Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data Analytics & AI · what data to extract

Data Extraction for a Systematic Review: Determining What to Extract

The items to be extracted are determined before extraction begins, from the review question and the planned analyses, and the list is longer than most teams anticipate. Cochrane’s methodological standard is that at least two people extract data independently, and the Handbook scales the team to the size of the review. An outcome is not a single field: it has five elements, and the omission of any of them produces a result that cannot later be interpreted.

Pubrica AcademyWritten by the medical writing and biostatistics teams Reviewed byPubrica biostatistics team Published9 October 2026 Last checked9 October 2026 Reading time12 minutes Sources29 peer-reviewed

Data extraction is the stage at which a systematic review ceases to be a search and becomes a dataset. Data not extracted at this stage must either be retrieved later, at several times the cost across dozens of papers, or forgone, which commonly results in a planned subgroup analysis being omitted from the manuscript without explanation. The question of what to extract has a procedural answer: everything that the planned analyses, the characteristics table, the risk-of-bias assessment and the reporting checklist will require, decided in advance and piloted on real papers before extraction begins.

Learning objectives

  • Construction of an extraction list from the planned analyses rather than from the available papers
  • Recording of all five elements of every outcome, so that each result remains interpretable
  • Sizing and structuring of the extraction team according to the Handbook’s guidance
  • Distinction between extracting too much and extracting the wrong items
  • Assessment of where automated extraction can currently assist and where it cannot
  • Design and piloting of a form that produces an analysable dataset rather than unstructured notes

Common misconceptions

  • “Effect estimates can be extracted first, with other fields added if required.” Returning to included papers for an unplanned field is the most expensive avoidable task in a review, and it is where errors cluster.
  • “One person extracts, a second checks.” Cochrane’s MECIR standard for data collection is at least two people working independently. Checking someone else’s completed form is a different and weaker procedure, because it anchors the second reader on the first reader’s interpretation.
  • “An outcome is a number.” Review authors should collect all five elements of an outcome. A mean without an indication of variability, a sample size, a timepoint or an instrument does not constitute an interpretable result.
  • “Everything should be extracted as a precaution.” An oversized form slows extraction, lowers accuracy on the fields that matter and fills the dataset with columns that are never analysed. Extraction volume should be determined by the planned analyses rather than by precaution.
  • “Automated extraction can take this over.” Automated tools each extract only a small number of data elements, ranging from one to seven, and well over half of the data elements needed for systematic reviews have not been explored at all.

Deriving extraction items from the planned analyses

The most reliable way to decide what to extract is to list every output the review will produce and then determine what each requires. There are five such outputs, which together generate almost the entire field list.

The characteristics of included studies table needs the descriptive fields: design, setting, country, recruitment years, participant characteristics, intervention and comparator as actually delivered, outcome instruments and timepoints, funding. The planned syntheses need the results data in a form the analysis can use — which, for each outcome and comparison, means a specific set of numbers rather than whatever the paper happens to report. The pre-specified subgroup and sensitivity analyses need the variables they are defined on; a planned subgroup by baseline severity is impossible if baseline severity was not extracted 4. The risk-of-bias assessment needs the supporting text, because a judgement with no quoted justification cannot be checked or defended. And the reporting checklist needs items such as funding source, conflicts of interest and registration identifier, which teams routinely forget because they are not analytic fields 8.

This exercise should be completed on paper before the form is built. It takes approximately an hour and determines whether one pass through the papers is required or two. Our page on framing the question with PICO covers where the list of planned analyses comes from in the first place.

Categories of data required in every review

Six categories are required, listed here in the order in which they belong on a form.

Provenance and eligibility. Study identifier, all reports of the same study, the eligibility decision and the criterion that settled it, registration identifier, language, publication status. The second of these is more important than it appears: multiple reports of one study, and one report containing several studies, are both common, and the unit of analysis must be the study, not the paper 1,9.

Methods. Design, unit of randomisation or allocation, sequence generation, blinding, sample size calculation, analysis population, handling of missing data, follow-up duration and attrition.

Participants. Eligibility criteria as the study stated them, recruitment setting and dates, numbers screened, randomised and analysed, and the baseline characteristics relevant to the planned subgroup analyses.

Interventions and comparators. What was delivered, at what dose or intensity, for how long, by whom, with what co-interventions, and against what — the comparator described as the study described it, not as the reviewer would summarise it 18.

