Original Research Article A Model to Identify Patients at Risk for Prescription Opioid Abuse, Dependence, and Misusepme_

Transcription

1 bs_bs_banner Pain Medicine 2012; 13: Wiley Periodicals, Inc. OPIOIDS, SUBSTANCE ABUSE & ADDICTIONS SECTION Original Research Article A Model to Identify Patients at Risk for Prescription Opioid Abuse, Dependence, and Misusepme_ J. Bradford Rice, PhD,* Alan G. White, PhD,* Howard G. Birnbaum, PhD,* Matt Schiller, BA,* David A. Brown, PhD, MPH, and Carl L. Roland, PharmD *Analysis Group, Inc., Boston, Massachusetts; Parexel International, Durham, North Carolina; Pfizer, Inc., Cary, North Carolina, USA Reprint requests to: Brad Rice, PhD, Analysis Group, Inc., 111 Huntington Ave., 10th Floor, Boston, MA 02199, USA. Tel: ; Fax: ; brice@analysisgroup.com. This study was sponsored by King Pharmaceuticals, Inc., which was acquired by Pfizer Inc. in March David Brown was an employee of Pfizer Inc. at the time the study was conducted and the article was drafted. Abstract Objective. The objective of this study was to use administrative claims data to identify and analyze patient characteristics and behavior associated with diagnosed opioid abuse. Design. Patients, aged years, with at least one prescription opioid claim during (n = 821,916) were selected from a de-identified administrative claims database of privately insured members (n = 8,316,665). Patients were divided into two mutually exclusive groups: those diagnosed with opioid abuse during (n = 6,380) and those without a diagnosis for opioid abuse (n = 815,536). A logistic regression model was developed to estimate the association between an opioid abuse diagnosis and patient characteristics, including patient demographics, prescription drug use and filling behavior, comorbidities, medical resource use, and family member characteristics. Sensitivity analyses were conducted on the model s predictive power. Results. In addition to demographic factors associated with abuse (e.g., male gender), the following were identified as key characteristics (i.e., odds ratio [OR] > 2): prior opioid prescriptions (OR = 2.23 for 1 5 prior Rxs; OR = 6.85 for 6+ prior Rxs); at least one prior prescription of buprenorphine (OR = 51.75) or methadone (OR = 2.97); at least one diagnosis of non-opioid drug abuse (OR = 9.89), mental illness (OR = 2.45), or hepatitis (OR = 2.36); and having a family member diagnosed with opioid abuse (OR = 3.01). Conclusions. Using medical as well as drug claims data, it is feasible to develop models that could assist payers in identifying patients who exhibit characteristics associated with increased risk for opioid abuse. These models incorporate medical information beyond that available to prescription drug monitoring programs that are reliant on drug claims data and can be an important tool to identify potentially inappropriate opioid use. Key Words. Abuse; Opioids Introduction Pain is a prevalent and costly condition. The National Health and Nutrition Examination Survey found that more than a quarter of Americans age 20 and older experienced pain lasting at least 24 hours during the last month [1]. To treat pain, physicians often turn to prescription opioids, which are among the most effective drugs for pain management, but also come with an increased risk of abuse, dependence, and misuse ( opioid abuse ) [2]. Retail sales of opioids more than doubled between 1997 and 2007, and Americans now consume 80% of the world s 1162

2 A Model Predicting Prescription Opioid Abuse prescription opioid supply despite comprising only 5% of its population [3]. During this same period, there was a commensurate increase in rates of opioid abuse. According to the National Survey on Drug Use and Health (NSDUH), the use of prescription opioids is the second most common type of illicit drug use (after marijuana), with 12.4 million Americans using prescription pain relievers for nonmedical purposes in 2009, compared with 11.0 million in 2002 [4]. Similarly, the number of emergency department (ED) visits related to prescription opioid abuse has more than doubled in recent years from 173,000 in 2004 to 416,000 in 2009 [5]. Opioid overdose is the cause of over 3,000 deaths per month and accounts for 39% of all drug-related ED visits [4 6]. Opioid abuse is associated with substantial costs, both to health care payers and to society. According to White et al., diagnosed opioid abuse patients incurred $20,546 more in annual health care costs than demographically similar controls, and Medicaid opioid abuse patients cost $15,183 in excess of controls [7]. At the societal level, Birnbaum et al. estimated opioid abuse costs $55.7 billion annually in the United States, with $25.0 billion in health care costs plus an additional $25.6 billion in workplace costs, and $5.1 billion in criminal justice costs [8]. As increasing health care costs become an ever greater concern, better identification and treatment of opioid abusers may represent an opportunity to generate cost savings and improve