Skip to main content
v2026.11,986 entries · CC-BY 4.0

Data Synthesis: Choosing a Method and Writing the Section

Data synthesis combines study results. See which method the data allow, per the Cochrane Handbook, and how to write the synthesis section.

Written and maintained by CASRAI Editorial Board

Published · Last updated

On this page

Data synthesis, in a systematic review or evidence review, is the step where the results of the included studies are combined or summarised to answer the review question. This page covers that meaning (the synthesis of data across studies), shows which synthesis method the available data allow, and sets out how to write the synthesis section of a protocol or manuscript. It does not cover synthetic data generation, which is a different topic.

Ask CASRAI · free to tryQuestion about this guide? Ask CASRAI answers with linked sources. 2 free questions a day, no account.

Ask your first question freeSubscribe

What is data synthesis?

Data synthesis is the combining of results from separate studies. The most formal version is meta-analysis, which the Cochrane Handbook defines as statistically combining results from two or more separate studies; most methods are variations on a weighted average of the effect estimates from the different studies. When the studies report incomplete or incompatible data, a meta-analysis may not be possible, and the Handbook describes alternative synthesis methods for that situation (Cochrane Handbook v6.5, chapters 10 and 12, read 2026-11-08).

Qualitative evidence is synthesised too, with different methods; see meta-synthesis and the ENTREQ reporting statement guide. For the wider relationship between a systematic review and its statistical component, see systematic review vs meta-analysis.

Choosing a synthesis method: what the data allow

The table follows Cochrane Handbook chapters 10 and 12. The method is decided by what the included studies report, and the Handbook advises planning alternative methods in the protocol.

What the included studies report Method What it tells you
Comparable effect estimates with usable variances Meta-analysis (a weighted average of study effect estimates) A pooled effect estimate; the most decision-relevant result (Ch. 10)
Effect estimates, but variances missing or incorrect Summarising effect estimates (median, interquartile range, range) The typical size and spread of effects (Ch. 12.2.1)
Only P values, with a shared directional hypothesis Combining P values (for example Fisher’s method, using one-sided P values) Whether at least one study shows an effect (Ch. 12.2.1)
Only the direction of effect (benefit or harm) Vote counting based on direction of effect, with a sign test Whether effects lean one way (Ch. 12.2.1)
Only whether results were statistically significant Not acceptable: vote counting by statistical significance or subjective criteria The Handbook states that vote counting based on statistical significance has serious limitations and is unacceptable (Ch. 12.2.2)

The alternative methods give less information for decisions than a meta-analysis, but the Handbook notes they can be better than a narrative synthesis that privileges some results without justification. Any synthesis should be accompanied by tables or visual displays (Ch. 12). For the structured narrative approach, see the narrative synthesis (SWiM) guide.

Purpose of the synthesis section

The data synthesis section of a protocol or manuscript tells the reader how the included results were combined, so that another team given the same studies could reach the same synthesis. In a protocol it commits the team to a plan before results are seen; the Handbook advises that alternative methods be planned in the protocol rather than chosen after the data are in.

Structure of the synthesis section

  1. State what will be synthesised. Name the outcomes and comparisons, and how studies are grouped for each.
  2. State the criterion for meta-analysis. Say when results will be pooled and when they will not. The Handbook advises not to meta-analyse when results vary widely.
  3. Name the effect measure and the model. A fixed-effect model assumes all studies estimate the same effect; a random-effects model assumes different but related effects, gives relatively more weight to small studies and gives a wider confidence interval when heterogeneity is present. The Handbook advises against choosing the model with a test of heterogeneity. For the weighting mechanics, see inverse-variance weighting.
  4. Describe how heterogeneity will be assessed and handled. The Handbook notes that the chi-squared test has low power with few studies and that I2 thresholds can be misleading. Responses it lists: check the data, do not meta-analyse if results vary widely, use pre-specified subgroup analysis or meta-regression, use random effects, and exclude outliers only with justification. Prediction intervals can present the spread between studies. See also heterogeneity in meta-analysis.
  5. Say how special data cases are handled. For example, studies with zero events add no information to a risk ratio or odds ratio, and the Peto method is less biased for rare events (Ch. 10).
  6. Plan the fallback. Name the alternative method (table above) for outcomes where pooling is not possible, and the tables or plots that will accompany it, such as structured tabulation, forest plots, albatross plots or effect direction plots (Ch. 12.3).
  7. Plan sensitivity analyses. State which decisions will be re-tested to check that conclusions are robust (Ch. 10.14).
  8. Cover missing data and reporting bias. State how missing data are handled (Ch. 10.12) and see publication bias.

