The effects of glucagon-like peptide-1 receptor agonists (GLP1-RAs) on alcohol-related outcomes: a systematic review and meta-analysis

Studies were eligible if they included adults (≥ 18 years) with or without diagnosed alcohol use disorder (AUD), including those with co-morbid obesity or type 2 diabetes mellitus (T2DM). Inclusion of adults with or without diagnosed AUD was intended to reflect real-world GLP-1RA use and assess both treatment and preventive effects, with subgroup analyses exploring population differences where feasible. The intervention was any GLP-1 RA-based therapy (e.g., semaglutide, liraglutide, exenatide, tirzepatide) compared to placebo, no treatment, or other pharmacological interventions (e.g., naltrexone, acamprosate, non-GLP-1 RA anti-obesity or anti-diabetes medications). Non-pharmacological interventions (e.g., behavioural therapies) were excluded as the primary comparator unless they were part of a control arm in a randomised controlled trial (RCT) or other controlled pharmacological study. Outcomes of interest for meta-analysis included alcohol consumption (total intake or drinks per drinking day, measured as standardised mean difference [SMD]), AUD incidence/recurrence (hazard ratios [HR] or odds ratios [OR]), heavy drinking days (OR), and alcohol craving (SMD, e.g., via Penn Alcohol Craving Scale). Secondary outcomes, such as quality of life related to alcohol use, weight change, and glycaemic control, were extracted but not pooled in this meta-analysis due to the insufficient data available. (Supplementary Table 2) Eligible study designs included randomised controlled trials (RCTs), cohort studies, case-control studies, and case series. Case series were included in the systematic review but excluded from the meta-analysis unless they provided suitable data for pooling (e.g., HR, SMD, OR). Observational studies required the use of propensity-score matching, within-individual designs, or similar robust methods (e.g., multivariable regression with adequate covariate adjustment) to control for confounding, thereby ensuring reliable estimates of effect. This criterion led to the exclusion of 24 studies during the full-text review due to inadequate control of confounders, as part of the 86 studies excluded for alcohol-related endpoints that did not meet the inclusion criteria, as detailed in the PRISMA flow diagram (Fig. 1). No studies using alternative robust methods, such as instrumental-variable approaches, were identified in the search, possibly due to the emerging nature of GLP-1RA research in AUD. Qualitative studies, case reports, cross-sectional studies without comparators, and observational studies without such confounder control methods were excluded. We acknowledge that this criterion may favour more recent studies that use advanced methods, such as propensity-score matching, which could potentially introduce selection bias.

We conducted a comprehensive search of PubMed, Embase, and Cochrane Library (Trials) from January 1, 2005, to May 3, 2025, starting with the FDA approval of exenatide (April 28, 2005), the first GLP-1 receptor agonist. These databases were selected for their extensive coverage of biomedical, pharmacological, and clinical trial literature relevant to GLP-1 receptor agonists (GLP-1RAs) and alcohol-related outcomes. The search used Medical Subject Headings (MeSH) and free-text terms for alcohol-related outcomes (e.g., “Alcohol-Related Disorders“[MeSH], “alcohol use disorder,” “alcohol* consum*,” “alcohol craving,” “heavy drinking”) and GLP-1RAs (e.g., “Glucagon-Like Peptide 1“[MeSH], “semaglutide,” “liraglutide,” “exenatide,” “tirzepatide,” “albiglutide,” “dulaglutide,” “efpeglenatide”). An example PubMed search string was: (“Alcohol-Related Disorders“[MeSH Terms] OR “Alcohol Drinking“[MeSH Terms] OR “Alcoholism“[MeSH Terms] OR alcohol use disorder OR AUD OR alcohol* consum* OR alcohol craving OR heavy drinking) AND (“Glucagon-Like Peptide 1“[MeSH Terms] OR “Receptors, Glucagon“[MeSH Terms] OR GLP-1 receptor agonist* OR semaglutide OR liraglutide OR exenatide OR tirzepatide OR albiglutide OR dulaglutide OR efpeglenatide). Reference lists of included studies and trial registries (ClinicalTrials.gov, WHO ICTRP) were searched for ongoing and unpublished randomised controlled trials (RCTs). No unpublished RCTs meeting the inclusion criteria (e.g., assessing GLP-1RAs for alcohol-related outcomes in adults) were identified. No language restrictions were applied.

