Joseph C. Cappelleri, PhD

We sought to evaluate opioid prescribing in an ambulatory setting among patients with noncancer chronic pain (CP). 

 We identified patients with at least 2 CP encounters at least 30 days apart in 2012 in the electronic health records of a community-based healthcare delivery system in northern California. We used logistic regression models to assess associations of receiving an opioid prescription with respect to number and type of CP conditions and patient demographics and characteristics. Odds ratios (ORs) with 95% confidence intervals (CIs) and the adjusted prevalence of receiving an opioid prescription were generated after controlling for important covariates. 

 A total of 120,481 patients with CP met eligibility criteria, with 58% receiving an opioid in 2012. The adjusted prevalence of receiving an opioid was highest for back/cervical pain (71%). The odds of receiving an opioid increased linearly with the number of CP conditions per patient (OR, 1.29; 95% CI, 1.25-1.33; 

 <.001). Men were generally more likely to receive an opioid than women, as were patients with noncommercial insurance, especially Medicaid (OR, 2.77; 95% CI, 2.56-3.01; 

In an ambulatory healthcare setting, opioid prescribing to patients with CP varied by type and number of pain conditions. Opioid prescriptions to men, those with back/cervical pain, and Medicaid beneficiaries were particularly prevalent. The identification of populations more likely to receive an opioid in the treatment of CP should be of interest to healthcare systems to ensure these drugs are being used appropriately and safely. 

Opioid prescribing to patients with noncancer chronic pain (CP) in an outpatient setting varied by the number and type of pain conditions per patient, as well as by other patient characteristics. 

		Opioid prescriptions for men, those with back/cervical pain, and Medicaid beneficiaries were particularly prevalent. 

		Patients prescribed more than 5 medications were more likely to receive an opioid prescription than those prescribed fewer than 5 medications. 

		The identification of populations more likely to receive an opioid in the treatment of CP is important information for healthcare systems working to ensure these drugs are being used appropriately and safely.

An estimated 25 million Americans—more than 10% of the adult population&mdash;experience pain on a daily basis.

 In the United States, opioids are one of the most commonly prescribed drugs to treat pain, with nearly 260 million prescriptions written in 2012 alone.

In recent years, controversy around opioid prescribing has risen due to concerns about misuse, abuse, and diversion leading to opioid-related overdoses and deaths.

In 2014, there were 18,893 opioid overdose-related deaths in the United States, for an average of 50 per day.

A systematic review conducted by the Agency for Healthcare Research and Quality in 2014, and a recent update to this review published by the CDC in 2016, showed limited evidence supporting the effectiveness of long-term opioid use for noncancer chronic pain (CP) lasting more than 3 months and indicated that risks for serious harms are dose-dependent.

 Based on this evidence—or lack thereof&mdash;the CDC put forth guidelines for prescribing opioids for CP, with recommendations for weighing the benefits and risks of treatment, establishing treatment goals with patients, and prescribing the lowest effective opioid dosage, among others.

Healthcare systems are challenged with ensuring that patients with CP receive appropriate treatment with positive clinical outcomes while keeping them safe from harm. Among barriers to adopting opioid practice guidelines, clinicians report inadequate training on pain management

 and time constraints for comprehensive opioid risk assessment and drug monitoring, especially in busy primary care practices.

 An understanding of populations that have a greater propensity for receiving opioid prescriptions may inform initiatives within healthcare systems to ensure that these drugs are used appropriately and safely. To our knowledge, no studies have been conducted to evaluate opioid prescribing across diverse CP conditions within a single healthcare setting. 

Previously, we established a cross-sectional cohort of adult ambulatory patients with CP in 2012 from a large community-based healthcare system in northern California to calculate the prevalence of CP overall and by various CP conditions, including back/cervical pain, arthritis/joint pain, neuropathies/neuralgias, headaches/migraines, and unclassified pain.

 In the present study, we used this cohort to examine opioid prescribing, as well as prescribing of other pain medications. We explored differences in opioid prescribing by the number and types of CP conditions per patient and by patient demographics and characteristics. 

This study was conducted using Sutter Health electronic health record (EHR) data from 2012. Sutter Health is a community-based open-network healthcare system in northern California that provides ambulatory (primary and specialty) care across 130 medical clinics to approximately 3 million patients annually (10 million ambulatory visits in 2012). The system also provides care across 24 acute care hospitals, with 200,000 inpatient admissions and 800,000 emergency department visits in 2012. Similar to many other healthcare systems in the nation, Sutter Health is a mixed-payer organization with no single formulary; as such, this setting is appropriate to study drug prescribing patterns in a clinical population. Its EHR (EpicCare) is integrated across all ambulatory care clinics and hospitals. This study was approved by Sutter Health’s Institutional Review Board, and all data were de-identified in accordance with Health Insurance Portability and Accountability Act standards. 