A short annotated example

Illustrative template, not a real review. The wording below shows the pattern for a protocol; the bracketed items are placeholders.

Where two or more studies report [outcome] on a comparable scale with usable variances, we will pool effect estimates using [inverse-variance random-effects] meta-analysis. We will not pool where clinical or methodological diversity makes results too variable to combine. We will describe between-study variability with a prediction interval, and investigate it with [pre-specified subgroups]. Where variances are unavailable, we will summarise effect estimates (median, interquartile range, range); where only direction of effect is reported, we will use a sign test on direction. We will not use vote counting based on statistical significance. We will run sensitivity analyses [excluding studies at high risk of bias] and present results in a forest plot or structured table.

  1. The first sentence sets the pooling criterion and the model before results are seen.
  2. The second states when pooling will not happen.
  3. The third and fourth name the fallback methods from the table.
  4. The last sentence names the excluded method and the display.

Common mistakes

  • Counting studies with a significant result against those without (vote counting by significance). The Handbook calls this unacceptable.
  • Choosing fixed-effect or random-effects after looking at a heterogeneity test.
  • Treating I2 cut-offs as rules, or relying on the chi-squared test when there are few studies.
  • Pooling studies whose results vary widely instead of explaining the variation.
  • Dropping outlying studies without a stated justification.
  • Writing a narrative synthesis with no table or display, so some results are privileged without explanation.
  • Leaving the fallback method out of the protocol and choosing it after the data are known.

Checklist

  • Outcomes, comparisons and study groupings are named.
  • Criterion for pooling versus not pooling is stated.
  • Effect measure and model are named, with a reason that does not depend on a heterogeneity test.
  • Heterogeneity assessment and responses are described.
  • Alternative method and accompanying display are planned for outcomes that cannot be pooled.
  • Sensitivity analyses and missing-data handling are stated.
  • Methods are reported so the synthesis can be reproduced; see writing a reproducible methods section.

Frequently asked questions

What is data synthesis?

In evidence review, it is combining or summarising the results of separate studies to answer a review question. Meta-analysis is the statistical form, a weighted average of study effect estimates; where studies cannot be pooled, the Cochrane Handbook describes alternatives such as summarising effect estimates, combining P values and vote counting on direction of effect.

What does synthesis of data mean?

It means the same thing: bringing together results from several studies or sources into one summary. It can be quantitative (meta-analysis or the alternatives above) or qualitative (see meta-synthesis).

Is a synthesis database the same as data synthesis?

No. Data synthesis is an analytic step applied to the results of included studies. The databases a review searches to find those studies are a separate stage; see PRISMA 2020 for how the whole process is reported.

Is data synthesis the same as synthetic data?

No. This page concerns combining existing study results. Synthetic data generation, the artificial creation of data, is a different subject and is not covered here.

Sources

  • Cochrane Handbook for Systematic Reviews of Interventions, v6.5, Chapter 10, Analysing data and undertaking meta-analyses: cochrane.org (read 2026-11-08).
  • Cochrane Handbook v6.5, Chapter 12, Synthesizing and presenting findings using other methods: cochrane.org (read 2026-11-08).

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · free to try

Ask about Data Synthesis: Choosing a Method and Writing the Section

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →