How to do data extraction for a systematic review
Short answer
Design a data extraction form from your protocol, pilot it on a few studies, then extract in duplicate (two reviewers independently, or one extracting and one verifying) and resolve discrepancies. Collect study characteristics, participants, interventions, comparators, outcomes and the numerical results you need for analysis, and keep a link from every value back to where it appears in the report.
Step by step
- 1
Decide what you need before you design the form
Start from your protocol: which outcomes, which effect measures, which subgroups and which characteristics you will use to describe the studies and to assess risk of bias. Collect what you will use; a very long form slows extraction and increases errors.
- 2
Build the form
Typical items include study identification (authors, year, report IDs), design and setting, funding and conflicts of interest, participants and eligibility, interventions and comparators as delivered, outcomes with their definitions and time points, and results: number analysed, events, means, standard deviations or effect estimates with confidence intervals.
- 3
Pilot the form on a few studies
Have the extractors use the form on a small, varied set of studies and compare their results. Fix ambiguous items, add missing options and agree on conventions, for example which time point to take when several are reported.
- 4
Extract in duplicate
The Cochrane Handbook recommends that more than one person extracts data from each report, at least for outcome data, to reduce errors. Compare the two extractions and resolve discrepancies by discussion, going back to the source.
- 5
Collate multiple reports of the same study
One study may appear in a journal article, a conference abstract and a registry entry. Extract per study, not per report, note which report each value came from, and watch for inconsistencies between reports.
- 6
Handle missing and differently reported data
When results are missing or unclear, contact the study authors. When data are reported differently, for example a standard error or a confidence interval instead of a standard deviation, use the conversion methods described in the Cochrane Handbook and record every conversion you made.
How Prolytiq does this step
Prolytiq organizes extraction as evidence tables: one row per study and arm, one column per variable you define. An AI extraction engine reads the full text and proposes values, each with the passage it came from, and reviewers confirm or correct them before they feed the statistics step. The proposals are a starting point, not a replacement for checking the source.
See the workflow →Frequently asked questions
Do I really need two people for data extraction?
Duplicate extraction is recommended by the Cochrane Handbook because single extraction produces more errors, particularly for outcome data. If you cannot extract everything in duplicate, at least have a second person verify the outcome data.
Should data extraction and risk of bias be done at the same time?
Often, yes. Both require reading the full report closely, so many teams do them in the same pass, while keeping the risk-of-bias judgments in their own structured tool.
Methodological sources
- Cochrane Handbook for Systematic Reviews of Interventions, chapter 5: Collecting data
- Cochrane Handbook for Systematic Reviews of Interventions, chapter 6: Choosing effect measures and computing estimates of effect
- PRISMA 2020 statement (Page et al., BMJ 2021)
Run your next systematic review in Prolytiq
From research question to publishable chart, with every decision logged and exportable for peer review.