Skip to content

>>> Blog

What to Put in a Data Extraction Form for a Systematic Review

10 min readLittview

Screening is over and forty studies made it through full text. Now each one has to be read again, and the same facts pulled out of it in the same way. Whatever the data extraction form forgets to ask, you will be reopening forty PDFs to find.

Here is the short answer. A data extraction form is the set of questions you answer for every included study, so that each one is described the same way. It records who extracted what from which report, whether the study still qualifies, how it was done, who or what it studied, what it examined, what was measured, the results your synthesis will use, and anything else worth keeping. Build it from the tables your review will publish, word each question so two people would answer it alike, and pilot it before the real extraction starts.

You will meet it as a data extraction form, a data extraction sheet or a data collection form; the three names mean the same thing, and a data extraction template is the form with its questions in and its answers still blank. The map below shows the eight sections most forms need, for an invented review: how do onboarding practices relate to whether new employees stay through their first year?

Data extraction form example: one review, eight sections

SectionWhat it recordsIn the onboarding review
Extraction detailsExtractor, date, study and reportsOne study, reported twice, linked
Eligibility checkStill in, or the reason it is outIn: new hires, quits measured
MethodsDesign, data collection, sampling, time frameSurvey, 12-month follow-up
Who or what was studiedPeople, organisations or cases, and the settingNew hires in logistics firms
What was examinedThe practice or experience, and any comparisonMentoring against a one-day induction
OutcomesWhat was measured, how, and whenQuits within 12 months, HR records
ResultsThe numbers or findings the synthesis usesQuit rates per group; interview themes
OtherFunding, conflicts, conclusions, notesFunded by the firm studied

Group the questions in the order papers usually present them, so an extractor reads each paper front to back once. The Cochrane Handbook gives the same advice, and a review of the guidance on extraction counts its chapter on collecting data among the most complete there is.

1.Extraction details and the eligibility check

The first two sections are the dull ones, and the ones you reach for when something goes wrong.

RecordWho extracted, the date, the form's version, a study ID and an ID for each report, then whether the study still meets the criteria, with the reason if not
For exampleStudy ONB-07: a journal article and a later conference paper on the same sample of new hires
Watch forTwo reports of one study counted as two studies. The study is the unit, not the report: link the reports before you synthesise (Cochrane Handbook, 5.2.1)

A study can fail the criteria on a detail screening missed. If it does, record the reason here: it is one of the near misses the PRISMA checklist asks you to cite.

2.Methods

How the study was done decides how much weight its results can carry, and which appraisal tool applies to it.

RecordThe design, how data were collected, how people or cases were sampled, the time frame, and the unit of analysis: person, work group, firm or school
For exampleCross-sectional survey · semi-structured interviews · case study of one plant · quasi-experiment across two cohorts · simulation study
Watch forA design label with nothing behind it. Your appraisal tool will ask what happened in the study, so record that, with a quote from the paper where it helps (Handbook, 5.3.2)

3.Who or what was studied

Describe the studies' samples well enough that a reader can judge whose experience the review speaks for.

RecordWho or what the study covered, how many, and the setting: country, sector, size of organisation. For people, always age and sex, unless the context makes them obvious
For exampleNew hires in logistics firms · first-year teachers · warehouses with under fifty staff · maintainers of open-source libraries
Watch forCharacteristics with no use. Collect one when it could change the findings or help a reader see where they apply (Handbook, 5.3.3)

4.What was examined

Write this section so that someone could repeat what the study looked at.

RecordThe practice, programme, exposure, concept or experience, in enough detail to repeat it, and what it was compared with, if anything
For exampleA mentoring programme, its length and who ran it · an AI demand-forecasting tool against the planners' previous method · a flipped classroom against lectures
Watch for"Usual practice" as the comparison. Record what usual practice was in that study

5.Outcomes

Outcomes get one row of questions each, because each can be measured in its own way.

RecordFor each outcome: whether it was measured, the measure or instrument, the metric, and when it was measured. For a scale, its range and whether a higher score is better
For exampleQuits within 12 months from HR records · intention to stay on a five-point scale · on-time delivery rate per quarter · dropout by the end of year one
Watch forAn outcome measured but not reported. Record that it was measured anyway: the gap matters when you judge bias from missing results (Handbook, Table 5.3.a)

6.Results

Results are where the form splits by kind of study, because numbers and themes are extracted differently.

Kind of studyResults fieldsIn the onboarding review
QuantitativePer group, outcome and time point: the numbers analysed and lost, summary data, the effect and its precisionQuit rate and headcount per group; an odds ratio with its interval
QualitativeEach finding in the authors' words, one illustration from the same paper, and how well it supports the finding"Left alone in the first week", with a new hire's quote

For numbers, copy what the paper prints and convert later, off the form. Record how many people dropped out of each group, with the reasons the paper gives.

