Meta-analysis and quantitative evidence synthesis

Meta-analysis covers the statistical part of a systematic review, the quantitative synthesis: starting from the data extraction table the client provides, the service carries out the effect size calculation, the random-effects models, the heterogeneity assessment, the publication bias analysis, the subgroups, the meta-regression and the sensitivity diagnostics. This applies if you have already done, or have under way, the literature search and data extraction from the primary studies, and need the analytical part with a methodology that holds up in peer review. the client provides the data extraction table with the data already collected; the service carries out the quantitative analysis and prepares the results for publication. The literature search and data extraction are not included: they are the earlier stage of the review and belong to the research team. What you receive is a rigorous quantitative synthesis, not the full systematic review from start to finish.

Because DerSimonian-Laird, still the default method in many teams, can underestimate uncertainty when there are few studies or high heterogeneity, which is the usual situation in practice. This applies to any random-effects meta-analysis, where the choice of estimator changes the confidence intervals and, sometimes, the conclusion. We prioritise random-effects models with REML estimation and the Knapp-Hartung adjustment when they suit the structure of the meta-analysis, not as an optional extra, because they handle heterogeneity and small numbers of studies better. When it is useful, the classic method is run in parallel and the difference is explained in the report, so the team can defend the methodological choice to reviewers. If the conclusions change from one method to the other, it is better to know before submitting the manuscript than to find out in review. Using the robust method by default is a decision about rigour, not an added sophistication.

Not necessarily: a high I² indicates heterogeneity between studies, and the point is not only to report it but to try to explain it; sometimes it also determines how the studies are pooled, or whether pooling them makes sense. This applies when the included studies differ more than chance would lead you to expect, which is common when synthesising evidence in practice. Heterogeneity is assessed with Q, I², τ² and their confidence intervals, and complemented with prediction intervals, which quantify the expected spread of the effect in a new study: a more honest reading than the confidence interval of the pooled effect when there is heterogeneity. When moderators are available, meta-regression and subgroup analyses explore possible clinical or methodological sources of that variability, instead of merely noting it. If heterogeneity is so high that pooling makes no sense, we say so. A high I² is not a defect to hide, but a signal that steers the analysis towards explaining where the variability comes from.

With a robustness diagnosis delivered as standard: detection of influential studies and outliers, sensitivity analysis and assessment of publication bias. This applies to any meta-analysis whose conclusion will support a clinical or policy decision, or a publication. For the influence of individual studies, we use leave-one-out analysis, which recalculates the effect removing each study in turn, and the Baujat plot, so that if a reviewer asks what happens when a study is removed, the answer is already in the report. For publication bias, we use contour-enhanced funnel plots, Egger's test and sensitivity methods such as trim-and-fill or Fail-Safe N, when the number of studies allows them to be interpreted. These methods detect and characterise the problem; they do not magically correct it: trim-and-fill is a sensitivity analysis, not proof that there is no bias. What they provide is knowing before review, not after it, whether the conclusion looks stable or sensitive to influential studies, asymmetry or publication assumptions.

No: a meta-analysis synthesises the evidence that exists; it does not improve it. If the primary studies are weak or biased, the pooled result inherits those limitations. This needs to be understood before interpreting any pooled effect, because a diamond on a forest plot can convey a false sense of solidity. That is why the analysis integrates the risk of bias assessment of the included studies, with ROB 2 for trials and ROBINS-E or ROBINS-I for observational studies, shown as traffic-light plots, so that the quality of the underlying evidence is visible alongside the result. Pooling studies at high risk of bias can produce an estimate that looks precise but is unreliable: more data do not correct a shared systematic bias. What the meta-analysis provides is a transparent synthesis that shows both the pooled effect and the quality of what has been pooled. Synthesising well includes making clear how far the original studies can be trusted, not just calculating their weighted average.

Meta-analysis quantitatively synthesises the numerical results of several studies to estimate a pooled effect for a specific question. Bibliometrics analyses the literature as an object, with its output, its authors, its thematic structure and its evolution, without combining the results of the studies. They do not answer the same question: meta-analysis asks "what pooled effect do these studies show for this intervention?"; bibliometrics asks "what does this research field look like, and how is it evolving?". The starting material also differs: meta-analysis starts from the data extraction table of the selected studies; bibliometrics starts from a broad bibliographic corpus whose search string the bibliometrics service itself builds. They can be run in sequence, for example with a bibliometric map to understand the field before narrowing down a synthesis, but one synthesises effects and the other maps the literature.

Yes: independent reanalysis of an existing meta-analysis with more robust methods is one of the uses of the service. This applies when you want to check whether the conclusions of a published meta-analysis are sensitive to other methodological decisions, for example before relying on it for a clinical or policy decision, or as part of a methodological replication. The reanalysis starts from the primary study data of the original meta-analysis and checks whether the pooled effect holds when robust estimators are applied, whether any influential study shifts the result, and whether publication bias or heterogeneity changes the reading. The condition is having the data extracted from the included studies, as with a new meta-analysis. What you receive is a reproducible assessment of how far the published conclusions hold up, with the differences documented, not a verdict on someone else's work beyond what the reanalysis shows.

The complete analysis, ready for publication: forest plots with clinical annotations, heterogeneity assessment with prediction intervals, publication bias and influence diagnostics, meta-regression where appropriate, risk of bias assessment with summary plots and the flow diagram following PRISMA 2020, all with the methods section written. This applies if you are going to submit to an indexed journal and need results that align with reporting standards and are ready for peer review. Reproducibility matters for a specific reason: the analysis is delivered with documented code and dependency management, so the meta-analysis can be updated when new primary studies are published, without recoding from scratch. The forest plot is not a pasted image but code that regenerates with the updated data. The research team carries out the search and the extraction and signs the review; the service provides the quantitative synthesis, its interpretation and its reproducible documentation.