Pharmacological Management of Opioid Use Disorder: A Systematic Review and Network Meta-Analysis: Prise en charge pharmacologique du trouble de l'usage des opiacés : revue systématique et méta-analyse en réseau

Graphical abstract.

This is a visual representation of the abstract.

Abstract

Objective

To assess the comparative benefits and harms of opioid substitution and antagonist therapies (OSATs) for opioid use disorder (OUD) using a network meta-analysis (NMA) of randomized controlled trials (RCTs).

Data Sources

Medline, EMBASE, PubMed, PsycINFO, Web of Science, Cochrane Library, Cochrane Clinical Trials Registry, and the National Institutes of Health Clinical Trials Registry were searched through 10 August 2025, for trials assessing the effectiveness or safety of any OSAT for people with OUD.

Data Extraction and Synthesis

Pairs of reviewers independently screened studies, extracted data, and assessed risk of bias. Frequentist random-effects NMAs were conducted, and certainty of evidence for treatment effects was assessed using the Grading of Recommendations Assessment, Development, and Evaluation approach.

Main Outcomes and Measures

Treatment retention (people remaining in follow-up at study end), opioid abstinence (during follow-up), criminal activity, all-cause mortality, and overdose-related death.

Results

We included 103 RCTs assessing 13 interventions across 26,624 people with OUD. Compared with placebo, moderate-certainty evidence demonstrated that heroin was most effective for treatment retention (Odds Ratio [OR] = 31.47; 95% confidence interval [CI] = 12.29,80.58). Heroin + methadone (OR = 8.00; 95% CI = 3.21,19.93) and high-dose methadone (HMET) (OR = 6.94; 95% CI = 3.76, 12.82) also ranked highly (moderate certainty). High- and low-dose buprenorphine (moderate certainty) and HMET (low certainty), improved abstinence as measured by urine drug screens. No intervention reduced criminal activity (moderate certainty), all-cause mortality (low certainty), or overdose-related death.

Conclusions and Relevance

Heroin-assisted therapy, alone or with methadone, and HMET were most effective for treatment retention. Buprenorphine and HMET improved abstinence, although certainty was lower. Evidence for mortality, overdose, and other patient-important outcomes remains limited. Heroin-assisted therapy exclusively enrolled people with heroin dependence and should not be initiated in those without prior use. Larger trials directly comparing methadone, buprenorphine, and heroin-assisted therapy, and enrolling patients with fentanyl use, are needed in the evolving toxic drug supply.

PROSPERO Registration

CRD42013006507.

Plain Language Summary

Which Opioid Treatments Work Best for Opioid Use Disorder?

Background:

Opioid use disorder (OUD) is a serious and ongoing health crisis in Canada and across North America. Medications such as methadone, buprenorphine, and naltrexone are commonly used to treat OUD. These treatments are meant to help people stay in care, reduce opioid use, and improve health and social outcomes. However, it is not always clear which treatments work best, or whether they reduce serious harms such as overdose or death.

What We Did:

We reviewed and combined results from 103 randomized controlled trials that studied medications used to treat OUD. Altogether, these trials included more than 26,000 people. We compared different opioid treatments to see how well they helped people stay in treatment, reduce opioid use, and avoid harms such as overdose, death, or involvement with the criminal justice system.

What We Found:

We found that heroin-assisted therapy (sometimes given together with methadone) and high-dose methadone were the most effective treatments for helping people remain in treatment. Importantly, studies of heroin-assisted therapy only included people who already had heroin dependence; these results do not mean heroin should be started in people who have never used it. Buprenorphine, at both low and high doses, helped reduce opioid use based on urine drug tests. However, none of the treatments clearly reduced the risk of death, overdose, or criminal activity. For many of these outcomes, the available evidence was limited.

Why This Matters:

Our findings support the continued use of methadone and buprenorphine as effective treatments for OUD. While some treatments help people stay in care or reduce opioid use, much less is known about how well they prevent serious harms. More high-quality clinical trials are needed, especially in the context of today's increasingly toxic drug supply, which includes fentanyl and other contaminants.

Introduction

Opioid use disorder (OUD) is a chronic relapsing condition marked by compulsive opioid use, withdrawal symptoms, and significant psychological, physical, and social impairment. ¹ In Canada, OUD contributes to approximately 21 deaths and 15 hospitalizations each day due to opioid-related poisonings,^2,3 with over 350,000 people currently living with the disorder. ⁴ In the United States, opioid-related fatalities rose from 49,860 in 2019 to 81,806 in 2022. ⁵ This rising burden is largely driven by an increasingly toxic and unpredictable unregulated drug supply, leading to surging rates of overdose, hospitalization, and death across North America. ⁶

