Screening without losing your mindমাথা ঠিক রেখে বাছাই

Why screening goes wrong

A search brings back a thousand records. Most are not what you want, and the temptation is to start reading at the top and decide as you go. Three weeks later, record 700 is being judged by a stricter person than record 7 was, you cannot say why record 312 went out, and an examiner asks how many were excluded for each reason and you have nothing. Three failures, one cause: the rules were in your head, and your head changed.

Before the first record

1. Write the criteria as questions with a yes or no answer. Population, intervention or exposure, comparison, outcome, study design, dates, languages. Then write the exclusion reasons as a short fixed list with a code: E1 wrong population, E2 no comparison group, E3 not empirical, E4 outcome not measured, E5 language. Put them in order, because a paper that fails two gets the first one it fails.
2. Give every record an id. Export each database search, note how many came from each, remove the duplicates, and number what is left R0001 onwards. An id is never changed and never reused. If you rerun the search later, the new records start where the old ones stopped.
3. Pilot on fifty. Two people, blind to each other, on the same fifty. Where you disagree, the criterion is unclear: fix the wording, not the decision, and start the real screen only after that.

Pass one: title and abstract

Fast and generous. The question is "could this be included?", not "is it included?". If in doubt, keep it. You are removing what is plainly out: the wrong topic, an editorial, a conference abstract with no data. Every exclusion still gets a code, because the codes are what the diagram is drawn from. Eight or nine in ten records usually leave here.

Pass two: full text

Slow and strict. Now the question is "does it meet every criterion?". Fetch the paper. If after a real attempt (the library, an email to the author) you cannot, that is "not retrieved", and it is its own count rather than a quiet exclusion. Each excluded paper gets one main reason, the first on your ordered list that it fails. One reason each is what makes the reasons add up to the total.

A second screener and one number

Cohen's kappa compares how often two screeners agreed with how often they would have agreed by chance, so 1 is perfect agreement, 0 is chance, and anything above about 0.6 is usually taken as good enough to carry on. Disagreements are settled by talking, or by a third person, and the settlement is written down beside the record.

The counts become the diagram

Where it goes wrong

• The criteria were tightened halfway through, and records 1 to 400 were never screened again under the new rule.
• Exclusion reason: "not relevant". Nobody, including you in six months, can say what that meant.
• One screener, no sample checked by a second, and no kappa. An examiner will ask.
• A paper excluded under two reasons, so the reasons add to more than the exclusions.
• The search was rerun before submission and the new records were merged into the old numbering.
• Reports counted as studies, so three papers on one trial became three studies.

Never delete a record. A record that is out is a record with a reason on it. Deleting is how the boxes stop adding up, and there is no way back.

Checklist

• Criteria and exclusion codes written and dated before screening began.
• Count from each database recorded before deduplication.
• Every record has an id that has not changed.
• Every excluded record has exactly one code.
• A second screener has done at least a sample, and kappa is reported.
• "Not retrieved" is counted separately from "excluded".
• Every subtraction in the diagram has been checked by hand.

The review room in the studio holds the records with their ids and a reason on every exclusion, and the workshop's PRISMA drawer draws the diagram out of those counts, so the boxes cannot disagree with the table.