Our study cohort was inclusive of adult patients with CP with a medical record in the EHR system. First, we identified patients 18 years or older with at least 2 

International Classification of Disease, 9th Revision, Clinical Modification

) diagnoses for a CP condition at least 30 days apart in 2012. The

 codes for chronic pain conditions, by definition, exclude acute pain conditions and were based on well-described criteria from recent health system—based studies of CP,

 A listing of these chronic pain diagnosis codes has been published previously.

We required patients to have at least 1 encounter of any type before 2010 to confirm prior contact with the health system and to further characterize comorbidities and medication utilization. We excluded patients with an encounter or problem-list diagnosis of malignancy (with the exception of nonmelanoma skin cancer) in the 2 years prior to 2012 to restrict patients from the analysis with cancer-related pain. We also excluded patients with surgery in the 3 months prior to the first CP encounter in 2012 to restrict patients from the analysis with acute postsurgical pain. 

Data were collected from the EHR, including information on prescribed medications and patient demographics (age, gender, and race/ethnicity) and characteristics (insurance type and comorbidities). Patient race/ethnicity was self-reported and collected as a part of routine clinical practice according to US Census standards. We categorized race/ethnicity as Hispanic (of any race), non-Hispanic white (NHW), African American, Asian, other (Pacific Islander, American Indian/Alaskan Native, multiple races reported, or race reported as “other”), and unknown. We further disaggregated the population’s 6 largest Asian subgroups as East/Southeast Asian (Chinese, Japanese, Korean, Filipino, or Vietnamese) and Asian Indian. Charlson Comorbidity Index (CCI) scores were calculated for each patient as a measure of overall disease burden.

 Insurance type was categorized as commercial (preferred provider organization/health maintenance organization), Medicare, Medicaid, Medicare/Medicaid dual eligible, or other/unknown (including self-pay). Patients were grouped into 5 non—mutually exclusive CP categories based on anatomical location and/or pathophysiology: arthritis/joint pain, back/cervical pain, neuropathies/neuralgias, headaches/migraines, and unclassified pain (including fibromyalgia, pelvic pain, abdominal pain, and general pain). 

The primary outcome measures were medications prescribed for pain, categorized as analgesics (opioid and nonopioid agents) and nonanalgesics (benzodiazepines, antidepressants, muscle relaxants, antiepileptics, corticosteroids, antimigraine agents, antirheumatic agents, and topical agents). See the 

) for a listing of drug classes and, where appropriate, subclasses. Because data are from an open-network health system, we did not have comprehensive pharmacy claims data on this population and instead used EHR prescribing data. From a previous study of this cohort,

 56% of all outpatient visits were to a primary care physician (family or internal medicine); the remaining 44% of encounters spanned 20 different medical specialties and service lines.

We used descriptive statistics to summarize continuous and categorical variables and logistic regression to assess associations between receiving an opioid prescription and patient demographics/characteristics.

 Models included receipt of an opioid in 2012 as the binary dependent variable. Independent variables included patient age, CCI score, total number of CP conditions by category per patient (ie, from 1-5), insurance type, and the number of other non—pain-related medications received during the study period. Because previous study results have shown differences in opioid use by gender and race/ethnicity,

 we also explored these differences. We combined gender and race/ethnicity as a single variable in regression models to allow for different slopes for associations between opioid prescribing and racial/ethnic group by men and women (eg, NHW men, NHW women, Hispanic men, Hispanic women). This has the same consequence as an interaction term but with simpler interpretation of model coefficients. We also included categorical dummy variables in regression models for CP categories: arthritis/joint pain, back/cervical pain, neuropathies/neuralgias, and headaches/migraines. A dummy variable for unclassified pain was excluded due to collinearity with other CP categories. 

We generated unadjusted and adjusted odds ratios (ORs) with corresponding 95% confidence intervals (CIs) for univariate and multivariate associations, respectively. We compared model coefficients within levels of the gender and racial/ethnic category by posthoc estimation of linear combinations.

 We derived the adjusted prevalence of receiving an opioid prescription from posthoc estimation of adjusted mean effects (holding constant the values of all other covariates).

We repeated statistical models to estimate the adjusted prevalence of receiving an opioid by the number of total CP conditions for each category (eg, back/cervical pain alone, back/cervical pain plus 1 additional condition or plus 2 additional conditions). Because multiple tests were performed, we used a level of statistical significance of 0.01, instead of 0.05, to mitigate type 1 error (ie, the probability of falsely rejecting the null hypothesis of no difference). All analyses were performed in Stata version 13.0 (StataCorp; College Station, Texas). 

