Methods

For each condition, how do GLP-1 receptor agonist containing medicines compare with each other, with placebo and with other treatments on benefits and harms? Conditions (populations) are analyzed separately: type 2 diabetes, overweight and obesity, weight maintenance, children and adolescents, type 1 diabetes, prediabetes, heart failure, chronic kidney disease, steatotic liver disease, sleep apnea, knee osteoarthritis, polycystic ovary syndrome, neurodegenerative disease, substance use disorders and antipsychotic associated weight gain.

Randomized controlled trials (parallel, crossover and cluster) of any GLP-1 receptor agonist containing medicine, including co-agonists (GIP and GLP-1, glucagon and GLP-1, triple agonists), oral peptides and small molecules, and fixed ratio combinations with insulin, against placebo, standard care, lifestyle intervention or any active comparator. No language or date limits.

Bibliographic databases and trial registries are searched automatically every week with versioned strategies; each run stores its exact query and counts. Records are deduplicated and linked to the trial they report, so that every report, supplement and registry posting of one trial is used together.

Every record is screened by two independent reviewers (full texts by language model reviewers under amendment A021, with disagreements adjudicated). The data of every included trial are extracted from all of its reports, supplements and registry postings by two independent extractors (by amendment A023, language model extractors working in separate sets under a fixed extraction manual with controlled vocabularies). Each value carries its source location (record, table or figure). A third, coordinating reviewer compares the two extractions field by field and adjudicates every disagreement against the source; rulings are logged and published.

Only aggregate data printed in trial reports and registries are used and shown. Full texts are never published on this site.

A network is defined by population, outcome, time window and estimand. Treatments enter as nodes at the level of dose, route and frequency, and in a second analysis with the doses of each medicine pooled. Analyses need a connected network of at least two trials; smaller or disconnected sets are reported descriptively, and disconnected parts are analyzed separately.

Continuous outcomes are analyzed as mean differences (change from baseline preferred); binary outcomes as odds ratios; time to event outcomes as hazard ratios, using complementary log log models with a log follow-up offset for event counts. Standard deviations are derived from standard errors, confidence intervals or P values where needed; imputed and converted values are flagged on each page.

Random effects network meta-analysis in a generalized linear model framework, fitted with multinma in Stan: g(θ_jk) = μ_j + δ_jk, with δ_jk ~ N(d_tjk − d_tj1, τ²), consistency d_ab = d_b − d_a, and covariance τ²/2 within multi-arm trials. Arm level and contrast level data are combined in one model.

Priors: normal priors for trial baselines and relative effects scaled by a reference standard deviation per outcome; half normal priors on τ for continuous outcomes; for binary and time to event outcomes, the log normal predictive distributions of Turner and colleagues, with the pharmacological versus placebo or pharmacological versus pharmacological type chosen from the network geometry. Four chains of 2,000 warmup and 2,000 sampling iterations; convergence requires R-hat below 1.01, bulk and tail effective sample sizes of at least 400 and no divergent transitions.

The same networks are fitted with netmeta (graph theoretical model): random effects with a common heterogeneity variance estimated by REML, with 95% confidence and prediction intervals, P-scores, node splitting (SIDE), the design by treatment decomposition of Q, the contribution matrix (netcontrib) and comparison adjusted funnel plots. For harms analyzed on the hazard scale, the frequentist check uses odds ratios of trials with 20 to 104 weeks of follow-up.

For every network, eligibility, outcome definitions and assessment, and the baseline values of prespecified effect modifiers and prognostic factors are compared across designs (the set of treatments each trial compares); imbalance is flagged on each page. Consistency is checked globally (between designs Q) and locally (node splitting); a nonsignificant test is not taken as proof of consistency in sparse networks. See effect modifiers.

SUCRA (Bayesian) and P-scores (frequentist) are secondary summaries, always shown with the effects themselves; rankings are never used alone to draw conclusions.

Searches run weekly, screening monthly, and every release refits all models on the cumulative data with the original priors and fixed seeds. No sequential testing is done; results are estimates, not tests. Large new trials, the first trial of a new treatment and regulatory safety communications trigger an earlier release.

The NMA lab refits any analysis in the browser with TypeScript ports written from the R source of netmeta (pairwise contrasts, the multi-arm reweighting, the graph theoretical estimator, DerSimonian and Laird and REML heterogeneity, P-scores, prediction intervals) and multinma (the same likelihoods, priors and random effects parameterization, sampled with a No-U-Turn sampler following Stan's adaptation). Validated against the published fits: the frequentist port reproduces Q in every network and τ and the effects to four decimals in all but one network; the Bayesian port agrees within Monte Carlo error.

The results on this site are interim. Risk of bias assessments and certainty ratings (CINeMA) are in progress and will be added to each estimate; until then, read each result with its number of trials, its prediction interval and the comparability notes on its page.

R version 4.6.0 (2026-04-24); multinma 0.9.1; netmeta 3.7.0; rstan 2.39.0.9000. Seed 20261027; 4 chains of 2000 warmup and 2000 sampling iterations. Every release records the checksums of its inputs. Code: github.com/choxos/LivingGLP1.