Healthcare costs and utilization associated with high-risk prescription opioid use: a retrospective cohort study

Data

We used QuintilesIMS patient-level administrative claims, which are derived from participating health plans across the United States, including commercial plans and those contracting with Medicare and Medicaid; however, the fee-for-service portion of the Medicare or Medicaid data was not present in the QuintilesIMS data. We included enrollees from the largest multi-region commercial plan. Other than demographic and enrollment information, the database also included diagnosis, procedure, medication, and cost information from inpatient, outpatient, and Emergency Department (ED) settings. Patient information was de-identified to comply with the Health Insurance Portability and Accountability Act.

Study design and subjects

This is a 2-year retrospective cohort study (2012–2013). We used 2012 data to construct baseline covariates and three indicators of high-risk opioid users; for each type of high-risk opioid use, we also assigned an individual to a four-level indicator representing the magnitude of their opioid use. We used 2013 data to define our outcomes. Among 1,267,605 enrollees with continuous medical and pharmacy enrollment in 2012 and 2013, we restricted to 893,835 (70.51%) individuals between 18 and 64 years of age and with known sex; among them, 191,405 (21.41%) enrollees had at least one prescription opioid claim in 2012.

Identification of high-risk opioid users

High-risk group membership was identified using 2012 claims. First, we used files provided by the Centers for Disease Control and Prevention to define the prescription opioids and benzodiazepines of interest [22]. These files contained information on strength per unit and a morphine milligram equivalent (MME) conversion factor at the National Drug Code (NDC) level. The MME provides a standardized measure across quantities and strength. Next, we used the 2012 medical and pharmacy claims to define three patient groups at elevated risks of adverse events from prescription opioids. We defined ‘chronic users’ as those consuming more than 100 MMEs per day for more than 90 consecutive days [23, 24]. Chronic opioid use is related to higher medical utilization and a greater likelihood of overdose death [20, 21, 25]. We defined ‘concomitant users’ as patients filling more than 30-days of concomitant opioids and benzodiazepines [23, 24]. Benzodiazepines are associated with approximately one-third of overdose deaths involving prescription opioids [26] and the odds of dying by overdose is four-fold higher among veterans with current benzodiazepine prescriptions [27]. Finally, we defined ‘opioid disorders’ as patients with one or more diagnosis codes representing opioid use disorders (Appendix 1) [9, 12, 14–16, 28, 29].

We used two approaches to examine a dose–response association between high-risk opioid use and healthcare utilization. First, we categorized high-risk users into tertiles based on their magnitude of high-risk opioid use (level 1–3); we included non-high-risk opioid users as a reference group (level 0). For ‘chronic users’, we determined the number of days an individual was considered a chronic user to define the magnitude of high-risk opioid use; for ‘concomitant users’, we used the number of days with both drugs on hand; for those with opioid disorders, we adopted the number of months diagnosed with opioid use disorders. Second, we counted the number of high-risk groups each enrollee belonged to (0–3).

Outcome variables

Outcomes were derived from 2013 claims data. We calculated total, medical, and pharmacy costs (we also calculated pharmacy costs associated with prescription opioids only) and four binary measures of utilization (being among the top 5% total cost users, being among the top 5% pharmacy cost users, having any hospitalization, and having any ED visit) as our outcome variables. The annual costs were derived from claims and represents the sum of allowed amounts, reflecting what the insurance plan paid for services.

Control variables

We derived control variables from the Johns Hopkins Adjusted Clinical Group (ACG v11.0) Risk Adjustment System, using both medical and pharmacy claims as inputs. The ACG system is a widely used morbidity measure [30–37]; it has been validated against costs [33, 35, 36], utilization [32, 38], and death [39]. The ACG system assigns all ICD codes to one of 32 Aggregated Diagnostic Clusters (ADGs). Each ADG is a morbidity group consisting of clinically homogeneous diagnosis codes with a similar expected need for medical resources. Based on their ADGs, age, and sex, individuals are assigned to one of 93 discrete ACG categories. The ACG system also assigns each NDC code to one of 67 Rx-defined morbidity groups (RxMG) based on the combination of active ingredient and route of administration [30, 40]. The ACG system also calculates the number of chronic conditions/active ingredients using the diagnosis codes/NDC an individual encountered. In addition, the ACG system generates a risk score for concurrent total costs, including independent variables such as diagnosis-based overall disease burden, high-impact chronic conditions, diagnoses representing a high likelihood of hospitalization, and acute conditions.

Statistical analysis

We first described the characteristics of all enrollees, enrollees with any opioid use, and high-risk opioid users. Then, we constructed statistical models to evaluate the impact of being a high-risk user and the magnitude of high-risk opioid use on the prospective healthcare costs and utilization.

