What Is Meta-Analysis? Definition, Steps and Examples
Meta-analysis is a statistical method that combines the results of several studies on the same question to produce one overall estimate. It is often the final step of a systematic review when the studies are similar enough to pool.
This guide explains what meta-analysis is, when to use it, gives six steps with an example at every step, explains effect sizes, fixed and random-effects models and heterogeneity, and works through an example with equations.
- The idea: what meta-analysis is and when it is suitable.
- The method: six steps, effect sizes and models.
- The results: a worked example, plots, bias and common mistakes.
Key takeaways
- Meta-analysis pools effect sizes from similar studies to give a more precise overall estimate.
- It usually follows a systematic review and a clear protocol.
- Choose a fixed-effect or random-effects model and assess heterogeneity with Q, I² and τ².
- Show results in a forest plot and check publication bias with a funnel plot.
What is meta-analysis?
What is meta-analysis in one line? It is a quantitative synthesis of study results.
- Definition: a statistical method of combining effect sizes from independent studies to estimate an overall effect.
- Benefit: greater precision and power than any single study.
- Relation to systematic reviews: meta-analysis is an optional step; see what is a systematic literature review.
- Origin: the term was coined by Gene Glass in 1976; see the overview of meta-analysis.
- Guidance: the Cochrane Handbook explains methods for health research.
When to use meta-analysis
Knowing what is meta-analysis also means knowing when it is the wrong tool. Use it only when pooling makes sense.
- Similar questions: studies address the same population, intervention and outcome.
- Comparable measures: outcomes can be put on the same scale.
- Enough studies: usually at least two, and preferably several more.
- Adequate data: each study reports effect sizes and measures of variability.
- Not suitable: when studies differ too much, use a narrative synthesis instead.
Step-by-step guide to a meta-analysis
Follow these six steps. Each has an explanation and an example.
- Define the question and protocol. Write a focused question and decide in advance on effect measures, models and subgroup analyses. Example: Does cognitive behavioural therapy reduce anxiety in adolescents?
- Search and select studies. Run a systematic search, screen with clear criteria and report a flow diagram. Example: 27 randomised trials included.
- Extract data and assess bias. Extract means, standard deviations and sample sizes, and assess risk of bias in each study. Example: Two reviewers extract data independently.
- Calculate an effect size for each study. Convert results into a common measure, such as a standardised mean difference or odds ratio. Example: Hedges’ g for each trial.
- Pool the results. Choose a fixed-effect or random-effects model, calculate the pooled estimate and examine heterogeneity. Example: Random-effects g = −0.55.
- Explore, test for bias and report. Check subgroups and meta-regression, test for publication bias and report following PRISMA 2020. Example: Funnel plot and Egger test.

Common effect sizes in meta-analysis
Effect sizes put results on one scale.
| Effect size | Used for | Example |
|---|---|---|
| Standardised mean difference (Cohen’s d, Hedges’ g) | Continuous outcomes measured on different scales | g = 0.50 is a medium effect |
| Odds ratio | Binary outcomes | OR = 1.71 |
| Risk ratio | Binary outcomes, easier to interpret | RR = 0.80 |
| Correlation (r, Fisher’s z) | Associations | r = .30 |
| Mean difference | Continuous outcomes on the same scale | −2.1 points |
Fixed-effect vs random-effects models
The model reflects what you assume about the studies.
| Feature | Fixed-effect | Random-effects |
|---|---|---|
| Assumption | One true effect for all studies | True effects vary between studies |
| Weights | Based on within-study variance | Within-study plus between-study variance |
| Confidence interval | Narrower | Wider |
| Typical use | Very similar studies | Most reviews in practice |
- Do not choose by heterogeneity test alone: decide the model by design and context.
- Report both: if choice matters, show a sensitivity analysis.
A worked meta-analysis example with equations
Three invented studies show how pooling works. Each reports a standardised mean difference g with variance v = 0.04.
| Study | g | v | Weight w = 1/v |
|---|---|---|---|
| A | 0.20 | 0.04 | 25 |
| B | 0.50 | 0.04 | 25 |
| C | 0.80 | 0.04 | 25 |
- Interpretation: the pooled effect is medium (0.50) and the studies differ moderately (I² ≈ 56%).
