Risk of opioid-related mortality associated with buprenorphine versus methadone: A systematic review of observational studies

2.1. Search Strategy

We conducted a systematic literature search of studies comparing buprenorphine and methadone among people with opioid use disorder from 1978 to February 6, 2021 in Medline, Embase, PsycINFO, and Web of Science databases. We also scanned the bibliographies of the included articles for additional references. We tailored our search strategy to each database, and search terms and keywords included those related to buprenorphine, methadone, opioid use disorder, and opioid-related mortality (see Supplementary Materials for Search Strategy). We restricted the search period to 1978 and onward because buprenorphine was first identified as a treatment option for treating addiction in 1978 [34]. There were no restrictions on the language of publication, and the search was updated on April 8, 2023.

2.2. Inclusion and Exclusion Criteria

We included observational studies (e.g., cohort or case-control studies) whose objective was to compare the risk of opioid-related mortality from buprenorphine with that from methadone in populations undergoing medication-assisted treatment for opioid use disorder. We excluded randomized controlled trials because they often have shorter follow-up period, resulting in an insufficient number of mortality cases to draw inferences [35]. We also excluded cross-sectional studies, letters to the editor, commentaries, editorials, and conference abstracts, as these studies do not contain the information needed to adequately assess their methodological quality.

2.3. Outcome Measures

The primary outcome of interest was drug-related mortality, including poisoning and fatal overdose events.

2.4. Study Selection

Two independent reviewers (JL and SO) screened titles and abstracts, and then conducted a full-text review of all articles retrieved from the databases for eligibility (study selection) based on specified inclusion/exclusion criteria. Disagreements between the two reviewers were resolved through consensus or adjudication by a third independent reviewer (DP).

2.5. Data Extraction

Two independent reviewers (JL and IF) extracted data using a pre-piloted, standardized data extraction form (see Supplementary Materials), with disagreements resolved through consensus or adjudication by a third independent reviewer (DP). Descriptive texts and quantitative data extracted included study characteristics, baseline participant characteristics, and outcome data. Study characteristics included study design, setting, data source, duration of follow-up, opioid use disorder assessment method, inclusion/exclusion criteria, and sample size. Baseline participant characteristics included distribution of age, sex, and history of psychiatric disorders (if applicable). Finally, the outcome data included a measure of association (e.g., incidence rate ratios [IRR], odds ratios [OR], or hazard ratios [HR]) and 95% confidence interval (CI).

2.6. Assessment of Risk of Bias

Two reviewers (JL and IF) independently assessed the risk of bias for all included studies. For these studies, we used the Risk of Bias in Non-Randomized Studies - of Interventions (ROBINS-I) tool to examine potential biases in seven domains [36]. These include (1) Confounding, (2) Selection bias, (3) Misclassification of intervention groups, (4) Deviations from intended intervention, (5) Missing data, (6) Outcome measurement error, and (7) Selective reporting. Based on the assessment of each domain, we assigned the overall risk of bias as low, moderate, serious, or critical, where the overall risk was determined by the highest risk assigned in any individual domain. In addition, we assessed the presence of other forms of biases that stem from inappropriate study designs in pharmacoepidemiology, including prevalent user bias (i.e., inability to capture early events due to prevalent users having survived initial stages of medication use), informative censoring (i.e., stopping follow-up at an arbitrarily set calendar date), and left truncation (i.e., initiation of follow-up at an arbitrary calendar date).

3.1. Search Results

The database searches identified 3109 publications (Figure 1). We removed 1330 duplicates and an additional 1697 studies upon review of the title and abstract. These studies were excluded because they were irrelevant to the research question or had non-observational study designs. Two reviewers independently examined the remaining 82 articles and excluded 75 of them (see Supplementary Materials for the list of articles for full-text review and reasons for exclusion). Our review included seven studies [37–43]. Since the number of studies was not sufficiently large for quantitative data synthesis, we conducted a systematic review without meta-analysis.

• Download figure
• Open in new tab

Figure 1.