Current Canadian and U.S. clinical guidelines recommend opioid substitution and antagonist therapies (OSATs), including methadone, buprenorphine, and naltrexone, as first-line treatments for OUD, and slow-release oral morphine (MORPH) as a second-line recommendation.^7,8 Prior network meta-analyses (NMAs) have sought to compare these OSATs; however, these reviews have only focussed on specific interventions or outcomes. Some have examined mortality during methadone and buprenorphine treatment using observational cohorts, ⁹ while syntheses of randomized controlled trials (RCTs) have solely compared methadone, buprenorphine, and levomethadyl acetate (LAAM). ¹⁰ More recent NMAs have primarily evaluated treatment retention among a limited set of medications ¹¹ or focussed specifically on long-acting injectable buprenorphine ¹² or heroin ¹³ formulations. Others have only assessed the effectiveness of opioid agonist therapy on mental health symptom outcomes ¹⁴ or evaluated behavioural interventions, i.e., contingency management, for people with OUD. ¹⁵

While these studies provide important insights, differences in intervention scope, network structure, and outcome selection limit comparability across existing NMAs. For instance, prior NMAs have largely emphasized treatment retention or short-term opioid use, with limited synthesis of outcomes such as criminal activity, overdose-related death, and all-cause mortality that reflect broader dimensions of patient recovery and social reintegration. ¹⁶ Criminal activity, in particular, is an important marker of social stability and engagement in care. ¹⁷ Moreover, much of this evidence predates the proliferation of fentanyl and its analogues, which have reshaped the clinical and risk environment for people with OUD since 2016.^18,19 Furthermore, few have formally rated the certainty of evidence (CoE) using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework, limiting the ability to draw strong clinical or policy conclusions.^20–22 Given the expanding range of pharmacologic options available in many jurisdictions and the evolving risk environment associated with synthetic opioids, an updated and comprehensive synthesis to assess the full range of OSATs across multiple clinically and patient-important outcomes is warranted.

To address these gaps, we conducted a systematic review and NMA of RCTs comparing 13 interventions across outcomes meaningful to patients, clinicians, and policymakers, and formally assessed CoE using GRADE.

Method

We followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) extension statement for reporting systematic reviews with NMAs (Supplemental Material S1, eMethods 1) ²³ and registered our protocol with PROSPERO (CRD42013006507). Full methods, including modifications to the published protocol, ²⁴ are detailed below.

Search Strategy and Selection Criteria

A comprehensive search was conducted from inception to 10 August 2025, in Medline, EMBASE, PubMed, PsycINFO, Web of Science, the Cochrane Library, the Cochrane Central Trials Registry, and the National Institutes of Health (NIH) Clinical Trials Registry, with no language restrictions (Supplemental Material S1, eMethods 2). Reference lists of included studies were also reviewed (Supplemental Material S1, eResults 1).

Eligible studies were RCTs with ≥2 arms comparing any OSAT with another OSAT, placebo or waitlist/usual care (WLUC) for management of OUD. The WLUC group included people who received a delayed or unspecified active intervention alongside psychotherapy or other non-pharmacological treatment. Waitlist and usual care comparators were pooled to preserve network connectivity and reduce sparsity across outcomes. Given our focus on comparative effectiveness and safety, we excluded pilot or dose-finding trials (i.e., those lacking appropriate comparators), as well as studies enrolling specialized populations (e.g., prisoners, pregnant women). No restrictions were placed on age or comparator type.

Title/abstract and full-text screening were conducted independently by two reviewers, with disagreements resolved by discussion. For RCTs with multiple publications, we included all outcome-relevant reports but selected the version with a larger sample for duplicate outcomes. Detailed screening and adjudication criteria are described in Supplemental Material S1: eMethods 3.

Data Abstraction and Risk of Bias Assessment

Pairs of reviewers independently extracted data on trial characteristics, study population, interventions, and outcomes using a standardized form. Outcomes of interest included: (i) treatment retention, defined as the number of participants who remained in follow-up at the end of the study period; (ii) opioid abstinence, assessed as the number of opioid-negative urine drug screens (UDSs) recorded throughout the study follow-up period and/or (iii) the number of people with opioid-negative UDS results at the end of follow-up (excluding study medication); (iv) criminal activity (e.g., involvement with the justice system); (v) all-cause mortality; and (vi) overdose-related death. These outcomes were selected based on clinical relevance, ²⁵ prior literature,^26,27 and because they were commonly reported across the included trials. ¹⁶ Outcome data were extracted as reported in the original trial publications. When available, intention-to-treat (ITT) denominators based on randomized participants were preferentially used; however, reporting of ITT versus completer analyses was inconsistent across older trials. This limitation was considered during risk of bias (RoB) and CoE assessments.

RoB was assessed using the Cochrane RoB instrument ²⁸ addressing randomization, allocation concealment, blinding (patients, providers, outcome assessors), and missing data. The original Cochrane RoB instrument was selected because it was prespecified in our registered protocol ²⁴ and allowed for consistent assessment across older trials with variable methodological reporting. Each domain was rated as low, high, or unclear RoB. Trials with >10% loss to follow-up or any domain rated as high or unclear were considered at overall RoB (Supplemental Material S1: eMethods 3). Although no universally accepted threshold exists, this threshold was selected pragmatically in recognition of the substantial retention challenges commonly observed in OUD trials. ²⁹ We visually depicted RoB for included RCTs using the robvis (Risk-Of-bias VISualtization) Shiny tool. ³⁰

