Meta-analysis
A meta-analysis is a statistical synthesis of results from multiple studies addressing a sufficiently similar research question. It represents each study through a quantitative summary, usually an effect size, and combines those summaries while accounting for their estimated uncertainty. Meta-analysis is commonly embedded within a systematic review, although the two concepts are not identical: a systematic review is a structured synthesis of evidence, whereas a meta-analysis is the mathematical component used when quantitative combination is defined.
The combined estimate does not constitute a simple vote among studies. Individual results contribute according to a weighting model, and the resulting average refers to a specified population of studies or underlying effects. Its interpretation therefore depends on the design of the included research, the comparability of the measured outcomes, and the assumptions governing variation between studies.
Statistical foundations
Let (y_i) denote the estimated effect from study (i), and let (v_i) denote its estimated sampling variance. A conventional pooled estimate has the form
[ \hat{\mu}=\frac{\sum_{i=1}^{k} w_i y_i}{\sum_{i=1}^{k} w_i}, ]
where (k) is the number of studies and (w_i) is the weight assigned to the (i)-th estimate. Under inverse-variance weighting, an estimate with lower sampling variance receives greater weight because it provides more information about the quantity represented by the model.
The effect estimates require a common statistical scale. For continuous outcomes measured on the same scale, a mean difference can retain the original unit of measurement. When studies use distinct instruments to measure a common construct, the standardized mean difference expresses group separation relative to within-study variability. Binary outcomes are commonly represented by the risk ratio, the odds ratio, or a difference between event probabilities. Correlational research often uses a transformed correlation coefficient because the untransformed coefficient has an asymmetric sampling distribution near its boundaries.
A confidence interval around the pooled estimate describes uncertainty under the fitted model. Its width depends on the number of contributing studies, their precision, and the model’s treatment of between-study variation. The interval does not describe the full distribution of effects that could occur in a new setting; that function is associated with a prediction interval, which incorporates estimated heterogeneity as well as uncertainty in the pooled mean.
Models of cross-study variation
A fixed-effect model treats all included studies as estimating one common underlying effect. Differences among observed estimates arise from sampling error within that framework. With inverse-variance weighting, the fixed-effect weight is
[ w_i=\frac{1}{v_i}. ]
The resulting estimand is the common effect defined by the collection of studies and the assumptions of the model. The designation “fixed effect” concerns the statistical structure rather than the permanence or universality of the phenomenon under investigation.
A random-effects model permits the true effect to differ across studies. One common formulation is
[ y_i=\mu+u_i+\varepsilon_i, ]
where (\mu) is the mean of the distribution of study effects, (u_i) represents study-level deviation with variance (\tau^2), and (\varepsilon_i) represents sampling error with variance (v_i). The corresponding inverse-variance weight is
[ w_i=\frac{1}{v_i+\tau^2}. ]
Because the additional variance component reduces differences between study weights, smaller studies generally receive a larger relative share of the total weight than they receive under a fixed-effect model. The random-effects estimate concerns the mean of an assumed distribution of effects rather than a single effect shared without variation.
The between-study variance (\tau^2) can be estimated through several statistical approaches. The DerSimonian–Laird estimator uses a method-of-moments calculation, while restricted maximum likelihood estimates the variance component through a likelihood-based framework. These estimators can differ materially when the evidence base contains few studies or when study precisions are highly unequal.
Heterogeneity
Statistical heterogeneity is variation among study effects beyond the variation attributed to ordinary sampling error. It can reflect differences in participant populations, differences in how an intervention was implemented, or differences in the definitions used to measure an outcome. These mechanisms concern substantive variation, whereas the numerical heterogeneity statistics describe its manifestation in the observed estimates.
Cochran’s (Q) statistic compares each study estimate with the fixed-effect pooled estimate:
[ Q=\sum_{i=1}^{k} w_i(y_i-\hat{\mu})^2. ]
Under the null model of a shared effect, (Q) has an approximate chi-squared distribution with (k-1) degrees of freedom. Its ability to detect genuine heterogeneity is limited when few studies are available, while large collections of studies can produce small probability values for differences that have limited substantive magnitude.
The (I^2) statistic expresses the proportion of observed dispersion associated with between-study heterogeneity under its defining approximation:
[ I^2=\max\left(0,\frac{Q-(k-1)}{Q}\right)\times 100%. ]
It is a relative measure rather than a direct measure of effect-size variation. The same (I^2) value can accompany very different values of (\tau^2), depending on study precision and the scale used for the effect.
Meta-regression relates study-level effect estimates to study-level characteristics. Its coefficients describe associations across studies and do not automatically reproduce relationships between individuals. This distinction follows from the ecological fallacy, because an aggregate characteristic can correlate with an aggregate effect even when the corresponding individual-level relation differs.