patient outcomes. More specifically, administrative claims data available to managed care organizations (MCOs) and state prescription drug monitoring programs (PDMPs) offer a wealth of information concerning pharmacy and medical resource use as well as demographic information that could be combined with the latest clinical research to design algorithms that more precisely identify individuals at risk for opioid abuse. Prior research has shown such models are feasible [9]. No study to date, however, has attempted to use the clinical literature on risk factors for opioid abuse to develop claims data-based models to identify and analyze patient characteristics and behavior associated with opioid abuse among a nationwide, privately insured population of opioid users [10 16]. This study s objective was to expand and improve on previous research by assessing the differences in patient characteristics among prescription opioid users with and without an abuse diagnosis from all 50 states during a recent time period ( ) in order to highlight and quantify the relative impact of various characteristics and risk factors associated with opioid abuse. Specifically, the goal of this analysis was to build a model that would be directly applicable to and implemented by payers using information readily available through claims data. The characteristics assessed include patient demographics, prescription drug use and patterns of use (e.g., filling at multiple pharmacies, early refills), comorbidity profile, and medical resource utilization (e.g., ED visits). In addition, this analysis incorporates data on family member characteristics that were unavailable in previous studies. In addition to providing further insight into the factors associated with opioid abuse in a privately insured population, a secondary objective of this model was designed with applications for future observational studies in mind, particularly those where risk of opioid abuse may be an important confounder. For example, as opioid formulations designed with abuse-deterrent properties launch over the next several years, models such as the one designed here could be used to assess and control for differences in the likelihood of abuse between patient populations. This issue is of special importance for robust retrospective study design, which may not otherwise be able to account for disproportionate or preferential prescribing of opioids with abuse-deterrent properties by physicians to patients they perceive to be at higher risk for abuse. Methods Data This study used de-identified administrative claims data from a privately insured population (Ingenix Employer Solutions, Eden Prairie, MN, USA). The database covers over 12 million lives during and contains information from 55 self-insured U.S. companies operating nationwide in a broad array of industries. The data include information on patient demographics, medical diagnoses, and prescription drug and insurance enrollment history for all beneficiaries (i.e., employees, spouses, and dependents). Study Design From the privately insured patient population, all patients with at least one opioid prescription filled during were identified (n = 1,552,489). All patients with an opioid prescription were divided into two mutually exclusive groups: diagnosed abusers (n = 9,755) and those with no such diagnosis, or non-abusers (n = 1,542,734). Opioid abuse was identified using the following International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes, diagnosed at any time in the patient s history: 304.0x (opioid-type dependence), 304.7x (combinations of opioid-type dependence with any other drug dependence), 305.5x (nondependent opioid abuse), and (poisoning by opiates and related narcotics excluding [poisoning by heroin]). The rationale for identifying abuse using the full time period in the data is based on the notion that abuse and addiction are chronic conditions, which are long-developing and present for a prolonged period of time [14]. In addition, while the approach of identifying opioid abusers in claims data is consistent with prior research [15,16], note that reliance on ICD-9-CM coding conventions may fail to identify patients who misuse or abuse opioids but do not receive a clinical diagnosis. 1163

3 Rice et al. Step 0 Total patients, (n = 8,316,665) Step 1 Patients with at least one opioid claim, (n = 1,552,489) Opioid non-abusers Opioid abusers Step 2 Patients not diagnosed with opioid abuse/dependence, [1] (n = 1,542,734) Patients diagnosed with opioid abuse/dependence, [1] (n = 9,755) Step 3 Patients ages years throughout the 12 months prior to the index opioid Rx [2] (n = 1,110,410) Patients ages years throughout the 12 months prior to the index opioid Rx [2] (n = 8,455) Step 4 Patients continuously eligible on the index date and throughout the 12 months prior (n = 815,536) Patients continuously eligible on the index date and throughout the 12 months prior (n = 6,380) Notes: [1] Opioid abuse was identified using the following ICD-9-CM codes