Systematic Reviews and Meta-Analyses: PRISMA 2020 and GRADE

By Benn Bluestein-Veyra, M.Sc. Organic Chemistry · Updated 21 October 2026

The question of systematic review vs meta-analysis comes down to method and statistics. A systematic review identifies, appraises and synthesizes all the evidence that meets stated eligibility criteria, using explicit methods aimed at minimizing bias.^[1] A meta-analysis is the quantitative step that combines the results of those independent studies into one estimate.^[2]

This reference follows a review from protocol to conclusion: registration, the PICO question, the search, the PRISMA 2020 statement, effect models, heterogeneity, funnel plots, GRADE certainty and AMSTAR 2 appraisal. Each tool is credited to the group that developed it.^[3][4][5]

Systematic review vs meta-analysis: how the review and the statistics divide the work

A narrative review is an expert summary of a field built from the author’s own selection of studies. The National Library of Medicine indexes a meta-analysis as a quantitative method of combining independent studies and distinguishes it from a review of the literature.^[2]

A systematic review can stand without a meta-analysis. When studies differ too much in design, participants or outcomes to pool, reviewers synthesize them in structured prose, and PRISMA 2020 covers the reporting of synthesis methods of every kind.^[3][6]

Murad and colleagues describe systematic reviews and meta-analyses as tools for consuming evidence: they appraise, synthesize and apply the primary studies beneath them. Their strength therefore follows the strength of the studies they include.^[7]

A registered protocol, a PICO question and a reproducible search define the review in advance

Registering the protocol in PROSPERO

Moher and colleagues argue that every systematic review should build on a protocol describing its rationale, hypothesis and planned methods. Their PRISMA-P 2015 statement provides a 17-item checklist, and a protocol on record lets readers detect later changes to methods and selective reporting.^[8]

PROSPERO, the international prospective register of systematic reviews, was developed by the Centre for Reviews and Dissemination at the University of York. Booth and colleagues describe its aims as reducing unplanned duplication and making the review process transparent, with registration ideally before formal screening begins.^[9]

Framing the question with PICO

Richardson and colleagues described the well-built clinical question in 1995, and its parts became the PICO frame: population, intervention, comparator and outcome.^[10] Joanna Briggs Institute guidance presents the review question in the same PICO format.^[11]

The GRADE working group asks that a question specify the setting, population, intervention, comparator and every important outcome, whether or not studies report it. The relative importance of each outcome is rated before the evidence is gathered.^[12]

Searching reproducibly and screening in duplicate

Rethlefsen and colleagues developed PRISMA-S, a 16-item extension for reporting literature searches, through a Delphi survey and consensus conference. It asks authors to report each component of the search completely enough to be reproduced.^[13]

Screening runs in two steps, titles and abstracts and then full reports, conventionally by two reviewers working independently.^[14] In a methodological systematic review, Waffenschmidt and colleagues found that single screening missed a median of 5% of eligible studies: 3% for experienced reviewers and 13% for less experienced ones.^[14]

The Cochrane Handbook for Systematic Reviews of Interventions describes each of these stages. Cumpston and colleagues introduced its 2019 edition as updated guidance for trusted systematic reviews.^[15]

Understanding the PRISMA 2020 statement and flow diagram

The PRISMA 2020 statement, published by Page and colleagues in the BMJ, replaced the 2009 statement with a 27-item checklist, an expanded checklist, an abstract checklist and revised flow diagrams (Fig. 1).^[3]

The flow diagram counts records at each stage: identification from databases and registers, removal of duplicates, screening, retrieval of reports and assessment of eligibility. It ends with the number of studies included, and reports excluded at full text are listed with reasons.^[3][6]

A companion explanation and elaboration paper sets out why each item is recommended and gives examples from published reviews.^[6] Extensions cover protocols, literature searches and network meta-analyses.^[8][13][16]

Extracting data and judging each study

PRISMA 2020 asks reviewers to report how data were collected from each report, including how many reviewers took part and whether they worked independently, and to list every data item sought.^[6]

Each included trial is then appraised for risk of bias. The Cochrane RoB 2 tool, published by Sterne and colleagues in 2019, structures that judgment for randomized trials, and the risk of bias assessment reference on this site covers it domain by domain.^[17]

Fixed-effect and random-effects models answer different questions

A fixed-effect model assumes that every study estimates one common true effect, so the spread of results reflects sampling error alone (Fig. 2). A random-effects model assumes that true effects vary between studies around a mean.^[18]

Borenstein and colleagues stress that the two models rest on fundamentally different assumptions, and that the chosen model sets how the summary statistics are estimated and interpreted.^[18] Both weight studies by inverse variance; the random-effects weight adds the between-study variance, tau², to each study’s own variance.^[18][19]

DerSimonian and Laird proposed a simple noniterative random-effects method in 1986 that carries the heterogeneity of effects into the pooled analysis.^[19] Riley, Higgins and Deeks explain that a random-effects summary estimates the average effect, and recommend reporting a prediction interval alongside it.^[20]

Forest plots, I² and tau² show how far study results agree

A forest plot shows each study as a square at its point estimate, sized by its weight, with a line for its confidence interval and the pooled estimate as a diamond beneath (Fig. 3). Lewis and Clarke traced the history of the display and its name in the BMJ in 2001.^[21]

Heterogeneity is variation in true effects between studies. The classic Q test detects it, and Higgins and Thompson noted that its result depends on the number of studies in the meta-analysis.^[22]