For qualitative studies, the form asks for findings instead of numbers. The JBI Manual sets out a precise recipe for its meta-aggregation method: a standardised data extraction tool, filled in two passes, the study's characteristics first and its findings second. Each finding is the authors' own interpretation, taken word for word, with at least one illustration from the same paper (a participant's words, a field observation) and a credibility level: unequivocal, credible or not supported.

7.Other information

The last section catches what a reader will ask about the study's independence, and what you need to remember later.

RecordFunding, the authors' conflicts of interest, the authors' key conclusions, any correspondence with them, and free notes
For example"Funded by the logistics firm studied" · "Emailed the first author for the 12-month counts"
Watch forFunding and conflicts left out because they seem minor. The Handbook calls them a particularly important piece of information

How to word each question

A form goes wrong less often in what it asks than in how it asks it. These six rules come from the Cochrane Handbook's steps for designing a form (section 5.4.3).

RuleInstead ofAsk
Closed questions, with a way outDesign: ______Survey · interviews · case study · quasi-experiment · other, specify
No blank answersAn empty boxNot reported, not applicable or cannot tell, whichever the question needs
Ask about the reportDid new hires get a mentor?Does the report say new hires got a mentor?
Quote what you judged"Low response rate"The paper's sentence, with its page
Keep the reported formatA percentage you worked outThe figure as printed
Put definitions next to the questionA separate instructions fileA short note under the question

"Cannot tell" earns its place. The Handbook notes that it flags the items worth an email to the study's authors. JBI's scoping review guidance adds a habit worth copying. Give each item a name, a description, a format and its allowed values, so that a definition, for instance, is copied word for word rather than summarised.

Pilot the form, then extract

A form looks complete until real papers meet it. Piloting lets that happen on three papers instead of forty.

StepWhat to do
PlanSketch the tables and figures the review will publish, and let them set how much to ask: too much wastes time, too little sends you back to the papers (Handbook, 5.4.3)
PilotEveryone who will extract fills in the form, independently, for at least two or three papers of each kind of source (JBI Manual, 10.2.9.2)
FixAsk what was missing, what was redundant, what was ambiguous, and how long each paper took (JBI Manual, 10.2.9.2). Pilot again after major changes
FreezeFinish changes in the pilot. A later change may mean going back to the papers already done, and a change to what you collect is a protocol change to report (Handbook, 5.4.3; JBI Manual, 10.2.9.1)
ExtractTwo people, independently, at least for results and for anything that needs judgement (Handbook, 5.5.2)
ReconcileCompare, settle disagreements by a rule set in the protocol, and keep the values as first extracted next to the agreed ones (Handbook, 5.5.5)

When two full extractions are too much, a lighter arrangement is one extractor and one checker who confirms every value against the paper. It has a price: the Handbook cites a study in which it left more errors than two independent extractions (Handbook, 5.5.2). JBI accepts it for large scoping reviews, reported as a limitation, and the guidelines for systematic reviews in software engineering describe the same pairing and say it may help when there are many papers (section 6.4.3).

Working alone on a thesis? The same guidelines suggest that your supervisor extracts a random sample of the studies and compares results with yours, or that you extract a random sample a second time and compare yourself with yourself.

Whatever you choose goes into the paper. PRISMA 2020 items 9 and 10 ask how many people collected data from each report, whether they worked independently, and which data items were sought; item 27 asks whether your template data collection forms are publicly available (the checklist, item by item).

In Littview, when two or more people extract, each paper gets an extractor and a verifier, and every correction keeps the extractor's original value beside it. The form itself is built in the extraction template builder, where the ready-made question sets come with their help text and every new list of choices starts with Other and Not stated.

FAQ

Three questions come up most.

Is there a PRISMA data extraction form? No. PRISMA 2020 is a reporting guideline: it asks you to report how data were collected and which items were sought (items 9 and 10), and whether the forms you used are publicly available (item 27). The form itself is yours to build.

Can I use Excel for data extraction? Yes. JBI lists Excel on a shared drive among its examples, which keeps everyone on one file, and suggests dropdown lists for closed questions. The Cochrane Handbook's comparison of tools fits simple electronic forms to reviews of about 10 to 20 studies, and review software to any size, especially past 20 studies or when the review will be updated (Table 5.4.a).

Does risk of bias go on the data extraction form? It can. The software engineering guidelines require a separate form when quality questions decide inclusion, because that is settled before extraction, and allow them on the same form when they feed the analysis. Either way, collect the details your appraisal tool needs while you have the paper open.

For this week: write down the two or three tables your review will publish, then list the fields each one needs. That list is the first draft of your form. Pilot it on three papers with whoever will extract with you, and rewrite every question that made either of you hesitate.