diagnosed at any time: 304.0, 304.7, 305.5, , , and [2] The index date was defined as the date of the most recent opioid Rx during (based on fill date). Figure 1 Sample selection flowchart The index date was defined as the fill date of the most recent opioid prescription during this period. Finally, further selection criteria were imposed, requiring patients to be ages and continuously eligible throughout the 12 months prior to the index date in order to ensure that the data contain all relevant drug and medical claims for the final sample of patients. The above selection criteria resulted in an analytic sample containing 6,380 abusers and 815,536 non-abusers (see Figure 1). Once the study population was selected, independent variables were constructed to measure a variety of patient characteristics and potential risk factors. These variables were derived from patient eligibility, pharmacy, and medical claims and were based on prior research by White et al. and others [9 13,15 20]. All independent variables were measured over the 12-month period prior to the index date. This period prior to an opioid prescription was of special interest because one of the study s objectives was to create a claims-based tool to control for risk of abuse in future retrospective studies. Such research typically measures patient characteristics between study groups during a 6- or 12-month baseline period, so this study was designed to mimic that approach. Continuous variables with right-skewed distributions (e.g., days of medical resource use, number of opioid prescriptions) were categorized into three ordinal buckets of patients according to each variable s distribution as follows: 1) patients for whom the value of the variable is zero (e.g., 0 days hospitalized); 2) patients in the top quintile; and 3) the remaining patients. 1 Statistical Analysis The probability that patient i was an opioid abuser is modeled as a logistic regression, ( ) Pr ( abuse i = 1) = F α+ β j X ij j 1164

4 A Model Predicting Prescription Opioid Abuse in which F(z) = e z /(1 + e z ) is the cumulative logistic distribution, X ij represents a matrix of j patient-specific independent variables, and the symbols a and b j represent parameters to be estimated within the model. The performance of various model specifications was assessed to identify potentially relevant variables (including interaction terms) and select the model with the best predictive power as measured by the c-statistic (a metric reflecting the area under the curve ranging from 0.5 [no predictive power] to 1.0 [perfect predictive model]). In cases where multiple models resulted in the same c-statistic, the model with a smaller set of relevant independent variables was chosen according to those variables that were 1) clinically relevant and 2) statistically significant (P < 0.01). All statistical analysis was completed using SAS 9.2 (SAS Institute, Cary, NC, USA). Results Descriptive Statistics The sample of 821,916 abusers and non-abusers were compared based on demographic information, prescription drug use, and comorbidity profile, as well as household and family characteristics. Table 1 reports differences in demographics between the two cohorts. Abusers were statistically significantly different (P < 0.01) from nonabusers on all measures, with abusers more likely to be male (52.5% vs 46.2%), and having different age group and geographic distributions. There were also significant differences in the amount and types of prescription drug use between the two cohorts (Table 2), with abusers filling approximately seven times as many prescription opioids in the year prior to index compared with non-abusers (13.3 opioid Rxs for abusers vs 1.9 for non-abusers). In addition, abusers were prescribed four times as many different types of opioids (3.7) as non-abusers (0.9) based on the National Drug Code (NDC), and nearly three times as many based on active ingredient, with abusers having prescriptions for 1.9 active ingredients, on average, vs 0.7 for non-abusers. Abusers also had opioid prescriptions filled at over three times as many pharmacies (2.4 pharmacies vs 0.7), received prescriptions from four times as many prescribers (3.2 prescribers vs 0.8), and were over nine times as likely to receive at least one early refill (38.4% vs 4.1%). Abusers were more likely than non-abusers to have received previous opioid prescriptions regardless of type Table 1 Demographics at index date* Prescription Opioid Abusers Prescription Opioid Non-abusers N % N % P Value Total patients 6, % 815, % Male 3, % 438, % < Age, mean (SD) 41.6 (13.5) 42.7 (14.5) < Age group < % 85, % , % 89, % , % 131, % , % 190, % , % 213, % % 104, % U.S. Census Bureau Division < East North Central 1, % 139, % Middle Atlantic 1, % 123, % Pacific % 97, % South Atlantic % 138, % West South Central % 120, % Mountain % 49, % New England % 47, % East South Central % 46, % West North Central % 53, % * Index date was defined as the date of the most recent opioid Rx during (based on fill date). P values were calculated using c 2 tests for categorical variables and using t-tests for continuous variables. U.S. Census Bureau Divisions were classified as follows: East North Central (IL, IN, MI, OH, WI), Middle Atlantic (NJ, NY, PA), Pacific (AK, CA, HI, OR, WA), South Atlantic (DC, DE, FL, GA, MD, NC, SC, VA, WV), West South Central (AR, LA, OK, TX), Mountain (AZ, CO, ID, NM, MT, UT, NV, WY), New England (CT, ME, MA, NH, RI, VT), East South Central (AL, KY, MS, TN), and West North Central (IA, KS, MN, MO, NE, ND, SD). SD = standard deviation. 1165

