How to build a passage-level evidence matrix for a literature review

You read a paper, highlight three useful paragraphs, and move on. A month later, the highlights are still there, but the reason you saved them is harder to recover. Was one a key finding, a limitation, or evidence against your argument? Which other papers reached a similar conclusion? Which ones disagreed?

Paper-level tags cannot answer those questions. A tag such as remote supervision tells you what a paper is about. It does not tell you which page contains the finding you need, why that finding matters to your review, or how the finding compares with evidence from other sources.

A passage-level evidence matrix records those connections. You save the relevant passage, record why it matters, keep it linked to its source, and compare it with passages from other papers under the same theme.

Agent Bayes supports this workflow directly. You can label a passage while searching or reading, reopen the exact page behind it, and attach it to a claim in your mindmap. The method also works in a spreadsheet. This guide shows you how to build the structure, classify the evidence, and move from saved passages to a synthesis you can check against the original papers.

The workflow combines two established practices. Thematic synthesis codes qualitative findings line by line before developing themes. The Framework Method places summarized evidence in a grid so researchers can compare cases across themes.

Applied to a literature review, these practices create three connected parts:

1. Sources: List every paper, report, or book chapter included in your review. Record its full citation, source type, research context, methods, and relevant limitations.
2. Passages: Save each passage you may use in the review. Record the source, exact page, relevant text, your labels, and a note explaining why the passage matters.
3. Matrix: Create one row per source and one column per theme. Summarize what each source reports about the theme and cite the relevant pages.

Use a short citation such as Chen 2024 for the source. Add the page when you refer to a passage, for example Chen 2024, p. 18. If two sources have the same author and year, distinguish them as Chen 2024a and Chen 2024b, just as you would in a reference list. Agent Bayes keeps the source link and page with the saved passage automatically.

Set up Sources, Passages, and Matrix

Create these three sheets before you classify passages from the full set of papers. Use the same short citation in each sheet so you can move from a matrix summary to the saved passage and then to the original source.

1. Record each source and its context

Give each source a short citation that you already recognize, such as Chen 2024. Record the full citation, source type, sample or corpus, setting, methods, and any appraisal notes relevant to your question.

2. Record each passage you may use

Create one row per meaning unitA stretch of text that expresses one coherent idea while retaining enough context to interpret it.. A meaning unit may be a sentence, several sentences, or a paragraph. Fixed line length matters less than semantic completeness.

Each row should include the short citation, exact location, passage text or faithful paraphrase, source voice, codes, and your memo. Keep verbatim quotation separate from paraphrase so you know which wording belongs to the source.

3. Compare sources in the matrix

Summarize the coded material in a framework matrixA grid with sources as rows, themes as columns, and page references in each cell.. Sources become the rows and topics or themes become the columns. Each cell holds a concise summary with citations to the relevant pages.

The Framework Method uses a comparable structure for primary qualitative data, with one participant per row and one code per column. For a literature review, replace the participant with a source.

Label individual passages. Compare sources by theme. Cite the pages behind each summary.

Step 1: decide what you need to learn from the papers

Write one question that tells you which passages belong in the matrix. This prevents the matrix from becoming another collection of interesting highlights. For example:

How do doctoral researchers describe the effects of remote supervision on their sense of belonging?

If you want to classify passages as supporting or challenging an argument, write that argument down too:

Delayed supervisory feedback weakens doctoral researchers' sense of belonging.

Without a written argument, the label supports is ambiguous. The same passage might support one claim, challenge another, and provide context for a third.

Also decide what counts as data. A qualitative evidence synthesis might code findings and results sections. An argument review may need theoretical discussion and conclusions. A methods review may code procedural descriptions. Record the choice. Do not move silently between participant quotations, author interpretations, cited background, and your own inference.

Step 2: pilot a codebook on a varied sample

Choose a small, deliberately varied group of included sources. Read them closely and test the fields you expect to use. The Cochrane Handbook's data-extraction guidance recommends pilot-testing forms on several reports because missing fields and ambiguous instructions usually appear only when real papers meet the form.

Use separate coding axes instead of one flat list:

This separation prevents a common attribution error. A paper may report a theory in its literature review and reject it later. The topic code may be correct while the source role and relation are wrong. A paper can mention a theory without believing it, so read enough surrounding text to identify whose position you are recording.

For every controlled code, write:

• A definition
• An inclusion rule
• An exclusion rule
• One positive example
• One difficult or borderline example

Version the codebook. When a pilot reveals that one code contains two ideas, split the code and revisit earlier rows. When two codes cannot be applied consistently, merge or redefine them. A changelog records those decisions.

Do not aim for the smallest possible codebook at any cost. Aim for the smallest codebook that preserves distinctions your question requires. Six reusable role labels may be enough for one review and too blunt for another.

Step 3: extract enough text to preserve meaning

Code the smallest passage that still makes sense when you return to it later. Keep a link to the page so you can read the surrounding text again.

For each passage, record at least:

Several codes may apply to one passage. Thomas and Harden applied multiple codes to many lines in their thematic synthesis, and the Framework Method also attaches codes to meaningful passages before charting. Multiple coding is useful when the overlap is real. It should not excuse an extraction unit so large that it contains several unrelated claims.

Keep context at two levels. Sources records the design, setting, and other details that affect interpretation. The page or section reference lets you reopen the surrounding text. A generated summary can help you navigate, but it is not source evidence and should not replace reading the page.

