PRISMA Flow Diagram: How to Report Your Search

What a Flow Diagram Actually Reports

PRISMA stands for Preferred Reporting Items for Systematic reviews and Meta-Analyses. It is a reporting guideline, which means it does not tell you how to do a review, only what you have to disclose about the review you did. The current version, PRISMA 2020, was published in 2021 and replaced the 2009 statement. The flow diagram is one item in a longer checklist, and it is the item readers look at first.

What it discloses is attrition. You started with a large pile of search results and ended with a small set of included studies, and the diagram accounts for the difference. That accounting is what makes a review checkable. Without it, a reader has no way to tell whether you screened four hundred records carefully or skimmed forty and stopped when the argument felt full.

One vocabulary note that trips almost everyone up. PRISMA 2020 distinguishes records, reports, and studies. A record is a database entry, a title and abstract in a results list. A report is the document itself, the full text you open. A study is the piece of research being described, and one study can be spread across two reports, a conference paper and a later journal article. The upper boxes count records, the lower boxes count reports and studies. Using the words loosely is the fastest way to produce a chart whose numbers cannot be made to agree.

Do You Actually Need One?

This depends entirely on the kind of review you are writing, so settle that first. A systematic review needs a flow diagram; it is close to non-negotiable, and reviewers will ask for it. A scoping review follows PRISMA-ScR, the scoping extension, which also expects a diagram of the selection process. A narrative or thematic literature review usually does not require one, and adding one does not upgrade the review. The practical differences are laid out in scoping vs systematic vs narrative review and in literature review vs systematic review.

There is a middle case worth naming, because it is where most dissertation students actually sit. You are writing a narrative review, no diagram is required, but you did run a structured search. Including a simple version of the chart, or even just the counts in a sentence, is a cheap way to show a marker that your selection was deliberate. Keep it honest about what it is. A diagram on a selective review is a record of your search, not a claim that the review was systematic.

The Four Stages, and the Number Each One Needs

The chart reads top to bottom. Each stage has records flowing in, records leaving, and a smaller number continuing to the next stage. Here is what each box is asking for.

Notice how many of those say “at the moment you”. That is the whole difficulty of this diagram. Not one of these numbers is hard to obtain while the work is happening, and several of them are effectively gone once you move on.

The Recording Habit That Makes This Easy

Open one spreadsheet before you run your first search and give it two tabs. The first tab is the search log, one row per database: database name, the exact string you ran, the date, any filters, and the number of hits. That tab is also the evidence for your search strategy, which is worth building deliberately rather than improvising; the method is in how to write a search strategy.

The second tab is the screening log, one row per record, with a decision column and a reason column. The reason column only needs filling in when you exclude at full text, and the reasons should be the mirror image of your eligibility rules. If you have not written those rules yet, do that before screening, not during it, using how to write inclusion and exclusion criteria. Criteria written midway through screening tend to describe the papers you have already decided to keep.

Two practical notes on duplicates. Deduplicate once, after you have merged every export into a single library, and record the count the software reports. Then check a sample by hand, because reference managers miss duplicates when a preprint and a published version carry different titles. If your library lives in Zotero,Mendeley, or EndNote, the deduplication behaviour differs between them, which is one of the differences covered in Zotero vs Mendeley vs EndNote.

A Worked Example, Box by Box

Here is a small review on remote work and employee wellbeing, with every number shown so you can see the arithmetic close.

Read it back as a chain and every step reconciles. 951 plus 17 is 968 identified. In the database column, 951 minus 268 duplicates minus 41 automation exclusions leaves 642 to screen, and 642 minus 558 excluded leaves 84 reports sought. The 17 records from other methods sit in their own column in the PRISMA 2020 template and join here, so 84 plus 17 is 101 sought. 101 minus 6 not retrieved leaves 95 assessed. 95 minus 73 excluded leaves 22 studies, and the four exclusion reasons themselves sum to 73. Nothing disappears without a stated cause. That property, not the shape of the boxes, is what a reviewer is checking.

Make the Arithmetic Add Up

The most common reason a flow diagram gets sent back is that the numbers do not reconcile. Before you paste the chart into your chapter, run these four checks in order. They take two minutes and catch nearly everything.