Among 1,784,114 adult patients with an ambulatory record in the health system, 120,481 (6.8%) met study eligibility criteria for CP, which is the estimated prevalence of these conditions during the study period.

 The most prevalent CP conditions by category were arthritis/joint pain (57%), back/cervical pain (49%), and neuropathies/neuralgias (40%) (

). Approximately 40% of patients had a total of 1 or 2 CP conditions. Patients were aged 56 years on average; the majority were women (66%) and NHW (63%). Approximately half of the cohort comprised commercially insured beneficiaries. Patients were prescribed a median of 11 unique (pain and nonpain) medications during the study period. 

Analgesic and Nonanalgesic Prescribing by CP Category 

Across all CP categories, most patients (92%) received an analgesic or nonanalgesic medication (

). Nonsteroidal anti-inflammatory drugs (NSAIDs) were the most commonly prescribed medications (68%), followed by short-acting (immediate-release) opioids (57%); 58% received a short- or long-acting (extended-release) opioid. 

Analgesics (opioid or nonopioid) (89%), any opioid (68%), and, specifically, short-acting opioids (67%) were most common in patients with back/cervical pain (Table 2). Patients received, on average, 2 analgesics during the study period; those with arthritis/joint pain and unclassified pain had the highest average number of prescriptions for analgesic medications (mean = 2.3 each for arthritis/joint pain and unclassified pain). 

Prescriptions for nonanalgesics were most common in patients with headaches/migraines (86.3%) (Table 2). Patients were prescribed, on average, 1.8 unique nonanalgesic medications; those with headaches/migraines and unclassified pain received the highest average number of nonanalgesic medications (mean = 2.6 each for arthritis/joint pain and unclassified pain). 

Opioid Prescribing by Number and Type of CP Conditions 

Among patients with a single CP type, the highest prevalence of receiving an opioid, after controlling for other covariates, was for back/cervical pain (71%) (

). For each CP category, the adjusted prevalence of receiving an opioid increased linearly as the number of CP conditions increased: arthritis/joint pain (OR, 1.39; 95% CI, 1.36-1.42; 

<.001), back/cervical pain (OR, 1.07; 95% CI, 1.05-1.09; 

 <.001), neuropathies/neuralgias (OR, 1.65; 95% CI, 1.61-1.69; 

 <.001), headaches/migraines (OR, 1.51; 95% CI, 1.47-1.56; 

 <.001), and unclassified pain (OR, 1.48; 95% CI, 1.44-1.53; 

Patient Demographics and Characteristics Associated With Receiving an Opioid Prescription 

We found several patient demographics and characteristics to be associated with receipt of an opioid prescription (

). After statistical adjustment, older patients (≥66 years vs 18-45 years), those with moderate chronic disease burden (CCI score = 2-3 vs 0), women, Asians (vs NHW), and commercially insured beneficiaries (vs all other insurance types) had lower odds of receiving an opioid. Patients with a greater total number of CP conditions had higher odds of receiving an opioid. 

Because of the potential correlation between age, CCI score, and insurance type (eg, older individuals tend to have more comorbidities and are more likely to be Medicare beneficiaries), we performed analyses to examine the effect of CCI and payer type on the relationship between age and receipt of an opioid. The presence or absence of these covariates had minimal effect on model coefficients within the age category (data not shown). 

In this cross-sectional analysis of EHR prescribing data from a community-based ambulatory healthcare system setting, we found differences in opioid prescribing to patients by number and type of CP conditions. Among patients with 1 CP condition, those with back/cervical pain had the highest prevalence of receiving an opioid, whereas those with neuropathies/neuralgias and headache/migraines had the lowest prevalence; these latter groups more frequently received nonanalgesic medications (eg, antiepileptic or antimigraine medications). The prevalence of receiving an opioid increased linearly with the number of CP conditions per patient, reaching more than 60% for those with 3, more than 70% for those with 4, and approximately 80% for those with all 5 CP conditions; however, few patients had conditions within 4 or all 5 CP categories (4.5% and 0.8%, respectively). 

Our study is consistent with previous studies showing that patients with back pain are more likely to receive an opioid than those with other pain conditions.

 Back pain is one of the most common reasons for a visit to a doctor’s office, composing approximately one-fourth of all encounters.

 Higher rates of opioid prescribing among patients with back pain may be due to the paucity of effective alternative pharmacological agents, particularly for back pain lasting more than 4 weeks.