Records were deduplicated using EndNote software. Two reviewers (SG and BS) independently screened titles and abstracts using Covidence software, followed by full-text review, with discrepancies resolved by discussion. The selection process was documented in a PRISMA flow diagram (Fig. 1), which details the identification of 2,655 records from a database search (Cochrane Library, PubMed, and Embase) up to May 3, 2025. After removing 1,217 records before screening (577 duplicates, 640 non-primary studies), 1,438 records were screened, 187 reports were sought for retrieval, 24 reports were not retrieved due to lack of response from primary authors, 163 reports were assessed for eligibility, and nine studies were ultimately included in the meta-analysis (3 RCTs, six observational studies) after excluding 154 reports (68 not relevant to GLP1-RA, 83 alcohol-related endpoints not matching inclusion criteria). Additionally, three studies were included in the systematic review but excluded from the meta-analysis due to incompatible outcomes (alcohol consumption not pooled) or study design (case series), as detailed in Section “Baseline characteristics”.

Two reviewers (SG and BS) independently extracted data using a standardised form, including study design, population characteristics (age, sex, BMI, AUD status), intervention details (type, dose, duration of GLP-1RA), comparator, outcomes, and risk of bias. Outcomes extracted included alcohol consumption (total intake or drinks per drinking day), with the latter reflecting country-specific standard-drink definitions (e.g., 14 g in the US, 12 g in Denmark); conversion to grams per day was not feasible due to inconsistent reporting, limiting direct volume comparability, AUD incidence/recurrence (as part of alcohol-related events in observational studies, not assessed in RCTs due to their short-term design focused on consumption and craving), alcohol craving, and heavy drinking days, defined as the mean number of days with ≥ 4 drinks for women or ≥ 5 drinks for men, with effects reported as standardized mean differences (SMDs). The data extraction form is provided in Supplementary Tablefor transparency. 1

Secondary outcomes, such as weight change and glycaemic control, were also extracted but not pooled in the meta-analysis due to insufficient data. A summary of secondary outcome findings is provided in Supplementary Table. Authors were contacted for missing data (e.g., effect sizes, standard deviations). 2

Two reviewers (SG and BS) independently assessed the risk of bias, with discrepancies resolved by consensus. RCTs were evaluated using the Cochrane Risk of Bias Tool 2 (RoB 2), and observational studies were assessed using the Newcastle-Ottawa Scale (NOS) [10, 11]. Case series were evaluated using a modified NOS for the systematic review, but were excluded from the meta-analysis due to study design. Publication bias was planned using funnel plots and Egger’s test for observational studies, while the trim-and-fill method was selected for RCTs [12]. Egger’s test was not planned to be conducted in case of small number of small number of studies (k < 10).

Meta-analyses were conducted using the meta package in R (version 4.3.1). Continuous outcomes (e.g., alcohol consumption, craving) were pooled as standardized mean differences (SMD), and binary outcomes (e.g., alcohol-related events including AUD incidence/recurrence in observational studies, heavy drinking days in RCTs) were pooled as hazard ratios (HR) or odds ratios (OR) where sufficient data were available, using a random-effects model with Restricted Maximum Likelihood (REML) estimation and Hartung-Knapp (HK) adjustment to account for small numbers of studies.

Heavy drinking days were reported in one RCT but not pooled due to insufficient data. Heterogeneity was assessed using prediction intervals (PI), τ², and the Q test p-value, with sources explored through subgroup analyses (e.g., by GLP-1 RA type for RCTs, and by outcome type and GLP-1 RA type for observational studies).

Subgroup analyses were systematically conducted to investigate heterogeneity and effect modification, including: (a) GLP-1RA type (e.g., Semaglutide, Liraglutide, GIP/GLP-1RAs) for RCTs to assess craving reduction differences and for observational studies to evaluate outcome-specific effects, (b) outcome type (e.g., AUD incidence/recurrence vs. SUD hospitalization) for observational studies, and (c) baseline characteristics (e.g., AUD diagnosis, hazardous drinking status) in RCTs where data permitted, such as for craving subgroup analysis. These analyses employed random-effects models with REML and tested subgroup differences using the Q statistic, with significance set at p < 0.05.