We used propensity score weighting to control for baseline differences in patient characteristics between high-risk and non-high-risk opioid users. Logistic regression was applied to derive a propensity score of becoming a high-risk user based on patient sex, four age categories (18–34, 35–44, 45–54, and 55–64), 32 diagnosis-based ADGs, 67 medication-based RxMGs, number of chronic conditions, number of active ingredients, and a concurrent risk score. An individual had three propensity score weights corresponding to each measure of high-risk use. We chose propensity score weighting because we could include all high-risk users, as might not occur with matching, and we wanted one interpretable overall effect, which might be impossible with stratification. We derived an average treatment effect of the treated weighting because we could estimate the average effect of treatment on the treated subjects, thus making comparisons between the actual outcomes of high-risk users and the expected outcomes under the counterfactual if they had not been high-risk users. This is especially useful when the study sample differs systematically from the overall population [41].

We examined the performance of our propensity score weighting based on comparisons of (1) high-risk users versus non-high-risk users with an opioid prescription and (2) high-risk users versus enrollees without an opioid prescription. We restricted our analyses to the first comparison because, after applying propensity score weighting, the average of the absolute standardized differences of all variables in the models reduced to less than 0.1; such differences were larger for the second comparison.

We used a generalized linear model (GLM) with a log-link function and gamma distribution to model costs, given the non-negative and positively skewed distribution of costs as well as a much higher proportion of people with very high costs [42–44]. We added $1 to all costs so that non-users could be included in the model. The log-link function provides an estimate of the proportional change in mean costs. We applied logistic regression for binary utilization indicators and included the same set of covariates from the propensity score model to control for residual confounding. We constructed both a crude and adjusted model with covariates and weights to explore the relationship between being high-risk opioid users and each outcome. For analyses of the dose–response association, we constructed a crude mode and an adjusted model with covariates since propensity scores were usually generated for binary outcomes.

As a sensitivity analysis, we also performed multivariate linear regression to test the robustness of our findings, since it is widely used to analyze costs, delivers intuitive results, and often performs similarly as other statistical models (such as a two-part lognormal model) with a sufficiently large sample size [8, 9, 33, 45]. These results yielded substantive similar findings and are reported in Appendix 2, but not discussed further herein.

Characteristics of the eligible participants

Of the three high-risk groups examined, individuals with opioid use disorders had the lowest total ($25,000+) and pharmacy (~$4400) costs concurrently and prospectively; chronic users had the highest total (~$30,000) and pharmacy costs (~$11,000), but the lowest medical costs (~$20,000). Notably, more than one-third of chronic users’ total costs were due to pharmacy costs, while pharmacy costs of those with opioid use disorders accounted for less than one-fifth of their total costs. In addition, the amount of pharmacy costs accounted for by prescription opioids varied across the three groups, with chronic users having the greatest proportion (more than 50%), compared to individuals with opioid use disorders (45%) and concomitant users (~20%).

Chronic users had lower rates of concurrent and prospective hospital (~15%) and ED (~28%) utilization, while those with opioid disorders had the highest rates of such utilization, although they declined over time. For example, among those with opioid disorders, the rates of hospitalization decreased from 29% concurrently to 18% prospectively, and the rates of ED visits decreased from 41% concurrently to 34% prospectively. Among the three high-risk groups, concomitant users had the highest morbidity across five ACG-based indicators, both concurrently and prospectively, while patients with opioid use disorders had the lowest morbidity prospectively.

The overlap across these three high-risk groups was low; the highest was 44.13% for chronic users being also concomitant users while the lowest was 5.04% for concomitant users being patients with opioid disorders. In the concurrent year, chronic users had 116 days being considered as chronic users, concomitant users had 150 days being concomitant users, and patients with opioid disorders had 3.7 months with a diagnosis of opioid disorders. On average, chronic users were included in 1.63 high-risk groups, concomitant users in 1.18, and patients with opioid disorders in 1.55.

Prospective healthcare costs associated with high-risk use

Total and pharmacy costs, but not medical costs, increased as the magnitude of high-risk opioid use increased. After adjustment, each additional level in magnitude showed a statistically significant association with 19%/79% increases in total/pharmacy costs among chronic users, 4%/14% increases among concomitant users, and 19%/27% increases among patients with opioid use disorders. Costs were also greater among individuals with multiple high-risk group membership; each additional membership was statistically significantly associated with a 17% increase in total costs, a 6% increase in medical costs, and a 64% increase in pharmacy costs.

Prospective medical utilization associated with high-risk use