Statistical Analysis

We conducted all analyses using Stata SE (StataCorp., Version 17). For outcomes reported by ≥10 trials, we computed NMA estimates using a frequentist random-effects model assuming a common heterogeneity parameter.^31,32 This threshold was selected pragmatically to reduce network sparsity and improve the stability and interpretability of indirect comparisons, acknowledging that no universally accepted minimum study threshold exists for NMA. All outcomes were binary. Treatment effects were presented as odds ratios (ORs) with 95% confidence intervals (CIs) for all outcomes except for abstinence assessed by UDS. We calculated risk differences (RDs with 95% CIs, reported as percentages) for abstinence assessed by UDS using baseline risk from the control arms of included trials. To estimate baseline risk, we calculated the proportion of opioid-negative UDS out of the total number of tests conducted per trial arm, treating each test as a trial-level observation. ³³ For outcomes lacking placebo arms (i.e., criminal activity), we used high-dose methadone (HMET) as the reference comparator to improve the clinical interpretability of comparisons because it was the most commonly represented active intervention in the network and reflects an established standard treatment for OUD.^7,8 For any outcome reported by <10 trials, we pooled effect estimates using conventional pairwise meta-analysis.

To preserve network connectivity and prevent sparsity, we classified most pharmacologic interventions with similar properties as equivalent (e.g., oral and injectable naltrexone, or different doses of MORPH). Methadone and buprenorphine were categorized as high-dose (HMET, HBUP) or low-dose (LMET, LBUP) based on clinically informed thresholds (≥60 mg/day for methadone, ≥16 mg/day for buprenorphine). ³⁴ For Suboxone^®, interventions were categorized based on buprenorphine dosage. Interventions were grouped primarily according to active pharmacologic agent and dose rather than route or formulation; consequently, long-acting injectable buprenorphine formulations were categorized within the HBUP or LBUP nodes, as applicable. To assess robustness of the treatment retention network, we planned sensitivity analyses: (i) excluding WLUC trials, (ii) excluding trials contributing long-acting injectable buprenorphine to the HBUP or LBUP nodes, and (iii) restricting analyses to trials at low overall RoB where feasible.

Pairwise meta-analyses were performed for comparisons with ≥2 RCTs. Heterogeneity was assessed using visual inspection of forest plots and the I² statistic, ³⁵ and small-study effects were explored using Harbord's test when ≥10 studies were available.^36,37 We assessed global incoherence using the design-by-treatment interaction model ³⁸ and local incoherence via side-splitting and incoherence factor plots.^19,20,39

We evaluated transitivity through visual inspection of boxplots of key effect modifiers (sex,^40,41 age^42,43) using the NMAStudio tool, ⁴⁴ and planned to conduct meta-regression if reporting allowed. Interventions were ranked using the surface under the cumulative ranking curve values, mean ranks and rankograms.^39,45 Further statistical details (e.g., commands, assumptions, and transformations) are provided in Supplemental Material S1: eMethods 3.

Assessing Certainty of the Evidence

We used the GRADE approach to rate the CoE for each NMA estimate,^20–22 incorporating both direct and indirect comparisons. Direct evidence was assessed using standard GRADE domains: RoB, inconsistency, indirectness, and publication bias. ²² Indirect estimates were rated based on the dominant lowest-order loop. ²¹ For network estimates, we began with the higher of the direct or indirect ratings, accounting for their relative contributions. We rated down for incoherence if statistically significant difference between direct and indirect estimates was detected.^20,21 Imprecision was assessed using the NMA estimate; we rated down when the 95% CI crossed the null and was sufficiently wide to reduce decision certainty (e.g., spanned a > 5-fold range of the point estimate). Certainty ratings were assigned at the comparison/outcome level, with RoB downgrading based on the studies contributing to each estimate rather than the overall RoB profile across the full review.

Summary of Results: Categorizing Interventions

We used a minimally contextualized approach^46,47 to categorize interventions by effectiveness and safety, based on the NMA effect estimate, 95% CI, and GRADE CoE. For beneficial outcomes (i.e., treatment retention, abstinence), interventions were grouped as: (i) “among the least effective”—not superior to placebo; (ii) “intermediate”—superior to placebo only; (iii) “better than intermediate”—superior to placebo and at least one intermediate intervention; and (iv) “among the most effective”—superior to placebo and at least one “better than intermediate” intervention. We applied a similar approach to categorize interventions for harms outcomes (e.g., all-cause mortality, criminal activity) as: (i) “no more harmful than placebo/HMET,” (ii) “more harmful than placebo/HMET but less harmful than some alternatives,” and (iii) “among the most harmful.” All categories were further stratified by the CoE (moderate/high vs. low/very low) relative to the reference treatment (placebo or HMET).

Results

We identified 25,918 records through databases and 169 through the NIH registry. After removing duplicates, 103 publications from 95 unique trials (n = 26,624) were included (PRISMA flow diagram: Supplemental Material S1, eResults 2). The median (interquartile range) of the mean participant age was 36.25 years (30.98, 38.73), 26% were female, and 46 RCTs (48%) were conducted in the United States, with over half based in North America (52, 51%). Most trials (n = 74, 72%) were outpatient-based; 31(33%) had a six-month duration and 25 studies (26%) followed people for a year or longer (Supplemental Material S1, eResults 3–4).

