We hypothesized that among older Medicare beneficiaries diagnosed with spinal pain, subjects who receive chiropractic care have a lower risk of filling a prescription for an opioid analgesic as compared to beneficiaries who do not receive chiropractic care. To test this hypothesis, we employed a retrospective cohort design to analyze Medicare administrative data collected from 2012 through 2016. Medicare is a health insurance program administered by The Centers for Medicare and Medicaid Services (CMS) of the US Department of Health and Human Services. US residents aged 65 or older are eligible for enrollment in Medicare. CMS aggregates health claims and associated administrative data and makes research datasets available to qualified research scientists. Medicare claims data include those for inpatient services (Part A), outpatient services (Part B), managed care plans (Part C), and prescription medications (Part D). The data source for this project was Medicare Parts B and D. The study was conducted according to the terms of a data use agreement between the principal investigator and CMS.
The study population was comprised of older Medicare beneficiaries, alive as of 12/31/16, living in any of the 50 US states and the District of Columbia, aged 65–99 years, continuously enrolled throughout the study period in Medicare Parts B and D. We excluded beneficiaries enrolled in Medicare Part C. The study sample was restricted to patients with office visits to a primary care physician and/or Doctor of Chiropractic for a primary diagnosis of spinal pain. A complete list of diagnosis codes used to identify and categorize spinal pain disorders may be viewed in Additional file 1: Appendix C. Diagnostic codes for non-allopathic lesions were excluded to help assure congruence between chiropractic and medical patient populations. We only included Part B claims with dates of service within calendar years 2012–2016, with payment amount greater than zero for a primary diagnosis of spine-related disorder. To enhance the validity of the recorded diagnosis, we restricted the study sample to beneficiaries with at least 2 such visits between 7 and 90 days apart. Thus, the study population included subjects with multiple office visits for spinal pain. To exclude patients with pathological pain (who would be likely to receive opioids but less likely to receive chiropractic care), we excluded beneficiaries with a primary diagnosis of cancer or receiving hospice care at any time during the study period of 2012–2016. We restricted place of service to office visits, and restricted provider specialty to family medicine, general practice, internal medicine, or chiropractic. For each subject, the first date associated with diagnosis of a spinal pain disorder was designated as the index date. The earliest possible index date was Jan 1, 2013. 2012 data were used to capture patient inclusion and exclusion criteria, and to calculate Charlson comorbidity scores. The latest possible index date was Dec 31, 2015, thus allowing for an observation period of 365 days for all subjects. Only the first chiropractic visit was used as a cohort inclusion criterion for Recipients; any subsequent visits did not change the subject’s primary cohort assignment or index date. We included only those subjects with Part D coverage at index date plus 365 days. We excluded all subjects with an opioid prescription fill that occurred before the index date. In our analyses of claims data, in accordance with CMS rules for analysis of health claims, cells with n < 11 were suppressed to prevent disclosure of protected health information.
Among those included in the study population we identified two cohorts of subjects: (1) Recipients of chiropractic services (Recipients) received both primary care and chiropractic care within 120 days of cohort inclusion. (2) Non-recipients received primary care but did not receive chiropractic care at any time during the study period. For the Recipients cohort, we accounted for immortal time bias by using first chiropractic visit only as a cohort inclusion criterion for Recipients; thus, subjects with an opioid prescription fill after their index date but before their first chiropractic visit were included in the Recipients cohort. We categorized spinal pain diagnoses as 1, 2, or 3 as indicators of progressively more unfavorable prognosis. [Additional file 1: Appendix C] We stratified the recipient population into three groups: (1) Early—patients whose first chiropractic encounter occurred within 30 days of the index date, (2) Delayed—first chiropractic encounter occurring within 31–90 days of the index date, and (3) Late—first chiropractic encounter occurring within 91–120 days of the index date. Thus, for purposes of this study, the terms Early, Delayed, and Late refer specifically to the timing of first chiropractic visit among Recipients of chiropractic care.
The principal outcome measure was incidence of opioid prescription fill, as recorded in Part D data. Prescription opioids were identified by the Centers for Disease Control list of opioid-containing analgesic medications and associated National Drug Codes [13]. Following aggregation of claims data and assembly of cohorts, we generated descriptive statistics by cohort on subject demographics, health status, category of spine pain diagnosis, and for Recipients, the timing of first chiropractic visit. Spine pain diagnoses were categorized as 1, 2, or 3 to broadly indicate progressively higher risk of poor outcomes.
We employed Cox proportional hazards modeling to evaluate risk of opioid prescription fill for up to 365 days following index. To assess the impact of receiving chiropractic care early in an episode of care, we sub-analyzed for risk of opioid prescription fill in the Early, Delayed, and Late groups of Recipients. We controlled for patient characteristics, including age, sex, race/ethnicity, state of residence, spinal pain diagnosis category, and health status at baseline as measured by Charlson comorbidity score. To reduce the effect of selection bias, we controlled for subjects’ propensity to utilize chiropractic care, using inverse probability of treatment weighting [14]. To help achieve statistical modeling that would be consistent across all measurements, both national and state-by state, propensity scores were binned into quintiles for use in the Cox proportional hazards models. The adjusted hazard ratios were estimated by including the propensity score quintiles within models as a single (categorical) variable. We did not retain data showing distribution of covariates between recipients and non-recipients before and after adjusting for propensity scores. Observational research on Medicare claims data affords a limited selection of variables for patient characteristics that can be used for propensity scoring, and as the population ages the demographics of the Medicare beneficiary population (and specifically among beneficiaries who use chiropractic care) have been slow to change.
We performed adjusted time-to-event analyses, generating hazard ratios to compare Recipients and Non-recipients regarding the risk of filling an opioid prescription. To analyze for geographic variations in outcomes, we generated hazard ratios by state. All statistical analyses were performed using SAS 9.4 (SAS Institute, Cary, NC).