Opioid Dose Trajectories and Associations With Mortality, Opioid Use Disorder, Continued Opioid Therapy, and Health Plan Disenrollment

This cohort study evaluates the association between opioid dose trajectory and mortality, opioid use disorder, continued opioid therapy, and health plan disenrollment.


Findings
In this cohort study of 3913 patients, a decreasing opioid dose trajectory was associated with a lower risk of opioid use disorder and continued opioid therapy compared with stable dosing trajectories, but also was associated with an increased risk of disenrollment.Decreasing dose was not associated with mortality in the year after the end of the trajectory period; however, an increasing dose trajectory was associated with an increased risk of mortality and opioid use disorder but had no association with continued opioid therapy or disenrollment.
Meaning These findings suggest clinicians and patients should carefully weigh the long-term risks and benefits of opioid dose increases and decreases compared with maintaining stable opioid dosing.

Introduction
2][3][4] For instance, the 2016 Centers for Disease Control and Prevention (CDC) guideline 4 recommended avoiding opioid dose escalation to 90 or more morphine milligram equivalents (MME), and carefully reassessing risks and benefits when increasing to 50 or more MME.These recommendations were supported by observational studies [5][6][7] showing that higher doses are associated with increased overdose risk.The guideline also recommends tapering to lower doses or discontinuing opioids, preferably in collaboration with patients using an individualized approach, 3,4 if patients are not benefiting from opioids or have other overdose risk factors.
Tapering or discontinuing opioids are clinical strategies that may be selected in hopes of preventing sustained opioid exposure and long-term risks of opioid use disorder (OUD), overdose, and death, and may improve function, pain severity, and quality of life. 8However, observational studies have identified potential harms of these practices, including increased risks of heroin use, 9 substance use disorders, 10 care termination, 11 overdose or suicide deaths, 12,13 and emergency department use and hospitalizations for overdose, withdrawal, and mental health crises. 10,14,15ditionally, tapering attempts, particularly without shared decision-making, could prompt patients to seek care elsewhere by changing insurance plans or physicians, 11,16 potentially disrupting care continuity for comorbid health conditions.
Researchers have studied tapering using predefined opioid dosing patterns that occur after periods of stable dosing. 14,15However, prior research also suggests that many patients in routine clinical care experience complex and variable dosing patterns over time. 17For example, a physician may temporarily increase the dose to treat acute pain, or abruptly discontinue opioids if there is evidence of nonprescribed drug use.This study's objective was to identify subpopulations of patients with distinct, clinically meaningful trajectory patterns over a 1-year period using group-based trajectory modeling (GBTM). 18,19We also aimed to compare baseline characteristics across trajectories and outcomes occurring after the trajectory including 1-year all-cause mortality, incident OUD, continued opioid therapy, disenrollment from the health plan, and overdose.

Methods
The data-only study was approved by the Kaiser Permanente Colorado institutional review board, with the other sites' boards ceding oversight to the Kaiser Permanente Colorado board, and was granted a waiver of informed consent in accordance with 45 CFR §46.Study reporting followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

Study Settings
We used data from 3 health systems in 2 states: an integrated insurance provider and health care delivery organization covering Colorado's urban and suburban regions (site 1), an integrated, safety-net health system serving an urban Colorado region (site 2), and a health care system serving a predominantly rural Wisconsin population (site 3).Data were extracted using a common data model 20 from their coordinated and quality-controlled secure warehouse research databases; member and patient data were derived from electronic health records (EHRs), health system  24,25 To establish dosing trajectories during the year after the index date (the trajectory period), we excluded patients who died, were hospitalized, or were institutionalized for 28 or more days, or stopped being enrolled or empaneled.

Identifying Opioid Dosing Trajectories
Within a large population, GBTM can create meaningful trajectory groups representing patterns of individual change in opioid dose over time, such as maintaining, increasing, or decreasing the dose.
Groups are interpreted as latent longitudinal strata, meaning that individuals within a group are assigned a probability of group membership and are assumed to follow the same trajectory over time. 26,27Using GBTM, we identified clusters of patients according to their probability of following a similar mean daily MME pattern, calculated in 30-day intervals, over the trajectory period.
We modeled MMEs over time using a β distribution, which allowed for a diversity of dose patterns for identified groups. 28For the β distribution, we scaled the MME values to between 0 and 1 by truncating mean daily values above 400 MME (>99th percentile) during the trajectory period and dividing the values by 400.We assessed model fit with bayesian information criterion and the log Bayes factor approximation and evaluated linear, quadratic, and cubic terms to best fit the data (eTable 1 in the Supplement).We examined models with 1 to 7 trajectory groups; 6-group and 7-group models failed to converge using quadratic terms, and the 6-group model using linear terms had a group that was too small (<5%).Thus, a 5-group model with quadratic terms was considered the optimal model, and all groups had posterior probabilities of 0.95 or greater.We characterized trajectory groups according to our clinical interpretation of the patterns observed in Figure 1: decreasing, slight decreasing, stable moderate dose, stable high dose, and high-dose increasing.We also considered the mean doses in the first and last month of the period, the dose trend, and the proportion who stopped opioids for 1 month or longer during the trajectory.For analysis, we compared the decreasing group and the high-dose increasing group with the other 3 groups (slightly decreasing, stable moderate dose, and stable high dose, together referred to as "stable") combined.