Treatment retention was the outcome reported across most trials (n = 76), for which we synthesized 13 total interventions, including: placebo (reference), dihydrocodeine (DIHYCO), HBUP, heroin- or diacetylmorphine-assisted therapy (HERO), heroin + methadone-assisted therapy (HERMET), HMET, levacetylmethadol or LAAM, laudanum/opium tincture (LAUD), LBUP, LMET, MORPH, naltrexone (NALT), and WLUC. Figure 1 depicts networks for each outcome. We excluded 22 publications from analysis because they compared different doses of the same intervention (n = 12), enrolled overlapping populations (n = 7), or insufficient dosing information to classify methadone or buprenorphine (n = 3) (see Supplemental Material S1, eResults 4).

Figure 1.

Network of eligible comparisons. (A) Treatment retention. (B) Abstinence—urine drug screens. (C) Abstinence—number of people. (D) Criminal activity. (E) All-cause mortality. Abbreviations: APLCB = placebo (reference treatment/comparator for networks A–C and E); DIHYCO = dihydrocodeine; HBUP = high-dose buprenorphine; HERMET = heroin + methadone assisted-therapy; HERO = heroin (heroin-assisted or diacetylmorphine-assisted therapy); HMET = high-dose methadone (or AAAHMET: reference treatment for Criminal activity - network D); LAAM = levacetylmethadol or levomethadyl acetate; LAUD = laudanum/opium tincture; LBUP = low-dose buprenorphine; LMET = low-dose methadone; MORPH = morphine; NALT = naltrexone; WLUC = waitlist or usual care.

Risk of Bias

Of 103 included RCTs, 88 studies (85%) were judged to be at high or unclear RoB, with all but one trial having high or unclear RoB in at least one domain. Fifteen studies (15%) were rated as low RoB overall. Additionally, 22 trials (21%) had >10% missing data, and 19 trials (18%) reported any form of industry funding (Supplemental Material S1: eResults 5–6).

Treatment Retention

Treatment retention was reported in 76 RCTs (n = 19,201). Of the 24 direct comparisons, 18 were informed by ≥2 studies, with most showing moderate to substantial heterogeneity (I² > 30%; Supplemental Material S1: eTable 1). No evidence of global incoherence was observed (P = .15; Supplemental Material S1: eFigure 1), and loop-specific incoherence was detected in only one loop (with LBUP, LMET and naltrexone; P = .044; Supplemental Material S1: eFigure 2).

Moderate-certainty evidence suggests HERO, evaluated in five trials enrolling participants with heroin use disorder who had generally failed prior standard treatments, was associated with the greatest improvement in treatment retention compared with placebo (OR = 31.47; 95% CI = 12.29, 80.58; RD = 62.4 per 100; Table 1, Figure 2). Although the confidence interval was wide, it remained strongly favourable to HERO and did not cross the null effect. We did not rate down further for indirectness because the five HERO trials in our network directly evaluated heroin-assisted therapy in the treatment-refractory population for whom the intervention is clinically intended. Interventions categorized as better than intermediate included HMET, HERMET, WLUC, LAAM, LMET, MORPH, and HBUP, with HMET (OR = 6.94; 95% CI = 3.76, 12.82) and HERMET (OR = 8.00; 95% CI = 3.21, 19.93), demonstrating the largest relative effects for treatment retention. Intermediate interventions-LBUP and NALT-were only superior to placebo. Low-certainty evidence suggested that LAUD and DIHYCO may be among the least effective options (Figure 3).

Table 1.

Network Meta-Analysis Results (Odds Ratio or Risk Differences with 95% CIs) Sorted Based on GRADE Certainty of Evidence for Comparisons of Active Interventions versus Placebo.