5 Rice et al. Table 2 Prescription drug use during 12 months prior to index date* Prescription Opioid Abusers Prescription Opioid Non-abusers N % N % P Value Total patients 6, % 815, % Number of opioid Rxs, mean (SD) 13.3 (13.1) 1.9 (4.5) < Number of unique opioid NDCs, mean (SD) 3.7 (3.7) 0.9 (1.4) < Number of opioids (by active ingredient), 1.9 (1.3) 0.7 (0.9) < mean (SD) Opioid use 5, % 383, % < By active ingredient Hydrocodone 3, % 253, % < Oxycodone 2, % 109, % < Buprenorphine 1, % % < Tramadol 1, % 54, % < Propoxyphene % 56, % < Codeine % 42, % < Fentanyl % 6, % < Morphine % 5, % < Hydromorphone % 6, % < Methadone % 2, % < Oxymorphone % % < Levorphanol 2 0.0% % Other 4 0.1% % By type Immediate-release 5, % 381, % < Extended-release 1, % 24, % < Proportion of opioid prescriptions that 0.12 (0.24) 0.01 (0.09) < were extended-release Opioid use without any other analgesic use 3, % 216, % < Non-opioid drug use 5, % 430, % < Antidepressants 3, % 174, % < Analgesics 2, % 245, % < Benzodiazepines 3, % 127, % < Relaxants 2, % 109, % < Hypnotics 1, % 66, % < Antipsychotics % 18, % < Stimulants % 27, % < Number of pharmacies filling opioid Rxs, 2.4 (2.3) 0.7 (0.9) < mean (SD) Number of opioid prescribers, mean (SD) 3.2 (3.5) 0.8 (1.3) < Early refills of opioid Rxs Any early refills 2, % 33, % < Proportion of refills filled early (among those with refills), mean (SD) 0.09 (0.16) 0.01 (0.06) < * The index date was defined as the date of the most recent opioid Rx during (based on fill date). P values were calculated using c 2 tests for categorical variables and using t-tests for continuous variables. Other opioids include meperidine, nalbuphine, pentazocine, and tapentadol. See Appendix A for active ingredients comprising drug classes. See Appendix B for a breakdown of immediate-release and extended-release opioids. Early refills were defined as any prescription opioid refill that occurred with greater than 25% of the days supply remaining on the previous prescription for the same active ingredient. All Rxs following the first Rx for each active ingredient were considered refills. NDC = National Drug Code; SD = standard deviation. (immediate- and extended-release formulations), although abusers were particularly more likely to have previously taken buprenorphine (26.3% vs 0.1%), methadone (5.8% vs 0.3%) (both substances commonly used as abuse treatment), 2 morphine (8.5% vs 0.7%), fentanyl (8.6% vs 0.7%), and oxycodone (40.8% vs 13.4%). Finally, abusers were more likely to receive other types of prescription medications such as antipsychotics (11.7% vs 2.3%). 1166

6 A Model Predicting Prescription Opioid Abuse Table 3 Comorbidities and resource use during 12 months prior to index date* Prescription Opioid Abusers Prescription Opioid Non-abusers N % N % P Value Total patients 6, % 815, % A. Comorbidity profile Charlson comorbidity index, mean (SD) 0.7 (1.6) 0.5 (1.3) < Chronic pulmonary disease % 58, % < Diabetes % 71, % Cancer % 49, % Liver disease % 18, % < Congestive heart failure % 16, % < Cerebrovascular disease % 16, % < Peripheral vascular disease % 13, % < Renal disease % 8, % < Peptic ulcer disease % 3, % < Myocardial infarction % 5, % < Hemiplegia % 2, % < HIV/AIDS % 2, % < Other selected comorbidities Chronic pain 4, % 479, % < Mental illness 3, % 134, % < Non-opioid drug abuse/poisoning 1, % 4, % < Tobacco use/dependence % 28, % < Skin infections % 50, % < Alcohol abuse/poisoning % 5, % < Hepatitis % 5, % < B. Resource use by place of service Outpatient 5, % 672, % < Emergency department 3, % 238, % < Inpatient 2, % 151, % < * The index date was defined as the date of the most recent opioid Rx during (based on fill date). P values were calculated using c 2 tests for categorical variables and using t-tests for continuous variables. Tobacco use/dependence was identified using medical diagnoses (ICD-9-CM codes: 305.1, 649.0, and V15.82) as well as pharmacy claims for smoking deterrents. SD = standard deviation. Table 3 reports the comorbidity profile and medical resource utilization for abusers and non-abusers. While there were some differences between abusers and nonabusers in the chronic comorbidity index, there were large differences between the two groups in other comorbidities. Namely, more than half of abusers (59.2%) received a mental health diagnosis in the 12 months prior to the index date, compared with only 16.6% of non-abusers. Similarly, abusers were more likely to receive a