Outcomes and results. The five elements of each outcome, covered in the next section, plus which outcomes were pre-specified in the study’s own protocol or registration where these are accessible 1,16.

Risk of bias, funding and conflicts. The judgement for each domain and the quoted text supporting it, the funding source, the role of any commercial funder, and declared conflicts.

The five elements of an outcome

This is the most useful rule in this article, and it derives directly from the Handbook: review authors should collect all five elements of an outcome 1. When an outcome is treated as a structured object rather than a single value, the common failures are avoided.

The elements are the outcome domain — the thing being measured, such as pain or depression; the specific measurement, meaning the instrument, scale or definition used; the specific metric, meaning the form the result takes, such as a value at a timepoint, a change from baseline, or an event occurrence; the method of aggregation, meaning how participant-level data were combined into a study result, such as a mean, a proportion or a rate; and the time point or period of measurement 1,24.

Two of these elements are where reviews most often fail. A metric mix-up, in which final values are combined with change-from-baseline scores without this being recognised, produces a pooled estimate that answers no question 25. And a missing timepoint turns a forest plot into a collection of results measured anywhere from two weeks to two years. Both are invisible once the number is in a spreadsheet cell, which is why the form must have a column for each element rather than one column called “outcome”.

For each outcome, it should also be recorded whether the study reported it in a usable form, in a form requiring conversion, or not at all. That third state is not an empty cell; it is a finding, and it informs the reporting-bias discussion. The conversions themselves are covered in extracting quantitative data and standardised mean difference estimation.

The five elements of an outcome, and what goes wrong when one is missing

ElementWhat it recordsConsequence of a missing value
Outcome domainThe construct being measured — pain, depression, mortality, infectionStudies measuring different constructs get pooled as though they measured one
Specific measurementThe instrument, scale or clinical definition usedNo way to judge whether two results are comparable, or to convert between scales
Specific metricThe form of the result — value at a timepoint, change from baseline, event occurrenceFinal values and change scores get combined, producing an estimate that answers no question
Method of aggregationHow participant data became a study result — mean, median, proportion, rateVariability measures get misread, and the wrong analysis method is applied
Time point or periodWhen the outcome was measured, or over what windowA forest plot silently mixes short-term and long-term results

Element structure as set out in the Cochrane Handbook for Systematic Reviews of Interventions, Chapter 5.

5elements of an outcome the Handbook says to collect
2people working independently, minimum, under Cochrane’s MECIR standard
1–7data elements a single automated extraction tool handles

Extraction requirements from the Cochrane Handbook for Systematic Reviews of Interventions, Chapter 5, on collecting data; the automated-tool range is from the review of natural language processing and machine learning approaches to data extraction cited there.

Extraction of source results data rather than derived values

For each outcome and each comparison, the numbers the paper actually reports should be extracted, at the most granular level available. For dichotomous outcomes, this means events and totals per group, not a risk ratio. For continuous outcomes, it means the mean, a measure of variability with its type named explicitly, and the number analysed per group, not a standardised effect computed during extraction 25. For time-to-event outcomes it means the hazard ratio with its confidence interval, the number of events, and the information needed to reconstruct it if the paper reports only a figure.

The reason is auditability. Every derivation performed during extraction is a step that a second extractor cannot verify against the source, and one in which an error is unlikely to be detected. Raw values should be extracted and their source location recorded (table 2, page 4, or the figure from which the value was read), with derivations performed in the analysis dataset where they remain visible. If a value came from a graph, a conference abstract, author correspondence or a registry rather than the paper, this should be recorded, because that provenance affects how much weight a reader should place on it 9.

Where variability is reported in an unusable form, what was reported should be recorded and flagged rather than silently converted. Where the number analysed differs from the number randomised, both should be recorded. Where a paper reports a result only in text and inconsistently with its own table, both values should be recorded and treated as a discrepancy; see handling discrepancies during data collection.

Over-extraction and inappropriate extraction

Both failures are common and appear similar from the outside. The diagnostic question is whether a field is required by a planned output. If no planned table, analysis, appraisal or checklist item requires a field, recording it constitutes note-taking rather than data collection.

Over-extraction has real costs. Long forms take longer per paper, which matters when the Handbook’s team guidance implies dozens of person-hours; attention degrades across a long form, so accuracy on the fields that matter falls; and a dataset with many unanalysed columns invites post hoc exploration that was never planned 1. Under-extraction costs more but later: a second pass through the included studies, usually under time pressure, with the original extractors no longer available.

Three fields warrant extraction even though they appear optional, because their omission is regretted more often than that of any others. Funding source and funder role, which the reporting checklists require and which inform the bias discussion. The study’s own registration or protocol identifier, which allows the reported outcomes to be compared with the planned ones 16,19. And the exact source location of every result, which is what makes verification possible at all.

Composition and size of the extraction team

Cochrane’s MECIR standard for data collection is to use at least two people working independently 1,2. Independently means both extract from the source into their own form without seeing the other’s entries. A single extractor with a second person checking the completed form is weaker, because the checker is anchored on the first reading and tends to verify rather than re-read 26.

The Handbook also scales the team to the size of the review. Small-scale reviews of fewer than 10 included studies are expected to use a small team of 2 to 3 data extractors; small-to-medium reviews of 10 to 20 studies, 4 to 6; and larger reviews more than 6 1. The purpose of the table is not a staffing rule for its own sake but a warning: extraction effort grows with the number of studies and the number of outcomes, and a team sized for a 10-study review is likely to miss its deadline on a 40-study review, or to compromise on rigour.

Two further structural decisions should be made deliberately. The first is whether extractors specialise (one on methods and risk of bias, one on results), which increases consistency within a field at the cost of cross-checking between fields. The second is who arbitrates, named in advance, so that disagreements are referred to a person rather than to a meeting.

Extraction team size, scaled to review size

Review sizeIncluded studiesData extractors
Small-scale reviewFewer than 10A small team of 2 to 3
Small-to-medium review10 to 204 to 6
Larger reviewMore than 20More than 6

Cochrane Handbook for Systematic Reviews of Interventions, Table 5.4.a. Cochrane’s MECIR standard for data collection is at least two people working independently regardless of review size.

Current capabilities and limitations of automated extraction

At present, automation assists with only a narrow portion of extraction. Automated tools each extract only a small number of data elements, ranging from one to seven, and well over half of the data elements needed for systematic reviews have not been explored at all in this literature 3. This is a useful capability but in no sense a replacement for a human extractor working to a form.

At present it is useful as a second check on the fields it handles, as a means of flagging likely locations of a value in a long paper, and as a consistency check across studies after human extraction. It is not useful for anything requiring judgement about what a study actually did, which includes most of the methods and risk-of-bias fields and all decisions about whether two studies measured the same construct 14,15.

If a tool is used, its output should be treated as one extractor’s entry rather than as the record: it enters the reconciliation process alongside the human entries, and the review reports which fields were machine-assisted and how disagreements with human extraction were resolved. A tool used in this way adds a check. A tool used as the primary extractor adds an unverified layer, and neither the review team nor the reader can determine where it erred.

Design and piloting of the extraction form

A well-designed extraction form has four properties. One field per concept, so that nothing has to be parsed later. Closed response options wherever a judgement is being recorded, with a free-text box for the supporting quotation beside it. An explicit distinction between “not reported”, “reported as none” and “not applicable”, which are three different states that a blank cell collapses into one. And a source-location field next to every extracted result.

The form should be built in a tool that produces an analysable dataset (a structured data capture system or review software rather than a word-processor table), so that extraction output enters the analysis without being retyped. Structured data capture justifies the setup time for any review with more than a small number of studies.

The form should then be piloted. Three to five included studies, deliberately chosen to be difficult (one poorly reported, one with multiple arms, one with several timepoints), should be extracted independently by all extractors and compared field by field. The pilot tests the form, not the extractors: every disagreement indicates a candidate for a clearer field label, a closed option list or an explicit rule. The form should be revised, piloted again if the first round revealed substantial problems, and then frozen with the date recorded. Changes after that point are amendments and should be logged as such.

One outcome block on an extraction form

Produces a number nobody can interpret

Outcome 1: Pain. Result: 3.2 vs 4.8, p=0.03. Notes: significant improvement in intervention group. Sample size: 84.

No instrument named. “Pain” measured on an unnamed scale cannot be compared with pain on any other scale, and cannot be converted.

No metric stated. Whether 3.2 is a final value or a change from baseline determines whether it can be pooled with the next study at all.

