Methodology · statistical analysis
Statistical analysis and interpretation
Axelium helps reviewers assemble datasets, run supported models, inspect plots and diagnostics, pin evidence, and document the decisions behind a result. The reviewer remains responsible for model choice, assumptions, and final interpretation.
Analysis workspace
The Analysis tab combines a conversational assistant, run history, results panels, plots, data views, and evidence pinning. Use it to ask for a planned analysis, inspect the dataset, run checks, and preserve the outputs that should support the report.
Ask can quote saved pooled estimates and 95% confidence intervals, study counts, heterogeneity and linked GRADE ratings from any review page, including a read-only review. It reports the scope used by that analysis, which can differ from the data currently marked validated. Reading a saved result does not run a new analysis. Automatic GRADE ratings remain provisional until finalised by a reviewer.
Common analysis types
| Analysis | Use when | Reviewer checks |
|---|---|---|
| Pairwise meta-analysis | Two or more comparable studies estimate the same effect. | Effect direction, model choice, heterogeneity, and study quality. |
| Sensitivity analysis | You need to test whether conclusions depend on one study or modelling choice. | Influential studies, alternate assumptions, and protocol justification. |
| Publication-bias checks | Enough studies exist to make asymmetry checks meaningful. | Small-study effects, heterogeneity, and whether tests are underpowered. |
| Subgroup or meta-regression | A prespecified covariate may explain variation. | Multiplicity, sparse groups, ecological bias, and over-interpretation. |
Ask the Analysis assistant for a subgroup analysis by a study tag, an extracted variable or a study detail such as the journal. Each group gets its own pooled estimate, and the results show the test for subgroup differences. Studies with no value for the variable are left out of the comparison, and the assistant says how many. Each group needs at least two studies, and up to ten groups can be compared. A meta-regression needs at least three studies with a numeric value for the covariate, and reports the slope and its p-value.
Pairwise pooled analysis
Pairwise meta-analysis pools study-level effects into a summary estimate. Fixed-effect models assume a common underlying effect; random-effects models allow effects to vary across studies. Report the pooled estimate, confidence interval, heterogeneity, and why the model fits the protocol and data.
Ratio measures (risk, odds, hazard and rate ratios) are pooled on the log scale and reported back on the ratio scale, with no effect at 1: the model fit summary, subgroup rows and forest plot show a hazard ratio of 0.57, not its log of −0.55. The forest plot draws ratios on a log axis marked at familiar values such as 0.5, 1 and 2. Pooled proportions are reported as proportions and pooled rates as rates.
Advanced synthesis workflows
Network meta-analysis and ecology/evolution synthesis have additional assumptions beyond a standard pairwise model. Use the dedicated guides before interpreting these outputs.
Ecology/evolution reviews default to a pairwise random-effects model and switch to a multilevel random-effects model when effects are dependent, nesting those effects within studies (estimated by REML with t-based inference), with an optional additive meta-regression on recorded moderators. Results are reported on the effect measure's natural scale with its null anchor, and certainty is narrative rather than clinical GRADE.
Custom analysis and the Analysis Workbench
When a requested method falls outside the curated toolset — bespoke model structures, methods from R packages Axelium does not curate, or your own lab's methodology — the statistics assistant can run custom R code in an isolated workspace. The run executes against a frozen, versioned export of your analysis dataset (with a data dictionary and provenance fingerprint), so custom results are comparable with curated runs on exactly the same rows.
There are two ways to run custom work. A single custom script runs inline: you ask for a method, approve it, and the result returns in the same turn. Larger projects — multi-file analyses, porting a third-party package, simulation pipelines — can be handed to the Analysis Workbench, which works through the task step by step in the background and keeps its workspace warm across delegations. While a Workbench task runs, the chat shows live progress — elapsed time, how many steps have completed, and a clear done, paused, or failed state — so a longer job is never an opaque wait.
- Every custom run requires your explicit approval before any code executes, and analysis code runs with network access disabled. Package installation from CRAN and related public repositories happens in a separate, controlled step.
- Outputs return in a structured format — estimates, tables, plots — with a complete package manifest (renv lockfile and session information) recorded for reproducibility.
- Results are second-tier: they display with a “User-defined” badge, never feed GRADE or generated reports, and can be frozen as re-runnable recipes that reproduce the promoted estimates on the frozen data.
NOTE · Custom results stay walled off
Risk of Bias 2.0
Risk-of-bias assessment can be drafted from source evidence, but the final judgement belongs to reviewers. Treat each domain judgement as a methodological decision: check the supporting evidence, edit where needed, and resolve reviewer disagreements before using the result in certainty assessment.
WARN · Do not let unreviewed appraisals drive conclusions
GRADE and Summary of Findings
Quantitative pairwise reviews can produce Summary of Findings rows across GRADE domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias. These rows should be checked and finalised by reviewers. A rating reads Draft, In review or Finalised: select Start review, or change a factor, and Finalise appears; it asks you to confirm, and the rating then counts towards the Analysis freeze milestone. A finalised rating shows who finalised it and when; reopening it asks why, keeps the reason in the outcome's history, and returns it to Draft. NMA uses a CINeMA-style certainty review path; ecology/evolution certainty is narrative by default.
The risk-of-bias domain reads the appraisal checklist chosen for the analysis — the one marked primary when several are configured, in the exact form each study was assessed with — falling back to RoB 2.0 for reviews that never chose one.
It weighs each study by its inverse-variance weight in the pooled analysis and counts only assessments a reviewer has confirmed: the domain is rated down once when studies at high risk of bias carry at least 40% of the confirmed weight, and twice at 70%. When a study contributes results at several timepoints, each result is matched to the assessment of its own timepoint. The rating says how many studies it rests on; for analyses run before study weights were kept, it also says when each study had to count equally instead.
Imprecision counts the events in the studies pooled for that outcome, and checks the pooled 95% confidence interval on the scale the result is reported on. A ratio is read as a ratio, with no effect at 1: a hazard ratio interval of 0.50 to 0.66 does not cross 1, while an interval of 0.82 to 1.35 does, and the outcome is rated down once for it. A minimal important difference, when one is set for an outcome, is given on the same scale, such as 0.80 for a ratio.
Evidence pinning and reporting
Pin plots, model summaries, dataset views, and important explanations to the Evidence board so the final report can cite the exact outputs the team reviewed. If you rerun a model after changing data or assumptions, pin the updated evidence and explain the change.