diagnosis for tobacco dependence (12.1% vs 3.5%) and alcohol abuse (10.1% vs 0.7%), and were more likely to have used other medical resources such as those from an ED (55.5% vs 29.3%) or inpatient hospital (40.3% vs 18.6%). Finally, the household characteristics of the cohorts were examined (Table 4). Abusers were more likely to have other family members with prescription opioid use (39.0% vs 28.6%) and with diagnosed abuse (2.6% vs 0.2%). In addition, abusers were more likely than non-abusers to have another family member with a mental health diagnosis (29.6% vs 18.1%). Logistic Regression Results Table 5 shows the results of the logistic regression estimating the probability that a patient is a diagnosed opioid abuser controlling for the effects of other variables included in the model. While, in general, the key differences between the abusers and non-abusers identified above remain, the following variables were identified as key characteristics (i.e., odds ratio [OR] > 2) [21]: prior opioid prescriptions (OR = 2.23 for 1 5 prior Rxs; OR = 6.85 for 6+ prior Rxs); at least one prior prescription of buprenorphine (OR = 51.75) or methadone (OR = 2.97); at least one diagnosis of non-opioid drug abuse (OR = 9.89), mental illness (OR = 2.45), or hepatitis (OR = 2.36); and having a family member diagnosed with opioid abuse (OR = 3.01). Each of these estimates was statistically significant at P < As mentioned earlier, buprenorphine (and to a lesser extent methadone) is used in the treatment of individuals with opioid abuse and therefore this is an example of a 1167

7 Rice et al. Table 4 Household characteristics during 12 months prior to index date* Prescription Opioid Abusers Prescription Opioid Non-abusers N % N % P Value At least one family member in the data 5, % 685, % Number of family members in the data, 2.1 (1.6) 2.1 (1.6) mean (SD) Household characteristics Prescription drug use Opioid use 2, % 233, % < Number of opioid Rxs, mean (SD) 2.9 (7.2) 1.1 (3.8) < Comorbidity profile Charlson comorbidity index, mean (SD) 0.4 (1.0) 0.4 (0.9) Mental illness 1, % 147, % < Cancer % 33, % Non-opioid substance abuse % 10, % < Opioid abuse % 1, % < * The index date was defined as the date of the most recent opioid Rx during (based on fill date). P values were calculated using c 2 tests for categorical variables and using t-tests for continuous variables. Household characteristics refer to behavior across all other members on the same health plan. Categorical variables refer to any opioid use or relevant diagnosis for any plan member. The number of opioid Rxs is summed across all other plan members, and the maximum Charlson comorbidity index of any family member is used for purposes of analysis. SD = standard deviation. relationship that is frequently interconnected as opposed to causal. We nonetheless included this as an independent variable in the model for two reasons. First, to the extent that the model can be used by payers to prospectively identify individuals with characteristics that suggest high risk for abuse, such a demonstrated relationship highlights the importance of including buprenorphine (even as past treatment) in the model. Second, removing buprenorphine from the model led to changes in the ORs and statistical significance of other independent variables, suggesting that it is picking up important unobserved patient characteristics. In addition to these key characteristics, some of the differences between abusers and non-abusers identified using the univariate comparisons in Tables 1 4 lose significance when controlling for other differences in characteristics. First, although abusers were more likely than non-abusers to receive prescriptions from multiple physicians, this difference was not statistically significant controlling for other differences (OR = 1.06, P = ). Second, although abusers were more likely to have received prior prescriptions for all types of opioids, prior use of propoxyphene (OR = 0.73, P < ) or hydrocodone (OR = 0.70, P < ) was associated with a reduced probability of abuse, controlling for other characteristics. The model had a c-statistic of 0.91, indicating that the model was able to correctly predict a higher probability of abuse for observed abusers in 91% of abuser non-abuser patient comparisons. Discussion This study examined the relationship between patient characteristics and diagnosed prescription opioid abuse. The study extends and improves upon previous analyses of this type, re-affirming factors previously associated with opioid abuse, and identifying several new predictors. For example, compared with White et al. [9], who analyzed a cohort of opioid users in Maine, this analysis includes patients from all 50 states, assesses a longer and more recent timeframe, utilizes a sample over seven times as large, and incorporates several new variables. Moreover, the predictive power as measured by the c-statistic suggests that models such as this may be an effective tool for payers to identify individuals at risk for opioid