Step 4: chart summaries into the matrix

Once Passages contains a useful body of coded material, create one matrix column per topic or descriptive theme. Summarize each source's relevant evidence in the matching cell and cite the pages behind the summary.

Here is a fictional example using the remote-supervision question:

The matrix supports two directions of reading. Read across a row to keep one source's findings together. Read down a column to compare what the sources report about one theme in different contexts.

Write compact cell summaries rather than copying long passages. Keep the full extract in Passages and the source comparison in Matrix.

Step 5: compare before you conclude

For each theme, write an analytic memo that answers five questions:

1. What pattern appears across the sources?
2. Which sources or passages do not fit it?
3. Do design, sample, setting, time, or source voice explain the difference?
4. Does the evidence justify a description, an interpretation, or a causal claim?
5. What information is missing?

Start with the exceptions. The matrix is most useful when one source does not match the general pattern. Gale and colleagues describe using analytical memos to compare within and between cases, examine deviant cases, and develop explanations from the summarized data.

Do not treat passage frequency as evidential weight. A detailed ethnography can produce many passages while a concise paper produces two. Examine the relevance, methods, context, and richness of the evidence. For a qualitative theme, importance also depends on what the passage helps explain, not only how often a phrase appears.

A collection of passages becomes a synthesis when you explain the similarities, the differences, and the sources that do not match the general pattern.

Step 6: record the evidence behind each claim

Before a synthesis statement reaches the review, record the evidence and reasoning behind it:

Those four confidence concerns come from GRADE-CERQual, a formal approach for qualitative evidence synthesis findings. A lightweight review should not borrow CERQual ratings without following its guidance, but the four questions make a useful check. Are the contributing studies methodologically adequate? Do they address your actual context? Is there enough relevant data? Does the finding fit the data across studies without ignoring important variation?

Treat quantitative null findings with particular care. A non-significant result may challenge a proposition, or it may be too imprecise to distinguish a meaningful effect from no effect. Altman and Bland's short note on why absence of evidence is not evidence of absence remains the right warning. Inspect the effect estimate, interval, design, and statistical power before applying challenges.

Add the review method your question requires

This matrix helps you organize and compare evidence. Use it when your question asks what authors or participants report, how concepts are framed, where explanations agree, or which contexts change a finding. It works particularly well with qualitative studies, mixed-method findings, conceptual papers, and arguments spread across a varied literature.

Some review questions require additional analysis. If you are asking whether an intervention works or how large an association is, add study-level outcome data, effect estimates, uncertainty, and an appropriate quantitative or structured narrative synthesis. The saved passages show exactly what each paper reports. The study-level results show how precise and relevant those reports are.

For a systematic review, use the matrix inside the review method you selected. The Cochrane guidance on qualitative evidence synthesis describes thematic synthesis, framework synthesis, and meta-ethnography as different methods for different purposes. The ENTREQ reporting statement also asks reviewers to name their synthesis method, define which parts of each study they analysed, and report how they coded and interpreted the material.

Quality checks before you write

Run these checks after the first matrix pass and again before drafting the review:

• Reopen every passage used by a central claim and read the surrounding source text.
• Confirm that every passage still points to the correct study and exact location.
• Check that reported prior work was not recoded as the current paper's finding.
• Review all passages under each code for consistent application.
• Revisit earlier rows after material codebook changes.
• Search deliberately for mixed, negative, and deviant cases.
• Keep study appraisal separate from passage coding, then use both when judging a claim.
• For a formal or high-stakes systematic review, use independent duplicate extraction where the chosen standard requires it. Cochrane recommends duplicate extraction for outcome data because extraction errors can change the synthesis.
• Report which sections you coded, who coded them, how disagreements were handled, and which synthesis method governed the analysis.

A link back to the source makes an interpretation easier to check. It does not make the interpretation correct. You still need to decide whether the studies can be compared, whether a theme applies across their different contexts, and whether the evidence supports the sentence you want to write.

A spreadsheet is enough to start

Create five sheets named Sources, Passages, Matrix, Claims, and Codebook. Use the same short citation for a source in every sheet. Pilot the structure on a varied handful of papers, revise the codebook, and only then expand extraction.

The first useful query is not complicated. Filter Passages by one topic, group the results by source, and write one summary per source in Matrix. Then read down the column and write a memo that explains the main pattern and the clearest exception.

If that workflow feels too slow in a spreadsheet, you now know what your research tool needs to preserve: exact source locations, multiple labels per passage, source context, codebook history, and links from each synthesis claim back to its evidence.

Agent Bayes can store the labeled passages

Agent Bayes connects saved passages directly to your research process. You can label evidence in research-agent results, project search, PDF search, and mindmap citations. The Labeled Items view keeps each saved passage with its labels and lets you search the collection or attach a passage to a mindmap node as a citation.

When you are ready to compare sources, summarize the selected passages in the matrix described above. Selecting several labels in Agent Bayes finds passages carrying any of those labels. If you need the narrower set of passages that carries every selected label, check the label combinations during the comparison pass.

This division keeps the workflow clear. Agent Bayes helps you find, label, reopen, and connect evidence to your argument. The matrix helps you compare sources and write the synthesis. You remain in control of the review question, the comparison, and the final claim.

Start with Sources, Passages, and Matrix. Before you use a claim in your review, follow its citations and page references back to the original text and check that the wording still fits the evidence.