 In general, opioid prescribing for all types of chronic musculoskeletal pain has increased 2-fold between 1980 and 2000 (8% to 16%).

These trends reflect an increased availability of newly marketed pain products in the 1990s and changes to state medical board opioid regulations around the same time, which relaxed opioid prescribing practices for CP.

 It is notable that more than half of patients with only unclassified pain—inclusive of conditions like fibromyalgia and pelvic pain&mdash;received an opioid; these patients had the highest prevalence of prescriptions for long-acting opioids (14.2%). Current guidelines do not support the use of most opioids in unclassified painful conditions where the harms may outweigh potential benefits, regardless of whether other treatments had been previously tried.

 The exception is tramadol, which has serotonin and norepinephrine reuptake inhibitory effects and is used in the treatment of fibromyalgia. 

In multivariate analyses, we found that older age and higher CCI score were negatively associated with receiving an opioid. Although older individuals may be expected to have a greater need for pain medications due to higher prevalence of painful conditions and greater overall disease burden,

 certain medications, including opioids, are not recommended in the elderly population due to increased risks.

 For example, opioid-related risks, such as respiratory depression, tend to be potentiated in older patients.

 Even minor side effects, including opioid-induced drowsiness or sedation, may have more serious consequences in this population. Similarly, older patients and those with a higher CCI may be more likely to have underlying liver or renal disease, which can impact opioid metabolism and lead to harmful effects.

We also found that for most racial/ethnic groups, men received an opioid prescription more often than women. This trend has been demonstrated in other studies of CP.

 We further showed that among men, there were no significant differences in the adjusted prevalence of receiving an opioid for NHWs, African Americans, or Hispanics; however, among women, African Americans were more likely than NHWs to receive an opioid. Notably, East/Southeast Asians and Asian Indians of both genders had the lowest prevalence of receiving opioid prescriptions. 

We are unaware of studies on this topic among Asians receiving treatment for CP in ambulatory settings, but the results of a small prospective cohort study from an inpatient postoperative setting showed that Chinese patients required lower opioid doses, yet were more likely to experience opioid-induced pruritus than a matched group of Caucasians.

 Differences in opioid metabolism have been described previously for Asians relative to Caucasians

; however, sociocultural factors may also explain disparities in opioid prescribing.

 Future studies are needed to understand whether physicians are less likely to prescribe opioids to Asians because of perceived risks or if Asians are less accepting of these medications. 

In our analysis, Medicaid beneficiaries were more likely than patients with other insurance types to receive opioids, even after controlling for other factors. CMS has reported that these beneficiaries are twice as likely to receive an opioid prescription than non-Medicaid patients.

 Another notable finding was that the number of nonpain medications prescribed was positively associated with an increase in opioid prescribing. This is contrary to expectations, but may be related to higher levels of comorbidity in patients who receive opioids. Further investigation of this finding is required.

To our knowledge, this is the first study to evaluate opioid prescribing across a range of conditions with distinct CP categories. We analyzed data from a large ambulatory care cohort and used rigorous methods to classify and identify patients with painful chronic conditions. Furthermore, data were derived from a mixed-payer organization with no single formulary, making it an appropriate setting for evaluating patterns in opioid prescribing. 

This was a cross-sectional analysis and causal inferences are restricted. Also, the study population was from a relatively small geographic area and generalizability to other parts of the United States is unknown; however, as a mixed-payer healthcare system, the setting is like many others in the nation. The findings of this study are reliant on the accuracy of, and frequency at which physicians document, information in the EHR. Because pharmacy data were not available on the majority of the population, we relied on prescribing data, meaning we cannot know what medications patients actually filled or consumed. We also cannot know whether nonanalgesic medications were prescribed to treat pain or comorbid conditions, such as epilepsy, depression, or anxiety. We did not specifically collect information on the type of provider who prescribed the drug, as the focus of this study was not on provider-level attributes of opioid prescribing. Lastly, we did not have information on over-the-counter medications, such as NSAIDs or acetaminophen, which were potentially used by patients to treat CP and may have influenced whether a prescription pain medication was warranted. 

Articles

This study identified populations with non-cancer chronic pain to determine which patients may be more likely to receive an opioid prescription in an outpatient setting.

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

To compare the incidence of serious gastrointestinal (GI) complications and associated medical costs in a population with either osteoarthritis (OA) or rheumatoid arthritis (RA) enrolled in Medicare plans with celecoxib formulary restrictions versus plans without such restrictions.

 This study was a retrospective cohort analysis of Medicare members in plans with and without celecoxib restrictions. Members diagnosed with OA or RA were identified and followed for 1 year.