Selection of studies for systematic review

3.2. Study Characteristics

Table 1 summarizes the characteristics of the six studies, which were conducted in Norway (N = 1) [37], the United Kingdom (UK) (N = 2) [38, 39], and Australia (N = 4) [40–43]. Four studies were retrospective cohort studies with a new-user design that included all patients who initiated OAT medications [38, 40, 42, 43], and three were administrative database studies that included only the OAT patients who died from exposure to either buprenorphine or methadone [37, 39, 41] (see Table 1 for the exact study design terminologies presented by the authors of the included studies). All studies defined exposure to buprenorphine or methadone as treatment for OUD, and they excluded buprenorphine or methadone that was indicated as analgesics or cough suppressants. Study cohort size ranged from 200 to 45664 patients. Drug-related mortality rates associated with each drug were calculated based on the number of deaths per 1,000 persons, the number of deaths per person-years of drug exposure, or the number of drug prescriptions. Five studies reported the number of deaths and person-years associated with the two medications [37, 38, 40, 42, 43].

3.3. Risk of Drug-Related Poisoning Deaths

Table 2 reports the effect estimates of the risk of drug-related mortality associated with buprenorphine and methadone. Three studies reported no difference between buprenorphine and methadone or only weak evidence in favour of buprenorphine [37, 42, 43]. The other four studies, on the other hand, reported a more favourable safety profile for buprenorphine (rate ratio ranging from 0.08 [95% CI = 0.01-0.48] in favour of buprenorphine to 6.23 [95% CI = 4.79- 8.10] against methadone) [38–41]. However, three studies restricted the study to only the patients whose cause of death involved buprenorphine or methadone, which may bias the findings and undermine the validity of results (see section on Risk of Bias for more detail) [37, 39, 41]. Therefore, we summarize below findings from four retrospective cohort studies with a new-user design [38, 40, 42, 43].

Of these studies with low risk of bias in selection of participants in the study, the estimated risks of drug-related poisoning deaths were heterogeneous across time. A UK-based study by Hickman et al. (2018) showed that being on buprenorphine treatment was associated with 92% lower rate of death compared to being on methadone treatment when the OAT patients were ‘on’ treatment (i.e., having active prescription of OAT medications) during the first four weeks (IRR = 0.08; 95% CI = 0.01-0.48) [38]. Similarly, an Australian study by Kimber et al. (2015) showed that methadone was associated with an almost 5-fold increased rate of fatal overdose compared to buprenorphine during the same time window (IRR = 4.88; 95% CI = 1.73-13.69) [40]. A more recent Australian study by Jones et al. (2022) showed that methadone was associated with nearly four times the higher rate of fatal overdose compared to buprenorphine during the first four weeks of being ‘on’ treatment, although the results were not statistically significant (IRR = 3.95; 95% CI = 0.89-17.51) [42]. However, after the four-week treatment initiation period, the rate of drug-related mortality while ‘on’ OAT differed across the three studies. The UK study above reported that being on buprenorphine treatment continued to have a protective effect against drug-related poisoning deaths compared to being on methadone treatment (IRR = 0.37; 95% CI = 0.17-0.79) [38]. On the other hand, the two aforesaid Australian studies showed that the effect of the two medications on drug-related poisoning deaths did not differ, with IRR = 1.18 (95% CI = 0.89-1.56) and IRR = 1.43 (95% CI = 0.94-2.17) observed by Kimber et al. (2015) and Jones et al. (2022), respectively [40, 42]. Relatedly, another Australian study by Larney et al. (2023) demonstrated that the risk of opioid overdose mortality was lower while being on buprenorphine treatment than on methadone treatment, although the results were not statistically significant (IRR = 0.67; 95% CI = 0.38-1.17) [43]. The authors of this study also observed similar results for patients across different age groups as well as patients with kidney disease and liver disease (see Supplementary Materials for measures of association done by subgroups).

