2.1 Rationale

Glucagon-like peptide-1 (GLP-1) receptor agonists began as glucose lowering drugs for type 2 diabetes. Members of the class are now approved in several jurisdictions for chronic weight management, for cardiovascular risk reduction, for kidney outcomes in type 2 diabetes with chronic kidney disease, for obstructive sleep apnea with obesity and for steatohepatitis, and are studied in heart failure, osteoarthritis pain, type 1 diabetes and neurodegenerative, psychiatric and substance use disorders. The class now includes daily and weekly injectable peptides, an oral peptide, oral small molecules, and molecules that also act on GIP, glucagon or amylin pathways. Trials are published and registered faster than conventional reviews can be updated, and published network meta-analyses differ in scope, node definitions, time points and certainty methods, so their conclusions are hard to compare and quickly outdated.

A living network meta-analysis compares every regimen with every other within a population, including regimens never compared head to head, and is updated as evidence accrues (7,8). Publishing the full inventory of studies and documents with all code and data lets others verify, reuse and extend each step.

2.2 Objectives

Primary objectives, for each population in populations.csv that meets the connectivity rule (Section 3.8.2):

1. To estimate the comparative benefits and harms of all GLP-1 receptor agonist containing regimens, defined by molecule, route, formulation, maintenance dose, frequency and dosing strategy, against each other and against placebo, standard care, lifestyle intervention and other active treatments.
2. To rate the confidence in every network estimate for critical and important outcomes, and to draw conclusions with the GRADE minimally contextualized framework.

Secondary objectives:

3. To maintain a public, continuously updated inventory of every randomized and nonrandomized study of these drugs, linking each study to all of its publications, registrations, protocols, statistical analysis plans, regulatory documents, results postings, supplements and shared code and data.
4. To synthesize nonrandomized evidence on rare and long latency harms.
5. To inventory, appraise and compare published systematic reviews and meta-analyses of these drugs.
6. To keep all of the above current, with dated, citable versions.

3.1 Eligibility criteria

3.2 Information sources

Embase is licensed by the guarantor’s institution, but the guarantor decided not to search it, to keep the manual searching needed at every release sustainable alongside the other manual sources; this is a resource decision, not a lack of access. For randomized trials its omission is mitigated by CENTRAL, which incorporates trial reports identified from Embase, by Europe PMC and OpenAlex coverage of the journal literature, and by the trial registries. Two gaps remain and are reported as limitations: nonrandomized studies of harms and conference abstracts indexed only in Embase. Embase is added by amendment if a team member runs it once per release, logged in search/manual/. Every source, date, access route and count is logged per run in search/runs/<date>/run_log.json; manual searches in search/manual/.

3.3 Search strategy

One versioned concept file (search/strategies/concepts.yaml, version 1.3.1) lists every class term, generic name, brand name and development code, a mechanism block, PubMed controlled vocabulary and the Chinese terms; R/search_strategy.R renders it into each source’s syntax, and every run stores the exact query sent (Appendix A reproduces the current rendering). No design, language, date or outcome filters are applied.

• PubMed: title and abstract terms for every name and code (PubMed supports truncation inside quoted phrases: on 2026-10-12 "GLP-1 receptor agonist*"[tiab] returned 8,109 records, the same as the singular and plural phrases combined); the MeSH heading and pharmacological action “Glucagon-Like Peptide-1 Receptor Agonists”; drug headings (liraglutide, exenatide, semaglutide, tirzepatide) and supplementary concepts (iGlarLixi has none and is covered by free text); the agonists and analogs subheadings of “Glucagon-Like Peptide 1”, for records indexed before the class heading existed; and the mechanism block (terms such as dual agonist, triple agonist and co-agonist, combined with GLP-1, GIP, glucagon, incretin or amylin). On 2026-10-12 the mechanism block retrieved 190 PubMed records missed by the version 1.0.0 terms, including trials of UBT251, zenagamtide (amycretin) and DR10624, which were then added as agents; with version 1.2.0 it still adds 185 records. Development codes are kept even when they return no records today, because a living search must find their first report. The query is partitioned by Entrez date so that every partition stays below the 9,999 record cap of ESearch, and the union is reconciled against the total count (book records are counted and listed).
• Europe PMC: TITLE_ABS terms with explicit word forms (no truncation inside phrases), all sources except patents.
• OpenAlex: for the baseline (amendment A022), the public snapshot (JSON Lines works) is streamed file by file, prefiltered for distinctive words and matched exactly on the title and reconstructed abstract, normalized as the search engine tokenizes them, with the same terms, mechanism block and ambiguous code rule; the baseline is current to the snapshot date (amendment A025). Updates use the search.title_and_abstract parameter with the same terms in chunks plus the mechanism block, unioned on work identifier. The terms and counts of every step are stored. No part of the snapshot is kept except the matched works. OpenAlex works that are not reports (peer reviews, datasets, supplementary files, paratext, software, reference entries, standards, book reviews, data papers) are left out of the inventory.
• ClinicalTrials.gov: intervention search for every name and code, and a general term search for every name, the class terms and the mechanism block, so that registrations naming the drug outside the intervention field are found; the two are unioned, all statuses. Codes that also name unrelated things are searched in the intervention field and, in the bibliographic sources, only together with the mechanism context terms.

The strategy was assessed with the PRESS 2015 checklist (9) on 2026-10-13 by two language model reviewers working independently (GPT-6 Astra and Fable 5.1) in place of a human information specialist, a deviation from the plan. Their assessments, the response to every finding and the live counts that verified each added term are in protocol/peer-review/press-2026-10-13/; terms were verified by API counts, not by testing in each interface. In place of a Chinese-reading team member, the Chinese names were compared with official Chinese listings cited by one reviewer and with the other reviewer’s knowledge, which was not live verified; the review team has not verified them. Disputed and unverified names stay marked as candidates in concepts.yaml, and a name with no records in the first CNKI search is investigated, not removed automatically. Both substitutions are reported as limitations; a later PRESS review by an information specialist is logged as an amendment. Sensitivity is checked against a benchmark set assembled independently of the harvest: pivotal trials listed in FDA and EMA labels and assessment reports for every approved agent and indication; publications listed in ClinicalTrials.gov results records; and trials included in published systematic reviews (Section 5). Relative recall is reported per source and overall; every benchmark trial not retrieved leads to a documented correction of the concept file.

3.4 Study records

3.5 Data items

Study level: identifiers and linked records; design; registration date versus start date; countries and sites; sponsor and funding; dates; eligibility criteria; run in; blinding; randomization ratio; planned and actual duration; estimands and the strategy for each intercurrent event (treatment discontinuation, rescue medication, death); analysis population; missing data method; cluster or crossover features; conflicts of interest.

Arm level: node (molecule, route, formulation, maintenance dose, frequency, dosing strategy), titration schedule, background therapy, lifestyle co-intervention category (EM11), numbers randomized, analyzed and completing, and baseline characteristics (the effect modifiers in effect_modifiers.csv plus age, sex, race and ethnicity, weight and waist).

Outcome level: definition as reported and its mapping to outcomes.csv; time point; estimand; statistic (adjusted mean change with standard error, mean and standard deviation, events and denominator with follow-up or exposure time, hazard or rate ratio with confidence interval); survival estimator and competing risk hazard type; number with known status at $t$ by arm; observation period; competing deaths without a preceding target event and person years at risk by arm (time until the first target event, competing event, censoring or cutoff, as defined in 1); any reported hazard ratio for the competing event; digitized reference arm curves, where published, for both first events, or the target cumulative incidence with event free survival (Aalen-Johansen cumulative incidence, or the Kaplan-Meier survival of each first event with the other censored), with the estimator of each curve, the published numbers at risk and any reported joint covariance of the curve ordinates (requested from authors, with their ordinate estimates, when not published) (a target curve alone is recorded only as a cumulative incidence); source location; extractors.

Nonrandomized studies: the items in harms_nrs.csv (design, data source and period, exposure and comparator definitions, time zero, latency, follow-up and censoring, confounders adjusted for and method, handling of immortal time and competing risks) and adjusted and unadjusted estimates.

3.6 Outcomes and prioritization

Importance. Ratings in outcomes.csv follow GRADE and were set by the guarantor; no clinical advisor or patient partner took part, which is reported as a limitation. They are locked at registration, before any outcome data are extracted, and change only under the rule for thresholds below.

