Meta-Analysis

Meta-analysis is the statistical procedure that combines effect estimates from several studies addressing the same question into a single, more precise pooled estimate. By weighting each study according to its precision, it extracts an overall answer that no single study could provide and reports the remaining uncertainty around it.

Definition

Meta-analysis is the quantitative combination of effect estimates from multiple studies into a weighted summary estimate, typically using inverse-variance weighting under either a fixed-effect model (one assumed common effect) or a random-effects model (effects assumed to vary across studies).

Scope

This entry covers the core mechanics of pooling: how individual study effects are weighted, the distinction between fixed-effect and random-effects models, and how the pooled estimate and its interval are read. It treats meta-analysis as a quantitative method within evidence synthesis and is a reference description rather than clinical guidance. The broader systematic-review process is covered in the related meta-analysis node under systematic reviews.

Core questions

• How are individual study results weighted when they are combined?
• What does the pooled estimate represent under a fixed-effect versus a random-effects model?
• How should the confidence interval around a pooled estimate be interpreted?
• When is it appropriate to pool studies at all?

Key concepts

• Inverse-variance weighting
• Fixed-effect model
• Random-effects model
• Pooled (summary) effect
• Confidence interval and prediction interval
• Forest plot

Mechanisms

Each study contributes an effect estimate (such as a risk ratio, odds ratio, or mean difference) together with its standard error. In inverse-variance weighting, more precise studies receive greater weight, and the weighted average is the pooled estimate. Under a fixed-effect model all studies are assumed to share one true effect, so weights depend only on within-study variance. Under a random-effects model, true effects are assumed to vary, so an estimated between-study variance is added to each weight, shrinking the influence of the largest studies and widening the confidence interval. The DerSimonian-Laird approach gave the classic moment-based estimator of that between-study variance; Riley and colleagues stress that the random-effects summary is an average effect whose interpretation, and the prediction interval around it, must reflect that effects differ across settings.

Clinical relevance

Pooled estimates from meta-analyses frequently sit at the top of evidence hierarchies and feed directly into guidelines and health technology assessment, so being able to read a forest plot and understand what its summary line means is part of evidence appraisal. This entry explains how the pooled estimate is produced and is not a basis for individual treatment decisions.

Evidence & guidelines

The conduct and transparent reporting of meta-analyses are governed by the Cochrane Handbook (Higgins & Green, 2008) and the PRISMA statement (Moher et al., 2009), which specify how the pooled estimate, model choice, and surrounding uncertainty should be presented.

History

The term meta-analysis was introduced by Gene Glass in 1976 for the quantitative synthesis of research findings. Its translation into clinical research was anchored by DerSimonian and Laird's 1986 random-effects framework, and later expositions such as Borenstein and colleagues (2010) clarified the conceptual difference between fixed-effect and random-effects pooling that still organises practice today.

Debates

What does a random-effects summary estimate actually mean?

Because the random-effects model averages over a distribution of true effects, its summary line is an average rather than a single common value; Riley and colleagues argue that a prediction interval, not just the confidence interval, is needed to convey the range of effects across settings.

Key figures

• Rebecca DerSimonian
• Nan Laird
• Michael Borenstein
• Larry Hedges
• Julian Higgins
• Richard Riley

Seminal works

• dersimonian-laird-1986
• borenstein-2010
• higgins-handbook-2008

Frequently asked questions

What is the difference between a fixed-effect and a random-effects meta-analysis?

A fixed-effect analysis assumes every study estimates the same single true effect, while a random-effects analysis assumes the true effect varies across studies and adds a between-study variance term, which usually widens the confidence interval.

Can any set of studies be combined in a meta-analysis?

No. Pooling is only meaningful when studies are similar enough in question, population, and outcome; when they are too diverse, combining them can produce a precise but misleading summary.