Evaluating the optimal duration of medication treatment for opioid use disorder

Abstract

Background and aims

Clinicians have little guidance on the ideal length of time patients should remain on medication treatment for opioid use disorder (MOUD) before being able to safely discontinue MOUD. This study estimated how the risk of all‐cause mortality changes with the duration of MOUD, controlling for patient characteristics that change the risk profile independent of duration of therapy.

Design, setting and participants

Retrospective cohort study using electronic health record data from the US Veterans Healthcare Administration. Veterans initiating MOUD with buprenorphine, methadone or extended‐release naltrexone from October 2010 to September 2020. Our analytic sample included 19 666 buprenorphine initiators, 8675 methadone initiators and 4007 extended‐release naltrexone initiators.

Measurement

Duration of MOUD was measured in days. Discontinuation was defined as a gap in any MOUD coverage exceeding 28 days, regardless of MOUD type initiated. The primary outcome was all‐cause mortality. We estimated multistate survival models allowing for the modeling of multiple states (i.e. on and off MOUD, death) without having to consider censoring or competing events, while adjusting for sociodemographic, clinical, prescription and facility and provider characteristics.

Findings

We observed approximately 226 000 person‐years of time at risk for discontinuation or pre‐discontinuation death, during which we observed 26 841 discontinuations (118.9 discontinuations per 1000 person‐years). We similarly observed a total of about 106 000 person‐years of post‐discontinuation follow‐up, during which we observed 3251 deaths (3.1 deaths per 1000 person‐years). We found the largest marginal gain in probability of 6‐year survival from an additional year on MOUD appears to occur around 2 years, as compared to 6 months on MOUD. Statistically significant gains continued through approximately 4–5 years of MOUD retention relative to 6‐month MOUD retention. After 4–5 years, the marginal gain from one additional year of MOUD was not statistically significant.

Conclusions

Among US veterans, the benefit of retention on medication treatment for opioid use disorder (MOUD) towards overall survival continues through at least 4 years of MOUD treatment. Quality metrics based on 6‐month MOUD retention may be insufficient.

INTRODUCTION

Opioid use disorder (OUD) reduces quality of life [1], increases risks to human life including overdose [2, 3] and infectious disease [4] and imposes higher societal costs [5]. Medications for opioid use disorder (MOUD) (buprenorphine, methadone and naltrexone) are effective at improving outcomes for persons with OUD. People with OUD who receive MOUD have fewer drug cravings, improved social functioning and better quality of life than those who do not [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. Although clinical practice guidelines from several organizations [18, 19, 20] advocate for universal, first line access to MOUD, there is conflicting advice about the optimal duration of therapy. MOUD treatment exceeding a year is associated with reduced emergency department visits, inpatient utilization and overdose, although risk of such adverse events remains high [18, 19, 20]. Yet, standards of care specify only 6 months as the minimum target for treatment duration [21, 22].

Clinicians are left with unclear guidance about when to counsel patients that it is safe to discontinue treatment. The American Society of Addiction Medicine National Practice Guideline for the treatment of OUD calls for further research on the optimal length of treatment with each type of MOUD [23]. The ambiguity may hinder patients' ability to fully participate in treatment plans. Only 50% to 60% of patients remain on MOUD for the current recommended minimum of 6 months following treatment initiation [24, 25, 26, 27, 28], with more than 25% saying they want even shorter durations [29]. Better information about potential differences in expected outcomes as treatment duration varies may help patients make more informed decisions about their care.

This study's objective was to quantify the relationship between MOUD therapy duration and risk of death among patients initiating MOUD. We estimated how the risk of all‐cause mortality changes with MOUD duration while controlling for patient characteristics that change the risk profile independent of duration of therapy (e.g. age). We, then, identified how survival probability increases with increased treatment duration for each type of MOUD initiated. The results of this study provide insights about optimizing treatment duration for clinicians and patients through empirical evidence about relative benefits and when improvements in survival are no longer expected with increased MOUD duration.