Certainty of Evidence	Classification	Intervention	OR [95% CI]b,c	SUCRA	Risk Difference, Per 100a,d
Treatment retention
Moderate to high	Among the most effective	HERO	31.47 [12.29, 80.58]	99.7	62.39 [53.67, 66.31]
	Better than intermediate: superior to placebo and at least one intermediate intervention	HERMET	8.00 [3.21, 19.93]	73.1	47.23 [28.05, 58.95]
		HMET	6.94 [3.76, 12.82]	68.2	44.72 [31.82, 54.21]
		WLUC	7.14 [3.03, 16.86]	67.6	45.23 [26.65, 57.34]
		LAAM	6.58 [2.72, 15.92]	62.9	43.72 [24, 56.73]
		LMET	6.02 [3.23, 11.23]	58.1	42.01 [28.2, 52.46]
		MORPH	5.88 [1.53, 22.62]	56.5	41.54 [9.74, 60.04]
		HBUP	4.45 [2.66, 7.45]	38.9	35.66 [23.44, 46]
	Intermediate: superior to placebo	LBUP	3.76 [2.17, 6.51]	28.8	31.82 [18.37, 43.52]
		NALT	2.06 [1.20, 3.52]	12.5	17.07 [4.03, 30.26]
Low to very low	May not be better than placebo (or may be among the least effective)e	DIHYCO	4.19 [0.85, 20.72]	42.3	34.31 [−3.36, 59.3]
		LAUD	4.09 [0.81,20.69]	40.8	33.76 [−4.32, 59.29)
Abstinence (number of opioid-negative urine drug screens)
Moderate to high	Among the most effective	HBUP	—	69.1	15 [3, 27]
		LBUP	—	60.7	14 [1, 26]
	Not better than placebo (or among the least effective)e	MORPH	—	60.6	14 [–7, 35]
		WLUC	—	19	1 [–20, 22]
Low to very low	May be among the most effective	HMET	—	76	16 [2, 31]
		NALT	—	53.2	12 [2, 22]
	May not be better than placebo (or may be among the least effective)e	LAAM	—	68.3	15 [–2, 33]
		LAUD	—	50.3	11 [–18, 40]
		LMET	—	31.4	7 [–8, 22]
Abstinence (people who stayed off drug use during study period, as measured through urine drug screens)
Low to very low	May not be better than placebo (or may be among the least effective)e	LAAM	3.15 [0.02, 438.50]	30.4	6.59 [−74.75, 9.97]
		NALT	2.14 [0.05, 91.99]	41.1	5.06 [−58.97, 9.88]
		LMET	2.22 [0.02, 218.79]	36.8	5.23 [−74.75, 9.95]
		LBUP	2.11 [0.03, 165.64]	38.5	5 [−68.74, 9.93]
		HMET	1.60 [0.02, 136.56]	50.7	3.51 [−74.75, 9.92]
		HBUP	1.49 [0.02, 103.16]	54.7	3.06 [−74.75, 9.89]
		WLUC	1.47 [0.01, 157.02]	54.3	2.97 [−81.74, 9.93)
		LAUD	0.98 [0.01, 146.14]	64.3	−0.18 [−81.74, 9.92)
		HERO	0.89 [0.01, 96.41]	70.9	−1.1 (−81.74, 9.88)
All-cause mortality
Low to very low	May not be more harmful than placeboe	MORPH	0.27 [0.01, 8.84]	72.9	−1.45 [−1.98, 13.28]
		LAAM	0.41 [0.02, 9.23]	64.6	−1.17 [−1.96, 13.85]
		HBUP	0.51 [0.14, 1.90]	69.7	−0.97 [−1.72, 1.73]
		NALT	0.64 [0.16, 2.65]	55.4	−0.71 [−1.67, 3.13]
		LBUP	0.65 [0.12, 3.46]	53.8	−0.69 [−1.76, 4.6]
		HERO	0.80 [0.14, 4.47]	45.1	−0.39 [−1.72, 6.36]
		HMET	0.81 [0.20, 3.32]	43	−0.37 [−1.59, 4.35]
		HERMET	1.23 [0.08, 20.03]	34.7	0.45 [−1.84, 27.02]
		WLUC	1.48 [0.18, 12.06]	24.5	0.93 [−1.63, 17.75)
Criminal activity
Moderate to high	Not more harmful than high-dose methadonee,f	WLUC	0.49 [0.09, 2.78]	32.9	−4.38 [−8.12, 12.57]
		LBUP	0.60 [0.06, 6.46]	41	−3.4 [−8.41, 29.98]
		NALT	0.71 [0.08, 6.45]	46.5	−2.44 [−8.22, 29.95]
		HBUP	0.72 [0.12, 4.17]	46.9	−2.35 [−7.83, 20.2]
		DIHYCO	0.72 [0.20, 2.52]	44.8	−2.35 [−7.06, 10.95]
		LMET	0.73 [0.32, 1.63]	42.8	−2.27 [−5.93, 4.88]
		LAAM	0.99 [0.44, 2.25]	63.3	−0.08 [−4.83, 9.2]
		HERO	1.05 [0.59, 1.87]	67.5	0.41 [−3.49, 6.61]

Abbreviations: CI = confidence interval; DIHYCO = dihydrocodeine; HBUP = high-dose buprenorphine; HERMET = heroin + methadone assisted-therapy; HERO = heroin (heroin-assisted or diacetylmorphine-assisted therapy); HMET = high-dose methadone; LAAM = levacetylmethadol or levomethadyl acetate; LAUD = laudanum/opium tincture; LBUP = low-dose buprenorphine; LMET = low-dose methadone; MORPH = morphine; NALT = naltrexone; OR = odds ratio; RD = risk difference (per 100); SUCRA = surface under the cumulative ranking curve; WLUC = waitlist or usual care.

Notes: ^aPositive RDs (RD > 0) indicate benefit for outcomes where higher is better (e.g., Treatment retention, Abstinence), and negative RDs (RD < 0) indicate benefit where lower is better (e.g., Mortality, Criminal activity).

^bFor outcomes where higher values are beneficial (Treatment retention, Abstinence), an OR > 1 or RD > 0 indicates a favourable effect of the intervention compared to placebo.