Cohort Characteristics
0][31][32] Demographic characteristics included age on the index date, sex, and race and ethnicity as reported by patients during visit registration and recorded in the EHR.Race and ethnicity, Medicaid coverage, and the calendar year of the index date were assessed because they could impact management decisions and the outcomes.All available race and ethnicity categories were described, and 2 categories were used for analytic modeling: racial and ethnic minority groups and non-Hispanic White, with 2.2% missing race and ethnicity added to the non-Hispanic White referent group.Medicaid coverage also served as a proxy for low-income status.Cancer, chronic or acute pain, 33

Outcomes Following Opioid Dose Trajectories
We examined 4 primary outcomes after the trajectory period ended: all-cause mortality, OUD, continued opioid therapy at 1 year, and disenrollment.Secondary outcomes were fatal and nonfatal overdose from any substance combined and from opioids.Study follow-up ended on December 31, 2019, to avoid COVID-19-related health care disruptions.We limited follow-up for mortality and disenrollment to 1 year because later outcomes were less likely to be associated with the opioid trajectories.We examined all available follow-up time for OUD and overdose because these outcomes could still be associated with dose trajectories after a year.We examined time-to-death regardless of enrollment because NDI data allowed us to identify deaths in any setting.OUD was identified using diagnostic codes in the HER, and overdose was identified using diagnostic codes supplemented with NDI cause-of-death (eTable 2 in the Supplement).Among patients who were enrolled and alive 1 year after the trajectory period ended, we examined any continued opioid therapy.The disenrollment outcome included patients who, for any reason, stopped being enrolled in the health plan (sites 1 and 3) or empaneled (site 2).

Statistical Analysis
Baseline characteristics were compared using 2-sided χ 2 tests for categorical variables and means and the Kruskal-Wallis tests for continuous variables.To examine associations of baseline characteristics with the trajectory groups, we used log-binomial models to compare the decreasing and high-dose increasing groups with the combined stable groups (slightly decreasing, stable moderate dose, and stable high dose).We tested for interactions between baseline characteristics and site and included all significant (P < .05)interactions in the final model.We developed regression models to examine associations between group membership and each posttrajectory outcome separately.Disenrollment was examined in a time-to-event analysis starting at the end of the trajectory period, with death as a censoring event.To avoid analyzing dose patterns caused by OUD and overdose, we analyzed time-to-incident OUD, overdoses, and opioid overdoses in patients who did not have each respective diagnosis before the index date.
Disenrollment and death were censoring events for OUD and overdose.All models compared the high-dose increasing and decreasing groups with the relatively stable groups.A log-binomial model was fit to examine the continued opioid therapy outcome, whereas Cox proportional hazards regression models were used to examine time-to-each of the other outcomes.Proportional hazards assumptions were confirmed by a review of survival graphs and proportional hazard global tests. 34dels were adjusted for all available baseline patient characteristics.We also tested interactions between group membership and site to determine whether trajectory groups had consistent associations with the outcomes across sites.
Sensitivity analyses excluded (1) patients with past-year cancer, to ensure cancer was not a factor in the observed associations, and (2) patients hospitalized during the trajectory period, to ensure limited inpatient opioid dispensing data was not leading to biased results for the mortality outcome.We used all available data for this population-based cohort study.Analyses were completed in SAS Studio statistical software release 3.7 (Enterprise Edition; SAS Institute, Inc) with the add-on procedure Traj by Jones 35 for GBTM.Data were analyzed from January 2020 to August 2022.