METHODS

Study design, setting and participants

Using electronic health record data, this retrospective cohort study was set in the Veterans Health Administration (VHA) of the United States (US) Department of Veterans Affairs (VA) from 1 October 2010 to 30 September 2020. Our target population was community‐dwelling US veterans diagnosed with OUD undergoing treatment with any MOUD. To be included, veterans had to have an OUD diagnosis and MOUD initiation during the study period and be 18 years or older at the diagnosis date. OUD diagnosis was identified using International Classification of Diseases (ICD) 9 and 10 codes, from a previous study [30]. Initiation was defined as first MOUD use following the OUD diagnosis that qualified them for the study and with no MOUD use in the prior 180 days, similar to previous studies [31, 32]. They entered the study at the initiation date (t = 0) and were followed until death or censoring at the end of study follow‐up, with assessment of whether they were still on any MOUD type or had discontinued at the time of death or censoring. We required neither a minimum duration of therapy nor a minimum follow‐up time as a criterion for inclusion, but the survival data methods used in our study precluded participants from initiating on the last day of the study where follow‐up time is zero days.

Participants who were not US residents, not users of VA at the time of initiation or no longer US residents at the time of discontinuation were considered out of scope for the study. User of VA was defined as having at least 2 visits to or paid for by the VA in the 365 days before the reference date, inclusive. Veterans were also excluded if they received MOUD from a non‐VA pharmacy, had a diagnosis for metastatic cancer, resided in a nursing home or were admitted to hospice care at any time during the study period. We also made two incidental exclusions of participants for whom we could not obtain a full year of history for baseline: those without a VA enrollment date, who enrolled less than a year before their first post‐diagnosis MOUD prescription or who filled their first post‐diagnosis prescription before 1 October 2011.

The study was approved by the Central Arkansas Veterans Healthcare System Institutional Review Board and uses the Strengthening the Reporting of Observational Studies (STROBE) reporting guidelines for cohort studies [33]. The analysis plan for this study was not pre‐registered; therefore, the findings should be considered exploratory.

Variables and data sources

We extracted study data from the VHA Corporate Data Warehouse (CDW), which includes electronic medical records, socio‐demographic characteristics and linked mortality data. All‐cause mortality was selected as the outcome instead of overdose mortality because OUD may increase the risk of death from other health conditions [34]. Death date was obtained from the VHA Mortality Data Repository. Duration of initial episode of MOUD therapy was the exposure of interest. Duration was measured as the time between initiation and discontinuation. Discontinuation was defined as in prior research [35, 36] to be a treatment gap greater than 28 days, where ‘gap’ is time between the last covered day from a previously documented fill or administration and the next documented receipt of any MOUD type. Switching to another MOUD type was not considered a discontinuation unless there was a treatment gap.

The type of MOUD initiated was included in the model as a categorical covariate. We did not explicitly capture exposure to additional types of MOUD (e.g. transitioning MOUD type after initiation). Other MOUD initiation characteristics were initiating location, prescribing provider specialty and prescribing provider credential. Socio‐demographic and geographic variables, including age, sex, race, ethnicity, Census region, rurality [37], employment status, marital status, VA priority group, social vulnerability index (SVI) scores [38], homelessness and justice‐involvement were assessed using information collected in a 365‐day review period at baseline and, except for sex, race and ethnicity again for discontinuation. Rurality and SVI were obtained by linking to veterans' ZIP codes. Comorbidity [39] and health service use also were assessed with a 365‐day review period, while prescription medications were assessed with a 30‐day review period because it is likely more recent medication exposure better indicates active treatment of the participant's comorbidities. Only the following had values that changed at discontinuation: age, marital status, employment status, VA priority group [40], rurality and social risk. All other covariates used values measured only at baseline.

Statistical analyses

Descriptive statistics were computed overall and for the subsets of the sample that experienced each state during follow‐up. Naïve and age‐sex adjusted Kaplan–Meier curves were constructed to illustrate the survival probabilities that would be computed when death is treated as a censoring event for discontinuation and likewise discontinuation as a censoring event when examining death after initiation as the outcome.

All other analyses were conducted using an illness‐death modelling framework, a simplified version of multi‐state survival models [41, 42]. Unlike the more common Cox proportional hazard model [43, 44], multi‐state survival models allow us to model different events a participant may experience without treating them as censoring or competing events. Further, these models do not require us to restrict our sample to participants who discontinued, avoiding the risk from immortal time bias that occurs when requiring one event to be observed before another and removing the need to use statistical corrections like inverse probability weighting to adjust for the probability that a subject survived to long enough to be observed in our sample. Because the model in this article is estimated parametrically, it does not carry the proportional hazards assumption. Instead, it has the usual parametric survival assumption that the functional form is correct and, as parameterized in this work, the additional semi‐Markovian assumption that transition probabilities in the future depend only on time spent in the current state [45]. The former assumption was checked during the model selection process described in Data S1. The latter assumption is untestable, but our specification of the model relaxes the full Markovian ‘memory‐lessness’ assumption that the future depends only on the present state with no reference to time in state.