No timepoint. This result will end up in a forest plot alongside results measured at a completely different follow-up.

Variability replaced by a p value. A p value cannot be converted into the standard deviation an analysis needs, and the type of any variability measure was never recorded.

One sample size for two groups. Per-group numbers analysed are what the analysis requires, and the randomised-versus-analysed distinction is lost.

“Significant improvement” is an interpretation, not data. It also presumes a direction of benefit that was never recorded.

No source location. The second extractor must search the paper to verify each value, so in practice verification seldom occurs.

Produces a result that stays interpretable

Outcome domain: pain intensity. Specific measurement: visual analogue scale, 0 to 100, higher is worse, as reported by the study. Specific metric: final value at the measurement timepoint (not change from baseline). Method of aggregation: arithmetic mean. Time point: 12 weeks after randomisation; the study’s nearest reported timepoint to our pre-specified 12 weeks was 12 weeks exactly. Intervention group: mean, variability measure with its type named explicitly as standard deviation, and number analysed, each entered in its own field. Comparator group: the same three fields. Numbers randomised per group recorded separately from numbers analysed per group, with the reason for any difference as stated by the study. Source location: Table 2, page 1042. Value read from: printed table (not estimated from a figure, not supplied by the author). Reported in a usable form: yes. Direction of benefit: lower score is better. Pre-specified in the study’s registration: yes, identifier recorded. Extractor: initials. Date: entered. Discrepancy with second extractor: none.

All five elements of the outcome are present, so the result can be matched correctly against other studies or correctly excluded from a pool.

The metric is stated explicitly, which is the field that prevents the commonest and least visible pooling error in meta-analysis.

Variability is recorded with its type named, so the analysis applies the right conversion rather than guessing.

Randomised and analysed numbers are separated, which is what lets an attrition or intention-to-treat sensitivity analysis actually be run.

Provenance is recorded — table, figure, correspondence or registry — so a reader knows how much weight the number carries.

The source location makes independent verification a ten-second task, which is the only way double extraction remains rigorous under deadline pressure.

Whether the outcome was pre-specified in the study’s own registration is captured, feeding the selective-reporting assessment rather than being reconstructed later.

The second block takes perhaps ninety seconds longer per outcome and removes the two errors that most often force a review to be reanalysed: mixing metrics and mixing timepoints. It also makes every number checkable by a second extractor without re-reading the paper.

Key points

Key points

The items to extract should be decided by listing the review’s planned outputs (the characteristics table, each planned synthesis, each pre-specified subgroup and sensitivity analysis, the risk-of-bias assessment and the reporting checklist) and extracting exactly what those outputs require. Every outcome should be treated as five elements rather than a number: domain, specific measurement, specific metric, method of aggregation and timepoint, all of which the Handbook advises collecting. Raw numbers should be extracted with their source location rather than as derived values, and the three distinct states of “not reported”, “reported as none” and “not applicable” should be recorded. At least two people should work independently, as the Methodological Expectations of Cochrane Intervention Reviews (MECIR) standard requires, and the team should be scaled to the review: 2 to 3 extractors for fewer than 10 studies, 4 to 6 for 10 to 20, and more than 6 above that. Automated tools currently handle only one to seven data elements each, with well over half of the elements a review needs unexplored, so any tool output should be treated as one extractor’s entry rather than as the record.

Application to your study

Application to your study

  1. Outputs should be listed before fields. Every table, analysis, subgroup, appraisal domain and checklist item that the review will produce should be listed, and the field list derived from them.
  2. Every outcome should be divided into five columns. Domain, measurement, metric, aggregation, timepoint — one field each, never a single column called outcome.
  3. Raw data should be extracted with their source recorded. Events and totals, means with a named variability type and per-group numbers analysed, plus the table or page the value came from.
  4. Three kinds of blank entry should be distinguished. Not reported, reported as none, and not applicable need separate response options, because they mean different things in the analysis.
  5. The team should be staffed to the Handbook’s scale, and the arbitrator named. At least two independent extractors, sized to the number of studies, with the person who resolves disagreements agreed in advance.
  6. The form should be piloted on the three most problematic papers. Suitable papers are those that are poorly reported, multi-arm or multi-timepoint; the form, rather than the extractors, should then be corrected.
  7. The final form should be frozen and dated. Any later change should be logged as an amendment, recording what changed and whether any results were already visible.

