How to do a systematic review: the PRISMA 2020 steps

What are the steps of a systematic review?

Six, and the order matters.

1. 1

   Frame the question

   One focused question (PICO, PEO or SPIDER), turned into written inclusion and exclusion criteria.

2. 2

   Protocol and registration

   Fix your methods before you see any results, then register on PROSPERO, OSF or INPLASY.

3. 3

   Search

   Several databases, registers and grey literature, with every search logged.

4. 4

   Screen

   Titles and abstracts, then full texts, by two people working independently.

5. 5

   Extract and appraise

   A piloted extraction form, and a risk-of-bias tool matched to each study design.

6. 6

   Synthesise and report

   Meta-analysis or narrative synthesis, reported with the PRISMA 2020 checklist and flow diagram.

The step people skip is the second. If you adjusted your criteria after the searches came back, you have a narrative review with extra paperwork, and an examiner can often tell from the dates.

How do you frame the question and register the protocol?

Everything downstream comes from the question: search terms, criteria, even the appraisal tool. PICO is the default in health research and fits badly outside it.

SPIDER comes from Cooke, Smith and Booth (Qualitative Health Research, 2012), who argued that PICO’s intervention-versus-comparison shape distorts qualitative questions. If your question won’t fit any framework cleanly, it is probably still too broad; our guide to the research problem statement helps with narrowing.

Take an illustrative PICO question: in adults with type 2 diabetes in low- and middle-income countries, do SMS or app reminders improve medication adherence compared with usual care? The criteria follow directly. Adults with type 2 diabetes. Any SMS or app reminder. Usual care as comparison. Adherence by any validated measure. Randomised trials only, from 2010 onwards. You will be asked at the viva why 2010, so decide for a reason.

The protocol (often five to ten pages, structured with the PRISMA-P checklist) fixes the question, criteria, databases, draft search string, screening process, extraction fields, appraisal tool and synthesis plan. Then register it. PROSPERO’s eligibility page limits it to reviews with outcomes of direct relevance to human health, and a review that has already progressed to data extraction is ineligible (check the page for the current wording before you register). For management, education or engineering topics, OSF Registries has a free systematic review template for any discipline, and INPLASY also accepts scoping reviews (it charges a fee, so check current rates). The trade-offs are set out on our systematic review support page.

How do you build and document the search strategy?

Use at least two or three databases suited to your field. Google Scholar alone won’t satisfy a reviewer.

• Health and nursing: PubMed/MEDLINE, Embase, CINAHL, Cochrane CENTRAL, and trial registers such as ClinicalTrials.gov and the Clinical Trials Registry – India (CTRI).
• Psychology and education: PsycINFO and ERIC, plus Scopus or Web of Science.
• Management, engineering, computing: Scopus and Web of Science, plus IEEE Xplore, ACM Digital Library or a business database, depending on your library’s subscriptions.

Each database gets its own version of the string, using controlled vocabulary where it exists (MeSH in PubMed, Emtree in Embase). For the diabetes example, a PubMed string might begin:

("Diabetes Mellitus, Type 2"[Mesh] OR "type 2 diabetes"[tiab]) AND ("Text Messaging"[Mesh] OR SMS[tiab] OR mHealth[tiab]) AND ("Medication Adherence"[Mesh] OR adherence[tiab])

Before trusting it, run it and confirm it retrieves five studies you are sure should be included.

Grey literature (registers, conference abstracts, reports, theses) catches what journals miss. For Indian topics, search Shodhganga; plenty of relevant work sits in theses that were never published. Add citation chasing on your included studies.

Log every search: database, platform, date, full string, number of records. PRISMA 2020 asks for the full strategy for every database, register and website, and the date each was last searched. Then deduplicate in a reference manager (see our Zotero vs Mendeley comparison) before screening.

How does screening by two reviewers work?

Twice over. First titles and abstracts, where you remove what is clearly irrelevant, then full texts judged against every criterion.