Our model was conceptualized with three states and permits participants to transition into each state a maximum of one time (Figure 1). All participants started in the ‘initiation episode’ state and could transition directly to either the ‘discontinued’ state (transition 1; discontinuation) or ‘dead’ state (transition 2; pre‐discontinuation death). Participants could also transition to dead through discontinued (transition 3; post‐discontinuation death). Parametric models for the transitions were selected using a pre‐defined process to determine an appropriate functional form for the baseline hazard and covariates in the models. Transitions were modeled separately to accommodate different hazard distributions for each transition. The model was fit using a clock‐forward approach for follow‐up time (i.e. all transitions share a common time zero at initiation). Details of the full process to fit and select our model may be found in Data S1.

FIGURE 1.

Observed transitions for veterans initiating medications for opioid use disorder (MOUD). This figure depicts the number of participants who started in the given state along with their transitions. For example, 32 348 initiated MOUD (as shown in the bottom left of the ‘initiation episode’ box) while 5324 remained on MOUD without transitioning to either the ‘discontinued’ or ‘dead’ states (32 348 participants initiating MOUD – 26 841 participants who discontinued MOUD – 183 participants who died before discontinuing MOUD). Numbers of the other states can be interpreted similarly.

Coefficient estimates for a multi‐state model are not directly interpretable. Instead, they are interpreted via post‐estimation predictions. This requires selecting covariate values to compute predictions at, for instance, ‘reference’ participants. We present predictions for four reference participants that were selected to capture expected mortality risk by age and sex: 30‐ and 50‐year‐old male and 30‐ and 50‐year‐old female. We present stratified results to minimize the effect of systematic differences by age and sex between the veteran and non‐veteran US population. Because the type of medication that a person initiated on was parameterized as a categorical variable and could alter their baseline risk, we assigned each drug covariate a value equal to its proportion within the sample to capture the treatment mix when performing our post‐estimation calculations.

In our post‐estimation analysis, we created stacked transition probability plots, which show the probability that a participant assessed at time 0 will be in one of the three states as of time t. We, then, predicted relative survival at increasing duration of treatment to identify if, and where, continuing MOUD no longer yields additional survival benefit. We made these predictions for treatment durations of 6 months and for 1 to 6 years in 3 month increments and then computed relative expected post‐discontinuation survival at 1 through 6 years with 95% CI. The 6‐year prediction horizon was selected after reviewing sample descriptives because 95% of participants had 6 or fewer years of follow‐up after discontinuation, and we did not want to use sparse data for the predictions. These survival predictions incorporate information from all transition models and so better capture the rich variation in experiences across groups that single‐equation models can only model with an interaction term. Predictions are presented graphically. Full model output is available on request.

Sensitivity analyses

The primary factor governing survival predictions is the estimated shape of the hazard function. We, therefore, conducted simplified sensitivity analyses that determined if the shape of the hazard was expected to change with modifications to the sample to confirm the robustness of the primary analysis. First, we excluded participants who switched MOUD type in their initiation episode. We, then, excluded participants who resumed MOUD to determine if the post‐discontinuation transition model was substantively different.

RESULTS

Our final analytic sample was 19 666 buprenorphine initiators, 8675 methadone initiators and 4007 extended‐release naltrexone initiators (Figure S1). Figure 1 shows the number of observed transitions between states (‘initiation episode’, ‘discontinued’, ‘dead’) and Table 1 summarizes characteristics as measured at initiation for veterans by state experience. Of all study‐eligible veterans, 60.8% initiated buprenorphine, 26.8% initiated methadone and 12.4% initiated extended‐release naltrexone. Regardless of state experience, veterans in the study were middle age (means between 49.2 and 53.4 years) and predominantly male (>92%), resided in urban areas (>85%), frequently had chronic physical and mental health conditions (e.g. >84% chronic pain, >52% anxiety, >67% depression). Overall, approximately 85% of participants had a year or more of post‐initiation follow‐up and 84% had at least a year post‐discontinuation. We observed approximately 226 000 person‐years at risk for discontinuation or pre‐discontinuation death for participants in the analytic sample, during which we observed 26 841 discontinuations (118.9 discontinuations per 1000 person years) and 183 pre‐discontinuation deaths (0.81 deaths per 1000 person‐years). We similarly observed a total of approximately 106 000 person‐years of post‐discontinuation follow‐up, during which we observed 3251 deaths (3.1 deaths per 1000 person‐years).