The restricted group had higher levels of nonselective nonsteroidal anti-inflammatory drug use (51% vs 40%; 

 <.001), and celecoxib use was double in the unrestricted group (16% vs 8%; 

 <.001). The incidence of a serious GI complication was slightly higher in the restricted group (5.4% vs 4.6%; 

 <.001). The adjusted mean serious GI complication—related cost for the restricted group was more than 15 times higher than that for the nonrestricted group ($1559 [95% confidence interval (CI) $1341-$1811] vs $101 [95% CI $87-$117]); adjusted mean arthritis-related medical costs were $5733 per year (95% CI $5097-$6448) for the restricted group and $3170 (95% CI $2816-$3569) for the unrestricted group.

 The restricted group had significantly less use of celecoxib, indicating that restriction was effective at reducing celecoxib utilization. Although limitations exist when comparing populations from different health plans, and the underlying causes of serious GI complications are multifactorial, the restricted group had a higher

	incidence of serious GI complications and higher costs related to serious GI complications and arthritis.

	This study compared the incidence of serious gastrointestinal (GI) complications and associated GI- and arthritis-related medical costs between Medicare plans with and without formulary restriction policies for celecoxib.

		 Plans with a formulary restriction policy had a significantly higher incidence of serious GI complications.

		 Plans with a formulary restriction policy had significantly higher serious GI-related medical costs.

		Plans with a formulary restriction policy had significantly higher arthritis-related medical costs.

	Nonsteroidal anti-inflammatory drugs (NSAIDs) are widely used in the treatment of osteoarthritis (OA) and rheumatoid arthritis (RA).

 The most common adverse events experienced by NSAID users are gastrointestinal (GI) related. Adverse GI events can range from very minor dyspepsia to life-threatening GI bleeding or perforation.

 The major risk factors for GI complications related to NSAID use include history of GI bleed, concomitant use of anticoagulants, use of 2 or more NSAIDs, history of peptic ulcer, high-dose NSAID, age >60 years, severe illness, Helicobacter pylori infection, and concomitant use of corticosteroids.

 Lanas et al indicated that individuals over the age of 75 years are considered to be at high risk and their risk equals that of persons with a history of peptic ulcer disease.

 It has been estimated that more than 90% of all NSAIDs prescribed are for individuals over the age of 65 years.

 It is well documented that the elderly experience significant and serious GI complications with NSAIDs.

	Celecoxib, a selective cyclo-oxygenase (COX-2) inhibitor NSAID, has been shown to be as effective as traditional nonselective NSAIDs (nsNSAIDs), 

 although the product labeling for all NSAIDs carries the same GI and cardiovascular warning. The majority of traditional nsNSAIDs have generic equivalents available, while there are no COX-2 generic equivalents. Many health plans, in an effort to control costs, have implemented utilization management techniques such as tier placement, prior authorization, or step therapy to decrease utilization of COX-2 agents.

 In 2004, Briesacher et al found that members in a 3-tier pharmacy benefit design with arthritis and serious GI comorbidity were less likely to use a COX-2 agent compared with those with a 1-tier pharmacy benefit.

 Johnson et al evaluated the impact of NSAID formulary restriction on medical costs.

 They showed that the more restrictive formularies led to higher hospitalizations for OA and RA members and higher total healthcare costs for RA members.

 Although these approaches have the ability to lower pharmacy costs by decreasing utilization, very little has been published to show the impact of the restrictions on medical outcomes. Since 2007, some Medicare Advantage and Prescription Drug Plans have used various formulary management strategies to restrict the use of celecoxib. Humana health plans have used a step-therapy policy with prior authorization restriction. That is, in order to get celecoxib covered through their prescription benefit, Humana Medicare members must have tried and failed with 2 prescription-strength nsNSAIDs on the preferred drug list in the previous 6 months or must satisfy at least 1 of the following secondary criteria: currently on a prescription-strength GI medication, anticoagulant or antiplatelet therapy, bisphosphonate, chronic oral corticosteroid, or antineoplastic agent. If neither of these criteria are satisfied, the member’s provider could file a prior authorization request.

	The purpose of this study was to compare the incidence of serious GI complications and associated medical costs in an elderly Medicare population with OA and/or RA enrolled in plans with restrictions on celecoxib versus plans without restrictions on celecoxib.

	This retrospective study used a cohort design to identify and track members with arthritis who were covered under a Medicare Advantage or Medicare Supplemental plan. Two groups were identified for comparison: (1) restricted celecoxib coverage and (2) unrestricted celecoxib coverage. The study design was approved by an institutional review board.

