Strata Academy

Systematic Review Data Extraction: Forms, Dual Review & PRISMA

Design extraction forms, pilot on 2–3 studies, resolve disagreements, and link extracted data to risk-of-bias and meta-analysis

• Extract from full text — abstracts miss arms, outcomes, and analysis plans.
• Pilot the form before scaling; ambiguous fields mean ambiguous inclusion criteria.
• Link each extracted outcome to ROB 2 / ROBINS-I / QUADAS-2 judgements.
• Report extraction methods in PRISMA Item 9.

1. Why extraction is not copy-paste from PDFs

After screening, each included study must be converted into structured data: population characteristics, interventions, comparators, outcomes (with time points), effect sizes, variance measures, and funding or conflicts. This step determines what enters meta-analysis, narrative synthesis, and GRADE tables.

Abstract-only extraction is a common student error. Abstracts often omit non-significant secondary outcomes, per-protocol analyses, or subgroup results that appear only in tables. Extraction from full text — including supplements — is the minimum standard for coursework marked at postgraduate level.

Poor extraction propagates through the entire review: wrong effect sizes in forest plots, misclassified study designs in subgroup analyses, and GRADE certainty statements that do not match the underlying data.

Note: If two reviewers extract independently and disagree on the primary outcome effect size, treat this as a signal to re-read the paper — not as noise to average away without discussion.

2. Building the extraction form

Start from your registered PICO and pre-specified outcomes. Each field on the form should map to an analysis planned in the protocol. Ad hoc fields added mid-review without protocol amendment weaken reproducibility.

Typical sections: bibliographic identifiers (DOI, trial registry ID), design and setting, population and eligibility, intervention and comparator details, outcome definitions and time points, effect estimates with CIs or SEs, sample sizes, missing data handling, funding, and conflicts of interest.

For meta-analysis, record enough detail to compute or verify effect sizes: means and SDs for continuous outcomes, events and totals for dichotomous outcomes, hazard ratios with CIs for time-to-event data. Note whether ITT or per-protocol analysis was reported.

• Use consistent units (e.g. mmHg, mg/dL) and document conversions
• Record multiple arms separately before choosing comparison pairs for synthesis
• Capture registration ID and compare pre-specified vs reported outcomes (selective reporting)
• Note imputation methods if authors used multiple imputation or LOCF
• Store PDF page numbers for audit during viva or supervision

Tip: Pilot on 2–3 diverse included studies. If reviewers cannot agree on field definitions, revise the form and update the protocol before extracting the rest.

3. Dual independent extraction

Cochrane recommends independent extraction by two reviewers for key data, with a third resolver for conflicts. This applies especially to outcome data, sample sizes, and risk-of-bias judgements that feed directly into synthesis.

Single extraction is acceptable only in rapid reviews with explicit limitation statements. For dissertation or publication targets, single extraction without justification will attract methodological criticism.

Blinding extractors to study authors and journal where feasible reduces prestige bias. Covidence, DistillerSR, and Rayyan support dual extraction workflows with conflict flags.

4. Linking extraction to risk of bias

Extract data needed for framework judgements in the same pass where possible. For ROB 2: randomisation method, allocation concealment, blinding, attrition flow, and outcome assessor blinding. For ROBINS-I: confounding control, selection, classification of interventions, and missing data.

Do not treat risk of bias as a separate checkbox exercise after extraction. The same missing ITT table that blocks effect size extraction often drives a high-risk judgement in Domain 3 (missing outcome data).

• ROB 2 Domain 5 — compare trial registry or protocol to published outcomes
• ROBINS-I — record confounders adjusted vs those mentioned in limitations
• QUADAS-2 — extract index test, reference standard, and patient flow on one form
• AMSTAR 2 — extract review-level methods (search dates, dual screening, RoB tool used)

5. From extraction sheet to meta-analysis

Before pooling, check clinical homogeneity: are populations, interventions, and outcome definitions similar enough that a common effect is meaningful? Statistical heterogeneity (I², τ²) is assessed after clinical plausibility — not instead of it.

Transform effect sizes consistently (log OR, log RR, mean difference). Document assumptions when converting SE from CI or p-value. Cochrane RevMan and R meta packages expect specific input formats — validate a subset manually.

Contact authors for missing data when pre-specified in the protocol. Document non-response as a limitation. Do not impute missing study-level data without sensitivity analysis and clear assumptions.

6. Reporting extraction in PRISMA

PRISMA 2020 Item 9 expects description of data extraction methods: number of reviewers, processes for resolving disagreements, and software used. Item 10 covers data items extracted.

If extraction methods changed after piloting, report the change transparently. Post-hoc changes without protocol update are a common reason for major revision at peer review.

Frequently asked questions

Can I extract data from abstracts only for a student systematic review?

Only for a deliberately scoped scoping or narrative exercise with supervisor approval. For a systematic review with synthesis or AMSTAR 2 appraisal, full-text extraction is expected. Abstracts frequently omit non-significant outcomes and analysis details.

How many studies should I pilot before full extraction?

Cochrane guidance suggests piloting on 2–3 included studies representing design diversity (e.g. one RCT, one cohort). Revise the form until independent extractors agree on field definitions.

What software should I use for data extraction?

Covidence, Rayyan, DistillerSR, or structured spreadsheets with version control all work. Choose one tool early, document it in PROSPERO, and export an audit trail for PRISMA.

Do I need dual extraction for study characteristics as well as outcomes?

Dual extraction is most critical for outcome data and risk-of-bias judgements. Study characteristics may use dual extraction or independent verification depending on review rigour — but single unverified extraction should be acknowledged as a limitation.

Interactive walkthroughs and quizzes load when JavaScript is enabled — the checklist and tables above are fully readable without it.