TABLE 1.

Characteristics of participants at initiation of medication treatment for opioid use disorder by experience of states: initiation, discontinuation and death.^a

Characteristics	State
	Initiation of MOUD, n = 32 348	MOUD discontinuation, n = 26 841	Death, n = 3434
Type of MOUD initiated (n, %)
Buprenorphine	19 666 (60.8)	14 683 (54.7)	1797 (52.3)
Methadone	8675 (26.8)	8333 (31.0)	1339 (39.0)
Extended‐release naltrexone	4007 (12.4)	3825 (14.3)	298 (8.7)
Age (y; mean, SD)	49.39 (14.09)	49.15 (14.09)	53.35 (13.55)
Sex (n, %)
Male	29 901 (92.4)	24 581 (92.6)	3293 (95.9)
Self‐identified race and ethnicity (n, %)
American Indian or Alaska Native	299 (0.9)	247 (0.9)	26 (0.8)
Black or African American	6164 (19.1)	5614 (20.9)	762 (22.2)
Asian, Native Hawaiian or other Pacific Islander	258 (0.8)	212 (0.8)	25 (0.7)
White, Hispanic or Latino	1236 (3.8)	1052 (3.9)	101 (2.9)
White, not Hispanic or Latino	22 026 (68.1)	17 735 (66.1)	2297 (66.9)
More than one race	285 (0.9)	239 (0.9)	29 (0.8)
Declined to answer	1273 (3.9)	1074 (4.0)	83 (2.4)
Unknown by patient	620 (1.9)	512 (1.9)	58 (1.7)
Not recorded	187 (0.6)	156 (0.6)	53 (1.5)
Marital status (n, %)
Never married	8179 (25.3)	7001 (26.1)	859 (25.0)
Married	9395 (29.0)	7340 (27.3)	802 (23.4)
Separated	2573 (8.0)	2222 (8.3)	283 (8.2)
Divorced	11 106 (34.3)	9360 (34.9)	1343 (39.1)
Widowed	986 (3.0)	838 (3.1)	132 (3.8)
Unknown	109 (0.3)	80 (0.3)	15 (0.4)
Employment status (n, %)
Employed for self or others	6011 (18.6)	4640 (17.3)	382 (11.1)
Other than employed	26 337 (81.4)	22 201 (82.7)	3052 (88.9)
Priority group b (n, %)
1	13 841 (42.8)	11 414 (42.5)	1223 (35.6)
2–3	5285 (16.3)	4354 (16.2)	516 (15.0)
4	2056 (6.4)	1850 (6.9)	385 (11.2)
5	8785 (27.2)	7380 (27.5)	1128 (32.8)
6	441 (1.4)	349 (1.3)	26 (0.8)
7A and 7C	452 (1.4)	353 (1.3)	39 (1.1)
8A–8G	1488 (4.6)	1141 (4.3)	117 (3.4)
US Census region residence (n, %)
Northeast	6660 (20.6)	5679 (21.2)	794 (23.1)
South	11 885 (36.7)	9675 (36.0)	1132 (33.0)
Midwest	6696 (20.7)	5636 (21.0)	796 (23.2)
West	7107 (22.0)	5851 (21.8)	712 (20.7)
Residential rurality (n, %)
Urban	27 539 (85.1)	23 156 (86.3)	3012 (87.7)
Large rural city/town	2834 (8.8)	2177 (8.1)	255 (7.4)
Isolated small rural town	1975 (6.1)	1508 (5.6)	167 (4.9)
SVI score c (mean, SD)
Socio‐economic status	0.44 (0.24)	0.44 (0.24)	0.44 (0.24)
Household characteristics	0.36 (0.26)	0.35 (0.26)	0.34 (0.26)
Racial and ethnic minority status	0.73 (0.25)	0.74 (0.25)	0.75 (0.24)
Housing type and transportation	0.64 (0.24)	0.65 (0.24)	0.66 (0.23)
Composite SVI	0.54 (0.24)	0.55 (0.24)	0.55 (0.24)
Prior year social risk (n, %)
Justice involved	3754 (11.6)	3357 (12.5)	394 (11.5)
Unhoused	12 742 (39.4)	11 464 (42.7)	1569 (45.7)
Prior year health service use (mean, SD)
Emergency department visits	2.51 (4.26)	2.69 (4.41)	3.45 (4.97)