Time windows. windows.csv gives each outcome a primary window (and for some outcomes secondary windows) per population, with a selection rule, a target week and limits (lower inclusive, upper exclusive; a blank upper limit means no upper limit). A window applies to an outcome and population only when the outcome’s population_ids include that population. Windows for the same outcome and population never overlap, so every observation belongs to at most one window, and each outcome and population has exactly one primary window. Under nearest_target, the observation closest to the target week within the limits is used, and ties go to the later observation; under longest_follow_up (W00 for event outcomes and harms; W25 for eGFR slope), the longest randomized follow-up at or above the lower limit is used. Outcomes measured per procedure (O36A, O36B) or per pregnancy (O120) have no week window. Each window is analyzed as its own network, with trial duration within the window as a meta-regression covariate. The narrow window sensitivity analysis keeps only observations within 4 weeks of the target. For type 2 diabetes the primary window for HbA1c and weight in kilograms (O01, O02) is 36 to 64 weeks (target 52), because newer agents reach their maintenance doses only after 16 to 24 weeks of escalation, so 26 week values understate their maintenance effects; the 20 to 36 week window (target 26) is a prespecified secondary analysis. Percent weight change (O10) uses 40 to 104 weeks (target 68); its short trial version O10B, used in P12 and P15, uses 12 to 52 weeks (target 24).

Composite definitions. Within any outcome identifier, trials are pooled only when their component sets and thresholds are identical; otherwise the outcome is analyzed in definition strata, each its own network. Kidney composites illustrate the rule: O22 (50 percent eGFR decline), O22B (40 percent), O22C (kidney failure, sustained doubling of serum creatinine, or kidney death), O22D (any composite including cardiovascular death) and O22E (composites including albuminuria or smaller declines without cardiovascular death) are never pooled with each other; a composite that includes cardiovascular death is O22D even when it also includes albuminuria or a smaller eGFR decline. The pre 2017 hypoglycemia mapping in O04 is a prespecified exception to this rule; mapped trials are flagged and excluded in a sensitivity analysis.

Event outcomes and harms. Time to event outcomes use reported (or validly reconstructed) log hazard ratios over the longest randomized follow-up. For an outcome with a competing event, complete status counts (rule below) enter only its subdistribution network and are never pooled with cause specific estimates. For an outcome with no competing event (at present only all cause mortality, O20), reported hazard ratios, Kaplan-Meier contrasts and complete status counts have hazard type not applicable and enter the same network, which is primary; for an outcome with a competing event a Kaplan-Meier contrast remains a cause specific input. A trial with more than one eligible input to a network contributes one, in this order: the reported hazard ratio, the cumulative incidence or Kaplan-Meier contrast, the arm counts; the alternatives are logged.

Binary harms and adverse events (model_scale log_hazard_cloglog) are analyzed in two prespecified analyses, each rated separately.

Known status. Status at time $t$ is known when the target event occurred by $t$, a documented competing death occurred by $t$ without a preceding target event (counted as event free; complete status counts are therefore subdistribution inputs for an outcome with a competing event, and not applicable inputs for O20), or observation confirms absence of the target event through $t$. Loss or administrative censoring before $t$ without an established target or competing event leaves status unknown. Participants with competing deaths remain in the randomized denominator.

• Hazard ratio analysis. A reported hazard ratio enters as a normal likelihood on the log hazard ratio. A reported cumulative incidence at a common time $t$ in each arm enters as the contrast of $\log(-\log(1 - p))$ between arms, a normal likelihood with a delta method standard error from the confidence intervals, only when $0 < p < 1$ with finite uncertainty. Arm counts enter a binomial likelihood with a complementary log log link and an offset of $\log t$ (years) only when status is known for every randomized participant at $t$, the longest follow-up at which that holds. In a sensitivity analysis, $t$ is instead the longest follow-up at which status is known for at least 95 percent of randomized participants in every arm; unknown participants remain in the denominator, count as event free, and the event counts by that $t$ enter the same likelihood as a sensitivity analysis of the subdistribution network (or of the not applicable network for O20), never of a cause specific network. There is no general bound or direction of the resulting bias, even with equal missing fractions, because missingness may depend on the outcome. In an illustrative calculation with missingness independent of the outcome and occurring at baseline, equal 5 percent loss changes a hazard ratio of 2.0 (risks 0.50 and 0.75) to about 1.93, while 5 percent loss from the control arm only changes 1.20 (risks 0.010 and 0.012) to about 1.26.
• Recorded harms analysis. All trials’ counts in the same complementary log log model with offset $\log t$, where $t$ is the nominal observation period in years. The reported quantity is the exponentiated contrast $Q = \{-\log(1 - r_B)/t_B\}/\{-\log(1 - r_A)/t_A\}$, where $r$ is the recorded event proportion; $Q$ is interpreted descriptively, without an incidence rate or hazard interpretation. No pooled absolute effect is derived from it; each trial’s recorded proportions and within trial risk difference are displayed with its observation period and follow-up completeness. Magnitude dependent certainty domains use the range 0.8 to 1.25 on $Q$, and the rating states the scale. Incomplete follow-up is extracted by arm and judged in RoB 2 domain 3, and a sensitivity analysis excludes trials with more than 20 percent incomplete follow-up or a difference above 10 percentage points between arms.