Higgins and Thompson proposed I², the proportion of total variation in study estimates that is due to heterogeneity, as a measure independent of the number of studies and the effect metric.^[22] Higgins and colleagues later explained I² for readers of Cochrane Reviews, suggesting values near 25%, 50% and 75% as rough guides to low, moderate and high heterogeneity.^[23]

Tau² is the between-study variance itself, on the scale of the effect measure. IntHout and colleagues note that tau² and I² are hard to read clinically, and that a prediction interval gives the expected range of true effects in similar studies.^[22][24]

IntHout and colleagues then studied Cochrane meta-analyses from 2009 to 2013. Among 479 statistically significant random-effects results with I² above zero, 72.4% had 95% prediction intervals that included no effect or an effect in the opposite direction.^[24]

Exploring heterogeneity with prespecified subgroups

GRADE guidance asks reviewers to explore heterogeneity with a small number of prespecified hypotheses about participants, interventions, outcomes and methods.^[25] Subgroup effects gain credibility when they are few, have a stated direction, come from comparisons within studies and show a low interaction p value.^[25]

Funnel plots and Egger’s test examine small-study effects

A funnel plot graphs each study’s effect estimate against its precision. Precise studies cluster near the pooled estimate and smaller studies scatter widely on both sides, tracing an inverted funnel when the evidence is complete (Fig. 4).^[26][27]

Egger and colleagues introduced a regression test of funnel plot asymmetry in 1997. In their sample, asymmetry was present in three of four meta-analyses that disagreed with a later large trial and in none of the concordant pairs.^[26] The authors advised caution when a meta-analysis rests on a limited number of small trials.^[26]

Sterne and colleagues recommend tests for funnel plot asymmetry only when a meta-analysis includes at least 10 studies. They list causes beyond publication bias, including true heterogeneity, data artifacts and chance, so asymmetry is read as a small-study effect whose cause needs investigation.^[27]

GRADE counts publication bias as one of its five domains. Guyatt and colleagues advise suspicion when the evidence comes from a number of small studies, most of them commercially funded.^[28]

GRADE rates the certainty of evidence behind a review’s conclusions

Guyatt and colleagues set out the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach in the BMJ in 2008, giving reviews and guidelines a shared way to rate a body of evidence (Fig. 5).^[4] A 2011 series in the Journal of Clinical Epidemiology explains each step.^[29]

Balshem and colleagues define four levels, high, moderate, low and very low, applied to a body of evidence for each outcome. In a systematic review the rating expresses confidence that the effect estimate is correct; randomized trials start at high and observational studies at low.^[29][30]

Five domains can lower certainty

• Risk of bias: limitations such as unconcealed allocation, absent blinding, loss to follow-up and selective outcome reporting, judged outcome by outcome.^[31]
• Inconsistency: unexplained differences in point estimates, little overlap of confidence intervals and a high I².^[25]
• Indirectness: participants, interventions or outcomes that differ from the question, including surrogate outcomes and comparisons never tested head to head.^[32]
• Imprecision: a confidence interval wide enough that the conclusion would change between its limits, or a total sample below the optimal information size.^[33]
• Publication bias: a body of evidence suspected of missing studies, often signaled by many small studies.^[28]

Three considerations can raise it

Guyatt and colleagues describe rating up mainly for methodologically rigorous observational studies: one level for at least a two-fold change in risk and two levels for at least a five-fold change. A dose-response gradient, or plausible confounding that would shrink an apparent effect, can also raise certainty.^[34]

GRADE presents the result in summary of findings tables, which show the certainty rating with the relative and absolute effects for each important outcome.^[29]

Appraising a published review with AMSTAR 2 and reading its conclusions

Shea and colleagues published AMSTAR 2 in 2017 to appraise systematic reviews that include randomized studies, non-randomized studies or both. The instrument has 16 items and rates overall confidence from weaknesses in critical domains, with no overall score.^[5]

Its suggested critical domains are protocol registration, an adequate search, justified exclusions, risk of bias in included studies, appropriate meta-analytic methods, risk of bias in interpretation and publication bias. Confidence is then rated high, moderate, low or critically low.^[5]

Umbrella reviews and network meta-analyses apply the same principles at larger scale

An umbrella review takes existing systematic reviews and meta-analyses as its unit of analysis. Aromataris and colleagues developed Joanna Briggs Institute guidance with a 10-question appraisal tool, and Fusar-Poli and Radua set out ten rules running from protocol to transparent reporting.^[11][35]

A network meta-analysis compares several interventions at once by combining direct and indirect comparisons. Higgins and colleagues describe how to test whether those sources agree, and Hutton and colleagues added a 32-item PRISMA checklist for reporting such reviews.^[16][36]

Questions to ask of any review’s conclusions

• Was a protocol registered before screening began, and do the reported methods match it?^[5][9]
• Does the PRISMA flow diagram account for every record, with reasons for full-text exclusions?^[3]
• Which model pooled the studies, and are I², tau² and a prediction interval reported?^[23][24]
• Was small-study bias examined, with enough studies to make a test informative?^[27]
• What GRADE certainty does each outcome carry?^[30]

A review that answers each question clearly offers conclusions a reader can weigh with confidence. The certainty rating attached to each outcome sets how firmly each conclusion stands.^[29]

Where to go from here

The hierarchy of evidence reference places systematic reviews among the other study designs, and the randomized controlled trials reference explains how the trials beneath a meta-analysis are built. The risk of bias assessment reference walks through RoB 2 and ROBINS-I, and the sham and placebo controls and clinical trial phases references complete the series.

For applied reading, the clinic guide to reading the evidence and the exosome clinical trials reference show these methods at work, alongside the remestemcel-L approval history, the potency assay reference and the quality references. Our team is glad to discuss any of the methods described on this page.