Group-Based Trajectories
Five opioid dosing trajectories were identified, characterized, and described (Figure 1 and  c Interactions presented considering site as an effect modifier with other variable effects displayed within site (eg, the relative risk of racial and ethnic minority groups vs non-Hispanic White at Site 1).Significant interactions for decreasing vs stable were found for age (P for interaction = .04),racial and ethnic minority groups (P for interaction = .02),cannabis use disorder (P for interaction < .001),tobacco (P for interaction = .01),cancer (P for interaction < .001),year of index date (P for interaction < .001);for increasing vs stable, Medicaid (P for interaction = .02).

Outcomes After the Trajectory Period
After the trajectory period, 165 cohort members (4.2%) died within a year, the majority from cancer or cardiovascular disease.In the year after the trajectory period, 3499 (89.4%) remained alive and enrolled, and among those, 2531 (72.3%) continued opioid therapy.Over a mean (SD) of 3.0 (1.3) years of follow-up, 401 (11.4%) had an incident OUD and 61 (1.6%) had an overdose (eTable 5 in the Supplement).

Discussion
In this cohort study of opioid dosing trajectories, we identified decreasing, high-dose increasing, and 3 stable trajectories.Although we cannot be certain what clinical strategies or circumstances led to the observed trajectories, the decreasing trajectory appeared consistent with tapering to a lower dose or discontinuing opioids, the high-dose increasing trajectory appeared consistent with dose escalation, and the stable groups appeared consistent with dose maintenance.
After the year-long trajectory period, decreasing was associated with a reduced incidence of OUD and continued opioid therapy compared with stable dosing.However, decreasing was not associated with a significant reduction in 1-year mortality or overdose, and one potential harm was health plan disenrollment.Compared with the stable groups, the high-dose increasing trajectory was associated with an increased risk of death and OUD, but was not associated with continued opioid therapy, disenrollment, or overdose.
Over the trajectory period, more than half of patients maintained opioid dose stability.
Compared with stable dosing, however, patients and physicians may be able to further prevent the longer-term development of OUD if they can complete a taper while maintaining health plan enrollment.Although this study cannot explain the association between decreasing dose and disenrollment, abrupt opioid discontinuations, rapid tapers, a lack of mutual agreement between patients and physicians, inadequate ancillary pain management, and inadequate attention to withdrawal symptoms could all contribute to patient dissatisfaction, disengagement, and disenrollment.Loss of insurance or disengagement from ongoing ambulatory care could subsequently have detrimental medical and mental health consequences for patients with chronic pain and comorbid health conditions. 36These results can help inform patients, physicians, and health plans or systems about expected outcomes after a dose reduction, but research is needed on the short-term benefits and harms of different opioid management strategies.Further research is needed on how to mitigate any unintended effects of tapers.Future research could also explore health system, insurance, or formulary effects on outcomes.For instance, some observed site differences could have been due to Colorado Medicaid policies limiting opioid tablets dispensed to 120 per 30 days and dose restrictions. 37r findings on the high-dose increasing trajectory support prior research and recommendations that caution about dose escalation, 4,25,38 but should be interpreted in context.
Despite accounting for measurable confounders, the positive association between increasing and mortality could be due to unmeasured factors related to the underlying risk of death among people selected for dose increases or the reasons dose increases were initiated, such as unremitting pain or poor health status.

Limitations
This study has limitations.Within-group heterogeneity is obscured using GBTM.For instance, the shape of the decreasing trajectory in Figure 1 does not necessarily reflect the pace at which all patients in that group reduce their doses.A minimum period of event-free survival is required to assess trajectories, underestimating adverse event rates.Thus, the association between trajectories and outcomes occurring during the trajectory period, rather than after the trajectory, could not be Although we excluded prior OUD or overdose to minimize reverse causation the analysis of those outcomes, we cannot exclude the possibility that a nonfatal overdose during the trajectory period influenced opioid management and, therefore, group membership.
GBTM has been applied to data across many Department of Veterans Affairs clinics, 19 but has not been extensively applied to multisite studies from different types of health systems; thus, we explored site interactions in all models.Low numbers from 2 sites precluded robust site-specific estimates and larger sample size is needed to definitively test the association between decreasing and overdose, given findings reported by Agnoli and colleagues. 14Additionally, overdoses not leading to a hospital visit or death could not be identified using our methods.Despite these limitations, the health systems studied serve 2 states with distinct prescribing policies, care delivery systems, and insurance products, enhancing generalizability.

Conclusions
Our findings suggest that physicians and patients should be informed of the longer-term benefits and risks of opioid dose increases and decreases compared with maintaining dose stability.

