Association of Medication Treatment for Opioid Use Disorder With Suicide Mortality

We identified all VA patients who were diagnosed with OUD at two or more visits on different days within the same year between 2003 and 2017 and received one or more days of MOUD (defined below). Patients entered the study cohort on the first day of MOUD and were followed for up to 5 years, or until the date of their death, or until the end of the predefined study period (December 31, 2017), whichever came first.

We collected demographic and diagnostic information as well as health care utilization and pharmacy data from the VA Corporate Data Warehouse. Because some patients receive portions of their care in the community, we also used linked VA “fee basis” (FB) files and Centers for Medicare and Medicaid Services (CMS) claims files. We identified causes of death by using the VA–Department of Defense Mortality Data Repository. The Mortality Data Repository links mortality data on veterans from the Centers for Disease Control and Prevention (CDC) National Death Index with the VA Corporate Warehouse (https://www.mirecc.va.gov/suicideprevention/Data/data_index.asp).

We categorized the causes of death using the CDC National Center for Health Statistics ICD-10 definitions. Because of the high risk of bias in attributing cause of death, we considered three mortality outcomes: suicide mortality, external-cause mortality, and all-cause mortality (13). External-cause mortality per the CDC definition includes death due to any accident or injury, either intentional or accidental. Common examples include unintentional injuries (e.g., motor vehicle accidents, drowning, and fires), suicide deaths, homicide deaths, and overdoses deaths regardless of intent (self-harm, accident, or undetermined). The outcomes are not mutually exclusive; for example, suicide is nested within external causes.

There were three main groups of covariates: variables related to MOUD treatment, patient demographic characteristics, and health care contacts. We derived patient covariates from the VA Corporate Data Warehouse as well as linked CMS and FB files.

We considered MOUD exposure to include the receipt of methadone, buprenorphine, or naltrexone. Methadone for OUD is typically dispensed in liquid formulation from federally licensed opioid treatment programs. Because these doses are not usually recorded as prescriptions, we used service delivery codes as a proxy for dosing. To assess for buprenorphine exposure, we included prescriptions of sublingual buprenorphine and of sublingual combination buprenorphine-naloxone formulations. Because the transdermal, buccal, and intravenous formulations of buprenorphine were only FDA approved for the treatment of pain rather than OUD during our study period (14), we excluded these formulations from our analysis. To assess for naltrexone exposure, we included oral as well as long-acting injectable formulations. We identified buprenorphine and naltrexone using a combination of prescription fill data (VA Corporate Data Warehouse) and pharmacy claim National Drug Codes (CMS and FB).

We developed several rules to create MOUD coverage periods. First, we assigned patients to a new MOUD coverage period if they received 1 or more days of MOUD after a period of 7 or more days without MOUD. Second, we continued MOUD coverage periods unless patients had more than 7 consecutive days without MOUD. Third, we adapted an approach developed by Pearce et al. (15) to account for MOUD coverage during inpatient or residential stays. Finally, because previous research suggests that MOUD starting periods (first 14 days on) and stopping periods (first 14 days off) are associated with a higher risk of mortality (10), we divided time into four MOUD treatment periods: starting, stopping, stable on, and stable off. The stable off period was the reference group for all analyses.

Demographic characteristics included age, gender, and race/ethnicity. Race/ethnicity was coded as Black Non-Hispanic, White Non-Hispanic, and Hispanic; those with missing race/ethnicity data were assigned to the “unknown or other” race/ethnicity group. To describe physical and mental health comorbidities, we used previously developed categories (16) representing separate counts of individual physical diagnoses in the Elixhauser comorbidity index and diagnostic categories from DSM-5, excluding the OUD category, as it is a defining characteristic of the cohort.

We identified each day a participant had any health care visits. We provide increased granularity by identifying non-MOUD substance use disorder visits, mental health visits, and all other visits. We also identified any inpatient mental health services use.