Conclusions about a binary harm or a time to event outcome use the primary hazard type network when it meets the connectivity rule; otherwise the other hazard type network, labeled with its hazard type; otherwise, for binary harms, the recorded harms analysis, labeled as such, and for time to event outcomes, descriptive reporting. For O20 the not applicable network is the only network: it is used when it meets the connectivity rule, otherwise results are reported descriptively.

Denominators are the number randomized; the safety population is used only when the randomized number is not reported, and those trials are flagged.

Competing risks. For every input to a hazard scale analysis, the estimator and the hazard type are recorded: cause specific (a Cox model with competing deaths censored, or a contrast between arms of one minus Kaplan-Meier estimates with competing deaths censored, which enters as a normal likelihood on the contrast of $\log(-\log S)$ at a common $t$, the difference in log cumulative cause specific hazards, where $S$ is the Kaplan-Meier estimate, with a delta method standard error from the reported confidence intervals or the Greenwood variance, only when $0 < 1 - S < 1$), subdistribution (a Fine and Gray model, or a contrast derived from cumulative incidence that counts competing deaths as event free, including arm counts with complete status), or not applicable (only for outcomes with no competing event, at present only O20; for every other outcome the hazard type follows the estimator even when no competing event occurred in the trial). Hazard type is part of the network key; cause specific and subdistribution estimates never share a network; a trial reporting inputs of both types contributes each to its own network. For a time to event outcome with a competing event the cause specific network is primary, because outcome trials report Cox models; for O20 the not applicable network is the only network and is primary; for binary harms the subdistribution network is primary, because counts and cumulative incidence estimate it. Subdistribution networks assume proportional subdistribution hazards: reported proportionality assessments are extracted and available curves inspected, and where material nonproportionality is found, horizon specific cumulative incidence contrasts are reported separately and no constant subdistribution hazard ratio is used to convert effects across horizons. One minus a Kaplan-Meier estimate with competing deaths censored is never used as an absolute or reference risk for an outcome with competing events.

Absolute risks for outcomes with competing events are cumulative incidences that account for competing deaths; a single event rate is never sufficient. Let $h_{1a}(u)$ and $h_{2a}(u)$ be the cause specific hazards of the first target event and of the competing event in arm $a$. Survival free of both events is $S_a(u) = \exp[-\int_0^u \{h_{1a}(v) + h_{2a}(v)\}\,dv]$ and the cumulative incidence of the target event is $F_{1a}(t) = \int_0^t S_a(u)\,h_{1a}(u)\,du$. For the treated arm the target hazard is multiplied by its cause specific hazard ratio and the competing hazard by the cause specific hazard ratio for the competing event when one is reported; otherwise the competing hazard is unchanged, which assumes no treatment effect on it and is stated. The reference hazards come from the covering reference arms of Section 3.8.14 by one of two routes.

