm. Accessed January 16, 2014. 147. Prescription Drug Monitoring Program Training and Technical Assistance Center. Technical Assistance Guide No.01 13, Calculating Daily Morphine Milligram Equivalents. Available at: http://www.pdmpassist.org/pdf/BJA_performance_measure_aid_MME_co n version.pdf. Accessed January 4, 2014. 148. Paulozzi LJ, Kilbourne EM, Shah NG, et al. A History of Being Prescribed Controlled Substances and Risk of Drug Overdose Death: Prescription History and Risk of Overdose. Pain Med 2012; 13 : 87 95. doi:10.1111/j .1526 4637.2011.01260.x. 149. Austin PC, Steyerberg EW. Interpreting the concordance statistic of a logistic regression model: relation to the variance and odds ratio of a continuous explanatory variable. BMC Med Res Methodol 2012; 12 : 82. doi:10.1186/147 1 2288 12 82. 150. Goodin A, Blumenschein K, Freeman P, Talbert J. Consumer/patient encounters with prescription drug monitoring programs: evidence from a Medicaid population. Pain Physician 2012; 15 : 75. 151. Brandeis University. State Profiles. Prescr Drug Monit Program Tech Train Assist Cent 2014. Available at: http://www.pdmpassist.org/content/state profiles. Accessed April 18, 2014. 152. McDonald DC, Carlson KE. Estimating the prevalence of opioid diversion PloS One 2013; 8 : e69241. doi:10.1371/journal.pone.0069241. 153. Fleming ML, Chandwani H, Barner JC, Weber SN, Okoro TT. Prescribers and Pharmacists Requests for Prescription Monitoring Program (PMP) Data: Does PMP Structure Matter? J Pain Palliat Care P harmacother 2013; 27 : 136 142. doi:10.3109/15360288.2013.788598. 154. Rigg KK, Kurtz SP, Surratt HL. Patterns of prescription medication diversion among drug dealers. Drugs Educ Prev Policy 2012; 19 : 145 155. doi:10.3109/09687637.2011.631197. 155. Volkow ND, Frieden TR, Hyde PS, Cha SS. Medication Assisted Therapies Tackling the Opioid Overdose Epidemic. N Engl J Med 2014. doi:10.1056/NEJMp1402780.
185 156. Atluri S, Sudarshan G, Manchikanti L. Assessment of the trends in medical use and misuse of opioid a nalgesics from 2004 to 2011. Pain Physician 2014; 17 : E119 128. 157. Fischer B, Keates A, BÃ¼hringer G, Reimer J, Rehm J. Non medical use of prescription opioids and prescription opioid related harms: why so markedly higher in North America compared to the rest of the world? Addict Abingdon Engl 2014; 109 : 177 181. doi:10.1111/add.12224. 158. Degenhardt L, Charlson F, Mathers B, et al. The global epidemiology and burden of opioid dependence: Results from the Global Burden of Disease 2010 study. Addiction 2 014: n/a n/a. doi:10.1111/add.12551. 159. Substance Abuse and Mental Health Services Administration. Drug Abuse Warning Network (DAWN). Drug Abuse Warn Netw 2011 Natl Estim Drug Relat Emerg Dep Visits . Available at: http://www.samhsa.gov/data/2k13/DAWN2k1 1ED/DAWN2k11ED.htm#high 4. Accessed April 15, 2013. 160. Boscarino JA, Rukstalis MR, Hoffman SN, et al. Prevalence of prescription opioid use disorder among chronic pain patients: comparison of the DSM 5 vs. DSM 4 diagnostic criteria. J Addict Dis 2011; 30 : 185 194. doi:10.1080/10550887.2011.581961. 161. Becker WC, Fiellin DA, Merrill JO, et al. Opioid use disorder in the United States: Insurance status and treatment access. Drug Alcohol Depend 2008; 94 : 207 213. doi:10.1016/j.drugalcdep.2007.11.018. 162. Edlund MJ, Martin BC, Fan M Y, Devries A, Braden JB, Sullivan MD. Risks for opioid abuse and dependence among recipients of chronic opioid therapy: results from the TROUP study. Drug Alcohol Depend 2010; 112 : 90 98. doi:10.1016/j.drugalcdep.2010.05.017. 1 63. Ghate SR, Haroutiunian S, Winslow R, McAdam Marx C. Cost and comorbidities associated with opioid abuse in managed care and Medicaid patients in the United States: a comparison of two recently published studies. J Pain Palliat Care Pharmacother 2010; 24 : 251 258. 164. Birnbaum HG, White AG, Schiller M, Waldman T, Cleveland JM, Roland CL. Societal costs of prescription opioid abuse, dependence, and misuse in the United States. Pain Med 2011; 12 : 657 67. doi:10.1111/j.1526 4637.2011.01075.x. 165. Cummi ngs JR, Wen H, Ko M, Druss BG. Race/ethnicity and geographic access to Medicaid substance use disorder treatment facilities in the United States. JAMA Psychiatry 2014; 71 : 190 196. doi:10.1001/jamapsychiatry.2013.3575.