- Fixed vs random: the fixed-effect interval would be narrower: SE = 1/√75 = 0.115, giving 0.27 to 0.73.
- Report: Hedges’ g = 0.50, 95% CI [0.16, 0.84], I² = 55.6%, τ² = 0.05.

Heterogeneity: how different are the studies?
Heterogeneity is the variation in results beyond chance.
- Q statistic: tests whether the variation is greater than expected by chance; it has low power with few studies.
- I²: the percentage of variation due to heterogeneity; rough guides are 25% low, 50% moderate and 75% high.
- τ²: the estimated variance of true effects between studies.
- Prediction interval: shows the range in which the effect in a new study is likely to fall.
- Explain it: investigate with subgroup analyses or meta-regression, not just report it.
Forest plots and funnel plots
Two plots summarise a meta-analysis.
- Forest plot: shows each study’s effect and confidence interval as a square and line, with the pooled estimate as a diamond.
- Funnel plot: plots effect size against precision; asymmetry may suggest publication bias.
- Egger test: a regression test for funnel plot asymmetry; best with ten or more studies.
- Trim and fill: an adjustment method that estimates missing studies, to be used as a sensitivity check.
- Software: R packages meta and metafor, Stata and RevMan.
Publication bias and quality of evidence
The result is only as good as the studies and the search.
- Publication bias: studies with significant results are more likely to be published.
- Search widely: include registers, theses and unpublished data.
- Risk of bias: use a tool such as Cochrane RoB 2 for trials.
- Sensitivity analysis: remove studies at high risk and check whether the result changes.
- GRADE: rate the certainty of the evidence for each outcome.
Common mistakes in meta-analysis
Avoid these problems.
- Pooling very different studies: apples and oranges. Example: check that population, intervention and outcome are comparable.
- Ignoring heterogeneity: one number hides variation. Example: report I², τ² and a prediction interval.
- No protocol: analyses change after seeing the results. Example: register and follow a protocol.
- Counting a study twice: multiple reports of the same trial. Example: link reports to one study.
- Over-interpreting subgroups: many tests produce false findings. Example: limit and pre-specify them.
Need help? See our order page.
Software for meta-analysis
Several tools carry out the calculations.
| Tool | Strengths | Cost |
|---|---|---|
| RevMan | Cochrane’s tool; easy forest plots | Free |
| R (meta, metafor) | Flexible, advanced models | Free |
| Stata | Commands for meta-analysis | Paid |
| Comprehensive Meta-Analysis | Menu-driven | Paid |
| JASP and jamovi | Point and click, free | Free |
- Report the software: and the version, with the model and the method used to estimate τ².
- Share the data: publish the extraction sheet so others can repeat the analysis.
Subgroup analysis and meta-regression
If heterogeneity is large, look for reasons.
- Subgroup analysis: compare pooled effects between groups of studies. Example: adults versus adolescents.
- Meta-regression: relate effect size to a study characteristic. Example: effect by duration of treatment.
- Pre-specify: decide the subgroups in the protocol to avoid fishing.
- Check the number of studies: a rough guide is at least ten studies per characteristic.
- Interpret with care: differences between studies are observational and do not prove cause.
A second example: pooling odds ratios
Binary outcomes are pooled on the log scale.
- Interpretation: the pooled odds ratio suggests a 31 per cent lower odds, but the interval includes 1.
- Why the log scale: odds ratios are skewed, but their logarithms are roughly normal.
Reporting a meta-analysis
Follow PRISMA 2020 and report these items.
- Effect measure: and why it was chosen.
- Model: fixed or random, and the estimator of τ².
- Heterogeneity: Q, I², τ² and the prediction interval.
- Plots: forest plot and funnel plot.
- Sensitivity analyses: and their results.
- Certainty of evidence: using GRADE.
Read the full PRISMA 2020 statement and see the PRISMA flow diagram.
Interpreting a pooled effect
A number alone does not tell the story.