Frequently asked questions

The outcome domain — the construct being measured, such as pain or mortality; the specific measurement, meaning the instrument, scale or clinical definition used; the specific metric, meaning the form the result takes, such as a value at a timepoint, a change from baseline or an event occurrence; the method of aggregation, meaning how participant-level data were combined into a study result, such as a mean, proportion or rate; and the time point or period of measurement. The Cochrane Handbook states that review authors should collect all five. Recording them in separate fields prevents the two errors that most often require a meta-analysis to be redone: combining final values with change-from-baseline scores, and pooling results measured at incompatible follow-up times.

Cochrane’s MECIR standard for data collection is at least two people working independently, with disagreements resolved by discussion or a third assessor. The Handbook also scales the team to the review: fewer than 10 included studies, a small team of 2 to 3 data extractors; 10 to 20 studies, 4 to 6; more than 20 studies, more than 6. “Independently” is the operative word. One person extracting while a second checks the completed form is a weaker procedure, because the checker is anchored on the first reading and tends to confirm it rather than re-read the source.

The raw numbers, at the most granular level the paper reports. For a dichotomous outcome, this means events and participants per group rather than a risk ratio; for a continuous outcome, the mean, a variability measure with its type named, and the number analysed per group rather than a standardised effect computed during extraction; for time-to-event data, the hazard ratio with its interval plus the number of events. There are two reasons. Any derivation performed during extraction is a step the second extractor cannot verify against the source, so its errors are invisible. In addition, raw data allow analysis decisions to be changed later (a different effect measure, or a different rule for handling variability) without returning to every paper.

The absence should be recorded as a value rather than left as a blank cell, distinguishing between not reported, reported as none and not applicable, three different states that a blank collapses into one. The available options should then be pursued in turn: the data may be found in other reports of the same study, in the trial registry entry, in a conference abstract, or in the regulatory record where one exists; where the paper reports a figure but no table, it should be recorded that the value was read from a figure; and where none of these succeeds, the corresponding author should be contacted with a specific request for the named outcome at the named timepoint. The review should report what was requested, how often, and how often a reply was received. A study that measured the outcome of interest but did not report it is a reporting-bias signal, not merely a missing row.

Not yet, as the primary extractor. The published evidence is that automated tools each extract only a small number of data elements, ranging from one to seven, and that well over half of the data elements a systematic review needs have not been explored in this literature at all. This supports machine assistance on a narrow set of fields: locating likely values in a long paper, acting as a second check on the fields a tool handles well, and flagging inconsistencies across studies after human extraction. If a tool is used, its output should be treated as one extractor’s entry that enters reconciliation alongside the human entries, and the review should report which fields were machine-assisted and how disagreements were resolved.

Detailed enough that a reader could tell whether two studies delivered the same thing, which in practice means the agent or procedure, the dose or intensity, the schedule, the duration, who delivered it and with what training, and which co-interventions were permitted. The comparator requires the same level of detail and usually receives less: “usual care” should be extracted as the study described it, not summarised, because usual care differs by health system and by decade, and pooling across an undescribed comparator pools across an unmeasured intervention. For complex and multi-component interventions, a structured intervention description framework keeps the fields consistent across studies and makes the gaps visible.

Yes, although it is the less damaging of the two errors. A long form takes longer per study, attention degrades across it so that accuracy falls on the fields that matter, and a dataset full of unanalysed columns invites post hoc exploration that was never planned and may be interpreted as data dredging. The test for any field is whether a planned output requires it: a table, a synthesis, a pre-specified subgroup, a risk-of-bias domain or a reporting checklist item. Three fields warrant retention even though they appear optional, because teams most often regret omitting them: funding source and funder role, the study’s own registration identifier, and the exact source location of every extracted result.

Before extraction starts, and after a pilot on real papers. The form should be built from the planned outputs, and every extractor should then complete it independently on three to five deliberately difficult included studies (one poorly reported, one with multiple arms, one with several timepoints), with the results compared field by field. Every disagreement is evidence about the form rather than about the extractor, and the remedy is usually a clearer label, a closed option list or an explicit rule. The form should be revised, piloted again if the first round revealed substantial problems, and then frozen with the date recorded. Later changes are legitimate but constitute amendments: what changed, when, and whether any results were visible at the time should be logged.