186 166. Adelmann PK. Mental and Substan ce Use Disorders Among Medicaid Recipients: Prevalence Estimates from Two National Surveys. Adm Policy Ment Health 2003; 31 : 111 129. doi:10.1023/B:APIH.0000003017.78877.56. 167. Substance Abuse and Mental Health Services Administration. Results from the 2011 National Survey on Drug Use and Health ., 2012. 168. Olsen Y, Daumit GL, Ford DE. Opioid prescriptions by U.S. primary care physicians from 1992 to 2001. J Pain Off J Am Pain Soc 2006; 7 : 225 235. doi:10.1016/j.jpain.2005.11.006. 169. Peck SB, Gilchr ist J, Clemans Taylor L. Clinical Inquiry: Is there a primary care tool to detect aberrant drug related behaviors in patients on opioids? J Fam Pract 2014; 63 : 162 164. 170. Simoni Wastila L, Ritter G, Strickler G. Gender and other factors associated with the nonmedical use of abusable prescription drugs. Subst Use Misuse 2004; 39 : 1 23. 171. Sullivan MD, Edlund MJ, Fan MY, DeVries A, Braden JB, Martin BC. Risks for possible and probable opioid misuse among recipients of chronic opioid therapy in commerci al and medicaid insurance plans: The TROUP Study. PAIN 2010; 150 : 332. 172. Cheatle MD. Psychological Dependence and Prescription Opioid Misuse and Abuse: Editorial. Pain Med 2014; 15 : 541 543. doi:10.1111/pme.12419. 173. Pasquale MK, Joshi AV, Dufour R, et al. Cost drivers of prescription opioid abuse in commercial and Medicare populations. Pain Pract Off J World Inst Pain 2014; 14 : E116 125. doi:10.1111/papr.12147. 174. Jones CM, Paulozzi LJ, Mack KA. Sources of Prescription Opioid Pain Relievers by Fr equency of Past Year Nonmedical Use: United States, 2008 2011. JAMA Intern Med 2014. doi:10.1001/jamainternmed.2013.12809. 175. Mailloux AT, Cummings SW, Mugdh M. A Decision Support Tool for Identifying Abuse of Controlled Substances by ForwardHealth Medi caid Members. J Hosp Mark Public Relat 2010; 20 : 34 55. doi:10.1080/15390940903450982. 176. Sehgal NM. Prescription opioid abuse in chronic pain: a review of opioid abuse predictors and strategies to curb opioid abuse. Pain Physician 2012; 15 : ES67 ES92.
187 177. Butler SF, Budman SH, Fernandez KC, Fanciullo GJ, Jamison RN. Cross Validation of a Screener to Predict Opioid Misuse in Chronic Pain Patients (SOAPP R): J Addict Med 2009; 3 : 66 73. doi:10.1097/ADM.0b013e31818e41da. 178. Cochran BN, Flentje A, Hec k NC, et al. Factors predicting development of opioid use disorders among individuals who receive an initial opioid prescription: Mathematical modeling using a database of commercially insured individuals. Drug Alcohol Depend 2014; 138 : 202 208. doi:10.101 6/j.drugalcdep.2014.02.701. 179. Cepeda MS, Fife D, Kihm MA, Mastrogiovanni G, Yuan Y. Comparison of the Risks of Shopping Behavior and Opioid Abuse Between Tapentadol and Oxycodone and Association of Shopping Behavior and Opioid Abuse. Clin J Pain 2013. doi:10.1097/AJP.0000000000000067. 180. Edlund MJ, Steffick D, Hudson T, Harris KM, Sullivan M. Risk factors for clinically recognized opioid abuse and dependence among veterans using opioids for chronic non cancer pain. Pain 2007; 129 : 355 362. doi:10.101 6/j.pain.2007.02.014. 181. Dufour R, Joshi AV, Pasquale MK, et al. The Prevalence of Diagnosed Opioid Abuse in Commercial and Medicare Managed Care Populations. Pain Pract Off J World Inst Pain 2013. doi:10.1111/papr.12148. 182. Texas Medicaid/CHIP Vendo r Drug Program. Available at: http://www.txvendordrug.com/providers/out of state.shtml. Accessed February 3, 2014. 183. Tamayo Sarver JH, Hinze SW, Cydulka RK, Baker DW. Racial and ethnic disparities in emergency department analgesic prescription. Am J Pu blic Health 2003; 93 : 2067 2073. 184. Heins JK, Heins A, Grammas M, Costello M, Huang K, Mishra S. Disparities in analgesia and opioid prescribing practices for patients with musculoskeletal pain in the emergency department. J Emerg Nurs JEN Off Publ Emer g Dep Nurses Assoc 2006; 32 : 219 224. doi:10.1016/j.jen.2006.01.010. 185. Becker WC, Starrels JL, Heo M, Li X, Weiner MG, Turner BJ. Racial differences in primary care opioid risk reduction strategies. Ann Fam Med 2011; 9 : 219 225. doi:10.1370/afm.1242. 1 86. Baldacchino A, Gilchrist G, Fleming R, Bannister J. Guilty until proven innocent: a qualitative study of the management of chronic non cancer pain among patients with a history of substance abuse. Addict Behav 2010; 35 : 270 272. doi:10.1016/j.addbeh.2 009.10.008.