abuse, which can allow for appropriate management and follow-up that could lead to improved patient outcomes and cost savings. The model and results were robust to changes in functional form and sample selection criteria. Specifically, to the extent that payers (e.g., MCOs) and state PDMPs plan to use administrative claims data, we estimated a variation of the model reported above to determine whether the results hold when allowing for less restrictive data requirements, such as including shorter periods of time in which to identify independent variables (i.e., 6-month vs 12-month study period), and a more narrow period of time during which to identify an opioid abuse diagnosis (such as requiring that the opioid abuse diagnosis be within 12 months of the index date (i.e., contemporaneous with the time period in which the independent variables are created) vs diagnosis 1168

8 A Model Predicting Prescription Opioid Abuse Table 5 Logistic regression results on the probability of being a diagnosed prescription opioid abuser Full Model (No. of Non-abusers = 815,536; No. of Abusers = 6,380) 50% Random Subsample (No. of Non-abusers = 407,747; No. of Abusers = 3,211) OR (95% CI) P Value OR (95% CI) P Value I. Demographics Male 1.35 ( ) ( ) Age group* (reference) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) < ( ) U.S. Census Division New England (reference) Pacific 0.94 ( ) ( ) Middle Atlantic 0.93 ( ) ( ) Mountain 0.88 ( ) ( ) East North Central 0.78 ( ) ( ) East South Central 0.76 ( ) ( ) South Atlantic 0.69 ( ) < ( ) West South Central 0.65 ( ) < ( ) West North Central 0.51 ( ) < ( ) < II. Prescription drug use Opioid use Number of opioid Rxs 0 (reference) ( ) < ( ) < ( ) < ( ) < pharmacies filling opioid Rxs 1.61 ( ) < ( ) < physicians writing opioid Rxs 1.06 ( ) ( ) Early refill of opioid Rxs 1.41 ( ) < ( ) < Use without other analgesic use 1.14 ( ) ( ) Use by active ingredient Immediate-release Buprenorphine ( ) < ( ) < Methadone 2.97 ( ) < ( ) < Oxymorphone 1.52 ( ) ( ) Fentanyl 1.24 ( ) ( ) Morphine 1.09 ( ) ( ) Hydromorphone 0.99 ( ) ( ) Oxycodone 0.94 ( ) ( ) Tramadol 0.91 ( ) ( ) Codeine 0.91 ( ) ( ) Propoxyphene 0.73 ( ) < ( ) < Hydrocodone 0.70 ( ) < ( ) < Other 1.97 ( ) Extended-release Levorphanol 1.98 ( ) ( ) Morphine 1.52 ( ) < ( ) < Fentanyl 1.43 ( ) < ( ) < Oxycodone 1.20 ( ) ( ) Tramadol 1.10 ( ) ( ) Oxymorphone 0.96 ( ) ( ) Non-opioid drug use Benzodiazepines 1.39 ( ) < ( ) < Stimulants 1.23 ( ) < ( ) Hypnotics 1.21 ( ) < ( ) Relaxants 1.19 ( ) < ( )

9 Rice et al. Table 5 Continued Full Model (No. of Non-abusers = 815,536; No. of Abusers = 6,380) 50% Random Subsample (No. of Non-abusers = 407,747; No. of Abusers = 3,211) OR (95% CI) P Value OR (95% CI) P Value III. Comorbidities and resource use** Charlson comorbidity index Peptic ulcer disease 1.17 ( ) ( ) Hemiplegia 1.08 ( ) ( ) Peripheral vascular disease 0.99 ( ) ( ) HIV/AIDS 0.99 ( ) ( ) Myocardial infarction 0.97 ( ) ( ) Chronic pulmonary disease 0.96 ( ) ( ) Congestive heart failure 0.95 ( ) ( ) Renal disease 0.90 ( ) ( ) Diabetes 0.86 ( ) ( ) Cerebrovascular disease 0.81 ( ) ( ) Liver disease 0.76 ( ) ( ) Cancer 0.51 ( ) < ( ) < Other comorbidities Non-opioid drug abuse 9.89 ( ) < ( ) < Mental illness 2.45 ( ) < ( ) < Hepatitis 2.36 ( ) < ( ) < Alcohol abuse 1.67 ( ) ( ) Tobacco use/dependence 1.45 ( ) < ( ) < Skin infection 1.04 ( ) ( ) Chronic noncancer pain 0.99 ( ) ( ) Resource use Inpatient hospitalization days 0 (reference) ( ) ( ) ( ) < ( ) < Emergency department days 0 (reference) ( ) < ( ) ( ) < ( ) < Outpatient days 0 (reference) ( ) ( ) ( ) ( ) IV. Family effects Number of opioid Rxs 0 (reference) ( ) ( ) ( ) < ( ) < Comorbidities Opioid abuse 3.01 ( ) < ( ) < Mental illness 1.16 ( ) < ( ) < Non-opioid drug/alcohol abuse 1.06 ( ) ( ) Cancer 1.03 ( ) ( ) Charlson comorbidity index 0.98 ( ) ( ) V. Interactions Age group substance abuse n/a (n/a) n/a (n/a) Ages substance abuse (reference) Ages substance abuse 0.78 ( ) ( ) Ages substance abuse 0.50 ( ) ( ) Ages substance abuse 0.53 ( ) ( ) Ages substance abuse 0.55 ( ) ( ) Ages substance abuse 0.62 ( ) ( )

10 A Model Predicting Prescription Opioid Abuse Table 5 Continued Full Model (No. of Non-abusers = 815,536; No. of Abusers = 6,380) 50% Random Subsample (No. of Non-abusers = 407,747; No. of Abusers = 3,211) OR (95% CI) P Value OR (95% CI) P Value Age group gender n/a (n/a) n/a (n/a) Ages male (reference) Ages male 1.29 ( ) ( ) Ages male 0.97 ( ) ( ) Ages male 0.89 ( ) ( ) Ages male 1.09 ( ) ( ) Ages male 0.76 ( ) ( ) Age group alcohol abuse n/a (n/a) n/a (n/a) Ages alcohol abuse (reference) Ages alcohol abuse 1.13 ( ) ( ) Ages alcohol abuse 2.03 ( ) ( ) Ages alcohol abuse 2.12 ( ) ( ) Ages alcohol abuse 1.74 ( ) ( ) Ages alcohol abuse 2.20 ( ) ( ) c-statistic * Age group was determined as index date, which was defined as the date of the most recent opioid prescription fill during U.S. Census Bureau Divisions were classified as follows: East North Central (IL, IN, MI, OH, WI), Middle Atlantic (NJ, NY, PA), Pacific (AK, CA, HI, OR, WA), South Atlantic (DC, DE, FL, GA, MD, NC, SC, VA, WV), West South Central (AR, LA, OK, TX), Mountain (AZ, CO, ID, NM, MT, UT, NV, WY), New England (CT, ME, MA, NH, RI, VT), East South Central (AL, KY, MS, TN), and West North Central (IA, KS, MN, MO, NE, ND, SD). Prescription drug use was assessed during the 12 months prior to index date, which was defined as the date of the most recent opioid prescription fill during Early refills were defined as any prescription opioid refill that occurred with greater than 25% of the days supply remaining on the previous prescription for the same active ingredient. All Rxs following the first Rx for each active ingredient were considered refills. Other opioids include meperidine, nalbuphine, pentazocine, and tapentadol. ** Comorbidities and resource use were assessed during the 12 months prior to index date, which was defined as the date of the most recent opioid prescription fill during Mental illness was identified using medical claims for mental disorders not associated with substance abuse (ICD-9-CM: , ) as well as drug claims for antidepressants or antipsychotics. Tobacco use/dependence was identified using medical diagnoses (ICD-9-CM codes: 305.1, 649.0, and V15.82) as well as pharmacy claims for smoking deterrents. Household characteristics refer to behavior across all other members on the same health plan. Categorical variables refer to the presence of any relevant diagnosis for any plan member. Charlson comorbidity index (CCI) refers to the maximum CCI of any family member. OR = odds ratio; CI = confidence interval. at any time in the patient s history). The model results were robust to each of these sensitivity analyses, with the key characteristics remaining unchanged, and with c-statistics remaining at or above In addition, given the large sample size of the analytic sample, a version of the model was re-estimated based on a randomly selected 50% sample (without replacement) (Table 5). While some variables lost statistical significance (P < 0.01) vs the full model (e.g., age group 50 59, Census Divisions East North Central and East South Central, diagnoses of diabetes), this validation produced similar results, with no change in predictive power (c-statistic = 0.91). This study has a number of limitations. First, as with all claims data analyses of opioid misuse, identification of abusers in this analysis is limited to patients with diagnosed abuse, dependence, or misuse (using ICD-9-CM codes). In particular, it does not account for other definitions of abuse/dependence/misuse, such as that of nonmedical use of prescription drugs adopted by the NSDUH. Although the diagnostic approach is the only option for claims data analysis, these codes do not distinguish between illicit heroin use and misuse of prescription opioids. As a result, the abusers category may contain individuals who were diagnosed with heroin abuse but were not abusing prescription opioids. This analysis sought to account for this possibility by requiring that all patients have at least one prescription opioid claim in the study period; however, we expect that many heroin abusers may also use prescription opioids. Moreover, it is expected that many patients who misuse prescription opioids are undiagnosed. As mentioned earlier, to the extent that some prescription opioid abusers remain undiagnosed, they are contained in the non-abuser category for this analysis. While it is unknown how the characteristics of undiagnosed abusers compare with diagnosed 1171

11 Rice et al. abusers, if undiagnosed abusers exhibit similar characteristics of diagnosed abusers (e.g., filling multiple opioid prescriptions, filling at multiple pharmacies), this would imply that the descriptive statistics and ORs estimated in this analysis understate actual differences between abusers and non-abusers. However, a useful extension of this analysis would be to assess the generalizability of these results to other populations, in which a more in-depth assessment of abuse is possible (e.g., interview) or who are covered by Medicaid and other payers. In addition, the model was only able to incorporate information available from such administrative data, in essence a closed model. For example, many abusers are able to obtain their opioid supply from the black market, or who pay cash for opioids at in-and-out clinics such as pill mills, as access to these alternative supply sources is not captured in claims data. More generally, it is important to keep in mind that the characteristics presented here are representative of abusers filling prescriptions at a pharmacy (the relevant population from the perspective of MCOs and PDMPs) and may not be reflective of abusers overall. Finally, the model identifies characteristics that are predictive of opioid abuse; however, this study did not attempt to determine whether the variables associated with abuse are direct causes or consequences of abuse. Conclusions As prescription opioid abuse becomes more widespread, developing