Sensitivity analyses were performed using the DerSimonian-Laird (DL) method to compare with REML + HK, confirming the robustness of results across models (Supplementary Table S6). Additional sensitivity analyses excluded studies with a high risk of bias or small sample sizes (n < 15 per arm). For RCTs, high risk of bias was defined as ‘high’ or ‘some concerns’ in key domains of the Cochrane Risk of Bias Tool 2 (RoB 2), such as blinding or outcome measurement (Supplementary Table 3). For observational studies, high risk was defined as a Newcastle-Ottawa Scale (NOS) score less than 7. (Supplementary Table 4) Certainty of evidence was assessed using the GRADE approach [13]. (Supplementary Table 5)

Following the systematic search and screening process, nine studies were included in this meta-analysis, comprising three randomised controlled trials (RCTs) and six observational studies, as detailed in the PRISMA flow diagram (Fig. 1). The search identified 2,655 records from databases (Cochrane Library, PubMed, and Embase) up to May 3, 2025. After deduplication and exclusions, nine studies were included in the meta-analysis, encompassing 2,740,637 participants (430 from RCTs, 2,740,207 from observational studies). Additionally, three studies were included in the systematic review but excluded from the meta-analysis due to ‘alcohol consumption not pooled,’ reflecting incompatible data formats (e.g., median instead of mean) or insufficient reporting (e.g., missing standard deviations) for pooling, or study design (case series).

Baseline characteristics, including participant numbers per group, study country, and eligibility criteria, are summarized in Table 1. The RCTs included Hendershot et al. (2025) [14], Klausen et al. (2022) [15], and Probst et al. (2023) [16], all of which were conducted in outpatient settings with participants diagnosed with AUD. The overall risk of bias was low to moderate, with Hendershot et al. (2025) [14] showing moderate risk due to lack of blinding (open-label design), Klausen et al. (2022) [15] exhibiting moderate risk from potential missing outcome data, and Probst et al. (2023) [16] having moderate risk due to self-reported outcomes, as detailed in Supplementary Table 3.

The observational studies included Wium-Andersen et al. (2022) [17], Lähteenvuo et al. (2024) [18], Wang et al. (2024) [19], Qeadan et al. (2025) [20], Farokhnia et al. (2025) [21], and Xie et al. (2025) [22]. Most observational studies used propensity-score matching (PSM) or within-individual designs to assess alcohol-related outcomes in patients treated with GLP1-RAs, often in populations with co-morbid T2DM or obesity. The baseline characteristics of the studies included in the meta-analysis are summarised in Table 1. RCTs involved younger participants (mean age 42–48 years), with a higher proportion of males (60–70%), and focused on AUD-diagnosed individuals. Observational studies had a broader age range (mean age, 46–66 years) and included both AUD and non-AUD populations, often with comorbid conditions such as type 2 diabetes mellitus (T2DM) or obesity (mean BMI, 32–35 kg/m², where reported). Intervention durations varied, with RCTs ranging from 9 to 26 weeks and observational studies spanning 1 to 8.8 years (median). GLP1-RA-based therapies included semaglutide, liraglutide, dulaglutide, exenatide, and GIP/GLP1-RAs (e.g., tirzepatide), with comparators in RCTs being placebo and in observational studies often being non-GLP1-RA treatments, DPP-4 inhibitors, or no treatment. Outcomes assessed in the meta-analysis included alcohol consumption (SMD), alcohol craving (SMD), and heavy drinking days (OR) for RCTs, and alcohol-related events (HR/aHR) for observational studies.

Risk of bias assessments were conducted for all included studies (Supplementary Tablesand). For RCTs, the Cochrane Risk of Bias Tool 2 (RoB 2) indicated a low to moderate overall risk, with concerns primarily related to the blinding of participants due to the nature of the interventions. Observational studies, assessed using the Newcastle-Ottawa Scale (NOS), demonstrated good quality, with scores ranging from 7 to 8 out of 9; however, some studies had limitations in comparability due to residual confounding. The certainty of evidence was evaluated using the GRADE approach (Supplementary Table), with moderate certainty assigned to observational study outcomes and low certainty to RCT outcomes, due to heterogeneity and limited sample sizes. 3 4 5

Three randomised controlled trials (RCTs)—Hendershot et al. (2025) [14], Klausen et al. (2022) [15], and Probst et al. (2023) [16]—involving 430 participants diagnosed with alcohol use disorder (AUD) were analysed to assess the effects of glucagon-like peptide-1 receptor agonists (GLP1-RAs) on alcohol-related outcomes. Outcomes included total alcohol consumption, drinks per drinking day, alcohol craving, and heavy drinking days, as defined in the short-term focus of Section “Data extraction”. A random-effects model with Hartung-Knapp (HK) adjustment was employed to account for the small number of studies and potential heterogeneity.