Psychiatric admissions	0.71 (1.35)	0.78 (1.42)	0.86 (1.59)
Inpatient admissions	1.18 (1.93)	1.29 (2.01)	1.72 (2.35)
Psychotherapy visits	11.45 (18.20)	12.81 (19.10)	12.03 (18.56)
Prior year overdose history (n, %)
Opioid	1046 (3.2)	878 (3.3)	171 (5.0)
Other, non‐opioid	1972 (6.1)	1734 (6.5)	278 (8.1)
Elixhauser comorbidities d (mean, SD)	4.31 (2.46)	4.42 (2.47)	5.50 (2.93)
Prior year substance use history (n, %)
Alcohol use disorder	15 698 (48.5)	13 785 (51.4)	1804 (52.5)
Tobacco use disorder	19 331 (59.8)	16 463 (61.3)	2231 (65.0)
Other drug use disorder	18 809 (58.1)	16 497 (61.5)	2134 (62.1)
Prior year mental health history (n, %)
Anxiety disorder	19 948 (61.7)	16 538 (61.6)	1791 (52.2)
Bipolar disorder	5021 (15.5)	4454 (16.6)	595 (17.3)
Depression	21 794 (67.4)	18 618 (69.4)	2507 (73.0)
PTSD	14 947 (46.2)	12 709 (47.3)	1475 (43.0)
Psychotic disorder	3109 (9.6)	2838 (10.6)	434 (12.3)
Prior year medical history (n, %)
Pain‐related health conditions	27 010 (83.5)	22 494 (83.8)	2963 (86.3)
Hepatitis C	6390 (19.8)	5707 (21.3)	1172 (34.1)
Hepatitis B	290 (0.9)	265 (1.0)	61 (1.8)
Endocarditis	195 (0.6)	160 (0.6)	51 (1.5)
HIV	488 (1.5)	450 (1.7)	82 (2.4)
Prior year opioid treatment history (mean, SD)
Average opioid dose (MME)	19.34 (53.13)	19.64 (56.43)	26.34 (51.36)
Prior month prescription history (n, %)
Antidepressant	17 060 (52.7)	14 471 (53.9)	1901 (55.4)
Benzodiazepine	2887 (8.9)	2350 (8.8)	414 (12.1)
Non‐benzodiazepine hypnotics	2850 (8.8)	2402 (8.9)	354 (10.3)
Other, non‐opioid analgesics	11 026 (34.1)	9348 (34.8)	1341 (39.1)
Skeletal muscle relaxants	4691 (14.5)	3860 (14.4)	490 (14.3)
Provider and facility characteristics
Provider credential (n, %)
MD/DO	19 961 (61.7)	15 547 (57.9)	1908 (55.6)
PA/NP	1795 (5.5)	1352 (5.0)	103 (3.0)
PharmD	465 (1.4)	412 (1.5)	32 (0.9)
Other/undetermined	10 127 (31.3)	9530 (35.5)	1391 (40.5)
Provider specialty (n, %)
Behavioral/mental health	18 538 (57.3)	14 643 (54.6)	1788 (52.1)
Primary care/internal medicine	1837 (5.7)	1274 (4.7)	136 (4.0)
Emergency medicine/hospitalist	286 (0.9)	211 (0.8)	27 (0.8)
Other/unknown	11 687 (36.1)	10 713 (39.9)	1483 (43.2)
Facility type (n, %)
VA medical center	29 355 (90.7)	24 701 (92.0)	3208 (93.4)
Outpatient health facility e	2131 (6.6)	1455 (5.4)	153 (4.5)
Other outpatient services	324 (1.0)	206 (0.8)	13 (0.4)
Other	538 (1.7)	479 (1.8)	60 (1.7)
OTP clinic (n, %)	7356 (22.7)	7123 (26.5)	1099 (32.0)

Abbreviations: MD/DO, medical doctor or doctor of osteopathy; MME, morphine milligram equivalents; MOUD, medication treatment for opioid use disorder; OTP, opioid treatment program; PA/NP, physician assistant or nurse practitioner; PharmD, doctor of pharmacy; PTSD, post‐traumatic stress disorder; SVI, Social Vulnerability Index; US, United States; VA, Veterans Affairs.