References

  1. Chapter 5: Collecting data. Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Cochrane. 2024;6.5. read 9 October 2026. cochrane.org
  2. Higgins JPT, Thomas J, Chandler J, Cumpston M, Li T, Page MJ, Welch VA (editors). Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Cochrane. 2024;6.5. read 9 October 2026. cochrane.org
  3. Jonnalagadda SR, Goyal P, Huffman MD. Automating data extraction in systematic reviews: a systematic review. Systematic Reviews. 2015;4:78. read 9 October 2026. cochrane.org
  4. Moher D, Shamseer L, Clarke M, Ghersi D, Liberati A, Petticrew M, et al. Preferred reporting items for systematic review and meta-analysis protocols (PRISMA-P) 2015 statement. Systematic Reviews. 2015;4(1):1. doi:10.1186/2046-4053-4-1 read 9 October 2026. doi.org/10.1186/2046-4053-4-1
  5. Page MJ, McKenzie JE, Kirkham J, Dwan K, Kramer S, Green S, Forbes A. Bias due to selective inclusion and reporting of outcomes and analyses in systematic reviews of randomised trials of healthcare interventions. Cochrane Database of Systematic Reviews. 2014;10:MR000035. read 9 October 2026. cochrane.org
  6. Tsafnat G, Glasziou P, Choong MK, Dunn A, Galgani F, Coiera E. Systematic review automation technologies. Systematic Reviews. 2014;3(1):74. doi:10.1186/2046-4053-3-74 read 9 October 2026. doi.org/10.1186/2046-4053-3-74
  7. Ip S, Hadar N, Keefe S, Parkin C, Iovin R, Balk EM, Lau J. A web-based archive of systematic review data. Systematic Reviews. 2012;1:15. read 9 October 2026. cochrane.org
  8. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi:10.1136/bmj.n71 read 9 October 2026. doi.org/10.1136/bmj.n71
  9. Mayo-Wilson E, Li T, Fusco N, Dickersin K. Practical guidance for using multiple data sources in systematic reviews and meta-analyses (with examples from the MUDS study). Research Synthesis Methods. 2018;9:2-12. read 9 October 2026. cochrane.org
  10. Page MJ, Altman DG, Shamseer L, McKenzie JE, Ahmadzai N, et al. Reproducible research practices are underused in systematic reviews of biomedical interventions. Journal of Clinical Epidemiology. 2018;94:8-18. doi:10.1016/j.jclinepi.2017.10.017 read 9 October 2026. doi.org/10.1016/j.jclinepi.2017.10.017
  11. Munn Z, Peters MD, Stern C, Tufanaru C, McArthur A, Aromataris E. Systematic review or scoping review? Guidance for authors when choosing between a systematic or scoping review approach. BMC Medical Research Methodology. 2018;18(1):143. doi:10.1186/s12874-018-0611-x read 9 October 2026. doi.org/10.1186/s12874-018-0611-x
  12. Mayo-Wilson E, Li TJ, Fusco N, Bertizzolo L, Canner JK, Cowley T, et al. Cherry-picking by trialists and meta-analysts can drive conclusions about intervention efficacy. Journal of Clinical Epidemiology. 2017;91:95-110. read 9 October 2026. cochrane.org
  13. Jelicic Kadic A, Vucic K, Dosenovic S, Sapunar D, Puljak L. Extracting data from figures with software was faster, with higher interrater reliability than manual extraction. Journal of Clinical Epidemiology. 2016;74:119-123. read 9 October 2026. cochrane.org
  14. Millard LAC, Flach PA, Higgins JPT. Machine learning to assist risk-of-bias assessments in systematic reviews. International Journal of Epidemiology. 2016;45:266-277. read 9 October 2026. cochrane.org
  15. Marshall IJ, Kuiper J, Wallace BC. RobotReviewer: evaluation of a system for automatically assessing bias in clinical trials. Journal of the American Medical Informatics Association. 2016;23:193-201. read 9 October 2026. cochrane.org
