How to Build a Literature Review

The type is a consequence of the question and the resources, not a stylistic preference. Choosing it first determines the search, the screening, the reporting and roughly how many months this will take.

If the question is a specific effect or association with a defined population, comparison and outcome, and the evidence base is bounded, the answer is a systematic review. It requires a protocol, an exhaustive documented search, at least two screeners, and reporting against PRISMA.

If the question is what evidence exists at all, or how a concept has been used, the answer is a scoping review, which maps rather than synthesises and reports against the PRISMA extension for scoping reviews (Tricco and colleagues, Annals of Internal Medicine, 2018).

If the purpose is to frame an original study or to survey a broad area for a reader, a narrative review is appropriate and does not require exhaustiveness. It does still benefit from a written search strategy, because that is what lets you say what you covered.

The section of a research paper that reviews the literature is none of the above. It is an argument for your study, it cites selectively and deliberately, and applying systematic review machinery to it is a common and expensive mistake.

Match the type to the resources honestly. A full systematic review with dual independent screening is a multi-person project measured in months, and starting one alone is the most common reason reviews are abandoned.

A short scoping run, done deliberately and thrown away afterwards, tells you whether the review is feasible at the size you have in mind. It is the cheapest decision point in the project.

Run a rough search in one database and look at the number of records, not at their content. Two hundred records is a review; twenty thousand is a different question that needs narrowing before anything else happens.

Read fifteen or twenty of the most relevant records and note the vocabulary they use. This is where the real search terms come from, and building a search from your own vocabulary rather than the literature's is the most common cause of a search that misses half the field.

Check whether the review already exists. Search the review literature and the registries: PROSPERO records systematic reviews in progress, and finding yours already registered by someone else is far better news now than in six months.

Narrow on a dimension you can defend

a period, a population, a design, a setting. Narrowing on convenience is visible to readers and is the boundary decision most often challenged.

Decide the boundary in writing before the real search starts, and date it. Scope creep during screening is what turns a three month review into a two year one.

The test of a search is not how many papers it finds but whether someone else could run it and get the same set. That standard is required for systematic reviews and is worth meeting for any review, because it is what makes the coverage claim credible.

Structure the query in concept blocks

one block per element of the question, synonyms joined with OR inside a block, blocks joined with AND. This is the structure every database supports and the one that can be adapted between them.

Use controlled vocabulary as well as free text where the database has it, such as MeSH in PubMed or Emtree in Embase, since indexing catches records whose titles and abstracts use none of your words.

Search more than one database, and choose them for coverage of your field rather than by habit. Add the sources that indexes miss: preprint servers, dissertation repositories, trial registries, and the agency or standards bodies that publish in your area.

Add citation chasing to the database search rather than treating it as an alternative. Backward and forward chasing from the anchor papers reliably finds records that no keyword search returns.

Record everything as you go

the exact string per database, the interface and date of the search, any limits applied, and the number of records returned. PRISMA-S sets out what a reported search should contain, and reconstructing this afterwards is far harder than logging it live.

Test the search against the anchor papers before running it in full. A search that does not return the works you already know are relevant is a search with a fault in it, and this two minute check is the single most useful piece of quality control available. Our search strategy template gives a copyable structure for the query, the criteria and the log.

Screening is where reviews take the most time and lose the most credibility. The discipline is simple: the criteria are written before the records arrive, and every exclusion at full text is recorded with a reason.

Write inclusion and exclusion criteria that another person could apply without asking you a question. Vague criteria produce inconsistent decisions and are the main source of disagreement between screeners.

Screen titles and abstracts against the criteria only, and be liberal: at this stage the cost of a wrongly included record is one full text read, and the cost of a wrongly excluded record is a hole in the review.

Screen full texts against the same criteria, and record the reason for every exclusion. That list of reasons is a required element of a PRISMA flow diagram and is the part people reconstruct badly from memory.

Use two independent screeners with a documented resolution procedure where the review type requires it. Where a second person is genuinely impossible, say so in the limitations rather than implying dual screening happened.

Tooling helps and does not substitute for criteria. Rayyan and Covidence are built for this, a reference manager such as Zotero or Mendeley will do for a small review, and a spreadsheet works if the numbers are recorded.

Reading fifty papers and writing prose afterwards does not work, because nobody holds fifty papers in mind at once. Extraction converts reading into a structure you can sort, and sorting is what produces the argument.

Design the extraction table before extracting

one row per study, one column per dimension you will need to compare. Population, setting, design, sample, method, measures, findings, limitations, funding is a reasonable default.

Appraise quality where the review type requires it, using an instrument appropriate to the design rather than an ad hoc impression. The Cochrane Handbook is the reference for the clinical case and its logic transfers.

Sort the table repeatedly. The themes, the disagreements and the empty regions become visible when the table is sorted by design, by population or by year, and that is also where a gap analysis begins. Our gap analysis guide takes it from there.

Keep the table as an artefact. It becomes the synthesis table in the paper, the evidence for the coverage claim, and the thing that makes an update possible without starting again.

Reviews take long enough that the literature moves underneath them. The last part of the process is maintenance and a clean handover to the writing, which is a different skill and has its own page here.

Set alerts on the search strings and the anchor papers the day the search is run, so that new records arrive continuously rather than in a panic before submission.

Re-run the full search before submission and record the date. Journals increasingly ask when the search was last updated, and an unstated search date is a question you will be asked in review.

Hand off to writing only when the table is complete and sorted. Writing from an incomplete table is how reviews become descriptive lists, which is the failure our guide to writing a literature review is about avoiding.