• Per-database totals sum to the identification total. If they do not, you probably re-ran one search after adding a filter and kept the old figure.
• Each stage balances. Screened minus excluded equals sought within the database column, then add the records found by other methods. Sought minus not retrieved equals assessed. Assessed minus excluded equals included.
• Exclusion reasons sum to the exclusion total. One record, one reason. Pick the first rule it failed rather than listing every rule it failed, or your reasons will overcount.
• Studies and reports are counted separately. If 22 studies appear in 25 documents, both numbers belong in the final box, and your reference list will have 25 entries while your synthesis discusses 22 studies.

Reporting Exclusion Reasons Well

Exclusion reasons are only required at the full text stage, and that is a relief, because nobody expects you to justify 558 individual title and abstract rejections. But the reasons you do give carry weight, and vague ones invite questions. “Not relevant” tells a reader nothing. “No wellbeing outcome measured” tells them exactly which of your criteria the paper failed.

Use between three and six reason categories, drawn word for word from your eligibility criteria, and give each a count. If one category swallows most of your exclusions, that is useful information rather than a problem: it usually means your search string is wider than your question, which you can note honestly in your limitations.

How to Draw It Without Losing an Afternoon

Do not build the boxes from scratch. The PRISMA website publishes templates for the 2020 diagram, including a version for reviews that searched databases only and a version with a second column for records found by other methods, such as citation chasing and organisational websites. Download the one that matches your search and type your numbers into it. There are also free tools that generate the diagram from a filled-in spreadsheet, which is worth a look if you would rather not fight with text boxes.

Keep the presentation plain. It goes in the methodology chapter as a numbered figure, referenced from the text, usually right after you describe screening. Same font as the rest of the document, no colour coding, no shadows. Check it exports legibly to PDF at actual size, because a diagram that is crisp on screen and blurred in the submitted file is a needless mark to lose. Where it sits relative to the rest of your method is covered in how to write a methodology section.

Five Ways This Goes Wrong

• Leaving the counts until the end. The one mistake that causes all the others. Screening decisions are easy to recall in aggregate and impossible to recall as numbers.
• Reporting one lump figure for all databases. PRISMA 2020 expects the count per source. A single total hides the possibility that one database contributed almost nothing, which is worth knowing.
• Deduplicating twice. Removing duplicates before and after merging exports produces two counts and a chart that cannot balance. Merge first, then deduplicate once, then record.
• Confusing reports with studies. A review of 22 studies across 25 papers is not a review of 25 studies. Collapse the duplicates into one study and say so.
• Treating the diagram as decoration. It is a figure your examiner may audit against your reference list. Numbers that do not match the list you actually cite is one of the common literature review mistakes that is genuinely hard to talk your way out of in a viva.

If You Have Already Lost the Numbers

Plenty of people arrive at this page after screening rather than before it. Not all is lost, and the recovery is mostly mechanical. Re-run each search string with the same filters and record today’s hit count and today’s date, noting in your method that the reported search date is the date of this rerun. Your reference manager can usually tell you the library size before and after deduplication if the groups still exist. The included set you already have, and the excluded full texts are often still sitting in a folder you can count.

What you cannot honestly reconstruct is the title and abstract exclusion count, because that number only ever existed in the act of screening. If a rerun gives you a defensible figure, use it and date it. If it does not, say plainly in your limitations that screening counts were recorded retrospectively. A stated limitation costs you very little. An invented number that a reviewer can disprove by re-running your string costs you a great deal more, and reviewers do re-run strings.

Do This in the Next Ten Minutes

Open a spreadsheet and make two tabs, search log and screening log. Put one row in the search log for every search you have already run, filling in whatever you can still recover, and leave the rest blank rather than guessing. From here on, the rule is one line per search and one line per decision, written at the moment it happens. The system does the remembering so you do not have to, and in nine months the diagram will be twenty minutes of typing instead of a lost weekend.

Once the chart is settled, the counting is over and the actual work begins: reading the included studies and working out what they say together. That is a different habit, and it starts with organising your sources and then synthesising them.