ways to increase awareness of the characteristics that are associated with abuse may identify candidates for screening, brief intervention, and referral to treatment (SBIRT) programs and help physicians make more informed prescribing decisions. In addition, developing predictive models of abuse may assist payers in establishing evidence-based methods of identifying at-risk individuals and communities. The results presented in this study suggest that predictive models using medical claims as well as drug claims data, such as those likely available to most payers, can identify several such characteristics. These results indicate that these models can be an important tool to identify potentially inappropriate opioid use. An important contribution of such models is that they incorporate medical information beyond that available to prescription drug monitoring programs that are reliant only on drug claims data. Notes 1. The two exceptions to this rule were for number of pharmacies and number of physicians, whose distributions did not allow such construction. These variables were constructed as 0/1 dichotomous variables, with the cut-offs chosen to indicate the bottom 90% and top 10% of patients for that variable. 2. In fact, some companies cover buprenorphine only for patients with an opioid abuse or dependence diagnosis (see, e.g., BlueCross BlueShield of Kansas City), although it is unknown how many of the plans in these data have such prior authorization policies. 3. Full model results, including underlying model, coefficients and sensitivity analyses are available upon request. References 1 National Center for Health Statistics. Health, United States: With Chartbook on Trends in the Health of Americans. Special Feature: Pain. Hyattsville, MD: U.S. Department of Health and Human Services; Volkow ND, McLellan TA. Curtailing diversion and abuse of opioid analgesics without jeopardizing pain treatment. JAMA 2011;305(13): Manchikanti L, Fellows B, Ailinani H, Pampati V. Therapeutic use, abuse, and nonmedical use of opioids: A ten-year perspective. Pain Physician 2010;13: Substance Abuse and Mental Health Services Administration (SAMHSA). Results from the 2009 National Survey on Drug Use and Health: Volume II. Technical Appendices and Selected Prevalence Tables (Office of Applied Studies, NSDUH Series H-38B, HHS Publication No. SMA Appendices). Rockville, MD: U.S. Department of Health and Human Services; Available at: 2k9NSDUH/2k9ResultsApps.htm#AppG (accessed July 12, 2012). 5 Substance Abuse and Mental Health Services Administration, Drug Abuse Warning Network, 2009: National Estimates of Drug-Related Emergency Department Visits. HHS Publication No. (SMA) , DAWN Series D-35. Rockville, MD: Substance Abuse and Mental Health Services Administration; Webster LR, Cochella S, Dasgupta N, et al. An analysis of the root causes for opioid-related overdose deaths in the United States. Pain Med 2011;12(suppl 2):S White AG, Birnbaum HG, Schiller M, et al. The economic impact of opioid abuse, dependence, and misuse. Am J Pharm Benefits 2011;3(4):e Birnbaum HG, White AG, Schiller M, et al. Societal costs of opioid abuse, dependence, and misuse in the United States. Pain Med 2011;12(4): White AG, Birnbaum HG, Schiller M, Tang J, Katz N. Analytic models to identify patients at risk for prescription opioid abuse. Am J Manag Care 2009;15(12): Augmon SC. Characterizing the emerging population of prescription opioid abusers. Am J Addict 2006;15(3):

12 A Model Predicting Prescription Opioid Abuse 11 Braker LS, Reese AE, Card RO, Van Howe RS. Screening for potential prescription opioid misuse in a Michigan Medicaid population. Fam Med 2009;41(10): Dasgupta N, Kramer ED, Zalman MA, et al. Association between non-medical and prescriptive usage of opioids. Drug Alcohol Depend 2006;82(2): 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(3): Dennis M, Scott CK. Managing addiction as a chronic condition. Addict Sci Clin Pract 2007;4(1): Ives TJ, Chelminski PR, Hammett-Stabler CA, et al. Predictors of opioid misuse in patients with chronic pain: A prospective cohort study. BMC Health Serv Res 2006;6: Sullivan MD, Edlund MJ, Fan MY, et al. Risks for possible and probable opioid misuse among recipients of chronic opioid therapy in commercial and Medicaid insurance plans: The TROUP study. Pain 2010;150(2): Turk DC, Swanson KS, Gatchel RJ. Predicting opioid misuse by chronic pain patients: A systematic review and literature synthesis. Clin J Pain 2008;24(6): Dunn KM, Saunders KW, Rutter CM, et al. Opioid prescriptions for chronic pain and overdose. Ann Intern Med 2010;152: Jamison RN, Butler SF, Budman SH, Edwards RR, Wasan AD. Gender differences in risk factors for aberrant prescription opioid abuse. J Pain 2010;11(4): Martyres RF, Clode D, Burns JM. Seeking drugs or seeking help? Escalating doctor shopping by young heroin users before fatal overdose. Med J Aust 2004;180: Chanock SJ, Manolio T, Boehnke M, et al. Replicating genotype-phenotype associations. Nature 2007;447: