A systematic review is not a longer literature review. A narrative review summarises what an author already knows; a systematic review asks a focused question and then finds, appraises, and synthesises every eligible study by methods decided in advance. That discipline is what makes its conclusions trustworthy, and it is also what makes it demanding to do well.
This guide walks through the full process: writing a protocol, searching reproducibly, screening and appraising studies, extracting data, and synthesising results, with or without a meta-analysis. It builds on our guide to writing a literature review, which covers narrative reviews, and it connects to our guides on methodology and data management, because a review is only as good as its methods and its records.
Three features separate a systematic review from a narrative one. First, the question is stated precisely, usually in terms of the population, intervention or exposure, comparison, and outcome, so that it is clear which studies are eligible and which are not. Second, the methods for finding and selecting studies are defined before the results are known, which limits the temptation to include only studies that support a preferred conclusion. Third, every included study is appraised for quality, and the synthesis takes that appraisal into account.
A meta-analysis is a further step, not a synonym for a systematic review. It combines the numerical results of comparable studies statistically to give an overall estimate. A systematic review can be completed without a meta-analysis, using a narrative synthesis, when the studies are too different to combine sensibly. Combining studies that measured different things in different ways produces a precise-looking number that means very little, and reviewers will challenge it.
Before starting, check whether a recent systematic review already answers your question. Searching a review register and the major databases for existing reviews can save months of work and may reveal a gap, such as an outdated search or a population not yet covered, that justifies a new review. If a good recent review exists, updating it or addressing its stated limitations is often a stronger contribution than starting again from scratch.
The protocol is the review's plan, written before the search begins. It states the question, the eligibility criteria, the databases and search terms, the screening and extraction procedures, the quality appraisal tool, and the synthesis method. Writing it first forces decisions while they are still neutral, and it gives the team a document to follow when disagreements arise later.
Registering the protocol in a public register, such as PROSPERO for health-related reviews, creates a dated public record of the plan. Registration protects the authors as well as the readers: if the plan changes during the review, the changes can be reported transparently against the registered version instead of looking like decisions made after seeing the results. Not every field has a dedicated register, but the principle holds everywhere — write the plan down, date it, and keep it.
Eligibility criteria deserve particular care, because they decide the entire evidence base. Criteria that are too narrow may exclude relevant studies and leave too little evidence; criteria that are too broad may combine studies that should not be combined. Pilot the criteria on a small set of studies before the full screening, and refine the wording where two reviewers disagree. A criterion that two careful reviewers interpret differently is a criterion that needs rewriting.
A systematic search must be comprehensive and reproducible. That means searching more than one database, using both controlled vocabulary and free-text terms, and recording the full search strategy for at least one database so that another researcher could repeat it. Limiting a search to a single database, or to studies the authors already knew, is the most common weakness in weaker reviews.
Search terms should cover each element of the question, with synonyms and spelling variants, and should be tested against a set of known eligible studies: if the search misses a study that should clearly be included, the terms need widening. Grey literature, such as theses, conference proceedings, and reports, should be considered, because relying only on published journal articles can bias the review towards studies with positive results. Whether grey literature is included should be decided in the protocol and reported, not decided after seeing what it contains.
Keep complete search records: the databases searched, the dates of the searches, the exact terms used, and the number of records retrieved from each source. These records feed directly into the flow diagram that reports how many records were found, screened, excluded, and included, and reviewers will ask for them if they are missing. A search that cannot be reconstructed cannot be trusted, however thorough it felt at the time.
Screening usually happens in two stages: titles and abstracts first, then full texts. The first stage removes clearly irrelevant records quickly; the second applies the eligibility criteria in full. Both stages should be done independently by at least two reviewers, with disagreements resolved by discussion or by a third reviewer. Independent duplicate screening is not bureaucratic overhead; it is the main defence against a single reviewer's blind spots and errors.
At the full-text stage, record a reason for every exclusion. A review that reports only the number excluded, without reasons, cannot be audited, and readers cannot tell whether eligible studies were lost. Common exclusion reasons include wrong population, wrong intervention, wrong outcome, wrong study design, and duplicate publication of the same study. Duplicate publication deserves attention: the same study reported in two papers must be counted once, or its results will be double-counted in the synthesis.
Agreement between reviewers should be reported, and persistent disagreement should prompt a review of the criteria rather than a forced compromise on individual studies. If two reviewers repeatedly disagree about a particular criterion, the criterion is ambiguous and should be clarified and reapplied. This is slower than a single reviewer working alone, but it is the reason systematic reviews are trusted more than one person's reading of the literature.
Every included study should be appraised for quality, because studies of different quality should not be given equal weight in the conclusions. Risk-of-bias tools assess specific features, such as how participants were allocated, whether outcome assessment was blinded, and whether results were reported completely. The right tool depends on the study design, and the choice should be stated in the protocol.
Quality appraisal is not a reason to discard inconvenient studies silently. Studies at higher risk of bias can still be included, but their influence should be examined, for example by a sensitivity analysis that shows whether the overall conclusion changes when they are excluded. A review that includes weak studies without comment, or excludes them without explanation, will be challenged by reviewers either way.