Patients were entered into a series of proportional hazard models at the time of MOUD initiation and were followed for 5 years or through the end of 2017, whichever came first. We entered data using the counting process format such that each row of data represented a fixed state. New observations represent a new state (e.g., transitioning from on MOUD to a stopping period). We calculated hazard ratios and 95% confidence intervals to compare risk for each of our three outcomes (suicide mortality, external-cause mortality, and all-cause mortality) for MOUD treatment periods (starting, stopping, stable on), using stable off MOUD as the reference. Our initial models were crude. We then adjusted for age, gender, and race. We then added previous-year mental health and physical health diagnoses into our model. We separately categorized previous-year mental and physical health conditions as none, one or two, and three or more. In this way, we included separate variables for number of prior mental health and physical health conditions. Finally, we added health care utilization for the 90 days prior to the current period. We note that because suicide is a rare outcome, the models needed to be parsimonious to avoid overfitting. To achieve this, we categorized race as White or other and age as >50 years or younger. We categorized days of health care contact into <5 days, 5–20 days, and >20 days. We developed a categorical variable for any inpatient stay. The more specific measures of health care utilization (e.g., mental health services contact days) were correlated with total number of contact days and were not included in the models. Thus, Table 1 provides finer granularity for all covariates and more details for health care utilization than were included in our modeling. Because individual MOUD agents were introduced at different times and suicide rates among veterans have generally been increasing over time, all models included an additional term for three time periods: 2003–2007, 2008–2012, and 2013–2017. Gender and race were fixed over time. Age and year were defined at the start of each MOUD period (row of data). Comorbidities were calculated annually, and we attributed the previous-year values to each MOUD period to ensure that they were preexisting conditions. Health care contacts were based on the 90 days before the current MOUD period.

We performed additional analyses to evaluate the robustness of our primary analysis. The first reduced the maximum follow-up time to 3 years instead of 5 years because patients’ use of MOUD tends to decrease over time, and some patients might have improved to the point where they no longer required OUD treatment. The second converted the data into monthly observations where MOUD was classified as either fully on, fully off, or partial. The outcome and covariates were the same as those used in the primary analysis, but here we used logistic regression. The reason for this data structure and statistical model was that it allowed us to calculate E-values to examine the sensitivity of our findings to observational bias. The E-value represents the minimum strength of association that an unmeasured confounder would need to have with both MOUD exposure and mortality to fully explain away effect estimates (17).

Because the three MOUD agents included in the analysis effectively represent very different treatment programs with varying amounts of required health care contact, we performed an additional analysis that evaluated the effect of each compared with periods off MOUD. We employed the identical methods and model from our primary analysis and added the same set of covariates in a stepwise manner. The only model specification change was to code MOUD status as either on or off, as there would not be adequate statistical power to evaluate transition periods. Patients were assigned annually to the MOUD compound used most, as patients rarely received treatment with multiple compounds in a year.

To examine the effect of each individual MOUD agent on mortality outcomes independently, we performed a similar series of proportional hazard models. For this set of analyses, we only compared the MOUD treatment period with the reference period off MOUD. This approach was necessary because each MOUD sample was less than half the total MOUD sample, so we did not have statistical power to examine transition period independently. We then repeated the analysis as in the primary analysis considering each of the three mortality outcomes (suicide, external, and all-cause) and considering potentially confounding patient covariates in a stepwise fashion (demographic characteristics, diagnosis, and mental health treatment) (17).

We performed all analyses in SAS, version 9.4 (SAS Institute, Cary, N.C.). We considered baseline covariates to be significantly different when p values were less than 0.05 and hazards to be significantly different when the 95% confidence interval did not cross 1.

The overall cohort included 61,633 patients who received MOUD. Roughly half of the patients received buprenorphine (N=30,950), and the other half were almost evenly split between methadone (N=15,446) and naltrexone (N=15,237) (Table 1). The study population was largely male and predominantly White, with a mean age of 46.5 years (SD=13.1). The sample was notable for a high level of mental health comorbidity and treatment: over 35% had three or more mental health conditions in addition to OUD. Patients had, on average, 28.2 mental health visits (SD=44.3) and 9.1 non-MOUD substance use visits (SD=17.8) in their first year of MOUD. Almost a third of patients (31.8%; N=17,683) had inpatient mental health or substance use disorder–related admissions. Patients were followed for a mean of 1,282.5 days (SD=613.1), during which time there were 6,608 deaths, including 2,374 due to external causes and 298 due to suicide. Patients received MOUD for a mean of 151.2 days (SD=133.0) during the first year of observation.