^cFor outcomes where lower values are beneficial (All-cause mortality, Criminal activity), an OR < 1 or RD < 0 indicates a favourable effect.

^dRisk differences were calculated using the pooled control event rate for Abstinence -Urine drug screens and transformed from odds ratios (other outcomes). Values are rounded to the nearest whole number or first decimal for interpretability.

^e“Not better/more harmful than placebo/high-dose methadone” includes interventions whose 95% confidence intervals cross the null value (OR = 1 or RD = 0), indicating statistical non-significance.

^fNo placebo in network, so interventions were compared to high-dose methadone.

Figure 2.

League table for treatment retention. Abbreviations: APLCB = placebo (reference treatment/comparator); DIHYCO = dihydrocodeine; HBUP = high-dose buprenorphine; HERMET = heroin + methadone assisted-therapy; HERO = heroin (heroin-assisted or diacetylmorphine-assisted therapy); HMET = high-dose methadone; LAAM = levacetylmethadol or levomethadyl acetate; LAUD = laudanum/opium tincture; LBUP = low-dose buprenorphine; LMET = low-dose methadone; MORPH = morphine; NALT = naltrexone; WLUC = waitlist or usual care.

Notes: Expressed in odds ratios with 95% confidence intervals, and colour-coded according to the GRADE certainty of evidence.

Figure 3.

Network meta-analysis results sorted based on GRADE certainty of evidence and effect estimates for comparisons of active treatments versus placebo: effectiveness outcomes. Abbreviations: CI = confidence interval; DIHYCO = dihydrocodeine; HBUP = high-dose buprenorphine; HERMET = heroin + methadone-assisted therapy; HERO = heroin (heroin-assisted or diacetylmorphine-assisted therapy); HMET = high-dose methadone; LAAM = levacetylmethadol or levomethadyl acetate; LAUD = laudanum/opium tincture; LBUP = low-dose buprenorphine; LMET = low-dose methadone; MORPH = morphine; NALT = naltrexone; OR = odds ratio; RD = risk difference (per 100); WLUC = waitlist or usual care.

Since age and sex were not consistently reported across all comparisons, meta-regression could not be conducted. Instead, we visually assessed potential effect modification using boxplots, which revealed no appreciable patterns (Supplemental Material S1: eFigures 3–4). Sensitivity analysis excluding WLUC arms and long-acting injectable buprenorphine formulations yielded results similar to the primary analysis (Supplemental Material S1: eFigures 6–13, eTables 3–4). Trials rated as low RoB were sparsely distributed across interventions and outcomes, precluding meaningful low-RoB restricted NMA sensitivity analyses (Supplemental Material S1: eResults 5–6).

Abstinence (UDSs)

Thirty-five RCTs (n = 213,532 UDS) assessed abstinence over the full trial duration. There was no evidence of global (P = .81) or loop-specific incoherence (Supplemental Material S1: eFigures 14–17, eTables 5–6). Moderate certainty evidence showed that HBUP (RD = 15%; 95% CI = 3%, 27%) and LBUP (RD = 14%; 95% CI = 1%, 26%) were among the most effective interventions for achieving abstinence (Table 1; Supplemental Material S1: eFigure 15, eTable 5). Conversely, MORPH (RD = 14%; 95% CI = –7%, 35%) did not perform better than placebo. Low to very low certainty evidence suggested that HMET (RD = 16%, 95% CI = 2%, 31%) and naltrexone (RD = 12%; 95% CI = 2%, 22%) may also be among the more effective interventions (Figure 3).

Abstinence (Number of People)

Fifteen trials (n = 3,724) assessed whether people remained abstinent from opioids throughout the trial period. Low-certainty evidence indicated that no intervention was superior to placebo in increasing the odds of abstinence (Table 1, Figure 3). No global (P = .62) or loop-specific incoherence was detected (Supplemental Material S1: eFigures 18–21, eTables 7–8).

Criminal Activity

Nineteen trials (n = 4,402) assessed post-treatment engagement in criminal activity, with no evidence of global (P = .20) or loop-specific incoherence. No intervention was found to be more harmful than HMET, with moderate CoE (Table 1, Figure 4; Supplemental Material S1: eFigures 22–25, eTables 9–10).

Figure 4.

Network meta-analysis results sorted based on GRADE certainty of evidence and effect estimates for the comparisons of active treatments versus placebo: harm outcomes.

Abbreviations: CI = confidence interval; DIHYCO = dihydrocodeine; HBUP = high-dose buprenorphine; HERMET = heroin + methadone assisted-therapy; HERO = heroin (heroin-assisted or diacetylmorphine-assisted therapy); HMET = high-dose methadone; LAAM = levacetylmethadol or levomethadyl acetate; LAUD = laudanum/opium tincture; LBUP = low-dose buprenorphine; LMET = low-dose methadone; MORPH = morphine; NA = not applicable; NALT = naltrexone; OR = odds ratio; WLUC = waitlist or usual care. Notes: ^aThere was no placebo in the Criminal activity network, so high-dose methadone was the reference comparator.

All-Cause Mortality

Twenty-four RCTs (n = 7,266) contributed to the mortality network, with no evidence of global (P = .95) or loop-specific incoherence. Low-certainty evidence demonstrated that no intervention may be more harmful at increasing the risk of death compared to placebo (Table 1, Figure 4; Supplemental Material S1: eFigures 26–29, eTables 11–12).

Overdose-Related Death

In pair-wise meta-analyses of nine trials (n = 2363), no intervention was superior to any other in reducing the odds of overdose-related death (Supplemental Material S1: eFigure 30).

Discussion

Our systematic review and NMA of 103 RCTs synthesizing 13 interventions for OUD provides updated comparative evidence across five clinically and policy-relevant outcomes. Moderate certainty evidence supported the effectiveness of several interventions, particularly heroin-assisted therapy, combined heroin + methadone therapy, and HMET, in improving treatment retention relative to placebo. High- and low-dose buprenorphine also improved abstinence based on UDS, though no intervention was effective in sustaining abstinence at the person-level. Across all other outcomes-mortality, overdose-related death, and criminal activity-no intervention demonstrated clear superiority, and most estimates were imprecise or based on low- or moderate-certainty evidence. These findings offer relevant insights for clinicians, policymakers, and health system planners tasked with evaluating the comparative value of OSATs beyond traditional pairwise comparisons.

Comparison with Other Studies

It is challenging to directly compare our findings with prior NMAs due to the inclusion of different sets of interventions ¹⁵ and/or outcomes. ¹⁴ Nonetheless, our findings that heroin-assisted therapy, alone or in combination with methadone, was the most effective intervention for treatment retention are consistent with prior RCTs and systematic reviews of supervised heroin among treatment-refractory populations.^13,48 For instance, the North American Opiate Medication Initiative trial in Canada ⁴⁹ and a Dutch RCT ⁵⁰ both demonstrated superior outcomes, including higher retention and improved physical and psychosocial health, among people receiving combined heroin + methadone compared to methadone alone.

Both Lim et al. (2002) ¹¹ and Farre et al. (2002) ¹⁰ identified methadone as the highest-ranked intervention for treatment retention. In contrast, our broader network incorporated heroin-assisted therapy and additional comparators not included in earlier NMAs. These differences in intervention scope and network structure likely explain variation in treatment rankings across reviews rather than representing contradictory evidence regarding the effectiveness of HMET, which remained among the most effective interventions in our analysis and continues to be recommended as first-line therapy in Canadian guidelines. ⁸ Similarly, a recent NMA focusing on long-acting injectable buprenorphine formulations reported improved retention and abstinence compared to transmucosal buprenorphine and methadone. ¹² Unlike these formulation-specific analyses, our network grouped buprenorphine according to pharmacologic agent and dose rather than route of administration. Nevertheless, our findings support the effectiveness of both high- and low-dose buprenorphine for abstinence as measured by UDS. In contrast, although slow-release oral morphine is recommended as a second-line option in Canadian guidelines, ⁸ we found no significant benefit for abstinence, with an overall low CoE.

We did not identify any interventions that significantly reduced the odds of all-cause mortality, overdose-related death, or criminal activity. This likely reflects the limited number of trials assessing these outcomes rather than absence of effect. Prior research^51–54 and syntheses of cohort studies ⁹ suggest potential benefits of heroin and HMET in reducing mortality and criminal involvement; however, high-quality comparative randomized evidence remains limited.

Strengths and Limitations

We conducted the first NMA to compare a broad range of OSATs for OUD, and to hierarchically present estimates across both clinically- and patient-important outcomes. Our approach included a comprehensive search, duplicate screening and data extraction, RoB assessment, and GRADE assessment of the CoE. The use of summary rankings and sensitivity analyses enhanced the interpretability and robustness of our findings.

Several limitations warrant consideration. Some comparisons relied solely on indirect evidence due to the absence of head-to-head trials. Although we planned meta-regression and low-RoB restricted sensitivity analyses, inconsistent reporting of effect modifiers and sparse distribution of low-RoB trials across intervention nodes limited feasibility. Furthermore, reporting of analytic populations was inconsistent across many older trials, with several studies not explicitly distinguishing ITT from completer analyses or consistently reporting analytic denominators. Although ITT denominators were preferentially extracted whenever available, the underlying literature did not permit reliable classification of studies according to analytic approach, precluding meaningful ITT-restricted sensitivity analyses and potentially introducing additional uncertainty into some treatment effect estimates. Additionally, some interventions, particularly heroin-assisted therapy, were evaluated in unblinded supervised treatment settings in which performance bias may have been unavoidable despite treatment retention being a relatively objective outcome. Furthermore, our networks included both short- and long-acting treatments, e.g., opium tincture and methadone respectively; however, given the consistency of rankings in sensitivity analyses, this did not meaningfully impact our findings. Furthermore, we combined waitlist and usual care arms due to network sparsity; however, these comparators may not be clinically equivalent across settings. Usual care includes active and potentially intensive treatment, whereas waitlist conditions reflect delayed access to OSAT or non-pharmacological support only. Although our sensitivity analysis excluding waitlist or usual care trials demonstrated similar findings to our primary analysis, we caution that residual heterogeneity in this intervention's intensity poses a limitation.