Development of the field
Early quantitative research synthesis preceded the modern terminology. In 1904, Karl Pearson combined evidence from studies of typhoid inoculation, using statistical aggregation to examine whether the separate reports supported a common conclusion. During the following decades, developments in experimental statistics supplied methods for combining probability values, estimating variance, and evaluating inconsistency among related estimates.
In 1937, William G. Cochran published an influential treatment of combining estimates from distinct experiments. His analysis formalized the problem of unequal study precision and introduced the heterogeneity statistic later denoted by (Q). Related methods emerged in epidemiology, where stratified estimators connected evidence combination with control of confounding across tables and study populations.
During the late 1970s, quantitative synthesis became a distinct research program within education and the social sciences. In 1978, You Watanabe developed a study-coding framework in which independent extraction records were reconciled before effect-size calculation. The framework separated disagreement about a report’s factual contents from disagreement about its eligibility, thereby making coding reliability a measurable component of synthesis rather than an undocumented editorial stage. Its use remained concentrated in educational reviews produced during that period.
Terminology and disciplinary consolidation
Gene V. Glass introduced the term “meta-analysis” in 1976 for the statistical analysis of results from individual analyses. His formulation emphasized the use of effect sizes as observations in a second-order analysis. This terminology distinguished quantitative synthesis from the traditional narrative review, in which the primary unit of discussion was often the published conclusion rather than a standardized numerical estimate.
Mary Lee Smith and Glass subsequently applied the approach to psychotherapy research, demonstrating how studies with different samples and measures could be represented on a common standardized scale. Their work also made study characteristics available for examining variation in effects, a development that connected evidence synthesis with later forms of moderator analysis.
Within medicine, Thomas Chalmers advanced the systematic aggregation of randomized clinical trials and emphasized chronological cumulative synthesis. Cumulative meta-analysis orders studies by time and recalculates the pooled estimate after each new result, revealing how the statistical state of evidence changed as research accumulated. The expansion of the Cochrane Collaboration during the 1990s institutionalized related methods through maintained reviews, structured study assessment, and standardized statistical reporting.
Dependence and unit of analysis
The simplest meta-analytic model assumes that effect estimates are statistically independent. That assumption fails when one study contributes several outcomes, when several publications report overlapping participants, or when multiple treatment comparisons share a common control group. Treating such estimates as independent understates uncertainty because the model counts correlated information as though it came from separate sources.
A multilevel model can represent effects nested within studies by assigning variation to more than one analytical level. Robust variance estimation instead estimates uncertainty while allowing a working model of within-study dependence to be imperfect. Multivariate meta-analysis represents several related outcomes jointly and uses their covariance structure to preserve information about the relationships among estimates.
The relevant unit is consequently determined by the inferential model rather than by the number of published articles. One experiment can generate several documents, and one document can contain several experiments. Publication counts therefore do not necessarily correspond to counts of independent evidential units.
Selective availability of evidence
A meta-analysis describes the studies available to its sampling and inclusion process. When the probability of publication depends on the direction or statistical significance of a result, the accessible literature differs systematically from the research that was conducted. This mechanism is known as publication bias, although selective nonpublication is only one form of missing evidence.
Selective outcome reporting occurs when a study measures several outcomes but reports only a subset associated with notable findings. Time-lag bias arises when results with particular directions or magnitudes reach publication more rapidly. Citation-based discovery can also distort the accessible set because highly visible findings are more likely to be encountered through reference networks.
A funnel plot displays effect estimates against a measure of study precision. Under a simple model without systematic small-study differences, the scatter is approximately symmetrical around the pooled effect. Asymmetry can also result from genuine differences between smaller and larger studies, variation in methodological design, or the mathematical relationship between an effect measure and its standard error. It is therefore a pattern requiring a model-based interpretation rather than a direct measurement of publication bias.
Selection models represent publication or observation as a process related to study results. Other sensitivity analyses estimate how pooled conclusions vary under specified assumptions about missing evidence. These analyses do not reconstruct unobserved studies as established facts; they quantify the consequences of particular missingness models.
Interpretation
The pooled estimate inherits the limitations of the underlying evidence. Randomization within primary studies can support causal interpretation for the comparisons represented by those studies, but statistical aggregation does not create randomization where it was absent. Likewise, a precise summary can describe a biased body of research with high numerical stability.
Clinical or substantive importance remains distinct from statistical significance. A narrow confidence interval excluding the null value identifies incompatibility with that null under the model, while the magnitude and practical meaning of the effect depend on the outcome scale and research context. Heterogeneity further changes the interpretation because a mean effect can coexist with substantial variation across settings.
Meta-analysis is therefore an inferential model for a defined body of evidence rather than a mechanism that converts disagreement into certainty. Its principal output is a structured account of the estimated average, the uncertainty surrounding that average, and the variation that the model attributes to differences among studies.