Comparative Risk of Opioid Overdose With Concomitant Use of Prescription Opioids and Skeletal Muscle Relaxants

Data Sources

We used data spanning January 1, 2000, to December 31, 2014, from Medicaid Analytic eXtract (MAX); January 1, 2012, to December 31, 2017, from a Medicare fee-for-service cohort of beneficiaries with diabetes, heart failure, and/or stroke; January 1, 2004, to December 31, 2019, from Optum© Clinformatics Data Mart; and January 1, 2003, to December 31, 2018, from IBM MarketScan Research Database. The databases contain deidentified and longitudinal administrative pharmacy and medical claims for commercially insured individuals (Optum© Clinformatics Data Mart and IBM MarketScan Research Database), publicly insured individuals younger than 65 years (Medicaid), and beneficiaries aged 65 years and older or younger than 65 years with disabilities and/or end-stage renal disease (Medicare) across the United States. Patient-level ZIP codes in Medicaid were linked to the 2017 American Community Survey from the US Census Bureau to assess measures of socioeconomic status.⁷ ZIP codes were also linked to the 2013 rural-urban continuum codes from the US Department of Agriculture Economic Research Service to distinguish between metropolitan and nonmetropolitan areas.⁸

Study Population

Eligible individuals were identified based on their first dispensing for a skeletal muscle relaxant at age 18 years or older (cohort entry date) immediately after at least 180 days of continuous health plan enrollment with no previous skeletal muscle relaxant dispensing and 90 days of continuous prescription opioid use (Figure 1). The skeletal muscle relaxants under study included baclofen, carisoprodol, cyclobenzaprine, metaxalone, methocarbamol, tizanidine, and chlorzoxazone or orphenadrine (evaluated as a single group). Continuous opioid use was defined by prescription fills for opioids with days’ supply that covered the 90 days immediately before and on the cohort entry date. Grace periods of up to 14 days were incorporated between opioid prescriptions to account for inaccuracies in the days’ supply estimation. Patients who experienced opioid overdose before or on the cohort entry date were excluded.

Figure 1 Active Comparator Cohort Study Design to Quantify the Comparative Risk of Opioid Overdose in Patients Who Initiated Skeletal Muscle Relaxants While on Long-Term Opioid Therapy

SMR = skeletal muscle relaxant.

Outcome Assessment and Follow-up

Opioid overdose was defined using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) and ICD, Tenth Revision, CM (ICD-10-CM) codes for an opioid overdose episode resulting in hospitalization or an emergency department visit (eTable 1, links.lww.com/WNL/C171). The ICD-9-CM codes used to define opioid overdose, together with codes for poisoning by heroin, have a specificity of 99.9%.⁹ The ICD-10-CM algorithm for opioid overdose unrelated to heroin poisoning has a positive predictive value of 79.4%.^9,10

We used sepsis as a negative control outcome to evaluate potential confounding by illicit opioid use, which is undercaptured in claims data. Patients who use illicit opioids may be more likely to be prescribed certain opioid-skeletal muscle relaxant therapies, and intravenous drug use associated with illicit opioids may lead to sepsis. Different rates of sepsis across the 7 treatment strategies could suggest that illicit opioid use may have biased the findings in the primary outcome analyses.

We performed a 30-day intention-to-treat analysis to evaluate the acute risk of opioid overdose after initiation of concomitant therapy and to avoid bias from potential informative censoring. Follow-up began the day after cohort entry until outcome occurrence, end of continuous enrollment, death, nursing home admission, or end of the 30-day period immediately after the cohort entry date. In the as-treated analysis, patients could contribute follow-up time beyond 30 days but were, in addition, censored at prescription opioid discontinuation, skeletal muscle relaxant discontinuation, initiation of a skeletal muscle relaxant from another group, or end of the 365-day period immediately after the cohort entry date. We allowed grace periods of up to 14 days between prescriptions for both opioids and skeletal muscle relaxants, and the medication discontinuation date was defined as 14 days after the final prescription’s days’ supply.

Covariates

We assessed demographic information, comorbidities, pain conditions, other prescription fills, previous prescription opioid utilization, and healthcare utilization over the 180 days immediately before and on the cohort entry date. We also used patient-level ZIP codes in Medicaid to capture socioeconomic status variables and type of metropolitan area.

Statistical Analyses

We used matching weights, an extension of propensity score–based weighting approaches, to adjust for covariates in this multigroup setting. Within each database, a multinomial logistic regression model was used to estimate the probability of treatment for each prescription opioid-skeletal muscle relaxant therapy as a function of the covariates. Each patient was subsequently assigned a matching weight, which is a ratio of the smallest conditional probability of treatment (of all probabilities estimated) and the conditional probability of receiving the treatment actually received.¹¹ The estimand of this weighting method (i.e., the subgroup of patients that would theoretically be eligible to receive any of the combination therapies under study) is asymptotically equivalent to that of 1:1 propensity score matching across the treatment groups provided that common support holds.¹¹

Within each database, we fit a Cox proportional hazards model weighted by matching weights to estimate the hazard ratios (HRs) and bootstrapped 95% CIs of opioid overdose for each treatment relative to a common reference group (i.e., cyclobenzaprine). To estimate a common HR of opioid overdose for each treatment group across all databases, we fit a Cox proportional hazards model stratified by database in the pooled data and weighted by matching weights. This approach allows the baseline hazard of opioid overdose to vary by database and assumes that the HR is constant across the databases. The p values were adjusted for multiple testing using the Bonferroni correction, which controls the probability of falsely rejecting 1 or more null hypotheses.¹² We calculated adjusted p values by multiplying each nominal p value by the number of comparisons (6).

Secondary Analyses

We conducted several additional analyses. Although in the same drug class, skeletal muscle relaxants have different indications for treatment. Baclofen is approved for spasticity (e.g., resulting from spinal cord diseases), whereas the other skeletal muscle relaxants are approved for muscle spasm (e.g., acute musculoskeletal conditions), and tizanidine is approved for both indications in the United States.¹³ We also observed that tizanidine initiators tended to be more similar to the baclofen initiators relative to the initiators of other skeletal muscle relaxants, and initiators of other skeletal muscle relaxants were quite similar to each other. We therefore conducted 2 subgroup analyses comparing: (1) baclofen and tizanidine and (2) carisoprodol, cyclobenzaprine, chlorzoxazone/orphenadrine, metaxalone, and methocarbamol. From the seven-group comparison, we also conducted a subgroup analysis restricted to patients without previous substance abuse as recorded by ICD codes for opioid dependence, opioid abuse, alcohol abuse, tobacco use, and other drug abuse in the baseline period to reduce potential confounding by unmeasured illicit opioid use. We separately stratified our analyses by baseline benzodiazepine use (binary); the median morphine milligram equivalents (MMEs) (above or equal to the median [high] or below the median [low] MMEs); oxycodone dispensing (binary); and hydrocodone dispensing (binary) on the cohort entry date to identify subgroups with an increased risk of opioid overdose. In Medicaid, we adjusted for socioeconomic status measures and rural-urban area to account for residual confounding that may not be captured by other demographic variables. We fit multinomial logistic regression models for each analysis to reestimate matching weights for each subgroup.

Standard Protocol Approvals, Registrations, and Patient Consents

This study was approved by the Institutional Review Board at Brigham and Women’s Hospital.

Data Availability

The data used in this study cannot be shared because of data use agreements for each of the 4 databases.