	All members with coverage through a Medicare Advantage or Medicare Supplemental health benefit plan who were continuously enrolled during an observation period 365 days prior to and after the index date and who were between 65 and 90 years of age on the index date were included in the study. Members who met either of the following 2 criteria were excluded from the study: (1) diagnosis of familial adenomatous polyposis during the preindex or follow-up period (defined as International Classification of Diseases, Ninth Revision, Clinical Modification [ICD-9-CM] code 211.3 in any diagnosis field on a medical claim) or (2) diagnosis of ankylosing spondylitis during the preindex or follow-up period (defined as ICD-9- CM code 720.0 in any diagnosis field on a medical claim).

	The restricted celecoxib group was identified from Humana’s Medicare population using medical, pharmacy, and eligibility data. The unrestricted celecoxib group was identified using the MarketScan Medicare Supplemental and Coordination of Benefits Database from Thomson Reuters. Unrestricted status was based on participating in a plan that did not have a prior authorization/step therapy requirement for celecoxib. Thomson Medstat associates assigned the unrestricted status specific to each member for each calendar year. To assign this indication, members were first grouped by drug plan and year. Details on celecoxib coverage by plan and year were determined using Fingertip Formulary (Fingertip Formulary LLC, Glen Rock, New Jersey), Epocrates (Epocrates Inc, San Mateo, California), or the plans’ Web sites. When information was not available, the proportion of celecoxib fills relative to all drug fills was used as a proxy for restriction. Members enrolled in plans with <1.5% fills for celecoxib were tagged as restricted. Only members enrolled in a plan with an unrestricted status during their entire study period were considered for inclusion.

	The study period was January 1, 2006, to December 31, 2008. Members who met all of the following criteria were included in the study: medical claim with an ICD-9-CM diagnosis code in any of the first 3 diagnosis fields for OA (715.x) or RA (714.x or v82.1) between January 1, 2007, and December 31, 2007. The service date of the first claim was defined as the index date. Those members who met the inclusion criterion were observed for 12 months prior to their index date and 12 months after their index date.

	Pharmacy and medical claims with service dates that occurred during the individual subjects’ 24-month observation period were used to identify the outcomes of interest for the study. The composite primary end point of a serious GI complication was defined as a GI bleed or perforation (

). Serious GI complications were further identified as upper or lower in nature. The relevant codes were selected based on published literature and expert opinion. The time until a serious GI complication was calculated as the number of days between the index date and service date on the medical claim.

	Serious GI complication—related total medical per member per year (PMPY) costs during the follow-up period included the total outpatient and inpatient medical costs where the primary or secondary diagnosis code was for a serious GI complication. Arthritis-related total medical PMPY costs in the follow-up period included the total outpatient and inpatient medical costs where the primary or secondary diagnosis code was for RA or OA. Both plan-paid and member out-of-pocket costs were included in the calculation of serious GI- and arthritis-related medical costs. Gastroprotective and gastrotoxic medication utilization was evaluated using the pharmacy claims for the member. Arthritis treatment utilization was also evaluated (

). An RxRisk-V score, which is an enhancement of the Chronic Disease Score,

 was calculated using pharmacy utilization during the 12 months prior to the index date. The unweighted version used in this analysis was calculated using a simple count of drugs in unique drug categories. The score can range from 0 to 27.

	Bivariate descriptive analyses were conducted on all study variables. The 

 test or Wilcoxon rank-sum methods were utilized for comparing continuous variables, and X

tests were utilized for the categorical variables. The baseline characteristics (demographic, clinical, and drug utilization patterns) were compared between the celecoxib-restricted and celecoxibunrestricted groups. Statistical significance was set at 

	Cox proportional hazard models were used to compare the restricted and unrestricted groups on the risk of serious GI complications. Time until a serious GI complication was the dependent (outcome) variable used in the model. The independent (predictor) variables used in the model were age, sex, region, OA/RA, use of medications or clinical conditions that increase risk of GI complications, and use of gastroprotective medications. Members were censored if they did not have a serious GI complication by the end of their follow-up period. Collinearity among the covariates was assessed, and based on the results there were no significant correlations between the covariates. Time-dependent covariates were included in the models to assess the proportional hazards assumption.

	Generalized linear models (GLMs) using a gamma distribution as their probability distribution and log-link as their link function were used to compare the medical costs of serious GI complications and arthritis for the groups. The usefulness of the GLMs for modeling the medical costs is their ability to handle nonconstant variance (heteroscedasticity) and avoid having to retransform log-transformed costs (a common strategy for handling highly skewed medical cost data).