The overall effect for total alcohol consumption across the three RCTs is shown in Fig. 2. The pooled SMD was − 0.24 (95% CI: -0.70, 0.23), indicating a slight, non-significant reduction in alcohol consumption favouring GLP1-RAs (p = 0.159), as determined by a random-effects model with Restricted Maximum Likelihood (REML) estimation and the Hartung-Knapp (HK) adjustment.

Heterogeneity was present, as indicated by the prediction interval (PI) of -0.90 to 0.43, with 54.1% of the observed variance (I²) reflecting variation in true effects between studies (τ² = 0.015, p = 0.113). For drinks per drinking day, the pooled SMD was − 0.23 (95% CI: -0.64, 0.19), indicating a slight, non-significant reduction (p = 0.169), with moderate heterogeneity (prediction interval = -0.79 to 0.33). (Supplementary Figure S1) The SMD for drinks per drinking day is based on country-specific definitions (e.g., 14 g in the US, 12 g in Denmark), with conversion to grams per day not possible due to data limitations. For alcohol craving (PACS), data were available from Hendershot et al. (2025) [14] and Klausen et al. (2022) [15], with a pooled SMD of -0.14 (95% CI: -2.84, 2.55, p = 0.156) and high heterogeneity (PI = -4.57 to 4.28, I² = 82.9%, τ² = 0.0763, p = 0.015). (Supplementary Figure S2) Potential self-reported outcome bias, referring to measurement error or recall bias in outcomes (e.g., alcohol consumption, craving) collected via participant self-reports, which may be influenced by social desirability or inaccurate memory, as assessed under the ‘bias in measurement of the outcome’ domain of the Cochrane Risk of Bias Tool 2 (RoB 2) (Supplementary Table 3), was noted in all RCTs. For heavy drinking days, data were only available from Hendershot et al. (2025) [14], with an OR of 0.84 (95% CI: 0.71, 1.00), indicating a non-significant reduction. This outcome was not pooled in the meta-analysis due to insufficient data, meaning it was reported in only one RCT, preventing the estimation of a pooled effect.

Six observational studies contributed 11 estimates assessing the effect of GLP-1RAs on alcohol-related events, which were pooled as hazard ratios (HRs) using a random-effects model with Restricted Maximum Likelihood (REML) estimation and the Hartung-Knapp (HK) adjustment. The pooled HR was 0.64 (95% CI: 0.59, 0.69, p < 0.001; I² = 19.1%, τ² = 0.003, p = 0.261), indicating that GLP-1RA use is associated with a statistically significant reduction in alcohol-related events. (Fig. 3) Alcohol-related events were defined as AUD incidence, AUD recurrence, SUD hospitalisation, or acute alcohol intoxication, with ‘AUD’ referring to incidence or recurrence unless specified (e.g., in subgroup analyses). In the included register-based study by Lähteenvuo et al., SUD hospitalisations were defined as a composite category encompassing both alcohol-related and non-alcohol-related substance use disorders; therefore, SUD hospitalisations inherently include alcohol-related admissions.

This review’s strength lies in its comprehensive inclusion of both RCTs and observational studies, providing a broad perspective on the effects of GLP1-RAs across diverse populations (N = 2,740,637). The use of random-effects models with Hartung-Knapp adjustment, detailed subgroup analyses, and funnel plot assessment enhances the robustness of the findings. The addition of separate AUD and SUD meta-analyses addresses concerns about pooling heterogeneous outcomes, improving causal interpretability and clinical relevance.