^a Values in the table are shown as either (n, %) to designate the number of participants in the column along with the percentage of all participants in the column this represents or as (mean, SD) to designate the value as the mean with the corresponding SD.

^b VA priority group status is a multi‐faceted measure of health and social need, determined by factors such as military service history, disability rating, income level and Medicaid eligibility. For example, priority groups 1 and 4 have high levels of disability (the former because of service‐connected injuries and the latter otherwise having catastrophic injuries preventing work). VA priority group status also helps to determine the co‐payments that veterans pay.

^c The SVI uses 16 US Census variables from the 5‐year American Community Survey to identify communities that may need support before, during or after disasters. The SVI has four themed measures of area risk: (1) socio‐economic status, (2) household characteristics, (3) racial and ethnic minority status and (4) housing type and transportation.

^d Elixhauser comorbidities are 38 different conditions that can be identified from administrative data and are used for severity adjustment.

^e Outpatient health facilities include Veteran Health Administration health care centers and community‐based outpatient clinics.

Our selected multi‐state models use a generalized gamma distribution for the discontinuation transition and log‐logistic distribution for the two transitions to the dead state. The models provided overall good fit for the data (Figure S2), with typical deviations at longer follow‐up as data becomes sparse. The stacked transition probability plots shown in Figure 2 show that, for example, males and females both have a higher probability of remaining in the ‘on MOUD’ state when initiating at 50 years old compared initiating at 30 years old. Other states and cross‐characteristic comparisons may be interpreted similarly. See Figure S3 for Kaplan–Meier plots.

FIGURE 2.

Stacked transition probabilities for initiation episodes from ‘on MOUD’ to ‘discontinued’ or ‘dead’ for 30‐ and 50‐year‐old male and female reference participants. Note: A ‘reference’ participant is a participant who serves as a point of comparison or benchmark within a study. In this figure, we have 4 reference participants, males who are 30 and 50 years of age (top two graphs) and females who are 30 and 50 years of age (bottom two graphs). Each of these reference participants have no other mortality risk factors (e.g. comorbid substance use disorders) beyond opioid use disorder. MOUD, medication for opioid use disorder.

Figure 3 plots the relative survival probability for the reference participants by different MOUD durations. As computed here, relative predicted survival probability is the ratio of the predicted probability of surviving to 6 years if a patient is treated for 1 year instead of 6 months, 2 years instead of 6 months and so on. They indicate that longer time on therapy increases expected relative survival, with larger gains in longer time horizons. Taking the highest‐risk reference participant (50‐year‐old male) as an example, continuing MOUD for 2 years is associated with a survival probability of 1.04. This means that the expected 6‐year survival probability is 4% greater than that of similar patients treated only 6 months, a 4% increase in the expected 6‐year survival probability. By contrast, our lowest‐risk reference participant (30‐year‐old female) is not projected to achieve a 4% greater survival probability until approximately 5 years of MOUD duration. Additionally, for the lowest risk reference participant, the marginal gains, improvements in relative survival for each additional year of therapy, start to diminish as duration increases, with no statistically discernible difference beyond 4 years. Similar plots showing relative predicted survival to 3 months and 1 year are shown in Figure S4.

FIGURE 3.

Relative predicted 6‐year survival probability by treatment duration for 30‐ and 50‐year‐old male and female reference participants.

Our sensitivity analyses excluded 1045 participants (3.2%) who switched MOUD type and separately excluded 13 664 participants (42.2%) who resumed therapy after discontinuation. These analyses indicate that the finding of increased relative survival with longer MOUD duration is not substantively changed by our treatment of switches and resumptions. The results are presented in the Sensitivity Analyses section of Data S1.