  16. Jones CW, Keil LG, Holland WC, Caughey MC, Platts-Mills TF. Comparison of registered and published outcomes in randomized controlled trials: a systematic review. BMC Medicine. 2015;13:282. read 9 October 2026. cochrane.org
  17. Stewart LA, Clarke M, Rovers M, Riley RD, Simmonds M, Stewart G, Tierney JF; PRISMA-IPD Development Group. Preferred reporting items for a systematic review and meta-analysis of individual participant data: the PRISMA-IPD statement. JAMA. 2015;313:1657-1665. read 9 October 2026. cochrane.org
  18. Hoffmann TC, Glasziou PP, Boutron I, Milne R, Perera R, Moher D, et al. Better reporting of interventions: template for intervention description and replication (TIDieR) checklist and guide. BMJ. 2014;348:g1687. read 9 October 2026. cochrane.org
  19. Dwan K, Altman DG, Clarke M, Gamble C, Higgins JPT, Sterne JAC, Williamson PR, Kirkham JJ. Evidence for the selective reporting of analyses and discrepancies in clinical trials: a systematic review of cohort studies of clinical trials. PLoS Medicine. 2014;11:e1001666. read 9 October 2026. cochrane.org
  20. Elliott JH, Turner T, Clavisi O, Thomas J, Higgins JP, Mavergames C, Gruen RL. Living systematic reviews: an emerging opportunity to narrow the evidence-practice gap. PLOS Medicine. 2014;11(2):e1001603. doi:10.1371/journal.pmed.1001603 read 9 October 2026. doi.org/10.1371/journal.pmed.1001603
  21. Gross A, Schirm S, Scholz M. Ycasd - a tool for capturing and scaling data from graphical representations. BMC Bioinformatics. 2014;15(1):219. doi:10.1186/1471-2105-15-219 read 9 October 2026. doi.org/10.1186/1471-2105-15-219
  22. Turner EH. How to access and process FDA drug approval packages for use in research. BMJ. 2013;347. read 9 October 2026. cochrane.org
  23. Squires J, Valentine J, Grimshaw J. Systematic reviews of complex interventions: framing the review question. Journal of Clinical Epidemiology. 2013;66:1215-1222. read 9 October 2026. cochrane.org
  24. Zarin DA, Tse T, Williams RJ, Califf RM, Ide NC. The ClinicalTrials.gov results database: update and key issues. New England Journal of Medicine. 2011;364:852-860. read 9 October 2026. cochrane.org
  25. Gotzsche PC, Hrobjartsson A, Maric K, Tendal B. Data extraction errors in meta-analyses that use standardized mean differences. JAMA. 2007;298:430-437. read 9 October 2026. cochrane.org
  26. Buscemi N, Hartling L, Vandermeer B, Tjosvold L, Klassen TP. Single data extraction generated more errors than double data extraction in systematic reviews. Journal of Clinical Epidemiology. 2006;59:697-703. read 9 October 2026. cochrane.org
  27. Jones AP, Remmington T, Williamson PR, Ashby D, Smyth RL. High prevalence but low impact of data extraction and reporting errors were found in Cochrane systematic reviews. Journal of Clinical Epidemiology. 2005;58:741-742. read 9 October 2026. cochrane.org
  28. Counsell C. Formulating questions and locating primary studies for inclusion in systematic reviews. Annals of Internal Medicine. 1997;127:380-387. read 9 October 2026. cochrane.org
  29. Richardson WS, Wilson MC, Nishikawa J, Hayward RS. The well-built clinical question: a key to evidence-based decisions. ACP Journal Club. 1995;123(3):A12-A13. doi:10.7326/ACPJC-1995-123-3-A12 read 9 October 2026. doi.org/10.7326/ACPJC-1995-123-3-A12

