How to Conduct a Systematic Literature Review

A systematic literature review is a review whose search and selection a stranger could repeat. A narrative review can be excellent and still be the author’s chosen path through the field. A systematic review makes the path visible: what you looked for, where, what you kept, and what you excluded. If you cannot write those rules before you search, you are writing a narrative review, and you should call it that.

Fix the question and the review type

The question comes first. Use the four-part check in How to formulate a research question, then decide which review can answer it. A systematic review answers a narrow question about what the evidence shows. A scoping review maps what kinds of work exist. A meta-analysis combines numbers, and only when the studies measured comparable things. Picking the type after you have a folder of PDFs is how eligibility rules start bending toward the papers you liked.

Write the eligibility rules before the search

Write inclusion and exclusion as tests you could hand to a second reader. Year range, language, publication type, the population or dataset, the method, and the outcome all belong here. “High quality papers” is not a rule, because two readers will not agree on it. “Peer-reviewed, English, empirical studies that compare at least two methods on a named task, published from 2018” is a rule. Record the date you froze it.

• Include a paper only if it can answer the question. A well-known paper outside the question is a citation for the introduction, not a row in the review.
• Decide what you will do with preprints and theses before you see one that supports your hunch.
• Decide whether one research group’s series of papers counts once or many times. Say so in the protocol.

Search more than one place, and write the string down

One database is a sample of that database’s coverage, not of the field. For computing, a typical set is a bibliographic database plus the venue indexes you know the community actually uses, and a snowball from the references of the papers you keep. For health, the expected set is different. Use the databases your readers will look for, and record the exact string, the date, and the number of hits for each one.

Build the string from the same slots as the question: the population or task, the method, and the outcome. Synonyms belong in the string. A string that only contains the phrase you personally use will miss the papers that used the other name for the same thing.

Screen in two passes, and keep the counts

First pass: title and abstract against the rules. Second pass: the full text of what survived. Write down why each full-text exclusion happened, in a short fixed list of reasons. Those counts are the flow diagram. The mechanics of the PRISMA 2020 boxes are in The PRISMA 2020 flow diagram, step by step. Do not invent a count at the end to make the figure look tidy. If two people screen, agree in advance how a disagreement is resolved, and do that the same way every time.

Extract the same fields from every paper

A spreadsheet with one row per included study stops the review becoming a stack of summaries. Minimum columns: bibliographic id, setting or dataset, what was compared, sample size, metric, main result, and a limitation you noticed. Add columns your question needs. Fill them while you read, not from memory a month later. If a cell is “not reported”, write that. An empty cell looks like you forgot.

Synthesise, do not serialise

The written review should answer the question, grouped by theme, method, or finding. A paragraph per paper, in the order you read them, is a bibliography with verbs. Where studies disagree, say what differs in their setup. Where they agree, say on what evidence, and how many studies that is. A gap you want to claim has to be visible in this table. If it is only visible in your memory of the field, it is not ready. That check is the same one in How to find a research gap.