This study was approved, with requirements for patient consent waived, by the California Committee for the Protection of Human Subjects and the University of California, Davis Institutional Review Board. The study protocol was pre-registered at AsPredicted.org (#133335).

Two data sources were utilized to conduct complementary and independent analyses. The first data source comprised controlled substance prescription records from California’s PDMP dispensed between January 1, 2016, and December 31, 2018. California’s PDMP contains statewide data on all outpatient Schedule II-IV prescriptions dispensed by California outpatient pharmacies. Prescription records included patient, prescriber, and pharmacy information. Each prescription record also included date dispensed, National Drug Codes, quantity, strength per unit, and days’ supply. Prescription records were linked to California death certificate data using patient date of birth, sex, name, and address. Linked data were then de-identified for analysis. The second data source was the OLDW with claims records from January 1, 2016, through December 31, 2021. The OLDW contains de-identified retrospective administrative data, including medical and pharmacy claims (with associated diagnosis codes) and eligibility information for commercial and Medicare Advantage enrollees. The database contains longitudinal health information on patients representing a mix of ages and US geographical regions.¹⁴

These two data sources have complementary strengths. The California PDMP captures nearly all outpatient controlled substance prescriptions in California statewide (12% the US population) regardless of insurance status. The OLDW is a national sample of patients with commercial insurance or Medicare Advantage with linked clinical claims in a sample representing over 20% of the commercial and Medicare Advantage markets nationally.

Separate cohorts were derived from each dataset (Fig. 1A/B), and included patients aged ≥ 18 years with at least one opioid prescription fill between July 1, 2016, and December 1, 2018 for the PDMP cohort, and between July 1, 2016, and December 1, 2021, for the OLDW cohort. Patients were eligible to enter the cohort if, on the first day of any given opioid fill, they had opioid coverage for at least 144 days (80%) of the prior 180 days (baseline period). Patients were excluded if they had any benzodiazepine or buprenorphine prescription, indicated for the treatment of opioid use disorder, during this 180-day baseline period. Patients with buprenorphine coverage during baseline were excluded to ensure that the sample comprised patients receiving opioids for chronic pain rather than opioid use disorder, though we did not exclude patients who received buprenorphine during the follow-up period.

The OLDW cohort dataset had additional cohort eligibility criteria. Only patients continuously enrolled in medical, pharmacy, and mental health coverage for 365 days prior to their index opioid fill date and for at least 30 days after their index fill date were eligible for inclusion. Additionally, we excluded patients with solid organ cancer diagnoses (International Classification of Diseases, Clinical Modification, Tenth Revision [ICD-10-CM] codes C00–C41, C43, C45–C96, D00–D09, D47.29, D46.9, E31.22, E31.23), any claims for hospice/palliative care, or ≥ 90 days of nursing, inpatient, or skilled nursing care during the 365-day period prior to the index opioid fill date.

For both datasets, upon cohort entry, patients were followed for the duration of the study period (PDMP study end: December 31, 2018; OLDW study end: December 31, 2021) until they met the definition for co-prescription or were censored due to patient death. Exclusive to the OLDW cohort, additional censoring events included health plan disenrollment, new solid tumor cancer diagnoses, initiation of hospice or palliative care, or receipt of ≥ 90 days of skilled nursing care. Follow-up was categorized into discrete 30-day (patient-month) observation periods starting on the date of each patient’s cohort entry. An example patient timeline is presented in Fig. 2.

The study outcome was incident benzodiazepine and opioid co-prescription, defined as the occurrence of ≥ 20-day supply of overlapping opioid and benzodiazepine prescriptions during any patient-month in the study period. As this definition is a shorter assessment window than the 90-day periods used in other studies,^15,16 we also conducted sensitivity analyses using a ≥ 60-day supply of overlapping opioid and benzodiazepine coverage during any 3 patient-months (90 days total) in the study period.

Time invariant independent variables common to both cohorts included patient demographics (age, sex), calendar year of cohort entry, and opioid prescribing patterns during the baseline period (180-day average of daily milligram morphine equivalents [MME], predominant opioid type, number of unique opioid prescribers).

Time-varying independent variables common to both cohorts included prior 180-day opioid prescribing characteristics (any long-acting formulation, any overlap of opioid prescriptions), prior 180-day additional controlled substance prescribing (any opioid and non-benzodiazepine sedative [including controlled sedatives such as zolpidem, zaleplon, eszopiclone, and carisoprodol in both cohorts, and also gabapentinoids in the OLDW cohort] prescription overlaps, any active psychostimulant prescription), and buprenorphine, indicated for the treatment of opioid use disorder, prescriptions during follow-up (none, initiation during follow-up and buprenorphine prescription in the prior 30 days, initiation during follow-up but no active prescription in the prior 30 days).

Prior work identified that opioid dose changes may be associated with adverse mental health outcomes^10,17 and therefore potential subsequent benzodiazepine use. To examine whether patients’ opioid dose trajectory was associated with incident co-prescription, we included changes in opioid dose, MME trajectory, in the preceding 60-day (“short term”) or 180-day (“long-term”) prescribing period, long-term opioid dose variability,¹⁸ and opioid discontinuation status. See Appendix 1 of the supplement for details on opioid trajectory variable construction and definitions.

Time invariant independent variables exclusive to the OLDW cohort included commercial vs. Medicare Advantage patient insurance status and area-level (Census block) demographics (rurality, median education level, median household income), patient history of anxiety, depression, psychoses, alcohol and substance use disorders using ICD-10-CM diagnosis codes, Elixhauser comorbidity indicators,¹⁹ and any overdose resulting in an emergency department or hospital visit occurring in the 365-day period prior to study entry.