The crude hazard ratio for suicide mortality during stable on MOUD periods was 0.45 (95% CI=0.32, 0.63) compared with stable off periods (Table 2). Stepwise adjustment for patient demographic characteristics, mental and physical health diagnoses, and health care utilization had a minimal effect; the hazard ratio after full adjustment was 0.45 (95% CI=0.32, 0.63). Hazard ratio estimates for suicide mortality for the MOUD starting and stopping periods were nonsignificant across all crude and adjusted models. Among covariates in the fully adjusted model, only White race (hazard ratio=4.16, 95% CI=2.63, 6.58) and the presence of three or more mental health conditions (hazard ratio=1.62, 95% CI=1.14, 2.30) had significant effects, with both acting as risk factors for suicide mortality.

The crude hazard ratio for external causes of mortality during stable on MOUD periods was 0.35 (95% CI=0.31, 0.40) compared with stable off periods (Table 2). After full adjustment, the estimate was nearly identical (hazard ratio=0.35, 95% CI=0.31, 0.40). In the fully adjusted models, hazard ratios for starting MOUD (hazard ratio=0.57, 95% CI=0.42, 0.78) and stopping MOUD (hazard ratio=1.24, 95% CI=1.01, 1.52) were significant, indicating decreased and increased risk, respectively.

The crude hazard ratio for all-cause mortality during stable on MOUD periods was 0.33 (95% CI=0.30, 0.36) compared with stable off periods (Table 2). The fully adjusted model had little effect on the estimate (hazard ratio=0.34, 95% CI=0.31, 0.37). In the fully adjusted models, hazard ratios for starting MOUD (hazard ratio=0.59, 95% CI=0.49, 0.72) and stopping MOUD (hazard ratio=1.14, 95% CI=1.00, 1.30) were significant, indicating decreased and increased risk, respectively.

The models with 3 years of follow-up produced nearly identical point estimates for the effect of MOUD: stable on periods were consistently associated with significant reductions in risk of death compared with stable off periods (results not shown). Transition periods had point estimates similar to the primary analysis values but failed to reach significance for the rarer outcome of suicide.

The results of the monthly logistic analyses agreed with those of the primary survival models. Fully on versus fully off MOUD provided odds ratios of 0.43 (95% CI=0.40, 0.47) for all-cause mortality, 0.48 (95% CI=0.43, 0.54) for external causes of mortality, and 0.60 (95% CI=0.45, 0.80) for suicide mortality. Partial MOUD provided similar estimates and likewise implied a reduced risk of death across all three outcomes relative to being fully off MOUD. The direction and magnitude of all covariates were similar to what was found in the fully adjusted primary analyses. The bias analysis suggested that unmeasured confounders would have to be very large to nullify the relationship between fully on MOUD periods and reduced mortality relative to fully off periods, with E-values of 3.87 for suicide mortality, 5.16 for external causes of mortality, and 5.33 for all-cause mortality.

When we examined the individual MOUD agents, very different patterns emerged (Table 3). Buprenorphine was associated with a decreased risk of suicide mortality in both the crude (hazard ratio=0.37, 95% CI=0.24, 0.56) and fully adjusted model (hazard ratio=0.34, 95% CI=0.23, 0.52). The crude estimate for suicide mortality while on methadone treatment suggested a decreased risk compared with off periods (hazard ratio=0.38, 95% CI=0.17, 0.85). While this finding remained after adjustment for patient demographic characteristics (hazard ratio=0.43, 95% CI=0.19, 0.95), the finding was no longer significant after additional adjustment for mental health conditions and contacts (hazard ratio=0.47, 95% CI=0.21, 1.08). Finally, naltrexone was not associated with decreased risk of suicide mortality in the crude (hazard ratio=1.45, 95% CI=0.77, 2.76) or fully adjusted models (hazard ratio=1.28, 95% CI=0.67, 2.44). We observed a similar pattern of results for external causes of mortality. However, for all-cause mortality, naltrexone did have a significant protective effect across models, albeit of a lesser magnitude than those observed for buprenorphine and methadone.