The current study used a longitudinal cohort design where active duty service members with chronic pain were identified through their healthcare records in the military data repository and other DoD sources compiled by SUPIC.25 MHS data up to the end of 2015 were included. Quasi-experimental methods were used to determine outcomes for identified service members who enrolled in VHA and based on clinical encounters registered in the VHA corporate data warehouse. Healthcare records related to outpatient visits and inpatient stays were used.
Active duty Army service members with chronic pain after an index OEF/OIF/OND deployment ending between October 1, 2007, and September 30, 2014, were included (N = 286,885). The relative timing of deployment end date, chronic pain diagnosis, NPT treatment, and VHA outcome measurement is shown in Figure 1. Chronic pain was operationalized as a recurrence at least 90 days apart within any of ten clusters of International Classification of Diseases (ICD-9) diagnoses known to be associated with pain (e.g., nontraumatic joint disorders, musculoskeletal disorders). Similar diagnosis clusters have been previously used to identify chronic pain in health record data.26, 27 The specific codes used for this study are described in detail elsewhere.26 In addition to diagnoses in a service member’s health records, any 60-days supply of opioids prescribed in a 3-month period or a 90-days supply in 12 months was taken as an indication of chronic pain.28 We determined days supply across non-injectable opioids prescribed in the MHS including codeine, dihydrocodeine, fentanyl, hydrocodone, hydromorphone, meperidine, methadone, morphine, oxycodone, oxymorphone, and tapentadol.
We excluded service members who were discharged from the military for reasons we classified as misconduct and service members who died while receiving care in the MHS, as they were unlikely to have substantial VHA records. To avoid cases where events occurred in reverse order, we also excluded service members who received VHA care before receiving NPT in the military. After applying exclusion criteria, 275,820 Army service members with chronic pain remained: 142,539 who received care in VHA and 133,281 who did not (see Fig. 2).
Independent Variable: Receipt of NPT in the MHS
For each service member, we determined if they received any NPT (yes/no) in the MHS after their index deployment. NPT were identified in the MHS data repository using ICD-9 diagnosis codes, Current Procedural Terminology (CPT) codes, and Healthcare Common Procedure Coding System (HCPCS) codes. NPT were defined as acupuncture/dry needling, biofeedback, chiropractic care, massage, exercise therapy, cold laser therapy, osteopathic spinal manipulation, transcutaneous electrical nerve stimulation (TENS) and other electrical manipulation, ultrasonography, superficial heat treatment, traction, other physical therapy, and lumbar supports. The specific codes used are described in detail elsewhere.29
Dependent Variables: Adverse Outcomes in the VHA
Main outcomes after service members enrolled in VHA were determined based on ICD-9 and ICD-10 diagnosis codes recorded in VHA healthcare records until fiscal year 2018 (see eTable 1 in the Supplement). Outcomes included diagnoses of alcohol and/or drug disorders (yes/no), poisoning with opioids, related narcotics, barbiturates, or sedatives (yes/no, separately for accidental and intentional poisoning), suicide ideation (yes/no), and self-inflicted injuries including suicide attempts (yes/no). The outcome for alcohol and/or drug disorders combined abuse and dependence and excluded tobacco use disorder.
Propensity Score Weighting
To account for differences between service members with chronic pain who received NPT in the MHS and those who did not, we used propensity score-weighted analyses. Propensity scores represent the probability of group membership, in our study membership of the group who received NPT, and were estimated using the following demographic, clinical, and military service characteristics recorded between the end of the index deployment and the last quarter of MHS utilization or the end of 2015, whichever came first: age, gender, race, marital status, rank/pay grade, fiscal year of index deployment, total days being deployed, days of deployment as of the index deployment, length of observation in the MHS, presence of any diagnoses for the following disorders: adjustment, depression, anxiety, or posttraumatic stress (PTSD), traumatic brain injury (TBI), alcohol use disorder (AUD), or substance use disorder (SUD), whether specialty services were received for mental health or substance use, use of prescription opioids (average daily morphine equivalents and days supply), sum of appointment days with low pain, sum of appointment days with moderate pain, sum of appointment days with severe pain, and number of inpatient days and hospital discharges. Characteristics that increase with increasing MHS observation time (e.g., days supply opioids, number of inpatient days) were normalized by the length of a service member’s observation in the MHS. Propensity scores were then used to determine inverse probability of treatment weights (IPTW), which were used in our final analyses to balance group differences.30 To avoid undue influence from extreme weights, we truncated the weights to 10, as 90% of our cohort had a weight less than 10.31 To account for potential differences between soldiers who do and do not enroll in VHA, we used a multinomial propensity score model to determine IPTW for four groups: (1) NPT and enrolled in VHA; (2) No-NPT and enrolled in VHA; (3) NPT and not enrolled in VHA; and (4) No-NPT and not enrolled in VHA.
Primary statistical analysis involved time-to-event analysis comparing the two groups that enrolled in VHA (NPT vs. No-NPT). Data were right censored. For each of the outcomes, we report a propensity score-weighted log-rank test for differences in the Kaplan-Meier survival curves and propensity score-weighted multivariable Cox proportional hazard models. We assessed both weighted and unweighted Kaplan-Meier survival curves (see Supplement) and found the weighted curves to support the proportional hazards assumption. The log-rank test combines results of χ2 tests of the probability of an event of interest between two groups across time.32 Cox proportional hazard models estimate the relative difference (NPT vs. No-NPT) in rates at which events occur across time, while accounting for covariates. Because the No-NPT group is weighted to balance the NPT group, the estimated model coefficient for the group variable represents the average adjusted difference among those exposed to NPT. As the NPT group and the No-NPT group were still significantly different after applying IPTW in age, the length of observation in the MHS, presence of TBI, and the number of inpatient days and the sums of appointment days with low, moderate, and severe pain (see Table 1), we included those variables as covariates in our final analyses, following Ridgeway et al.33 As MHS alcohol and drug use diagnoses were available, we limited analyses for that particular outcome to only those service members who had not been diagnosed with alcohol and/or drug abuse or dependence while in the MHS, essentially focusing on new-onset alcohol and drug use disorders in VHA (n = 86,773 with NPT, n = 18,789 No-NPT).
Descriptive statistics showed that 26,103 (9.5%) active duty service members with chronic pain received NPT before they were diagnosed with chronic pain. Running our primary analyses without these service members did not substantially change our results (data not shown), and these service members were retained in our analytic cohort.
As a secondary analysis and to address potential alternative interpretations of our results, we added additional covariates from each service member’s VHA healthcare records to the Cox proportional hazard models. Specifically, we added length of observation in the VHA, exposure to NPT in the VHA, and days supply of opioids in the VHA, based on the same specifications that were used to determine these variables in the MHS.
All analyses were done in R version 3.5.3 and IPTW were determined with the function ‘mnps’ from the R package ‘twang’.34 All propensity score-weighted analyses were done with functions from the ‘Survey’ package and P < 0.05 was considered statistically significant. Approval for this study was granted by the Brandeis University Committee for Protection of Human Subjects, the Stanford University and VA Palo Alto Health Care System Institutional Review Boards, and the Human Research Protection Program at the Office of the Assistant Secretary of Defense for Health Affairs/ Defense Health Agency (OASD/DHA). The DHA Privacy and Civil Liberties Office executed an annual Data Sharing Agreement.