Data Source
The MHRN is a consortium of research centers located within 13 large integrated health care systems, many of which also have affiliated health insurance plans and serve over 12.5 million individuals across 15 states with diverse populations in the United States. All MHRN sites maintain a Virtual Data Warehouse consisting of electronic health record (EHR) and insurance claim data for all enrolled members or patients. Data on encounters, pharmacy fills, diagnoses, laboratory tests and demographics are organized using standardized definitions across sites and are quality checked locally (26).
The current study involved 10 MHRN systems. These sites were 6 Kaiser Permanente sites (Georgia, Washington, Northwest, Hawaii, Northern California, Southern California), Henry Ford Health System, Essentia Health, Baylor Scott and White Healthcare and Health Partners. Institutional Review Boards at each site approved the study protocol for this project.
Study Population
Individuals were included if they met the following criteria: adults aged 18–70 years (as of January 1, 2016) with a diagnosis of MDD (ICD-9 296.2-296.39/ICD-10 F32-F33.9), BD (ICD-9 296.0x, 296.1x, 29.40-296.89/ICD-10 F30-F31.9) or schizophrenia including schizoaffective disorder (ICD-9 295.x/ICD-10 F20.x, F25.x) documented at least two times by mental healthcare provider in 2015 or 2016 (at least 1 diagnosis had to occur in 2015). Patients who had diagnoses in more than 1 of these categories were categorized hierarchically: schizophrenia>BD>MDD. For example a patient with schizophrenia and MDD would be classified in the schizophrenia group and a patient with only MDD would be classified in the MDD group. Eligible individuals had to have continuous health plan membership throughout 2015 and 2016 (but could have a gap in enrollment records of ≤30 days, as administrative gaps can occur as a result of delays in membership data processing and thus are not indicative of membership interruptions/disenrollment). Individuals with any cancer or metastatic cancer diagnoses (ICD-9 140-165, 170-172, 174-176, 179-199, 200-208, 238.6/ ICD-10 C00-26.9, C30.x, C37-C41.9, C43.x, C45-C45.7, C45.9, C46-C58,, C60-C76.8, C7A.x, C7B,x, C80.x, C81-C85.99, C86.x, C88.x, C90-C96.9, D03.x, D45, D47.Z9,) during this same time period were excluded.
Controls were identified using the same criteria as described above except that they had no documented mental illness diagnoses during 2015 or 2016. Matching was done separately for each group (e.g., schizophrenia controls were selected and removed from the pool of controls, then BD controls, followed by MDD controls). Controls for each group were matched on age (in 4-year bands), sex and Medicare status using stratified random sampling. Matching cases to controls was 1:2 for schizophrenia diagnosis and 1:1 each for BP and MDD diagnoses. These ratios were based on what numbers were required to find an adequate number of controls for each group. Only approximately half of the available MDD cases were selected (randomly) because there were not a sufficient number of controls available.
Measures
Non-cancer chronic pain diagnoses documented on at least 2 dates in 2016 were extracted for the matched samples. The chronic pain conditions extracted included: back pain, neck pain, limb/extremity pain, arthritis, fibromyalgia/widespread muscle pain, headache, orofacial/ear/temporomandibular pain, abdominal/bowel pain, chest pain, urogenital/pelvic/menstrual pain, fractures/contusions/sprains/strains and other painful conditions [which included sickle cell disease, complex regional pain syndrome, systemic lupus erythematosus, acquired deformities (excluding spinal disorders), spinal cord injury and neuropathic pain]. The list of ICD codes used for identifying pain conditions are available online (https://github.com/MHResearchNetwork/MHRN-Central/blob/master/WP_MHRN_SMI_painOpioids.zip).
Prescription opioid medication dispensings were also extracted for the matched samples. We were specifically interested in chronic opioid use, defined by prescriptions dispensed that covered at least 70 days in any 90-day period or 6+ dispensings in 2016. This definition was based on prior studies conducted at one of the MHRN sites (27, 28). The list of NDC codes used for identifying opioid medication dispensings are also available online (https://github.com/MHResearchNetwork/MHRN-Central/blob/master/WP_MHRN_SMI_painOpioids.zip).
We also examined sociodemographic (age, sex, race/ethnicity, neighborhood socioeconomic status) and clinical characteristics of the study population using data from 2016. Overall medical comorbidity burden was calculated using the Charlson Comorbidity Index Score (CCIS). This score contains 19 categories of comorbidity, with each category weighted based on the adjusted risk of 1-year post-discharge mortality. The overall comorbidity score reflects the cumulative increased likelihood of 1-year post-discharge mortality; the higher the score, the more severe the burden of comorbidity (29). Total health care utilization (hospitalizations, ED visits and other in-person outpatient encounters) was based on summarized data from the last 6 months of 2016. Because we were interested in counting utilization days, multiple outpatient encounters documented on the same day counted as one encounter. Preliminary data comparisons across sites were made by the study team to investigate site variation and to ensure accuracy of the data before creating aggregated estimates. This preliminary comparison found very little site variation, supporting the stability of the aggregated estimates.
Analyses
The primary goals of our analyses were to examine whether having a diagnosis of MDD, BD or schizophrenia/schizoaffective disorder was associated with receipt of a chronic pain diagnosis and then subsequent chronic opioid prescription dispenses. For initial bivariate models, we used t-tests for continuous variables and Pearson χ2-tests for categorical data. Multivariate analyses were conducted to evaluate (1) the odds of receiving a chronic pain-related diagnosis and (2) the odds of receiving opioids, by separate mental illness diagnosis category compared with matched controls, controlling for age, sex, Medicare status, race/ethnicity, income, medical comorbidities, healthcare utilization and chronic pain diagnoses. Results of the models were reported as adjusted odds ratios (ORs) with 95% confidence intervals (CIs).