While we stratified methadone and buprenorphine by dose, limited data and resulting network sparsity precluded doing so for other interventions. However, ignoring dose effects may affect validity. ⁵⁵ Additionally, the selected dose thresholds were clinically informed, as no universally accepted cut-points exist. Alternative thresholds would redefine intervention nodes, alter the underlying network structure, and consequently influence treatment rankings. Furthermore, as interventions were grouped according to pharmacologic agent and dose rather than formulation, our buprenorphine nodes likely reflect predominantly transmucosal evidence and therefore should not be interpreted as formulation-specific effectiveness estimates for long-acting injectable buprenorphine. Although sensitivity analyses excluding trials contributing long-acting injectable formulations yielded similar findings, our primary network was not designed to evaluate formulation-specific differences in pharmacokinetics, adherence, treatment delivery, or comparative effectiveness. ¹² Additionally, although our review included trials published after 2016, most were conducted prior to or during the early fentanyl era. Emerging evidence suggests that treatment retention with OSATs may now be lower, ⁵⁶ and withdrawal syndromes more complex, due to benzodiazepine contamination, ⁵⁷ potentially diminishing OSAT effectiveness ⁶ and raising concerns about generalizability.

Although we focused on clinically relevant, ²⁵ patient-important, objectively measured outcomes, we were unable to assess other patient-centered harms, such as infectious complications (e.g., endocarditis), medication side effects, and burdens of daily supervised dosing, ⁵⁸ which were rarely reported. Finally, the CoE was low or very low for several outcomes, particularly abstinence and safety endpoints, due to few trials and imprecise estimates.

Implications for Practice

Our findings that heroin-assisted therapy, alone or combined with methadone, ranked among the most effective interventions for treatment retention, presents complex implications for clinical practice and policy. ⁵⁹ Importantly, trials assessing heroin-assisted therapy were conducted exclusively in participants with heroin use disorder,^49,60–63 many of whom had failed to respond to prior standard treatments, and therefore may not be generalizable to the broader population of individuals with OUD. These findings should not be interpreted as support for initiating heroin in individuals without prior heroin use. Furthermore, only five heroin trials contributed to the network, resulting in wide confidence intervals around the NMA estimate. Moreover, heroin-assisted therapy is administered intravenously, carrying risks of injection-related complications (e.g., cellulitis, respiratory depression) that require supervised administration and ongoing clinical monitoring.^13,64 Prior reviews have characterized heroin as a high-intensity, resource-intensive specialist intervention intended for treatment-refractory populations rather than a first-line therapy for OUD. Accordingly, within Canada and North America more broadly, heroin-assisted therapy remains restricted to highly selected populations and specialized treatment settings,^13,65 with its effectiveness among people who inhale fentanyl or other opioids remaining unclear.

Heroin was absent from our network for abstinence measured by UDS—the intended treatment response for OSATs, thereby limiting its comparative assessment across all outcomes. While NMA estimates and rankings are calculated separately for each endpoint, people with OUD experience the full range of clinical outcomes. Prescribed treatment regimens should therefore consider effectiveness and safety across multiple outcomes. Although HMET demonstrated consistent effectiveness for both retention and abstinence in our synthesis of clinical trials and is widely available in Canada and across North American settings, clinical decision-making between methadone and buprenorphine also involves considerations not fully captured in our analyses, including safety profile, treatment accessibility, patient preference, comorbidity burden, and feasibility of treatment delivery.^8,66 Current Canadian guidelines recommend both methadone and buprenorphine as first-line treatment options for OUD, ⁸ while recent Canadian observational evidence suggests methadone may achieve improved treatment retention in some settings with similar mortality outcomes between the two OSATs. ⁶⁷

Implications for Research

Our review highlights important gaps in the evidence base. The limited number and size of heroin-assisted therapy trials, highlights the need for larger, rigorously designed studies to confirm its comparative effectiveness. Few trials assessed patient-important outcomes (e.g., quality of life, adverse events, burden of supervised dosing), limiting the relevance of available evidence to patients and policymakers. Most trials were conducted before widespread fentanyl use, raising questions about generalizability to the current toxic drug supply. To better assess these rare but meaningful markers of treatment success,^26,27 future research should prioritize larger trials and high-quality observational studies with longer follow-up, ensure better reporting of patient-centered outcomes, and consider dose effects across all interventions.^68,69

Conclusions

In this systematic review and NMA of 103 trials, heroin-assisted therapy-alone or in combination with methadone was among the most effective interventions for improving treatment retention, though these trials enrolled only people with heroin dependence and should not be interpreted as evidence to initiate heroin-assisted therapy in those without prior use. HMET was also highly effective. Both high- and low-dose buprenorphine, along with HMET, improved abstinence as measured by UDSs, though no interventions reduced all-cause mortality, overdose, or criminal activity. Overall, CoE was low for many outcomes, reflecting persistent evidence gaps. These findings support continued use of HMET as a first-line treatment while highlighting the need for further research in the context of an evolving and increasingly toxic drug supply.