 The covariates used in the models were the same as those used in the proportional hazard model. Members with zero costs (related to serious GI complications or arthritis) during the follow-up period had $1 added to their costs. Models provided the adjusted estimates of mean cost for each of the groups. SAS version 9.2 (SAS Institute Inc, Cary, North Carolina) was used for all statistical analyses.

	A total of 172,118 members met the inclusion criteria, with 86,011 members in the restricted group and 86,107 in the unrestricted group. 

 shows the results of the application of the inclusion/exclusion criteria for the groups.

 shows the baseline characteristics for the groups. The majority of members in both groups were female, with approximately 33% of both groups being male. Large differences were seen in the geographic regions between the restricted and unrestricted groups. The majority of members in the restricted group were located in the South and Midwest regions (70% and 23%, respectively). A more even distribution was seen in the unrestricted group between the Midwest, South, and West regions (29%, 37%, and 28%, respectively). The mean age for each of the groups was similar (75 years), although found to be statistically significantly different (

 <0.01) due to the large sample sizes of the groups. The restricted and unrestricted groups showed an equal distribution of members with a diagnosis of OA (88%), RA (5.3%), and OA with RA (6%). The unrestricted group had more members with a prior history of a GI bleed (4.4% vs 2.1%). There were slightly more members with a history of H pylori infection in the restricted group (0.6% vs 0.4%). The unrestricted group had a slightly higher RxRisk-V score (5.9 vs 3.8).

	With respect to the utilization of gastroprotective medications, more members in the unrestricted group had a claim(s) for a proton pump inhibitor (PPI) (37% vs 32%; 

 <.01), whereas more members in the restricted group had a claim(s) for a histamine H2-receptor antagonist (H2RA) (13% vs 8%; 

 <.01). Compared with the unrestricted group, the restricted group had more claims for drugs recognized as gastrotoxic: nsNSAIDs (51% vs 40%) and oral glucocorticosteroids (33% vs 29%). More utilization of bisphosphonates, antiplatelets, anticoagulants, and heparin was seen in the unrestricted group. As expected, fewer members used celecoxib in the restricted group compared with the unrestricted group (8% vs 16%). The unrestricted group also utilized more nonbiologic and biologic disease-modifying antirheumatic drugs.

	The members in the restricted group had a slightly higher risk for a serious GI complication compared with the unrestricted group (5.4% vs 4.6%; 

]). This translates into a higher relative risk (1.17 [95% confidence interval (CI) 1.12-1.21]) in the restricted group, although the absolute difference in risk was small (0.8%). While we agree that the clinical relevance of this finding is debatable due to the small difference, the difference was statistically significant, and given the prevalence of arthritis in the Medicare population, it could translate into significant outcome and cost differences between the 2 groups, as examined in subsequent sections. While no differences existed in lower GI complications between the groups, the incidence of upper GI complications was higher in the restricted group (1.0% vs 0.7%; 

 <.01). Both groups were similar with respect to time until a serious GI complication occurred. The average number of days until a serious GI event occurred was 167 days from the index date.

	Cox proportional hazard models were also used to predict the risk of serious GI complications. After controlling for demographic and clinical characteristics and medications that impact the risk for a serious GI complication, the restricted group still showed a higher risk for a serious GI complication (hazard ratio 1.69; 95% CI 1.60-1.78), confirming the results of the unadjusted results. Covariates associated with an increased risk were age, prior GI bleed, 

 diagnosis, RxRisk score, and sucralfate use. Those associated with a decreased risk were OA (vs OA and RA combined), and PPI, H2RA, glucocorticosteroid, oral bisphosphonate, antineoplastic, antiplatelet, anticoagulant, heparin, and celecoxib use (

). A sensitivity analysis was performed by excluding members with a prior GI bleed. The results from the proportional hazards were similar (hazard ratio 1.63; 95% CI 1.54-1.73).

	A second Cox proportional hazards model evaluating any gastroprotective therapy use and level of exposure was applied. Gastroprotective therapy use was defined as a member using a PPI, H2RA, sucralfate, or misoprostol between the index date and the end of their follow-up period. Gastroprotective therapy use was higher in the unrestricted group than the restricted group (36% vs 28%; 

 <.0001). The level of exposure was the number of claims made by the member during the period. The results of the Cox regression model indicated that users of any versus no gastroprotective therapy (dichotomous predictor) had more than 2.2 times the risk for a serious GI complication of nonusers (

). However, the increase in the number of gastroprotective therapy refills was associated with a 20% decrease in the risk for a serious GI complication.