The rates of drug-related mortality between buprenorphine and methadone varied across treatment status as well. When the OAT patients were ‘off’ treatment (i.e., discontinued or no longer on buprenorphine or methadone), the UK study showed no significant differences in the rate of death following buprenorphine discontinuation compared to that following methadone discontinuation during the first four weeks (IRR = 0.78; 95% CI = 0.36-1.66) and lower rate of death in favour of buprenorphine after the initial four-week period (IRR = 0.23; 95% CI = 0.12-0.48) [38]. However, the Australian study by Kimber et al. (2015) concluded that the rate of death following methadone discontinuation was lower than that following buprenorphine discontinuation (IRR = 0.50; 95% CI = 0.29-0.86) during the first four weeks and no difference after the initial four-week period (IRR = 1.12; 95% CI = 0.96-1.32) [40]. In addition, the Australian study by Jones et al. (2022) revealed no difference in the rate of fatal overdose when ‘off’ treatment both during the first four weeks (IRR = 1.55; 95% CI = 0.94-2.55) and after the initial four-week period (IRR = 0.82; 95% CI = 0.65-1.03) [42].

3.4. Risk of Bias

We conducted the risk of bias assessment for each study using the ROBINS-I (Table 3; see Supplementary Materials for a more detailed assessment of risk of bias). Out of the seven studies identified, three studies were assigned a moderate risk of bias [38, 42, 43], two were assigned a serious risk of bias [37, 40], and two were assigned a critical risk of bias [39, 41]. One domain that led to an increase in the risk of bias was the ‘bias due to confounding’, which resulted from residual confounding due to failure to adjust for important confounders [37, 39–41]. One study used fractional polynomial regression models to assess the presence of residual confounding and then the rule-out method to establish how strong the residual confounding would have to be to explain the effect sizes [40]. The authors of this study concluded that the association is unlikely to have been influenced by residual confounding, but they did not adjust for many key confounders (e.g., history of overdose or psychiatric morbidities) nor take into consideration potential interactions between them. Another study used a propensity score method with inverse probability weighting to control for confounding [38], but the potential for unmeasured confounding could not be excluded. Two studies also attempted to adjust for time-varying confounding (i.e., covariates whose values change over the course of follow-up) through regression adjustment [38, 42]. Of note, these covariates could be in the causal pathway; thus, adjusting for them possibly introduced over-adjustment bias [44]. Other, more appropriate techniques to control for time-varying confounding such as marginal structural Cox proportional hazards models were not considered [45], except in Larney et al. (2023) [43].

Another domain that led to an increase in the risk of bias was the ‘bias in selection of participants in the study’. Two studies included only the patients who experienced drug-related poisoning death and then retrospectively ascertained exposure to buprenorphine or methadone [37, 41]. These studies failed to account for person-time accumulated by the patients who were prescribed OAT for OUD but did not experience overdose death over the course of follow-up. In addition, due to left truncation and associated challenges in identifying OAT initiation, the possibility of prevalent user bias could not be excluded for these database studies. One study included all cases of drug-related poisoning death, including the patients who have not been prescribed OAT medications [39]. The mortality cases include those that involved buprenorphine or methadone from outside the healthcare system as well as deceased individuals who never utilized the healthcare system. Inclusion of such cases biases the relationship between drug-related poisoning mortality and medications prescribed as OAT within the healthcare system. Four retrospective cohort studies were new-user studies that followed OAT patients until overdose death or censoring, which mitigated the risk of bias in selection of participants into the study [38, 40, 42, 43]. However, one study censored patients after 12 months of OAT discontinuation [38]. According to the authors, this was done “to exclude cumulative dilution of mortality risks”, but it may have excluded treatment re-initiation or deaths that occur during the ‘off’ treatment period. Therefore, selection bias due to informative censoring is possible.

All identified studies defined exposure to buprenorphine or methadone based on records from administrative databases. Four studies used the time-varying exposure definition with grace periods to control for potential exposure misclassification [38, 40, 42, 43], but the addition of residual effect period does not completely eliminate misclassification of exposure time window. Hence, these studies were judged to be at moderate risk of bias in classification of interventions. Three studies did not identify with which of buprenorphine or methadone the patients initiated the OAT or how long they have stayed in the OAT prior to study entry [37, 39, 41]. Among these three, one study also classified whether mortality occurred during or after discontinuation of OAT (i.e., diversion) using database records of drug prescription [41], which further elevated the risk of exposure misclassification. Therefore, the above three studies were judged to be at severe risk of bias in classification of interventions. Finally, five studies were ascribed a moderate risk of bias in ‘selection of reported results’ due to the absence of a pre-specified protocol [37–41], and two were ascribed low risk due availability of a pre-specified protocol and analysis consistent with it [42, 43].