Figure 1 .
Figure 1.Opioid Dose Trajectories According to Group-Based Trajectory Model 160

Figure 2 .
Figure 2. Adjusted Associations Between Increasing and Decreasing Opioid Dose Trajectories and Outcomes

JAMA Network Open | Substance Use and Addiction Mortality
, Opioid Use Disorder, Continued Opioid Therapy, and Health Plan Disenrollment Tenth Revision), pharmacy dispensations (National Drug Codes), and utilization.To ensure data consistency, we used distributed SAS code and explored differences in covariate frequencies and outcome rates across sites.To ensure complete and consistent capture, vital status and cause-ofdeath were derived from patient identifiers linked to the National Death Index (NDI)-Plus.21 JAMA Network Open.2022;5(10):e2234671. doi:10.1001/jamanetworkopen.2022.34671(Reprinted) October 5, 2022 2/13 Downloaded From: https://jamanetwork.com/ by a University of Washington Libraries User on 02/10/2023 pharmacies, and external medical or pharmacy insurance claims.Data included demographic information, social history, diagnoses (International Classification of Diseases, Ninth Revision and nonfatal overdose, mental health disorders (eg, depression, anxiety, and bipolar) and tobacco, alcohol, opioid, cannabis, and other (sedative, hypnotic, or anxiolytic; cocaine; other stimulant; hallucinogen; or other psychoactive) substance use disorders were assessed in the year before the index date.Prescribed and dispensed benzodiazepines and stimulants were examined in the 6 months before the index date.Covariates were identified using International Classification of Diseases codes, social history fields, National Drug Codes, or other available data fields (eTable 2 in the Supplement) and when present were coded as yes; covariates not present were coded as no.Thus, other than race and ethnicity, there were no missing data.

Table 1 .
Baseline Demographic and Clinical Characteristics of the Multisite Study Cohort Other race is a category in the Virtual Data Warehouse which may reflect patient selections at the time of insurance enrollment or visit registration.
a b Diagnoses were assessed 1 year before the index date.cIndividuals could have more than 1 substance use disorder.Tobacco was assessed using diagnoses and social history fields.Other substance use disorders including sedative, hypnotic, or anxiolytic; cocaine; other stimulant; hallucinogen; or other psychoactive substance use disorders.dDispensations were assessed 6 months before the index date.JAMA Network Open |

Table 2 .
Opioid Dose Trajectories by Site and Dose Characteristics Over the Trajectory Period Abbreviation: MME, morphine milligram equivalents.

Table 3 .
Adjusted Associations Between Baseline Characteristics and the Opioid Dose Decreasing Trajectory (Group 1), Increasing Trajectory (Group 5) vs the Relatively Stable Dose Trajectories (Groups 2-4) a a Log-binomial models.bOther substance use disorders include sedative, hypnotic, or anxiolytic; cocaine; other stimulant; hallucinogen; or other psychoactive.

JAMA Network Open | Substance Use and Addiction Mortality
, Opioid Use Disorder, Continued Opioid Therapy, and Health Plan Disenrollment All diagnoses were recorded in routine clinical care, which could have led to inaccuracies.Data on education, income, and other potential confounders were unavailable.
If dose decreases are indicated, clinicians should consider how to mitigate the risk of disenrollment.If dose increases are indicated among patients on high doses, patients should be monitored for the development of opioid use disorder.Research on short-term effects of opioid dose changes is also needed.Model Fit Criteria Used to Select the Number of Trajectory Groups eTable 2. International Classification of Diseases Codes (ICD-9/10) and Other Codes Used to Identify Outcomes and Covariates eFigure 1. Flow Diagram for the Study Cohort From 3 Health System Sites eTable 3. Baseline Characteristics of Eligible Patients Included in the Analyses Compared With Eligible Patients Not Included in the Analyses eTable 4. Baseline Demographic and Clinical Characteristics by Each of the 5 Trajectory Groups eTable 5. Study Outcomes in the Overall Cohort and in Each of the 5 Trajectory Groups eFigure 2. Excluding Patients with Cancer at Baseline, Adjusted Associations Between Decreasing and Increasing Opioid Dose Trajectory vs Stable Groups and Outcomes (Sensitivity Analysis) eTable 6. Number of Deaths in the Study Cohort Used for the Primary Analysis Compared With the Cohort Excluding Patients Hospitalized During the Trajectory Period (Sensitivity Analysis), by Trajectory Group eTable 7. Adjusted Associations Between Decreasing and Increasing Opioid Dose Trajectory vs Stable Groups and 1-year Mortality in the Study Population Used for the Primary Analysis and After Excluding Individuals Hospitalized During the Trajectory Period (Sensitivity Analysis)