This guide assumes you have already completed a systematic review and hold a defined set of included studies. Pooling studies identified any other way is not defensible.
1. Decide what you are pooling
Every included study must report the same outcome, measured comparably, in comparable populations. Where they do not, that outcome cannot be pooled — and saying so is a legitimate result.
2. Extract effect sizes
Choose the effect measure the outcome demands:
- Continuous outcomes: mean difference where the scale is shared, standardised mean difference where it is not
- Binary outcomes: risk ratio, odds ratio or risk difference
- Time-to-event: hazard ratio
- Correlational: Fisher's z transformed correlation
Extract in duplicate. Transcription error is the most common source of a wrong pooled estimate, and it is invisible once the analysis runs.
3. Choose fixed or random effects
A fixed-effect model assumes every study estimates one true effect, and differences are sampling error alone. A random-effects model assumes the true effect varies between studies.
In practice, random effects is appropriate for most reviews of real-world literature, because populations and protocols genuinely differ. State the choice and the reason for it before you see the result.
4. Assess heterogeneity
Report I², tau² and the Cochran Q test. I² above roughly 50% suggests substantial heterogeneity that needs explaining rather than ignoring.
Explain it through pre-specified subgroup analysis or meta-regression. Subgroups invented after seeing the data are exploratory and must be labelled as such.
5. Examine publication bias
With ten or more studies, produce a funnel plot and run Egger's test. With fewer, say that the assessment was not possible rather than presenting an uninformative plot.
6. Test the robustness
Run sensitivity analyses: excluding high risk-of-bias studies, excluding the largest study, and switching the model. If the conclusion survives all three, say so. If it does not, that is the more important finding.
7. Report it properly
A forest plot with individual and pooled estimates, confidence intervals and weights. Then the software and package used, with version numbers — results differ between implementations.
Our data analysis service runs meta-analyses in R, STATA and Comprehensive Meta-Analysis, with reproducible scripts supplied.