DISCUSSION

In a large, national sample of veterans initiating one of three MOUD types, we estimated the relationship between treatment duration and expected medium‐term survival probability, adjusting for patient characteristics that alter risk. Several MOUD retention metrics endorsed by the National Quality Forum [46] and the OUD Cascade of Care [21, 47] suggest a minimum MOUD duration of 6 months. Yet, this is not empirically derived and in the presence of increased risk of opioid‐related adverse patient outcomes seen after MOUD discontinuation [18, 19, 48, 49], some have recommended against MOUD discontinuation at any time point [50]. Our aim with this study was to identify, if possible, the point beyond which additional days of treatment did not provide significant survival benefits.

Using potential therapy durations of up to 6 years, we found that the suggested 6‐month minimum MOUD is likely insufficient, regardless of the patient's individual mortality risk. This is a clinically critical finding. Even for our reference participants that represent patients at low mortality risk, we demonstrated a potential benefit up to 4 years of therapy over the 6 years examined, but with the gain in relative survival decreasing with each additional year of therapy. For the higher baseline risk groups, we did not observe a levelling, but there also does not appear to be a statistically significant difference in gain in relative survival past 4 years. This may be due, in part, to the small number of participants who had treatment duration longer than 4 years, limiting our ability to make more accurate predictions at the upper end of treatment duration. We found no point at which additional years of MOUD resulted in lower survival probability compared to the current standard of care. Our predictions were made for reference participants who had no additional health risk factors (e.g. comorbid substance use disorders), so predicted survival probability tended to be high regardless of MOUD duration. Therefore, we expect these are conservative estimates of potential survival benefit from longer MOUD duration.

Additionally, the largest survival benefit from longer MOUD duration accrued to the highest risk group we examined. This finding indicates patients at high mortality risk may benefit from MOUD treatment of much longer duration, possibly beyond the 6 years we examined. This offers promise that mortality gaps could be closed by targeting patients at high mortality risk for longer MOUD duration and merits future study with other populations known to be at high risk. Currently, we have treated risk factors only as adjustments to baseline risk, thereby implicitly assuming that the effect of time is constant once baseline differences are accounted for. Future work should evaluate other heterogeneous treatment effects that may moderate the relationship between MOUD duration and survival, such as individual‐ (e.g. comorbidities) and geographic‐level (e.g. SVI) factors.

Because our intent was to fit models that could inform the clinician treating a patient at the start of MOUD, when duration of therapy would not yet be known we made several study design decisions that deviate from the existing literature. We did not impose a minimum treatment duration. Neither did we restrict our sample to only participants who had discontinued therapy. The ability to be more inclusive with the sample was a result of our novel application of a multi‐state survival model to MOUD duration, the first application we are aware of in the substance abuse literature. We also allowed switching between different MOUD types in this analysis, a choice we examined in a sensitivity analysis that did not result in substantive differences in conclusions. Further work should explore the characteristics of patients with especially short durations of therapy. Finally, our large sample with available follow‐up over many years allowed us to look at differences in outcomes relative to 6 months of MOUD for durations longer than 18 months, previously the longest horizon examined.

With 50% to 60% of patients discontinuing MOUD by 6 months [24, 25, 26, 27, 28], the current 90% retention goal is already difficult to achieve [21]. If the recommended duration is increased from 6 months, it may be tempting to conclude that meeting performance standards will only become more challenging. However, this supposition ignores potential improvements by increasing shared decision‐making with patients. Some patients already have a strong commitment to continuing MOUD therapy because of their own health beliefs, for example, fear of relapse to illicit opioid use [51]. Other patients may focus decision‐making more on health system barriers (e.g. logistical hurdles), medication adverse effects, lack of efficacy in treating chronic pain or a belief that ‘drug free’ is the only true recovery and so MOUD is a bridge to that goal [52]. By providing patients with better information about increased expected relative survival time, our findings can be used to guide discussions about expected benefits with longer MOUD duration that patients can incorporate in their own value system.

Limitations

The veteran population presented in this study is quite different from the general US population with OUD. Most notably veterans are disproportionately male and older, and they also face a higher risk of death from other causes. The way in which we have presented results here is designed to minimize any effect on generalizability. Our results are stratified by age and sex, and we set all risk factors to zero when computing predictions. Further, we have reported results as relative gains in survival within age and sex group rather than absolute gains.