Chapter 4 of the Cochrane Handbook makes two people working independently mandatory for eligibility decisions, and desirable (not mandatory) for the title-and-abstract pass. Journal reviewers generally expect both. Neither reviewer sees the other’s decisions until both finish; conflicts go to discussion or a third person, often the supervisor. Pilot on 50 to 100 records first and tighten any criterion you disagreed on.

The real problem for most scholars is that there is no review team. Ask a fellow scholar in your department and screen their review in return. Name them in the methods. Claiming dual screening that didn’t happen is misreporting, and it can unravel in one viva question.

Record one reason for every full-text exclusion. Those counts go into the flow diagram.

How do you extract data, assess risk of bias and synthesise?

Design the extraction form first: study details, setting, sample, intervention or exposure, comparison, each outcome and how it was measured, and the results you will synthesise. Pilot it on two or three studies before using it on the rest.

Write one line of supporting text from the paper for each risk-of-bias judgement, so you can defend a rating without searching PDFs in front of the panel.

A meta-analysis pools effect sizes, and it is only meaningful when studies are similar in population, intervention and outcome measure. Otherwise, write a narrative synthesis that groups studies by outcome or theme and explains the patterns. The SWiM guideline (Campbell and colleagues, BMJ, 2020) says what to report when you don’t pool. I’d plan for narrative synthesis and treat a meta-analysis as something the data may allow; pooling five dissimilar studies adds little.

What goes in the PRISMA 2020 flow diagram and checklist?

PRISMA 2020 (Page and colleagues, BMJ, March 2021) replaces the 2009 statement. The flow diagram has three phases. Identification: records from databases, registers and other sources, minus those removed before screening, mostly duplicates. Screening: records screened and excluded, reports sought, reports not retrieved, reports assessed for eligibility and those excluded with reasons. Included: the number of studies in the review. The four-phase diagram with a separate “Eligibility” band belongs to 2009. The PRISMA website has four templates (new or updated review, with or without other sources such as citation searching).

An invented example shows the arithmetic. Four databases give 2,640 records and registers 48, so 2,688. Removing 1,012 duplicates leaves 1,676 to screen; 1,540 are excluded, so 136 reports are sought. Six can’t be obtained, 130 are assessed, 112 excluded (41 wrong population, 29 wrong intervention, 34 not randomised, 8 no adherence outcome). That leaves 18 studies. Examiners check these sums.

The 27-item checklist spans seven sections: title, abstract, introduction, methods, results, discussion and other information (registration, funding, competing interests, data availability). Some items have sub-items, and there is a separate 12-item abstract checklist. Fill in the page numbers against your final version.

Scoping reviews, which map what research exists instead of answering a focused question, are reported with PRISMA-ScR (Tricco and colleagues, 2018): 20 essential items and 2 optional ones.

How long does a systematic review take, and where does it fit in a PhD?

Longer than a supervisor who wants it “by Monday” thinks. Borah and colleagues (BMJ Open, 2017) found a mean of 67.3 weeks from registration to publication across 195 PROSPERO reviews, done by teams averaging five authors. A scholar working largely alone shouldn’t expect to be faster.

My own rough plan for a focused review is a year: two months for question and protocol, one for searching, three or four for screening, two or three for extraction and appraisal, two or three for synthesis and writing. A search returning 8,000 records instead of 1,500 changes everything.

It makes a good first paper. It uses reading you must do anyway, gives your literature chapter a defensible method, and can be finished before data collection. In the thesis it usually becomes chapter two, sometimes alongside a shorter chapter on theory and context; see our guide to literature review formats.

• Present the question and protocol at a DC or RAC meeting before registering, so the question is settled before screening starts.
• Keep one master screening file. The flow diagram numbers must match it.
• Write the methods as you go. Six months later you won’t remember why you dropped Embase.
• If more than about a year passes before submission, re-run the searches and report the new date.