How to read the results
Each analysis page answers one question: for people with a given condition, how do the treatments compare on one outcome over one period of time? This guide explains each part of a page, first in plain language and then for researchers.
Three things to check first
- The question. The condition, the outcome, the time window (for example weeks 20 to 32) and the estimand (whether results count people who stopped the medicine) are at the top of every page. Two pages about the same medicine can answer different questions.
- The size of the effect and its uncertainty, not only whether it is "significant". A medicine that lowers weight by 2 kg on average is not the same as one that lowers it by 15 kg, even if both are clearly better than placebo.
- How much evidence there is. The number of trials and participants, and whether the trials were comparable, are shown on each page. A comparison resting on one small trial is less certain than its interval may suggest.
What is a network meta-analysis?
Most trials compare a medicine with a dummy treatment (placebo) or with one other medicine. A network meta-analysis combines all of these trials at once, so medicines that were never tested against each other can still be compared through the trials they share. Think of it as working out who is taller when some people were only ever measured next to a common friend.
A network meta-analysis estimates all pairwise relative effects in a connected network of randomized comparisons, combining direct and indirect evidence under the transitivity and consistency assumptions. Here, Bayesian random effects models (multinma, Stan) are the primary analysis with a frequentist graph theoretical model (netmeta) as a check.
What does "living" mean?
New trials of these medicines are published every month. A living review keeps searching, adds each new trial as it appears and updates the results, so the answers here stay current instead of going out of date.
Searches run on a fixed schedule; newly eligible trials pass dual screening, dual extraction and adjudication before each release. Every release refits all networks with fixed seeds, and the protocol, amendments and data are versioned in the public repository.
How to read the effects plot
Each row is one medicine compared with the reference (usually placebo). The dot is the best estimate; the thick line shows where the true effect most likely lies; the dashed line shows the range a new trial might find. If a line does not cross the vertical line of no difference, the medicine very likely differs from the reference.
Rows show posterior medians with 95% credible intervals (thick) and 95% prediction intervals (dashed) from the Bayesian random effects model, with the frequentist estimate and 95% confidence interval (diamond) underneath. Ratios are drawn on a log axis.
Credible and prediction intervals
The credible interval is the range where the average effect most likely lies, given all the trials. The prediction interval is wider: it is where the effect in a new, similar group of people would likely fall, because results vary from trial to trial.
The 95% credible interval summarizes the posterior of the mean relative effect. The 95% prediction interval adds between trial heterogeneity (tau) and describes the effect expected in a new study setting.
How to read the league table
Every cell compares the medicine in its row with the medicine in its column. Colored cells show comparisons where the evidence points clearly one way.
Cells give posterior medians and 95% credible intervals of the row treatment versus the column treatment on the natural scale. Shading marks intervals excluding the null.
Rankings and SUCRA
Rankings show how often each medicine came out best, second best and so on across many simulated versions of the evidence. They are a summary, not a verdict: two medicines with very close results can rank differently by chance.
Rank probabilities come from posterior draws; SUCRA is the surface under the cumulative ranking curve (0 to 1). Rankings ignore the size of differences and their uncertainty, so read them only alongside the effect estimates.
Heterogeneity
Trials of the same medicines rarely find exactly the same result, because the people, doses and durations differ. Heterogeneity measures how much the results disagree beyond what chance would explain.
Tau is the between study standard deviation of relative effects (random effects model, informative prior from Turner or Rhodes where available). I squared and Cochran Q come from the frequentist model.
Inconsistency
When two medicines were compared both directly in head to head trials and indirectly through placebo, the two answers should agree. Inconsistency checks whether they do.
Global inconsistency is assessed with the design by treatment interaction Q statistic; local inconsistency by node splitting (direct versus indirect evidence per comparison).
Are the trials comparable?
Indirect comparisons are fair only if the trials behind them enrolled similar people and measured outcomes the same way. This section checks the things that can change how well a medicine works, such as starting blood sugar or body weight.
Transitivity is assessed across designs for eligibility, outcome definition and assessment, and the baseline distribution of prespecified treatment effect modifiers and prognostic factors (protocol/tem_pf.csv), with thresholds for clinically important imbalance.
Baseline characteristics
These are the people who took part in each trial when it started: their average blood sugar, weight, age and so on. Large differences between trials can make comparisons less fair.
Per trial means or percentages weighted by arm size for each effect modifier or prognostic factor that applies to this outcome and population; a stated stand in (median for a mean, derived share) is used when the exact statistic is not printed.
Estimands
Trials answer slightly different questions. Some ask what happens when people are offered a medicine, whether or not they keep taking it. Others ask what would happen if everyone took it as prescribed. These are kept apart because they give different numbers.
Results are grouped by estimand (ICH E9 R1): treatment policy, hypothetical, while on treatment and composite strategies for intercurrent events. Networks never mix estimands.
Run your own analysis
You can rerun any network here in your browser: switch medicines or trials off and on, choose the reference and see how the results change. Nothing is sent anywhere; the calculation runs on your device.
The lab fits contrast based models to the network's trial data: a frequentist random effects model (weighted least squares with multi arm covariance, DerSimonian Laird tau) and a Bayesian random effects model (collapsed Gibbs sampler with a slice step for tau, half normal prior). Binary data use log odds ratios with a 0.5 continuity correction for zero cells. Published results remain those of the R pipeline.
What these results cannot tell you
- How one person will respond. The results are averages across trial participants.
- Whether a medicine is right for you. That depends on your health, other medicines, preferences, cost and access. Talk to a clinician.
- Effects in people who were not in the trials, or over longer periods than the trials lasted.
- Until risk of bias and certainty ratings are complete, how much to trust each estimate. These interim results show the evidence, not a graded recommendation.
Unfamiliar word? See the glossary. For the full statistical methods, see Methods.