PRISMA 2020 and the systematic review, start to finish

A systematic review is defined by reproducibility. Somebody else should be able to take your protocol, run your searches, apply your criteria and finish with the same included studies.

PRISMA 2020 is the reporting standard that makes that possible. It does not tell you how to conduct the review; it tells you what you must report so a reader can judge and repeat it.

Write the protocol before you search

The protocol is what separates a systematic review from a thorough literature review. It commits you to your criteria in advance, so decisions cannot be shaped by the results you start seeing.

• Research question, structured with PICO, SPIDER, PEO or whichever framework fits your design.
• Inclusion and exclusion criteria, specific enough that two people would apply them identically.
• Databases and other sources to be searched, with the date range.
• Screening procedure — how many reviewers, and how disagreements are resolved.
• Data extraction fields.
• Risk-of-bias tool and synthesis approach.

Building a search string that works

A good search has two properties in tension: sensitivity (finding everything relevant) and precision (not returning ten thousand irrelevant records). Sensitivity matters more in a systematic review, because a missed study is a flaw in the review while an irrelevant one is merely screening work.

Build the string concept by concept, then combine.

1. Break the question into its core concepts — usually two to four.
2. For each concept, list every synonym, alternative spelling and abbreviation authors use.
3. Add the database's controlled vocabulary — MeSH in PubMed, Emtree in Embase, thesaurus terms elsewhere.
4. Combine synonyms within a concept using OR.
5. Combine concepts using AND.
6. Apply truncation carefully — therap* catches therapy, therapies and therapeutic.
7. Translate for each database. Syntax genuinely differs; the same string will not run everywhere.

PubMed example

("remote work"[tiab] OR "telework*"[tiab] OR "work from home"[tiab] OR telecommut*[tiab])
AND
("job satisfaction"[tiab] OR "employee satisfaction"[tiab] OR "work engagement"[tiab])
AND
("Job Satisfaction"[Mesh] OR "Teleworking"[Mesh])

Filters: 2015:2026[dp], English[lang]

Screening in two passes

Screen titles and abstracts first, then full texts. Two independent reviewers is the standard; if you are working alone, screen a sample twice yourself and report the agreement.

Record a reason for every full-text exclusion. The flow diagram requires them, and reconstructing them afterwards is painful and unreliable.

• Use Rayyan, Covidence or a structured spreadsheet — all decisions must be recoverable.
• Calculate Cohen's kappa for inter-rater agreement; above 0.60 is generally acceptable.
• Be inclusive at title and abstract stage. It is cheaper to exclude at full text than to miss a study.
• Resolve disagreements by discussion, with a third reviewer for anything unresolved.

The flow diagram must reconcile

The PRISMA flow diagram is where reviewers look first, because arithmetic errors there suggest the underlying records are not properly kept.

• Records identified from each database, listed separately.
• Records from other sources — citation searching, grey literature, contact with authors.
• Duplicates removed.
• Records screened, and records excluded at screening.
• Full texts sought, and any not retrieved.
• Full texts assessed, and exclusions with reasons and counts.
• Studies included in the review, and in any meta-analysis.

Risk of bias: use the right tool

Synthesis: narrative or meta-analysis?

Pool studies only when they are similar enough that a pooled estimate means something. If the populations, interventions or outcome measures differ substantially, a meta-analysis produces a number that is precise and meaningless.

Where pooling is not appropriate, SWiM (Synthesis Without Meta-analysis) gives a structured way to report narrative synthesis, which reviewers increasingly expect over an unstructured summary.

Questions this raises