Frequently asked questions

A statistical method that combines the results of several studies that address the same question, weighting each by its precision, to give a pooled estimate with a confidence interval. It is usually part of a systematic review, which supplies the studies through a documented search. See the guide on what a meta-analysis is for a worked example.

A systematic review is a structured way of finding, selecting and appraising all the relevant studies on a question, following a protocol. A meta-analysis is the statistical combination of their results. A systematic review may include a meta-analysis, or it may synthesize results in words when pooling is not appropriate. A meta-analysis without a systematic search is not a systematic review.

Technically two, but whether a pooled result is useful depends on the number of participants and events, how similar the studies are and what you want to learn. Subgroup and meta-regression analyses need many more, and funnel plot tests generally need about ten or more. The guide on the number of studies sets out the reasoning.

Choose on the question and on whether the studies can be expected to estimate one common effect. In most real reviews the true effects differ, so a random-effects model is usual, with a prediction interval to show the range. Do not choose the model by the result of a heterogeneity test. The guide on fixed-effect and random-effects models explains the trade-offs.

It is the share of the variation in study results that reflects real differences rather than chance. It does not measure the size of the differences, and it depends on the precision of the studies. Report tau-squared and a prediction interval alongside it, and treat the rough labels for low, moderate and high values with caution.

It is the range in which the true effect in a new study would be expected to fall. It is wider than the confidence interval for the average effect when studies differ, and it shows whether a result is likely to hold in a new setting. It needs at least three studies and is very wide with few.

Search widely, including registries and grey literature, and then look at the funnel plot and use an asymmetry test suited to your effect measure, if you have about ten or more studies. Asymmetry shows small-study effects, which can have other causes. Sensitivity analyses such as trim-and-fill show how much a conclusion depends on missing studies.

It depends on the outcome. Risk ratios and odds ratios suit binary outcomes, mean differences and standardized mean differences suit continuous ones, hazard ratios suit time to event, and correlations suit associations. Choose before looking at the results, and convert to absolute effects when presenting to decision makers.

No. They are close when the outcome is rare, but the odds ratio lies further from 1 when the outcome is common. Reading an odds ratio as a risk ratio then overstates the effect. The guide on odds ratios and risk ratios shows how to convert using the baseline risk.

Zero events make ratios undefined and need a deliberate approach. Options include Mantel-Haenszel or Peto methods, mixed models and Bayesian models, with a sensitivity analysis comparing them. Adding 0.5 to cells is common but can bias the result. Report how many studies had zero events and what was done.

Many journals require or encourage it, and it is good practice, because it fixes your plan before the results are known. PROSPERO accepts many health-related systematic reviews, and the Open Science Framework accepts others. Registration must happen before data extraction is far advanced, so do it early.

At least two or three that suit the topic. For health interventions, MEDLINE, Embase and CENTRAL are the usual minimum, with discipline databases added for fields such as psychology, education or economics, plus trial registries and citation chasing. Record the dates and strategies.

Independent double screening and double extraction are the usual standard because errors are common when one person works alone. Reduced approaches exist, particularly in rapid reviews, and should be stated as limitations. Resolve disagreements by discussion or a third reviewer.

A structured judgment of how likely it is that each study's result is distorted by flaws in its design, conduct or reporting. Use a tool suited to the design, such as RoB 2 for randomized trials or ROBINS-I for non-randomized studies of interventions, and show the results by study and by domain.

A system for rating the certainty of a body of evidence for each outcome, as high, moderate, low or very low. It starts from the study design and considers risk of bias, inconsistency, indirectness, imprecision and publication bias, and can rate certainty up for large effects and other factors.

The main reporting guideline for systematic reviews. PRISMA 2020 has a checklist of 27 items, a flow diagram and an abstract checklist, and there are extensions for specific designs. It tells you what to report. It does not tell you how to do the review.

Yes, but confounding and design differences matter a great deal. Pool adjusted estimates where possible, assess risk of bias with a tool such as ROBINS-I, report to MOOSE as well as PRISMA, and be cautious about causal conclusions. Heterogeneity is usually larger than in trials.

A method that compares several treatments at once by combining direct and indirect evidence. It rests on the assumptions of transitivity and consistency, which should be assessed and reported. Rankings such as SUCRA must be read with effect estimates and the certainty of evidence.

When the aim is to map what research exists, clarify concepts or identify gaps, and not to estimate an effect. It does not usually appraise risk of bias or pool results. PRISMA-ScR is the reporting guideline, and PROSPERO does not accept scoping review protocols.

R, with packages such as metafor, meta and netmeta, covers the widest range of methods and makes the analysis reproducible. Stata and RevMan serve many standard analyses. Whatever you use, record versions and options, because defaults differ between programs.

Check that the journal publishes reviews of your type, read recent issues, note requirements for registration and reporting, verify indexing and legitimacy directly, and rank candidates before submitting. Avoid choosing on impact factor alone. The guide on choosing a journal has a checklist.

Answer every comment in a point-by-point letter, say what changed and where, run requested analyses and label them as post hoc, and disagree politely with evidence where you must. Return a tracked and a clean version by the deadline.

A review that is kept up to date by searching regularly and adding new studies as they appear, usually for a question where evidence is changing fast and decisions matter. It needs a plan for how often to search, when to update the analysis and when to retire the review.

Yes, and many students do. A systematic review with meta-analysis is a large project, so agree the scope with your supervisor, register a protocol, plan for double screening if possible and allow time for searching and screening, which take longer than most people expect.

A confidence interval comes from a frequentist analysis and is defined by how often such intervals would contain the true value over repeated studies. A credible interval comes from a Bayesian analysis and gives the range holding a stated probability of the true value, given the model and prior. Both depend on assumptions.

We support the stages of an evidence-synthesis project, from protocol and search strategy through screening, extraction, analysis, risk of bias and certainty ratings to manuscript preparation and revision. Scope is agreed with you in writing. [OWNER VERIFICATION REQUIRED] See the services pages for details.

Authorship follows the criteria of the journal and of bodies such as the ICMJE, which require substantial intellectual contribution and accountability. Ghost-writing is not acceptable. The roles of everyone who contributed, including any support service, should be agreed in advance and acknowledged as the journal requires. [OWNER VERIFICATION REQUIRED]

Our policy on the use of AI tools is set out on the AI use policy page, and any use is disclosed as the journal requires. Responsibility for the content stays with the authors and the team. [OWNER VERIFICATION REQUIRED]

Your data and manuscript are treated as confidential, and we do not share them without your agreement. The privacy policy describes how contact details and files are handled. [OWNER VERIFICATION REQUIRED]

Send a short description of your question, the type of review, the stage you have reached, your target journal and your deadline through the contact page. We reply with questions or a proposed scope and quote. [OWNER VERIFICATION REQUIRED]

It varies with the number of records, the team and the methods. A systematic review with meta-analysis commonly takes many months from protocol to submission, and a rapid review can be shorter with stated limitations. Searching and screening usually take the most time. Give us your deadline and we will say what is feasible.

No. No honest service can, because acceptance is the decision of the journal's editors and reviewers. We can help to prepare a manuscript that meets reporting standards and to respond to comments, and we describe the limits of what we can promise on the terms page.