	The unadjusted mean serious GI complication medical costs PMPY were $146 and $29 for the restricted and unrestricted groups, respectively. The unadjusted arthritisrelated medical PMPY costs were higher for the restricted group ($1436 vs $764; 

 presents the results from the generalized linear regression models for serious GI complication—related and arthritis-related medical costs after controlling for age, sex, geographic region, underlying clinical conditions, and drug utilization. The adjusted mean PMPY costs for serious GI complications were $1559 vs $101 (

 <.0001) for the restricted and unrestricted groups, respectively, across the entire arthritis patient population. A second GLM model was used to compare the serious GI complication costs between the restricted and unrestricted groups for the subset of arthritis members with a serious GI complication only. This model showed that among those with a serious GI complication, the adjusted serious GI complication PMPY costs were 4 times higher in the restricted group compared with the unrestricted group ($4025 vs $1028, 

 <.0001). The model for arthritis-related medical PMPY costs showed that the adjusted costs for the restricted group were 80% higher than those for the unrestricted group.

	This study compared the incidence of serious GI complications and disease-related expenditures in 2 groups of members covered under a Medicare Advantage or Medicare Supplemental plan who were diagnosed with OA or RA. The groups differed in terms of whether or not the health plan that administered the drug coverage benefit had a restriction (ie, step therapy edit) in place for celecoxib. Overall the groups were similar in terms of age and sex. However, large differences were seen in the geographic regions represented in each of the groups. The analyses indicated that region did not have an impact on serious GI event rates, but there was an impact on the costs associated with treating serious GI complications and arthritis. This could be due to variations in treatment patterns and reimbursement rates between regions. Both groups were also similar with respect to the prevalence of OA, RA, or both conditions. As would be expected, the level of celecoxib utilization was much lower in the restricted group.

	Several other studies in Medicaid populations have demonstrated that prior authorization programs have a significant impact on reducing the utilization of COX-2 inhibitors with an expected increase in nsNSAID utilization.

 However, to our knowledge, no study has examined the association between restrictions and serious GI complications in a Medicare population. This association is important because for most health plans, age greater than 65 years is one of the risk factors indicating that the restriction should not be enforced (ie, members should be able to receive celecoxib). In spite of having this risk factor, only 8% of the restricted group received celecoxib, which is cause for concern and raises questions about factors influencing prescriber behavior. Future studies could focus on the outcomes observed in those individuals who are directly impacted by a formulary restriction policy to determine what alternative therapies, if any, are used if the decision is made to avoid or bypass the step therapy or prior authorization process.

	Another key finding that may have an impact on the incidence of GI complications was the difference in utilization of PPIs. The unrestricted group had higher levels of PPI use. It was confirmed that the restricted plan also has some formulary restrictions on PPIs, which might explain some of this difference. Other studies have shown that PPIs are used more often with COX-2 inhibitors than nsNSAIDs.

 Although this relationship was not directly analyzed in this study, it could be an underlying reason why PPI use was higher in the unrestricted group. Interestingly, the unrestricted plan members also utilized more of nonbiologic and biologic disease-modifying antirheumatic drugs, even though the incidence of RA in both groups was similar. The increased use and resulting effectiveness of these agents could be an underlying factor for the  lower arthritis-related medical costs in the unrestricted group.

	First, the study was observational, so it was unable to identify a causal relationship between formulary restriction policy and serious GI complications and treatment

	costs. Second, cumulative medication exposure was not accounted for when evaluating gastrotoxic/gastroprotective medication exposure. There could be significant differences between the therapeutic categories included in the analysis in terms of the cumulative amount of exposure required to impact the risk for a GI complication. For example, 30 days of continuous NSAID therapy may put a member at higher risk than 30 days of continuous antiplatelet therapy; however, based on presence of claims they are given equal weighting. Third, another clinical limitation is that other GI bleed risk factors were unaccounted for (eg, residing in a skilled nursing facility, where the likelihood of being sedentary may be higher and thus put the member at a higher risk for a serious GI complication). Fourth, the study was unable to account for confounding by health plan benefit structure. The restricted group consisted of members enrolled in a Medicare Advantage plan, whereas the nonrestricted group consisted of members enrolled in a Medicare Supplemental plan. The types of individuals enrolled in these types of plans could differ in terms of benefit structure. Lastly, members enrolled in a capitated (at-risk) benefit were not excluded. Providers commonly submit encounter claims only for their capitated members. The lack of reliable payment information for these members may have affected the cost analysis.

A formulary restriction policy in a Medicare population was associated with lower celecoxib utilization; however, higher gastrointestinal- and arthritis-related medical costs were observed.

Joseph C. Cappelleri, PhD

Articles

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

Impact of Celecoxib Restrictions in Medicare Beneficiaries With Arthritis