188 187. Brandeis University PMP Center of Excellence. Using PDMPs to Improve ., 2013. Available at: http://www.pdmpexcellence.org/sites/all/pdfs/washi ngton_nff_final.pdf. 188. National Association of Boards of Pharmacy (NABP). NarxCheck. NarxCheck 2012. Available at: http://www.narxcheck.com/. Accessed April 30, 2014. 189. Forrester MB. Temporal and geographic patterns in opioid abuse in Texas. J Addi ct Dis 2012; 31 : 93 99. doi:10.1080/10550887.2012.665694. 190. Wagenaar AC, Maldonado Molina MM, Ma L, Tobler A, Komro KA. Effects of legal BAC limits on fatal crash involvement: analyses of 28 states from 1976 through 2002. J Safety Res 2007; 38 : 493 499 . 191. Robert Wood Johnson Foundation. LawAtlas, the Policy Surveillance Tool. Public Health Law Res . Available at: lawatlas.org. Accessed May 5, 2014. 192. Frankel GEC, Intrater H, Doupe M, Namaka M. Opioid misuse in Canada and critical appraisal of abe rrant behavior screening tools. World J Anesth 2014; 3 : 61 70. 193. Turk DC, Swanson KS, Gatchel RJ. Predicting opioid misuse by chronic pain patients: a systematic review and literature synthesis. Clin J Pain 2008; 24 : 497 508. doi:10.1097/AJP.0b013e3181 6b1070. 194. Cavallo P, Pagano S, Boccia G, De Caro F, De Santis M, Capunzo M. Network analysis of drug prescriptions. Pharmacoepidemiol Drug Saf 2013; 22 : 130 137. doi:10.1002/pds.3384. 195. Jonas AB, Young AM, Oser CB, Leukefeld CG, Havens JR. OxyConti nÂ® as currency: OxyContinÂ® use and increased social capital among rural Appalachian drug users. Soc Sci Med 1982 2012; 74 : 1602 1609. doi:10.1016/j.socscimed.2011.12.053. 196. Fraser MW, Hawkins JD. The social networks of opioid abusers. Int J Addict 1984 ; 19 : 903 917. 197. Schneeberger A, Mercer CH, Gregson SAJ, et al. Scale free networks and sexually transmitted diseases: a description of observed patterns of sexual contacts in Britain and Zimbabwe. Sex Transm Dis 2004; 31 : 380 387. 198. Proceedings of the American Medical Association Organized Medical Staff Section. 2013. Available at: http://www.ama assn.org/resources/doc/omss/a13 proceedings.pdf. Accessed February 14, 2014.
BIOGRAPHICAL SKETCH Chris Delcher received his Doctor of Philosophy from th e Department of Epidemiology in August 2014. He was a member of the Institute for Child Health Policy and a fellow in the Department of Health Outcomes and Policy at the University of Florida. H e received his Master of Science degree from the University o f North Carolina at Chapel Hill in Environmental Science and Engineering . H e was previously employed as an epidemiologist for the Virginia Department of Health in Richmond (where he published the first of his 12 publications) and a s a health analyst for Vi rginia Health Information. He is an international consultant on building surveillance systems for the U . S . Centers for Disease Control and Prevention and often travels to Haiti to help build local epidemiologic capacity. His interest in national and inter national public health developed from his experience as a Peace Corps volunteer in El Salvador.