• Piecewise integration requires both cause specific hazards over the analyzed horizon with valid joint uncertainty, from every covering reference arm, on reported intervals whose boundaries at or before the analyzed horizon are identical across all covering arms, a boundary after the horizon counting as the horizon; otherwise this route is unavailable. Reported intervals are those of reported events and person time (the last may end after the analyzed horizon, and its rates then apply up to the horizon) or, for digitized curves, the intervals between the times of the published numbers at risk (with the last interval ending at the analyzed horizon when the curve extends to it). Inputs on each interval are either the events and person time at risk for both the target and the competing first event, or digitized curves from the same arm and risk set that identify both hazards: cumulative incidences of mutually exclusive first events (Aalen-Johansen curves of both events, or of the target event with event free survival), or Kaplan-Meier curves of each first event with the other censored. A target curve alone does not qualify, and a mortality curve that includes deaths after a nonfatal target event cannot substitute for the competing first event curve. Hazards are constant within each interval; on an interval from $t_s$ to $t_e$ of length $\Delta$ (in years) they are: from counts, the event count of each type over the person time at risk; from Aalen-Johansen curves, with $S = 1 - F_1 - F_2$ (or $F_2 = 1 - S - F_1$ when event free survival is published instead), $H = -\log\{S(t_e)/S(t_s)\}/\Delta$ and $h_k = H\,\Delta F_k/(\Delta F_1 + \Delta F_2)$, requiring $0 < S(t_e) < S(t_s)$; from a Kaplan-Meier curve of event $k$ with the other event censored, $h_k = -\log\{S_k(t_e)/S_k(t_s)\}/\Delta$, requiring $0 < S_k(t_e) < S_k(t_s)$. These forms are exact under the stated constancy.

  Within arm uncertainty. For reported first event counts with valid person time, the working normal approximation for the log hazards from the piecewise exponential likelihood has variance approximately $1/d_k$ ($d_k$ the count of event $k$ on the interval) and zero covariance across events and intervals, under its standard assumptions of independent participants and noninformative censoring; a zero count for either event makes the route unavailable. Here $y_a$ is the vector of estimated log hazards, one coordinate per event and common interval, in that order, and for counts $V_a$ is diagonal with those variances. These routes keep the working normal likelihood for counts, unlike Section 3.8.14, so that count and curve arms share one multivariate normal likelihood. For curve inversion, event counts are not inferred from curve increments and numbers at risk, because within interval censoring leaves them unidentified; a reported or author supplied covariance $\Sigma$ of the ordinates that enter the formulas is required, for both events and every interval end used, together with the authors’ estimates of those ordinates when $\Sigma$ is author supplied (digitized values are inverted only with a published covariance at the same times); with event free survival the ordinates are the target cumulative incidence and the event free survival. With $g$ the map from these ordinates to the arm’s vector of log hazards (both events, every interval) and $J$ its Jacobian, the within arm covariance is $V_a = J \Sigma J^{\mathsf{T}}$.

  Pooling. Each arm’s vector $y_a$ follows $y_a \sim \mathrm{N}(\mu + Z u_a, V_a)$, where $\mu$ holds the reference log hazard of each event on each common interval, $u_a = (u_{a1}, u_{a2})$ are arm level random effects, one per event, shared across that event’s intervals (a proportional deviation of the arm’s hazard function), $Z$ maps them onto the vector, and $u_{ak} \sim \mathrm{N}(0, \tau_k^2)$ independently across arms and events, with priors PR15 on $\mu$ and PR16 on $\tau_k$ (priors.csv). With one covering arm, $\tau_k$ is not estimated and $u_{\mathrm{new}}$ is not drawn: each draw of the reference log hazards is a draw from $\mathrm{N}(y_a, V_a)$. With two or more covering arms, each draw of the reference log hazards is $\mu + Z u_{\mathrm{new}}$, with $u_{\mathrm{new},k} \sim \mathrm{N}(0, \tau_k^2)$ independently for the two events at that draw’s $\tau_k$. The hazards are the exponentials of these log hazards and are integrated jointly with the formulas for $S_a$ and $F_{1a}$; the displayed reference cumulative incidence is the median of these integrated draws and its uncertainty comes from the same draws. Scenario values remain the empirical 25th and 75th percentiles of the covering arms’ own cumulative incidences (Section 3.8.14); their uncertainty comes from drawing each arm’s vector from $\mathrm{N}(y_a, V_a)$ independently across arms, exponentiating, integrating each arm’s cumulative incidence and recomputing the percentiles in each draw. When any covering arm lacks qualifying inputs, positive finite hazards or a finite covariance, this route is unavailable and another eligible route is kept; no constants are added and no interval or arm is omitted.