Appraisal should also be done independently by two reviewers, with disagreements resolved as in screening. The results are usually summarised in a table and, where appropriate, a figure showing the risk of bias across studies. This transparency lets readers judge for themselves whether the review's conclusions rest on strong or fragile evidence, which is exactly the judgement a systematic review exists to support.
Data extraction should use a piloted form that captures everything the synthesis needs: study characteristics, participant details, intervention or exposure details, outcome definitions and results, and funding sources. Piloting the form on a few studies reveals missing fields and ambiguous items before the full extraction, when fixes are still cheap. Extracting data that turn out to be needed later, after the team has moved on, is slow and error-prone.
Extraction, like screening and appraisal, is best done independently by two reviewers, with disagreements resolved by discussion. Numerical results need particular care: means, standard deviations, event counts, and effect estimates must be copied exactly, with attention to units, time points, and subgroups. An error in a single extracted number can distort a meta-analysis, so a second check of the numbers that drive the synthesis is time well spent.
Where data are missing or unclear, contact the original authors where possible and record the attempt. If data cannot be obtained, the synthesis should say so and, where appropriate, examine whether the missing data could change the conclusion. Missing data are a limitation to report, not a detail to hide, and reviewers will ask about studies whose results could not be used.
Synthesis begins with describing the included studies: how many, where and when they were done, who took part, and what they found. This descriptive synthesis is valuable on its own, because it shows the shape of the evidence, including gaps where no studies exist. A table of study characteristics is the standard way to present it, and it should be complete enough that a reader need not return to the original papers for basic facts.
A meta-analysis is appropriate when the studies are similar enough in question, population, intervention, and outcome measurement that combining them is meaningful. The analysis should account for differences between studies, examine heterogeneity rather than hiding it, and explore whether results differ by subgroup or study quality. A single pooled number without any examination of heterogeneity is rarely convincing, and a high degree of unexplained heterogeneity may mean that a narrative synthesis is the honest choice.
Publication bias should also be considered: studies with positive or striking results are more likely to be published than studies with null results, which can distort a synthesis. Funnel plots and related checks can suggest whether small studies with unwelcome results appear to be missing, though these checks are limited when few studies are included. Whatever the synthesis method, its limitations should be stated plainly, following the same honesty expected in the limitations section of an original paper.
Reporting should follow the PRISMA statement, which sets out what a systematic review report must contain, from the title and abstract through the search, selection, appraisal, synthesis, and limitations. The PRISMA flow diagram, showing records found, screened, excluded with reasons, and included, is not optional decoration; it is the audit trail of the whole review. A review without a flow diagram will be returned for revision by most journals.
Common flaws to check before submission include a search that cannot be repeated from the report, eligibility criteria applied inconsistently, quality appraisal reported but ignored in the conclusions, and a meta-analysis that combines studies too different to combine. Another frequent flaw is overstating the conclusions: a review of small, low-quality studies cannot support strong recommendations, and the discussion should reflect the strength of the evidence actually found.
Finally, remember that a systematic review is a substantial project, often taking many months, and it should be planned as such, with a team, a timeline, and a division of labour. Reviews attempted alone, in a hurry, or without a protocol are the ones most likely to be rejected or, worse, published with conclusions the evidence does not support. Done properly, a systematic review is one of the most useful contributions a researcher can make, because it tells the field what is known, what is not, and what should be studied next.
| Stage | Key task | Output |
|---|---|---|
| Question and protocol | Define the question and eligibility criteria; register the plan | Written, dated protocol |
| Search | Search several databases with a recorded strategy | Search records and flow counts |
| Screening | Two reviewers, two stages, with exclusion reasons | List of included studies |
| Quality appraisal | Assess risk of bias with a design-appropriate tool | Appraisal table and summary |
| Data extraction | Piloted form, independent duplicate extraction | Complete dataset for synthesis |
| Synthesis | Narrative synthesis, meta-analysis, or both | Overall findings with heterogeneity |
| Reporting | Follow PRISMA, including the flow diagram | Complete, auditable report |
These authoritative external resources explain the standards and guidance mentioned in this article:
How is a systematic review different from a literature review?
A literature review narrates what is known, usually from studies the author selected. A systematic review states a focused question in advance, searches reproducibly, screens and appraises studies by preset criteria, and synthesises the results. The preset methods are what make its conclusions auditable.
Do I always need a meta-analysis?
No. A meta-analysis is only appropriate when the included studies are similar enough to combine meaningfully. When studies differ too much in population, intervention, or outcome measurement, a narrative synthesis is the honest choice, and a pooled number would mislead.
Why register the protocol?
Registration creates a dated public record of the plan, so changes made during the review can be reported transparently. It protects both authors and readers against decisions that appear to have been made after seeing the results.
How many reviewers should a review have?
At least two, working independently at screening, appraisal, and extraction, with disagreements resolved by discussion or a third reviewer. Single-reviewer reviews miss eligible studies and make errors that duplicate work catches.
How long does a systematic review take?
Often many months, depending on the size of the literature and the team. The search, screening, and extraction stages are the most time-consuming. Planning a realistic timeline and team at the protocol stage prevents a review that stalls halfway.
What is the PRISMA flow diagram?
A diagram showing how many records were found, screened, excluded with reasons, and finally included. It is the audit trail of the review, and most journals require it. A review without one will usually be returned for revision.