As with any study relying on pharmaceutical records, we can only approximately measure the discontinuation date from the prescription runout date. Because of this, we may have misclassified some patients as being on therapy when they had already discontinued. Fills of buprenorphine within VA had to be 30 days or less up until the coronavirus disease 2019 pandemic, a period largely omitted from this study. Therefore, if we have misclassified veterans as being on therapy, it is likely by no more than 30 days for durations of therapy measured at up to 6 years. Although we excluded veterans who initiated therapy outside of VA, we were not able to capture veterans who continued therapy with a payer other than VHA. Therefore, some veterans may have been misclassified as being discontinued when they had simply changed payers. Veterans covered by VA receive their prescription medications for a co‐payment of no more than $5, and most of the veterans in our sample are covered at a benefit level that provides for free prescriptions from VA. We believe it is unlikely that many of the discontinuations in our sample were misclassified because veterans would have little incentive to choose an alternate payer for their medications.

Unfortunately, the results here do not establish that additional days of therapy cause improvements in survival because of the inherent limitations of the observational design. Although our models adjust for a wide array of observable risks that may affect the relationship between MOUD duration and survival, many factors are unmeasurable in electronic health record data (e.g. motivation). We also made no attempt to account for non‐random treatment assignment to different MOUD types, a limitation offset by the choice to not model the MOUD types as separate cohorts and compare differential survival by type.

Despite these limitations, our findings are consistent with other literature favoring treatment durations longer than the current 6‐month guideline from authors who examined associations with reduced emergency department visits, inpatient utilization and overdose [18, 19, 20]. Our findings, regardless of the observational design limitations, are of vital significance given ethical concerns associated with randomizing participants to shorter MOUD durations.

CONCLUSIONS

The findings of this study support others who have concluded that the current minimum needs to be revisited. The benefit of MOUD retention toward overall survival continues through at least 4 years of MOUD treatment. However, gains in survival with each additional year appear to diminish, or at least become not statistically discernible, somewhere between 4 and 5 years of MOUD retention. Patients and providers will value these gains differently and our work can help guide discussions on how to optimally choose duration of therapy based on expected survival while accommodating individual values.

AUTHOR CONTRIBUTIONS

Corey J. Hayes: Conceptualization (lead); funding acquisition (lead); investigation (lead); methodology (equal); project administration (equal); resources (lead); supervision (lead); visualization (equal); writing—original draft (lead); writing—review and editing (equal). Rebecca A. Raciborski: Conceptualization (equal); formal analysis (lead); methodology (lead); project administration (equal); supervision (equal); visualization (lead); writing—review and editing (equal). Mahip Acharya: Data curation (equal); methodology (equal); software (equal); writing—review and editing (equal). Nahiyan Bin Noor: Data curation (lead); software (lead); writing—review and editing (supporting). Edward V. Nunes: Conceptualization (supporting); methodology (supporting); supervision (supporting); validation (supporting); writing—review and editing (equal). T. John Winhusen: Conceptualization (lead); funding acquisition (lead); investigation (lead); project administration (supporting); resources (supporting); supervision (supporting); writing—review and editing (equal).

DECLARATION OF INTERESTS

None.

DISCLAIMER

The content does not represent the views of the US Department of Veterans Affairs, the National Institutes of Health or the US Government.

Supporting information

Data S1. Supplementary Information.

Figure S1. Study Flow Diagram.

Figure S2. Goodness of Model Fit Measured by Nelson‐Aalen Cumulative Hazard and Cox‐Snell Residuals.

Figure S3. Kaplan Meier Survival Curves by MOUD Initiation Type, Overall and Age‐Sex Adjusted.

Figure S4. Relative Predicted 3‐Month and 1‐Year Survival by Treatment Duration for 30‐ and 50‐Year‐Old Male and Female Reference Participants.

Hayes CJ, Raciborski RA, Acharya M, Noor NB, Nunes EV, Winhusen TJ. Evaluating the optimal duration of medication treatment for opioid use disorder. Addiction. 2026;121(4):922–933. 10.1111/add.70211

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the Veterans Health Administration. Restrictions apply to the availability of these data, which were used under license for this study. Data are available from the Veterans Health Administration with the permission of the Veterans Health Administration.