What is meta-analysis? It is the method of statistically combining the results of independent studies that address the same research question to produce a single pooled effect estimate. It sits at the top of the evidence hierarchy and has become one of the fastest-growing designs in postgraduate theses and SSCI/SCI articles. This guide summarises every step of a meta-analysis conducted to the PRISMA 2020 standard.
What is meta-analysis, and how does it differ from a systematic review?
A systematic review is the process of searching, screening and synthesising the literature against pre-specified criteria; a meta-analysis is the statistical pooling of the numerical findings of the eligible studies at the end of that process. Every meta-analysis requires a systematic review, but not every systematic review has to end in a meta-analysis — a qualitative synthesis is a legitimate endpoint.
The PRISMA 2020 flow: four stages
- Identification: databases such as Web of Science, Scopus, PubMed and ERIC are searched with a pre-registered (PROSPERO) strategy; every search string and date is reported.
- Screening: duplicates are removed; title-abstract screening is conducted by at least two independent reviewers, with agreement (Cohen's kappa) reported.
- Eligibility: full texts are assessed against the inclusion/exclusion criteria; a reason is recorded for every excluded study.
- Included: the final pool of studies is coded with a data extraction form: sample, measures, the statistics to be converted into effect sizes, and a quality score.
Choosing the effect size and the pooling model
The common metric is dictated by the statistics the primary studies report: Hedges' g (Cohen's d with a small-sample correction) for mean differences, Fisher z-transformed r for associations, and odds ratios / risk ratios for proportions. For the pooling model, the default choice today is the random-effects model, because the fixed-effect assumption — that all studies share a single true effect — is rarely defensible in practice.
| Index | What does it tell you? | Interpretation threshold |
|---|---|---|
| Q statistic | Is between-study variance beyond chance? | p < .05 → heterogeneity present |
| I² | What share of total variance is genuine heterogeneity? | 25% low · 50% moderate · 75% high |
| τ² (tau-squared) | Variance of the true-effects distribution | Absolute magnitude; near 0 means homogeneous |
| Prediction interval | Where would the effect of a new study fall? | Caution if the interval spans 0 |
Publication bias: the studies you never see
Studies with significant results are more likely to be published, which can make a meta-analysis overstate the true effect. The standard battery of checks: funnel plot symmetry, Egger's regression test, the trim-and-fill adjustment, and sensitivity analyses for small-study effects. Reviewers now expect at least two of these methods to be reported side by side.
Which software: R, RevMan or CMA?
The metafor and meta packages in R are the most flexible option and produce publication-quality graphics; they are free, and placing the syntax in your thesis appendix gives full reproducibility. RevMan dominates Cochrane-style health reviews, while CMA suits those wanting point-and-click convenience. At Celsus, our default is R, precisely because of its transparency.
A good meta-analysis earns its value not from how many studies it pools, but from the discipline with which it justifies what it excludes.
Frequently Asked Questions
How many studies does a meta-analysis need at minimum?
Technically, two studies can be pooled; but for heterogeneity and publication bias analyses to be meaningful, at least 10 studies are generally recommended. With fewer studies, foreground the prediction interval and sensitivity analyses.
Is PROSPERO registration compulsory?
Most health-field journals require pre-registration; in education and the social sciences it is not compulsory, but it is strongly recommended for transparency and to prevent duplicate reviews.
Can my thesis combine quantitative and qualitative synthesis?
Yes. Studies unsuitable for numerical pooling can be reported through narrative synthesis; the distinction must be shown explicitly in the PRISMA flow diagram.
What meta-analysis services does Celsus provide?
Search strategy design, dual-reviewer screening support, effect size calculation and conversion, analysis in R with metafor, forest and funnel plots, and PRISMA 2020-compliant reporting — all delivered with reproducible code.