Reporting a meta-analysis, section by section

Why reporting quality matters

Readers, editors and guideline panels cannot judge a review that does not say what was done. Studies of published systematic reviews have repeatedly found that many omit key elements: the search strategy, the methods of risk-of-bias assessment, the handling of heterogeneity and the certainty of evidence. Incomplete reports prevent replication, hide weaknesses and make the review less trustworthy than the underlying work may deserve.

PRISMA 2020 is the main reporting guideline for systematic reviews, with and without meta-analysis. It consists of a checklist of 27 items, an expanded explanation and elaboration document, a flow diagram template and a separate checklist for abstracts. Several extensions cover specific designs, for example network meta-analysis, diagnostic test accuracy, individual participant data and scoping reviews. Meta-analyses of observational studies are also reported with MOOSE. The guideline is about reporting, not about conduct. It does not tell you how to do a review, only what to say.

The structure at a glance

The table summarizes the main sections of the PRISMA 2020 checklist. Use the official checklist for item wording and sub-items, which are more detailed than the summary here.

Title and abstract

The title should say that the work is a systematic review, and, where appropriate, a meta-analysis, so that it can be found and classified. Add the population and intervention, or the topic, in plain words.

The abstract is often the only part read. A structured abstract should state the objectives, the data sources and the date of the last search, the eligibility criteria, the methods for assessing risk of bias and synthesizing results, the number of studies and participants, the main results with effect estimates and intervals, the limitations, the interpretation, the funding and the registration number. The PRISMA 2020 abstract checklist lists these elements, and they fit within the word limits of most journals if each is brief.

Introduction

Explain why the review is needed in the light of existing evidence, including earlier reviews and what they left unanswered. Many reviews duplicate existing ones, and an introduction that demonstrates novelty or a need for an update helps. State the objectives explicitly, using PICO or an equivalent structure. Keep the introduction short. It should lead the reader to the question and no further.

Methods: the section that must be complete

The methods section carries most of the checklist and most of the failures. A reader should be able to repeat the review from it. The main elements are as follows.

• Protocol and registration. State where the protocol was registered or published, and describe any deviations from it, with reasons.
• Eligibility criteria. Population, intervention or exposure, comparator, outcomes, study designs, setting, time period, language and publication status.
• Information sources. Every database, register and other source searched, with the date of each search.
• Search strategy. The full strategy for each database, preferably in a supplement, following PRISMA-S.
• Selection process. How many reviewers screened, whether independently, how disagreements were resolved and whether automation tools were used.
• Data collection. The extraction process, the data items, how missing data were handled and whether authors were contacted.
• Outcomes. The primary and secondary outcomes, time points and rules for choosing among multiple results.
• Risk of bias assessment. The tool, the number of assessors and how they were calibrated.
• Effect measures. The measure for each outcome, with the direction.
• Synthesis methods. Which studies were eligible for each synthesis, how data were prepared, the statistical model (fixed-effect or random-effects), the estimator for between-study variance, the method for the confidence interval, how heterogeneity was assessed, the planned subgroup and sensitivity analyses, and the software with version and packages.
• Reporting bias assessment. The methods for missing results, such as funnel plots and asymmetry tests, with the studies required.
• Certainty assessment. The approach, usually GRADE.

A frequent failing is to name the model but not the estimator and interval method. "A random-effects model was used" is not enough. Say, for example, that restricted maximum likelihood was used for tau-squared with the Hartung-Knapp adjustment, and that a prediction interval was calculated.

Results

The results section should begin with the flow of studies, then describe the studies, then give the results of each synthesis.

• Study selection. The PRISMA flow diagram, with numbers at each stage and reasons for full-text exclusions, and a list of studies that nearly met the criteria.
• Study characteristics. A table with the key features of each study, including those that will serve for subgroup analysis.
• Risk of bias. The assessment for each study and each domain, shown in a table or plot, with a summary.
• Results of individual studies. For all outcomes, a summary statistic with a confidence interval for each study, usually displayed in a forest plot.
• Results of syntheses. The pooled estimate with its confidence interval and, for random-effects analyses, the prediction interval, together with heterogeneity statistics. Results of subgroup, meta-regression and sensitivity analyses.
• Reporting biases. The results of the assessment for missing results.
• Certainty of evidence. A summary of findings table giving, for each main outcome, the effect, its certainty and the reasons for any downgrading.

Report numbers, not only significance. A statement that a result was significant, without the estimate and interval, is not informative. Give the number of studies and participants behind each result.

Discussion

Start with the main findings and place them in context of other evidence. Discuss the limitations of the evidence, such as risk of bias, inconsistency and imprecision, and the limitations of the review process, such as incomplete searching, language restrictions or the use of single screening. Then state the implications for practice, policy and research. Do not go beyond the evidence: a pooled association does not show causation, and a review of low-certainty evidence should not produce a strong recommendation. Where there are gaps, say what kind of study would fill them.

Registration, funding, data and code

The last group of items covers registration and protocol, sources of financial or non-financial support and the role of funders, competing interests of the authors, and the availability of data, code and other materials. Many journals now expect the extracted data, the analytic code and the search strategies to be shared. Posting them in a repository or as supplementary material is good practice and makes the work easier to update. Declare any use of AI tools in writing, screening or analysis, in line with the journal policy.

Frequent reporting failures

• No full search strategy, or strategy for only one database.
• No date of the last search.
• Flow diagram that does not add up, or with no reasons for exclusion.
• Risk of bias assessed but not shown by study.
• A statistical model named without the estimator and interval method.
• No prediction interval alongside a random-effects estimate.
• Subgroup analyses that were not prespecified, presented as if they were.
• Conclusions that ignore the certainty of evidence.
• No protocol registration, or unreported deviations from it.
• No statement on data and code availability.

Using the checklist in practice

Complete the checklist while writing, not at the end, and cite the page or section for every item. Journals often ask for it at submission, and a checklist filled in after the fact usually shows gaps. Have someone who did not write the paper read the methods section and try to describe the review back. If they cannot, something is missing. Choose the right extension: network meta-analyses, diagnostic accuracy reviews, scoping reviews and reviews of individual participant data each have their own guideline, used in addition to or in place of the main one. The EQUATOR Network library lists the guidelines for other designs.

Tables and figures a reader expects

A meta-analysis is read through its tables and figures, and a standard set serves most reviews. The PRISMA flow diagram shows how records became included studies. A characteristics table lists each study with its design, population, intervention, comparator, outcomes and follow-up. A risk-of-bias figure shows judgments by study and domain. Forest plots show each study's estimate with its interval and weight, the pooled estimate, and the heterogeneity statistics, with the axis labeled for the direction of benefit and on a log scale for ratio measures. A summary of findings table presents absolute and relative effects with the certainty of evidence. Funnel plots appear when there are enough studies. Every figure should have a legend that explains it without the text, and every table should say what the numbers are. Make the figures legible at journal column width, and provide the data behind them in a supplement so that readers can recreate the plots. Keep terminology consistent between text, tables and figures, so that an outcome called one thing in the abstract is not called another in the forest plot.

How we can help

We can prepare the manuscript to PRISMA 2020 and the relevant extension, complete the checklist with page references, prepare the flow diagram, summary of findings tables, supplements and data or code packages, and review a draft for reporting gaps. [OWNER VERIFICATION REQUIRED] The relevant services are manuscript editing and publication support.

Frequently asked questions