Move from a collection to comparable evidence
A reference manager can tell you which publications you have saved. Analysis asks different questions: what was investigated, how was a result produced, to whom might it apply, and what limitations did the researchers identify? A matrix places those features in consistent columns so that similarities and important differences do not depend on memory.
The matrix never replaces the full text. Every cell is a compression and may lose context. Include a precise locator—page, section, table or figure—and a stable key that opens the source or its library record. Mark your interpretation as such; otherwise a promising personal inference can later be mistaken for the author’s conclusion.
Give each column one unambiguous job
ID and full reference prevent mix-ups. Research question captures the study’s purpose in one sentence. Design identifies an experiment, interview study, review, document analysis or other approach. Sample/material records both size and selection. Context limits location, period, discipline or setting.
Main finding should be narrower than the source, never more causal. Locator points back to the supporting passage. Limitations can distinguish the authors’ report from your critical reading. Relevance states which sub-question or chapter may use the source. An optional workflow field can read “awaiting full text,” “extracting,” “checked against original” or “excluded—reason recorded.”
Visible blanks are safer than invented values. If the sample size is absent, write “not reported.” If only an abstract is available, say “full text pending” rather than filling method cells from inference. Missingness then becomes a work queue instead of hidden uncertainty.
Extract one source in a seven-step pass
- Assign a stable ID that will also appear in notes and draft comments.
- Verify authors, title, year, venue and DOI or durable URL against the source.
- Read methods and results rather than relying on abstract and conclusion.
- Enter question, design, material and context without inserting your evaluation.
- State the finding tightly enough to preserve population, direction and uncertainty.
- Record the locator and limitations before moving to another publication.
- Compare the row with the full text, then change its status to checked.
A team needs a small data dictionary. Define whether “sample” means planned, recruited, included or analysed participants and how missing information is coded. Two reviewers can independently extract a subset and discuss differences. That calibration improves consistency, although it cannot make a weak judgement automatically valid.
Filter for patterns without flattening differences
Group rows by sub-question, population, design or outcome direction. Do not simply count publications that sound positive or negative. Check whether they examine the same construct, period and conditions. An interview theme and an experimental estimate answer different questions; forcing both into one tally creates false comparability.
A useful synthesis paragraph makes a bounded claim, identifies supporting and diverging evidence, and then explains design or context differences. Use the matrix to plan that paragraph, but cite the original publications. Colour may indicate workflow status; it should not quietly become an unexplained quality score.
New categories will emerge as reading continues. Add a column only when it matters to the research question across several sources. Define it in the data dictionary and revisit earlier rows. Otherwise later publications contain richer data than earlier ones and the apparent pattern may reflect extraction drift.
Keep the matrix and reference library aligned
Review duplicates, broken links, absent locators and ambiguous status values at regular intervals. Any source used analytically should be findable in both the reference library and the matrix where appropriate. Not every source read must appear in the final text; if the method requires an exclusion trail, retain a concise, content-based reason.
Save dated versions and note major schema changes. Sensitive notes about unpublished data do not belong in a broadly shared sheet. Before submission, export a readable copy and check that links, special characters, wrapped text and long cells have survived.
The claim–source matrix continues from study-level evidence to draft-level claims. The source log template records reading, quotation and paraphrase during writing. Find more resources in the research methods hub, and consult the evidence-review foundation for tracing report findings.