Time-varying independent variables exclusive to the OLDW cohort included past 180-day prescriptions for serotonergic anxiolytics, including selective serotonin reuptake inhibitors, serotonin and norepinephrine reuptake inhibitors, and buspirone.

For further details on variable construction see Appendix 2 of the supplement.

We constructed separate time-varying Cox proportional hazards regression models for each cohort to examine the relationships between incident co-prescription and the independent variables, both time invariant and time varying. Subsequent 30-day prescribing periods (patient-months) were the units of analysis. Overall model fit was assessed using the Akaike information criterion.

Unlike PDMP models, OLDW models also included the OLDW-exclusive independent variables listed above but otherwise followed the same modeling process as that for the PDMP cohort. To account for the effects of non-benzodiazepine anxiety pharmacotherapy, OLDW models included an interaction term for patients with a diagnosis of depression or anxiety at baseline who also filled prescriptions for serotonergic anxiolytics during follow-up.

Data preparation and analyses were conducted using SAS 9.4.

Of the 617,946 patients in the PDMP cohort, 61,168 (9.9%) developed incident co-prescription during the follow-up period with an incidence rate of 4.6 per 1000 patient-months. In the OLDW cohort of 223,885 patients, 20,425 (9.1%) developed co-prescription during the follow-up period with an incidence rate of 3.9 per 1000 patient-months.

Patients with co-prescription were more likely to be female and have higher average MME during baseline (Table 1). Patients with co-prescription were also more likely to be prescribed long-acting opioids, stimulants, have overlapping opioid prescriptions, and have overlapping opioid and non-benzodiazepine sedative prescriptions in the previous 180-day period. Patients with co-prescription were also more likely to experience dose increases over prior short-term (60-day) and long-term (180-day) periods but were less likely to have stable opioid dose trajectories. Patients in the OLDW cohort with co-prescription were more likely to have prior diagnoses for anxiety, depression, substance use disorder, and psychosis; they were also more likely to receive a prescription for serotonergic anxiolytics in the previous 180-day period. Descriptive results for Elixhauser comorbidities are presented in the supplement (Appendix 3).

Adjusted analysis of the PDMP cohort revealed that male sex (adjusted hazard ratio [aHR]: 0.70, 95% CI: 0.69–0.71) and older age (relative to age 18–39 years) were associated with reduced risk for co-prescription (Table 2). A higher average baseline daily opioid dose, especially in the highest dose group of > 150 MME (aHR: 1.74, 95% CI: 1.67–1.81), and having 4 or more opioid prescribers (aHR: 1.09, 95% CI: 1.06–1.12) during the baseline period were associated with increased risk for co-prescription.

Long-term (180-day) decreases in 30-day opioid dose (aHR: 0.81, 95% CI: 0.78–0.85) and discontinuation of opioids, particularly sustained long-term discontinuation (aHR: 0.28, 95% CI: 0.25–0.30), were associated with decreased risk for co-prescription. Alternatively, overlapping opioid fills (aHR: 1.33, 95% CI: 1.30–1.36), highly variable opioid dosing (aHR: 1.15, 95% CI: 1.12–1.17), and both short-term (60-day) (aHR: 1.20, 95% CI: 1.17–1.23) and long-term (aHR: 1.13, 95% CI: 1.11–1.16) increases in opioid dose were associated with increased risk for co-prescription. Receipt of any buprenorphine prescription during follow-up (as compared to non-initiation) was associated with increased risk for co-prescription, though the strength of association differed slightly depending on whether patients had (aHR: 1.68, 95% CI: 1.49–1.89) or did not have (aHR: 1.35, 95% CI: 1.18–1.56) an active buprenorphine prescription during the previous 30 days. Overlapping fills of opioids with (a) non-benzodiazepine sedative prescriptions (aHR: 1.69, 95% CI: 1.65–1.72) and (b) prescription psychostimulants (aHR: 1.45, 95% CI: 1.40–1.50) were also associated with increased risk for co-prescription.

Adjusted analysis of the OLDW cohort (Table 3) revealed mostly similar associations but with some differences; patients with 4 or more prescribers (aHR: 0.86, 95% CI: 0.81–0.92) and patients with long-acting opioids in the previous 180 days (aHR: 0.90, 95% CI: 0.86–0.95) had reduced risk for co-prescription (versus increased risk in the PDMP cohort). Unlike in the PDMP cohort, long-term high variability in dose (aHR: 0.98, 95% CI: 0.94–1.03) and increasing long-term dose trajectory (aHR: 0.99, 95% CI: 0.95–1.03) were not associated with co-prescription.

The OLDW cohort model also assessed associations between patients’ clinical characteristics and co-prescription. A history of drug overdose at baseline was associated with increased risk for co-prescription (aHR: 1.25, 95% CI: 1.10–1.42). Additionally, baseline anxiety (aHR: 2.16, 95% CI: 2.06–2.27) and depression (aHR: 1.16, 95% CI: 1.09–1.23) diagnoses without prior fills for serotonergic anxiolytics were risk factors for co-prescription. However, patients with anxiety (aHR: 0.63, 95% CI: 0.59–0.67) or depression (aHR: 0.85, 95% CI: 0.79–0.92) had reduced risk for co-prescription if they also had prior fills for serotonergic anxiolytics.

Adjusted hazard ratios for independent variables not presented in Tables 2 and 3 of the manuscript can be found in the supplement (PDMP: Appendix 4; OLDW: Appendix 5).

Sensitivity analyses with the longer-term definition for co-prescription, a 60-day overlap of opioids and benzodiazepines in a 90-day period, revealed nearly identical findings to our primary analyses. Incidence rates were approximately 1.9 per 1000 person-months in both cohorts; descriptive findings are presented in Appendix 3 and 6. Differences between model estimates for the two outcomes were unremarkable in both cohorts (presented in Appendix 4 and 5).