Ask Mallard

Systematic review and meta-analysis

A systematic review answers a question by finding, appraising and where appropriate combining the studies that already exist, under a protocol written before the search.
NHMRC EVIDENCE LEVEL · 2009Depends on how the study is designed · InterventionProvisional design classification. Study quality and certainty of findings require separate appraisal.
Why this level?Specify the research question: intervention, diagnostic accuracy, prognosis, aetiology, or screening intervention. Other question types are outside this hierarchy. NHMRC 2009 source

A review of at least two Level II studies can be Level I within the relevant question category. Reviews of lower-level studies retain those levels; mixed designs require assessment of the studies contributing to each outcome. Review quality is appraised separately.

Where the NHMRC 2009 hierarchy names this design, it places it here:

Research questionLevelWhen
InterventionIReview of at least two Level II studies for the specified outcome. NHMRC assigns Level I to this combination.
InterventionIII-2Review of at least two Level III-2 studies for the specified outcome. A review of lower-level studies retains their level.

What question does this help answer?

What the whole body of evidence says, rather than what one study said. It is also the honest way to find out that the evidence does not answer the question, which a narrative summary tends to hide.

Systematic review and meta-analysis are separable

The review is the search, the selection and the appraisal. The meta-analysis is the statistical combination, and it is optional. Studies that are too different in population, intervention or outcome should be reviewed and described without being pooled; a pooled number that averages incomparable things is worse than no number.

The PRISMA 2020 flow

The PRISMA 2020 flow for a systematic review, without counts. Records identified from databases and registers lose duplicates before screening; records are screened and some excluded; reports are sought and some not retrieved; reports are assessed for eligibility and some excluded with reasons; the rest are the studies included.The PRISMA 2020 flow, without the counts a review fills inRecords identified from databases andregistersRemoved before screening: duplicates andother reasonsRecords screenedRecords excludedReports sought for retrievalReports not retrievedReports assessed for eligibilityReports excluded, with reasonsStudies included in the reviewAdapted from the PRISMA 2020 statement cited on this page. A review reports the number at every box.
Conceptual diagram. It shows the structure of the idea, not data from any study.

The protocol comes first

Registering the protocol, in PROSPERO or equivalent, before the search is what separates a systematic review from a literature summary assembled around a conclusion. Eligibility criteria, outcomes, the synthesis method and the subgroup questions are all decided in advance, because every one of them can be steered by knowing what the studies found.

Key assumptions and design requirements

  • The search covers more than one database and is reproducible, with the full strategy reported.
  • Screening and data extraction are done by two people independently, with disagreement resolved by a stated process.
  • Risk of bias is assessed per study with a named tool, and it feeds the interpretation rather than sitting in an appendix.
  • The synthesis method, fixed-effect or random-effects, is chosen for a reason and stated with it.
  • Small-study effects and publication bias are examined when the number of studies makes it meaningful.

Heterogeneity is a finding, not an obstacle

Studies differ because the world differs, so some heterogeneity is expected. The test for heterogeneity has low power with few studies, so a non-significant result is not evidence of consistency, and I-squared describes what PROPORTION of variability is not chance rather than how much variability there is. Large heterogeneity is a reason to ask why the studies disagree, which is often the most useful part of the review.

Why Mallard may recommend it

Mallard may point here when a description asks what is already known, or proposes to pool published results rather than collect new data.

How it is sized

There is no sample size calculation for a systematic review in the usual sense: the number of studies is whatever the literature contains. Mallard does not produce a headline N for this design. Where a review is planned prospectively, the relevant question is what precision the available evidence could support, which is closer to a feasibility assessment than a power calculation.

What to watch out for

Combining studies does not remove their biases, it averages them, and a meta-analysis of flawed studies produces a precise flawed answer. Heterogeneity is not a nuisance to be tested away: a small p-value for heterogeneity is not required before taking real differences between studies seriously, and a large I-squared does not by itself forbid pooling.

Clinical example

Question: Does early mobilisation after cardiac surgery reduce length of stay? Design: Register a protocol, search at least two databases with a documented strategy, screen in duplicate, appraise each trial for risk of bias, and pool only within groups of trials whose populations and mobilisation protocols are close enough that a single average means something.

Recent studies that used this
Searching PubMed…
Related Learn pages
Sources & further reading
  1. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews (opens in a new tab) · BMJ · doi 10.1136/bmj.n71
  2. Measuring inconsistency in meta-analyses (opens in a new tab) · BMJ · doi 10.1136/bmj.327.7414.557
PRISMA 2020 is a reporting guideline: it governs what a completed review discloses, and it is not a manual for conducting one. The second reference is the source of the I-squared statistic and explains what it does and does not measure.
Last source review: 2026-09-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.