However, several limitations must be acknowledged. The small number of RCTs (n = 3) and their short durations (9–26 weeks) limit their power to detect significant effects, particularly in subgroup analyses by GLP-1RA type (Supplementary Figures S3, S4, S5), which involved only 2–3 studies, thereby reducing statistical power and interpretability. The SUD analysis (k = 2) is underpowered, and single-study data limit conclusions regarding intoxication. The subgroup analysis by GLP-1RA type reflects individual trial effect sizes without pooling, due to the single-study limitation per subgroup, which limits the ability to assess between-group differences. Observational studies, although larger, are still subject to residual confounding, despite the use of propensity-score matching (PSM) or within-individual designs in all six studies, as required by the eligibility criteria. Although these methods controlled for measured confounders (e.g., age, sex, BMI, diabetes status), unmeasured variables such as socioeconomic status, mental health comorbidities (e.g., depression, anxiety), or lifestyle factors (e.g., smoking, diet) may not have been fully accounted for, potentially influencing the observed significant decrease in alcohol-related events (HR: 0.64, 95% CI: 0.59, 0.69). The requirement for observational studies to use PSM or similar methods resulted in the exclusion of 24 studies (Fig. 1), which may have introduced selection bias by favouring more recent studies with advanced methodological designs. Publication bias for separate analyses was assessed via visual inspection of funnel plots (Supplementary Figures S14, S15), with no Egger’s test performed due to k < 10. Accordingly, the apparent protective associations from observational studies should be interpreted as hypothesis-generating rather than definitive evidence of a treatment effect.

Additionally, the assessment of publication bias for RCTs relied on a visual inspection of a funnel plot, which included only three studies (Supplementary Figure), thereby limiting the reliability of detecting asymmetry and necessitating a cautious interpretation. Publication bias for separate analyses was assessed via visual inspection of funnel plots (Supplementary Figures,), with no Egger’s test performed due to k < 10. Furthermore, the inability to conduct a statistical comparison of subgroup effects by GLP-1RA type (e.g., Semaglutide, GIP/GLP-1RAs) in observational studies, due to single-study subgroups, limits the robustness of the claim about ‘strongest effects.’ Similarly, the lack of statistical comparison for RCT subgroups by GLP-1RA type (e.g., Semaglutide) restricts the ability to confirm the ‘strongest effect’ claim beyond descriptive differences in SMDs. Secondary outcomes (e.g., weight change, glycaemic control) were extracted but not pooled due to insufficient data (Supplementary Table). S7 S14 S15 2

The high heterogeneity in the RCT meta-analysis for alcohol craving (PI = -4.57 to 4.28) and the low heterogeneity in observational studies (PI = 0.55 to 0.75) suggest variability in study populations, GLP-1 RA types, and outcome definitions.

Future research should prioritise larger, longer-term RCTs to confirm the effects of GLP1-RAs on alcohol consumption and craving, with a focus on specific GLP1-RA types like Semaglutide and Tirzepatide, which showed promising results in subgroup analyses. Mechanistic studies exploring the neurobiological effects of GLP1-RAs on alcohol reward pathways, such as dopamine signalling in the ventral tegmental area, could further elucidate their therapeutic potential [27]. Additionally, studies in AUD populations without co-morbid metabolic conditions are needed to isolate the direct effects of GLP1-RAs on alcohol-related behaviours. Future research should focus on larger RCTs to validate GLP-1RA efficacy in AUD and explore the mechanisms underlying their effects. Ongoing trials (Supplementary Table 7) will provide further RCT evidence to confirm these findings and address current limitations [28–37]. Current evidence, primarily from observational studies, does not yet support guideline recommendations, which should await robust RCT data.

The observational findings of this review, demonstrating a significant reduction in alcohol-related events (HR: 0.64), are consistent with those reported by Singal et al. (2025) [23], Subhani et al. (2024) [24], and Moraes et al. (2025) [25], which highlight GLP-1RAs’ potential in real-world settings. However, methodological differences distinguish these studies. Subhani et al.’s systematic review, which relies on narrative synthesis due to the limited availability of RCT data, aligns with our non-significant RCT trends (p = 0.159). However, their focus on consumption outcomes (e.g., heavy drinking days) contrasts with our event-based approach. Moraes et al.’s meta-analysis, pooling eight studies, reports a stronger HR (0.56, I² = 63.6%) for events, potentially reflecting their inclusion of liver-specific endpoints and higher heterogeneity, compared to our lower PI. Singal et al.’s commentary, citing observational and RCT evidence, underscores the dual benefit of AUD-ALD.

The potential of GLP1-RAs to reduce alcohol-related events warrants consideration in clinical guidelines, particularly for patients with AUD and co-morbid metabolic conditions. However, policymakers should balance the promising observational data with the need for RCT confirmation, ensuring that any adoption of GLP1-RAs for AUD is supported by robust safety and efficacy data [38]. The separate AUD and SUD analyses suggest GLP-1RAs may be particularly effective for chronic AUD outcomes, guiding targeted use in practice.

Not applicable.

The authors declare no competing interests.

Available upon reasonable request.