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 9). Data Extraction for a Systematic Review: Determining What to Extract. Pubrica. https://pubrica.com/academy/systematic-review/what-data-to-extract-for-systematic-review/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Extraction field list and form skeleton

The form should be built from this skeleton, deleting fields that no planned output requires and adding those that the review question requires. The three-state blank options and the source-location fields may appear excessive but save the most time.

DATA EXTRACTION FORM — REVIEW: ____________________
Form version: ____  Frozen on: __________  Piloted on ____ studies
Extractor initials: ______  Date: ______  Second extractor: ______

For every field below, the permitted blanks are:
  NR = not reported    RN = reported as none    NA = not applicable
Never leave a cell empty.

=== 1. PROVENANCE AND ELIGIBILITY ===
  Study ID (first author, year): ______________________
  All reports of THIS study (list every record): _______________
  Does this report contain more than one study?  [ ] no  [ ] yes: ____
  Eligibility decision: [ ] include  [ ] exclude  Criterion that settled it: ____
  Trial registration identifier: ______________  [ ] none found
  Protocol available?  [ ] yes, where: ______  [ ] no
  Language: ________  Publication status: ________________

=== 2. METHODS ===
  Design: ______________________________________________
  Unit of allocation: __________  Unit of analysis: __________
  Sequence generation (quote): _________________________
  Allocation concealment (quote): ______________________
  Blinding, of whom (quote): ___________________________
  Sample size calculation reported?  Target n: ______
  Analysis population: [ ] ITT  [ ] per protocol  [ ] other: ______
  Missing data handling (quote): _______________________
  Follow-up duration: ______  Attrition per group: ______

=== 3. PARTICIPANTS ===
  Condition and how ascertained: _______________________
  Study's own inclusion criteria (quote): ______________
  Study's own exclusion criteria (quote): ______________
  Setting / care level: ____________  Country: ____________
  Recruitment start and end dates: ____________________
  Numbers: screened ____  randomised ____  analysed ____
  Baseline characteristics needed for PLANNED SUBGROUPS:
    Variable: __________  Intervention: ______  Comparator: ______
    Variable: __________  Intervention: ______  Comparator: ______

=== 4. INTERVENTION AND COMPARATOR ===
  Intervention as delivered: ___________________________
    Dose / intensity: ______  Schedule: ______  Duration: ______
    Delivered by, with what training: __________________
    Co-interventions permitted: ________________________
  Comparator AS THE STUDY DESCRIBES IT (quote, do not summarise):
    ___________________________________________________
  Number of arms: ____  Arms used in this review: ____________
  Arms excluded and why: ______________________________

=== 5. OUTCOMES — REPEAT THIS BLOCK PER OUTCOME PER TIMEPOINT ===
  Outcome domain: _____________________________________
  Specific measurement (instrument, scale, range, direction): ______
  Specific metric: [ ] final value  [ ] change from baseline
                   [ ] event occurrence  [ ] other: ______
  Method of aggregation: [ ] mean  [ ] median  [ ] proportion
                         [ ] rate  [ ] other: ______
  Time point / period: ______  Matches our pre-specified point? [ ]
  Pre-specified in the study's own registration?  [ ] yes  [ ] no  [ ] cannot tell

  DICHOTOMOUS DATA
    Intervention: events ____ / analysed ____
    Comparator:   events ____ / analysed ____
  CONTINUOUS DATA
    Intervention: mean ____  variability ____  TYPE OF VARIABILITY ______
                  number analysed ____
    Comparator:   mean ____  variability ____  TYPE OF VARIABILITY ______
                  number analysed ____
  TIME-TO-EVENT DATA
    Effect measure ____  value ____  CI ____  events ____

  Source location (table / page / figure): ____________________
  Value obtained from: [ ] printed table  [ ] text  [ ] read off a figure
                       [ ] author correspondence  [ ] registry  [ ] regulatory record
  Reported in a usable form?  [ ] yes  [ ] converted, method: ______
                              [ ] no — what is missing: ______
  Discrepancy with second extractor?  [ ] none  [ ] yes, logged as: ______

=== 6. RISK OF BIAS, FUNDING, CONFLICTS ===
  Tool used: ____________________  Assessors: ______
  Domain: __________  Judgement: __________  SUPPORTING QUOTE: ______
  Domain: __________  Judgement: __________  SUPPORTING QUOTE: ______
  Funding source: ______________________________________
  Role of funder as stated: ____________________________
  Declared conflicts of interest: ______________________

=== 7. NOTES FOR THE REVIEW ===
  Anything the review must report about this study: ____________
  Author contacted?  [ ] no  [ ] yes, date ______  Reply: ______

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica builds the extraction form from the planned analyses rather than a generic template, pilots it on the review’s own problematic papers, and conducts double extraction with a named arbitrator and a logged reconciliation. The deliverables comprise the form, the dataset in an analysable structure and a record of every discrepancy and its resolution, which the reporting checklist requires authors to describe and which most teams otherwise have to reconstruct.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

On receipt of a protocol and included-study list, Pubrica provides an extraction form piloted on the review’s own papers.

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.