# Meta-analysis

Source: https://onco.cc/terms/meta-analysis/  
OnCo record `meta-analysis` (Term). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

A meta-analysis combines estimates from separate studies with a stated statistical method; it can improve precision, but it cannot repair studies that answer different questions or share the same biases.

## Summary

A meta-analysis statistically combines estimates from studies addressing a sufficiently compatible question. It can be part of a systematic review, which explicitly finds, selects and appraises studies; pooling is not compulsory. The result depends on the studies, their biases, effect definitions, model and weights. A precise pooled mean need not describe every setting or an individual person.

## Fields

- Kind: Term
- Last checked: 2026-11-14
- Also known as: meta-analyses; meta analysis; pooled estimate; random-effects meta-analysis; I-squared; tau-squared

## Notes

- A common-effect model targets one underlying effect, while a random-effects model describes a distribution of study effects and its mean. Weights follow the stated estimator, not simply the largest participant count. Clinical and methodological differences, bias and heterogeneity beyond sampling error need investigation; a random-effects calculation does not remove them or make incompatible studies comparable.
- A confidence interval for the mean effect differs from a prediction interval for an underlying effect in another comparable study under the model. Neither is a personal prognosis interval. I-squared estimates the proportion of variability in study estimates associated with heterogeneity rather than sampling error; it is neither the percentage of people helped nor the percentage of studies that are wrong. Tau-squared is between-study effect variance on the chosen scale, such as squared score points or squared log risk ratios. These estimates can be uncertain, especially with few studies, and no I-squared cutoff alone makes pooling safe.
- Fictional forest example: five studies compare treatment A minus control on the same 0 to 100 symptom scale at day 30. Their effects are -8, -6, -4, -2 and 0 points, each with standard error 2 and weight 20 per cent, so the square areas are equal. The mean is -4.0 points. With a classical moment variance estimate, Q = 10, I-squared = 60 per cent and tau-squared = 6 squared score points. A normal 95 per cent mean interval is about -6.8 to -1.2; the stated 2009 prediction method, using a t distribution with three degrees of freedom, gives about -13.0 to +5.0. The study intervals, mean diamond and wider prediction interval answer different questions. Five studies are few and these intervals are approximate; a negative mean does not imply a negative effect in every setting.
- Retain search dates, eligibility, unique study identities, risk of bias, treatment sequence, event definition, starting clock, horizon, effect scale, model, weights, uncertainty and sensitivity analyses. Abstracts, full reports and follow-up of the same participants are linked reports rather than independently weighted trials. Availability related to findings can leave evidence missing. Pooling observational associations does not create randomization or establish one person's treatment benefit. PRISMA promotes transparent reporting; completing its checklist does not certify validity.
- Fictional weighted example: two comparable three-year risk differences are -4 and -8 percentage points with supplied weights 0.75 and 0.25. Their weighted mean is -5 points; these stipulated weights are not claimed precision estimates. Correct arithmetic would still not justify pooling if one study starts at randomization and another selects responders at a later landmark. The eligibility, clock and target question must be compared first.
- Before pooling, link abstracts, full reports and later follow-up of the same study using registry identifiers, recruiting centres, recruitment dates, treatments, participant counts and baseline characteristics. Different author lists or sample sizes need not mean different trials: a later cutoff or selected analysis cohort can change the reported number. Conversely, one report can describe more than one study. Cochrane Handbook chapter 5 makes the study, rather than the publication, the unit of interest. Retain useful details from secondary reports without counting the same participants again as independent replication.
- For the fictional five-study example, Q has four degrees of freedom. The DerSimonian and Laird moment estimate gives tau-squared = 6; each random-effects weight is 1/(4 + 6), normalized to 20 per cent. Each study's normal interval extends 3.92 points either side of its estimate. The pooled standard error is the square root of 2 points. In the specified 2009 method, -4 plus or minus 3.1824463 times the square root of (6 + 2) gives -13.0013 to +5.0013 points, using three degrees of freedom. These choices reproduce this example; they are not a universal recommended estimator.

## Sources

- Higgins, Thompson and Spiegelhalter: a re-evaluation of random-effects meta-analysis (JRSS A 2009): https://doi.org/10.1111/j.1467-985X.2008.00552.x
- Cochrane Handbook chapter 10: analysing data and undertaking meta-analyses: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-10
- PRISMA 2020 statement: reporting systematic reviews (BMJ 2021): https://doi.org/10.1136/bmj.n71
- Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses: https://doi.org/10.1016/j.jclinepi.2016.04.014
- Cochrane Handbook chapter 5: identifying and linking multiple reports of one study: https://training.cochrane.org/handbook/current/chapter-05

## Connected records

- terms: [Confidence interval](https://onco.cc/terms/confidence-interval/), [Publication format and peer-review status](https://onco.cc/terms/publication-and-peer-review-status/), [Randomised trial](https://onco.cc/terms/randomised-trial/), [Real-world evidence](https://onco.cc/terms/real-world-evidence/)

---
JSON: https://onco.cc/api/v1/entities/meta-analysis.json