Meta-Analysis

A meta-analysis is the statistical method that combines the quantitative results of several independent studies addressing the same question into a single summary estimate. By pooling effect sizes and weighting each study by its precision, it can produce a more precise overall estimate than any one study and can examine how much the effect varies across studies.

Definition

A meta-analysis is a statistical procedure for combining the effect estimates of multiple independent studies of the same question into a weighted summary estimate, where each study is weighted according to the precision of its result.

Scope

This entry covers the statistical core of quantitative evidence synthesis: effect sizes and their weighting, fixed-effect versus random-effects models, the summary estimate and its confidence interval, and the closely related question of heterogeneity. It treats meta-analysis as a methodological topic carried out within a systematic review, not as clinical guidance.

Core questions

• What is the best single estimate of an effect given several independent studies?
• Should the studies be treated as estimating one common effect or a distribution of effects?
• How consistent are the studies, and how precise is the pooled result?

Key concepts

• Effect size (e.g. odds ratio, risk ratio, mean difference)
• Inverse-variance weighting
• Fixed-effect model
• Random-effects model
• Summary estimate and confidence interval
• Forest plot
• Heterogeneity

Mechanisms

Each study contributes an effect estimate (such as an odds ratio or a standardised mean difference) and a measure of its precision. Meta-analysis combines these by weighting each estimate, typically by the inverse of its variance, so that larger and more precise studies count for more. A fixed-effect model assumes every study estimates one common true effect and that differences are due only to sampling error. A random-effects model, most commonly the method of DerSimonian and Laird, assumes the true effect varies across studies and adds a between-study variance component, producing wider intervals and giving smaller studies relatively more weight. The pooled estimate and its confidence interval are usually displayed in a forest plot. Because pooling assumes the studies are similar enough to combine, the degree of heterogeneity is assessed alongside the summary.

Clinical relevance

Meta-analyses provide many of the summary effect estimates cited in clinical guidelines and health-technology assessments, and their results can be more stable than those of individual trials. Interpreting them requires understanding the model used and the heterogeneity present. This entry describes how summary estimates are computed and read; it is reference material for appraising evidence, not advice for treating an individual.

Epidemiology

Meta-analysis is applied across clinical trials, observational epidemiology, diagnostic accuracy research, and the social sciences. In medicine it is most often conducted within Cochrane and other systematic reviews, and software for meta-analysis is widely available. The random-effects model of DerSimonian and Laird is among the most frequently used pooling methods in the biomedical literature.

Evidence & guidelines

Meta-analyses reported within systematic reviews follow PRISMA 2020 reporting standards (Page et al., 2021), which include items on the synthesis methods, heterogeneity, and certainty of evidence. These are reporting standards, not treatment recommendations.

History

Statistical combination of studies goes back to early twentieth-century work by Pearson and Fisher, and Gene Glass introduced the term meta-analysis in 1976 in education research. In clinical medicine the random-effects model of DerSimonian and Laird (1986) became the standard approach to pooling, and the method spread rapidly with the growth of the Cochrane Collaboration and dedicated software. Borenstein and colleagues (2010) later clarified the conceptual difference between fixed-effect and random-effects models.

Debates

Fixed-effect or random-effects model?

The two models embody different assumptions about whether studies share one true effect or estimate a distribution of effects; the choice affects weighting and the width of the confidence interval, and the random-effects model can be unstable when studies are few.

When is pooling appropriate at all?

Combining studies that differ substantially in populations, interventions, or design can yield a precise but misleading summary, so the degree of heterogeneity bears directly on whether a pooled estimate is meaningful.

Key figures

• Gene Glass
• Rebecca DerSimonian
• Nan Laird
• Larry Hedges
• Julian Higgins
• Michael Borenstein

Seminal works

• dersimonian-laird-1986
• borenstein-2010
• higgins-2003-i2

Frequently asked questions

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

A fixed-effect model assumes all studies estimate one common true effect and weights them by precision alone. A random-effects model assumes the true effect varies across studies, adds a between-study variance term, and therefore gives wider confidence intervals and relatively more weight to smaller studies.

Can any set of studies be meta-analysed?

No. Pooling assumes the studies are similar enough in question, population, and design to share a meaningful summary. When they are too heterogeneous, a single pooled estimate may be precise but misleading, and a qualitative synthesis may be more appropriate.