Data Extraction for Meta-Analysis:When Identical-Looking Outcomes Differ
You've extracted the numbers from the included papers and built the table. Then, at pooling time, the doubt creeps in — "are the values in this study and that one really measuring the same thing?" — and you end up reopening every paper. The real difficulty of extraction for meta-analysis isn't copying numbers. It's spotting the things that look the same but aren't, and confirming correspondence between studies. This article organizes the practical work around five checks.
Published: 2026-07-22
Meta-analysis extraction builds a correspondence table
If an ordinary review's extraction table is a list of information from each study, a meta-analysis extraction table has a stricter job: it must hold numbers that genuinely correspond, because they are the raw material for pooling.
So every cell carries an implicit question: is this value truly comparable with the values in the same column from the other studies? We break that question into five checks.
Check 1: outcome definition
Two studies can both report "mortality" and mean different things — all-cause versus disease-specific. Two "improvement" outcomes measured on different scales can't be pooled either.
Extract how each study defined its outcome from the Methods and list the definitions side by side. Match definitions, not labels — that's the first verification step.
Check 2: measurement time point
"Post-treatment outcome" means something different at 4 weeks than at 1 year.
Always give time points their own column, and decide as a team which time point to pool at (or whether to analyze time points separately). Timing information tends to be scattered through the text — sometimes it appears only on the x-axis of a figure. Check every figure and table as well.
Check 3: group definitions
The composition of the intervention and control groups needs correspondence checking too. A study where "control = placebo" and one where "control = usual care" are making different comparisons.
Recording dose ranges and durations here pays off later, when you're probing heterogeneity and discussing subgroups.
Check 4: analysis population
The denominator behind a value differs depending on whether it comes from an intention-to-treat analysis, a per-protocol analysis, or completers only — and within a single study, the denominator can change from table to table.
Record which analysis population each value came from, together with the source table number. A simple "Table 3, ITT" saves real time at verification and peer review. Agree on a denominator policy as a team beforehand.
Check 5: effect size format
Finally, the statistical format. Is the reported value a mean difference or a ratio? Is the spread an SD (standard deviation), an SE (standard error), or a confidence interval?
Confusing SD with SE is the classic serious extraction error in meta-analysis. Converting between formats involves statistical judgment, so don't attempt conversions on your own — consult a statistician or your supervisor.
Divide the work: AI drafts, humans verify everything
Extraction built around these five checks can still be drafted by AI. Ask it to "extract the outcome definition, time point, group definitions, analysis population, and effect size from this paper in this format," and the first pass of the table comes together fast.
But because these values feed the pooling, verification covers every data point, against the original text, figures, tables, and supplements. The AI draft is only the opening move, never a substitute for verification. When you find an error, correct the draft and mark it verified.
Keep the extraction record with the paper
In Paperfy, each included paper's AI summary, figures, and original PDF share one page, so retracing "where you looked in the source" along the five checks is fast.
The summary is hand-editable, so you can add confirmed notes to the paper's page — "Outcome: all-cause mortality (Methods p.4) / Time point: 12 months / Denominator: ITT (Table 2)". When doubt hits during pooling, that layer of notes between the extraction table and the original text cuts your recovery time dramatically.
Correspondence checks for your extraction table
- Outcomes were matched by definition, not by label.
- Time points and group definitions each have their own column.
- The analysis population and source table number are recorded together.
- Effect size formats (SD vs. SE, etc.) were checked, with an expert consulted for any conversion.
Verification notes that erase the time lost to re-checking
With Paperfy, all five verification marks stay on the paper's page — one click back to the original whenever pooling doubt strikes.
Meta-analysis
Explore This Topic
Read the full cluster for this workflow, from first steps to practical use.
Read Next
These guides continue the same workflow.