- Size: is the effect small, medium or large, and is it important in practice?
- Precision: does the confidence interval include no effect?
- Consistency: do the studies point in the same direction?
- Quality: are the studies at low risk of bias?
- Applicability: do the participants resemble the people you care about?
Limitations of meta-analysis
Meta-analysis is powerful but not perfect.
- Garbage in, garbage out: poor studies give a poor pooled result.
- Publication bias: missing studies can inflate effects.
- Heterogeneity: too much variation makes one number misleading.
- Aggregation bias: study-level data may hide patient-level differences.
- Over-precision: a narrow interval may hide uncertainty from bias.
These limits are why a meta-analysis should come from a careful systematic literature review.
Why meta-analysis matters: what is meta-analysis good for?
Understanding what is meta-analysis good for helps you decide whether to use it.
- Precision: combining studies narrows the confidence interval around the effect.
- Power: small studies that miss an effect alone can detect it together.
- Consistency: it shows whether results agree across settings and populations.
- Exploration: it can reveal why results differ, through subgroups.
- Decision making: guideline panels use pooled estimates to recommend treatments.
What is meta-analysis? A short history
Knowing what is meta-analysis and where it came from explains its place in research today.
| Year | Event |
|---|---|
| 1904 | Karl Pearson combines results from several studies of typhoid inoculation |
| 1976 | Gene Glass coins the term meta-analysis |
| 1993 | The Cochrane Collaboration is founded to prepare and update health reviews |
| 2009 | The first PRISMA statement is published |
| 2021 | PRISMA 2020 replaces the 2009 version |
Today, what is meta-analysis? It is a standard tool in health, education and social science.
What is meta-analysis in a student dissertation?
Many students ask what is meta-analysis expected to include in a dissertation.
- Scope: a small, focused meta-analysis of five to ten studies can be enough.
- Supervisor: agree the effect size, model and software early.
- Data: prepare an extraction sheet with means, standard deviations and sample sizes.
- Transparency: report every decision and share the sheet as an appendix.
- Limits: state that a small number of studies limits the precision and tests for bias.
Frequently asked questions about what is meta-analysis
What is meta-analysis in simple terms?
Meta-analysis is a statistical method that combines the results of several studies on the same question into one overall estimate. Each study is converted to an effect size, such as a standardised mean difference or an odds ratio, and the studies are weighted by their precision so that larger, more precise studies count more. The method gives a more precise answer than any single study and shows how much results vary. It is usually part of a systematic review and requires studies that are similar enough in population, intervention and outcome.
What is the difference between a systematic review and a meta-analysis?
A systematic review is the whole process of finding, selecting, appraising and summarising studies on a question, using methods set out in advance. A meta-analysis is one statistical technique within that process, used to combine numerical results. Every meta-analysis should be based on a systematic review, but not every systematic review includes a meta-analysis, because studies may differ too much to be pooled. In that case authors summarise the evidence narratively, in tables and in graphs, and explain why they did not combine results.
What does I-squared mean in a meta-analysis?
I-squared, written I², estimates the percentage of variation in results across studies that is due to real differences rather than chance. It is calculated from Cochran’s Q statistic and its degrees of freedom. As a rough guide, about 25 per cent is low, 50 per cent is moderate and 75 per cent is high heterogeneity, but you should interpret it with context and with its confidence interval. If I² is high, explore reasons through subgroup analysis or meta-regression, and consider whether pooling is appropriate at all.
How many studies do I need for a meta-analysis?
Technically, two studies are enough to calculate a pooled estimate, but results from very few studies are imprecise and heterogeneity is hard to assess. Many guidelines suggest several studies for a stable estimate, and tests for publication bias, such as the Egger test, work best with at least ten. If you have only a few studies, report the pooled result cautiously, explain the limits and consider a narrative synthesis. What matters most is that the studies are similar enough in question and measure to be combined.
Need expert help? Our experienced writers can handle the search, screening and write-up for you. Place your order in two minutes, or read how a systematic literature review works first.
Systematic Literature Reviews
Need a meta-analysis done right?
Effect sizes, heterogeneity and forest plots, reported to journal standards.