• Exponential approximation. A piecewise fit whose common reported interval is a single interval from 0 is the exponential approximation. When the piecewise route is unavailable, the exponential approximation is also available if every covering arm supplies both event counts and person time over one period that starts at 0 and reaches or passes the analyzed horizon (an arm that reports finer count intervals contributes their sums); each arm’s constant rates over its own period give its log hazard coordinates on one interval from 0 to the horizon, pooled with the model above. In both cases the route is used only when constant hazards are supported, as judged below, and it is labeled as such. Here $d_k$ is the arm’s count of event $k$ over its period (variance approximately $1/d_k$; a zero count for either event makes the route unavailable), $\lambda_k$ is the exponential of the log hazard of event $k$, and $F(t) = \lambda_1/(\lambda_1 + \lambda_2)\,(1 - e^{-(\lambda_1 + \lambda_2)t})$ is computed for each draw, with $t$ in years.

• Cause specific hazard ratios: the reference risk shown with the estimate is $F_{1a}(t)$ of the reference node computed by the same route as the treated risk (piecewise integration, otherwise the exponential approximation), so that both come from one source; when neither route is available, only relative effects are reported.

• Subdistribution hazard ratios: the treated risk is $1 - (1 - F_{\mathrm{ref}})^{\mathrm{HR}}$, where $F_{\mathrm{ref}}$ is the reference risk at the analyzed horizon: the reported Aalen-Johansen cumulative incidence; otherwise the piecewise integration; otherwise the exponential approximation under supported constant hazards. Conversion requires supported proportional subdistribution hazards; constant cause specific hazards are required only when $F_{\mathrm{ref}}$ comes from the exponential approximation, which then carries its label. When no $F_{\mathrm{ref}}$ is available, the relative only rule of Section 3.8.14 applies.

Rates are events per person year at risk: the total time until the first target event, competing event, censoring or analysis cutoff, whichever comes first, over the same interval as the events. The number randomized times the mean follow-up is used only when that mean is exactly this time at risk; general study follow-up, treatment exposure and median follow-up are not substitutes. When valid person time is unavailable, rates are not estimated by division, and another supported route or the relative only rule applies.

Both judgments (material nonproportionality, and support for constant hazards) are made by two reviewers before relative effects are modeled and logged with reasons, with the alternative as a sensitivity analysis: constant hazard support from the reference arm cumulative incidence curves where published, otherwise from cumulative incidences or event rates reported at two or more time points (when neither exists the exponential approximation is unavailable; any other eligible route is kept, including a reported Aalen-Johansen cumulative incidence for subdistribution conversion, and relative effects only are reported when no applicable route qualifies), and proportionality from reported tests and the published arm curves; nonproportionality is material when a reported test rejects proportionality at the 5 percent level or the arms’ cumulative incidence curves cross after the first quarter of follow-up.

Procedure related outcomes (O36A, O36B) are risks per procedure on the odds ratio scale. Recurrent hypoglycemia (O04B) uses only reported log rate ratios with their standard errors.

Estimands. Each reported result is classified by its strategy for every intercurrent event (treatment policy, hypothetical, while on treatment, composite, principal stratum) following ICH E9(R1) (10). Primary networks use treatment policy results only. Hypothetical (trial product) results form separate networks when enough trials report them; estimands are never substituted for one another in the same network.

Effect measures. Mean differences in natural units; standardized mean differences for O24B (physical functioning) and O24C (total HRQoL), each in its own network; hazard, odds and rate ratios as set by model_scale. Absolute effects are computed for every binary and time to event outcome where a valid conversion exists (Section 3.8.14).

3.7 Risk of bias in individual studies

• Randomized trials: RoB 2 for every result in a synthesis, with the cluster and crossover variants (11). The effect of interest follows the causal question of the network: the effect of assignment for treatment policy networks and the effect of adhering for hypothetical networks. Two reviewers assess independently and reach consensus.
• Nonrandomized studies: ROBINS-I V2 for cohort designs, recording the exact release used (12); ROBINS-I 2016 with documented adaptations for case control and self controlled designs (13). ROBINS-E (14) is not used because these studies estimate effects of interventions. Results at critical risk of bias are excluded from quantitative synthesis and described.
• Missing evidence: ROB-MEN for network estimates (15) and ROB-ME for pairwise syntheses (16), informed by registered trials without results and by registered outcomes not reported.

Judgments, signaling question answers and supporting quotations are stored as data.

3.8 Data synthesis

3.9 Meta-biases

Risk of bias due to missing evidence is judged with ROB-MEN for network estimates and ROB-ME for pairwise ones (15,16), using registered trials without results, registered but unreported outcomes, discrepancies between registered and published primary outcomes, and comparison adjusted funnel plots when at least ten trials inform a comparison with a common reference (36).

3.10 Confidence in cumulative evidence

Every network estimate of every critical and important outcome is rated with CINeMA in six domains (37,38), separately for the Bayesian and the frequentist estimates of the same evidence; where the two ratings differ, the lower is reported. Contribution matrices must represent the studies and network behind the estimate being rated. For hazard scale estimates the contribution network uses, for each trial, the input that entered the primary model: the reported log hazard ratio; or the cumulative incidence or Kaplan-Meier contrast with its delta method variance (no continuity correction); or a complementary log log estimate from the arm counts, adding 0.5 to every cell of a trial with a zero event or all event arm. The correction is used only for contributions, which depend on precision weights and network geometry, and never enters an effect estimate; every trial in the primary estimate is kept and shared arm covariance is preserved. When valid contributions still cannot be computed, within study bias and indirectness are judged qualitatively across the full evidence set, with the supporting studies and reasons recorded. Restricted sensitivity analyses (for example odds ratios within 20 to 104 weeks of follow-up) are rated separately and never substitute their rating for that of the primary estimate. Within study bias, reporting bias and indirectness are judged once per evidence set from these contributions; imprecision, heterogeneity and incoherence are judged on each framework’s own estimates. Numeric rules:

• Within study bias and indirectness: risk of bias (low 1, some concerns 2, high 3) and indirectness (low 1, moderate 2, high 3) are averaged over studies weighted by their contribution to the estimate; no concerns below 1.5, some concerns from 1.5 to below 2.0, major concerns at 2.0 or above.
• Reporting bias: the ROB-MEN judgment (low risk: no concerns; some concerns; high risk: major concerns).
• Imprecision: against the range of little or no difference set by the threshold (absolute scale at the horizon; the ratio range 1/1.25 to 1.25 for rare outcomes, for rate outcomes synthesized from reported rate ratios and for estimates under the relative only rule of Section 3.8.14; the range 0.8 to 1.25 on $Q$ for the recorded harms analysis): no concerns when the 95 percent interval lies entirely within the range or entirely beyond one threshold; some concerns when it crosses one threshold; major concerns when it crosses both.
• Heterogeneity: the same rules applied to the 95 percent prediction interval, counting only crossings that the confidence or credible interval does not already make.
• Incoherence: the 95 percent interval of the direct minus indirect difference (node split or SIDE; for ratio scales, the difference of log effects compared with plus or minus log 1.25, or with the threshold converted to the relative scale at the reference risk) is compared with plus or minus the threshold: no concerns when it lies entirely within, major concerns when it lies entirely outside, some concerns otherwise. When incoherence cannot be estimated (no closed loops or no identifiable inconsistency parameters), the limitation is recorded and the judgment is made from transitivity and network structure, with the reason stated.

Each domain lowers confidence by one level for some concerns and two for major concerns, starting from high, without double counting across domains. Thresholds in outcomes.csv were set by the guarantor as operational judgments and are locked at registration, before any outcome data are extracted; they are not validated minimal important differences, and the basis of each is not documented, which is reported as a limitation. A sensitivity analysis repeats the imprecision, heterogeneity and incoherence judgments of critical outcomes with every threshold multiplied by 0.5 and by 2: absolute thresholds directly, and ratio thresholds on the log scale (a ratio range of 1/1.25 to 1.25 becomes 1/1.118 to 1.118 and 1/1.563 to 1.563), including relative only estimates. A later change proposed by clinical advisors or patient partners is an amendment that applies only to outcomes with no data yet extracted, recorded with the before_or_after_results flag; otherwise the proposed thresholds are reported as a sensitivity analysis. Every judgment and its reason are stored as data.

Conclusions follow the GRADE minimally contextualized framework (39–41). Nonrandomized harms evidence is rated with GRADE using ROBINS-I, starting at high certainty and rating down for risk of bias (42).