Is meta-regression just ordinary regression on all participants?

No; aggregate-data meta-regression usually analyzes study effect estimates.

Can it prove why studies differ?

Not by itself; the associations are observational across studies.

Does a large combined participant count solve a shortage of studies?

No; study-level moderators still have only the available study-level variation.

AI Glossary / Statistics and probability / Intermediate

Meta-regression

Definition

Meta-regression is a statistical method that relates study-level effect estimates to characteristics of the studies contributing to a meta-analysis. These characteristics are often called moderators. The method can investigate whether observed differences in effects are associated with features such as study setting, intervention intensity, or follow-up duration, while accounting for the precision of the effect estimates and potentially residual between-study variation.

What it means

Ordinary meta-analysis often summarizes effects with a single average. Meta-regression adds a question: do study characteristics help describe why the estimates differ? The response variable is typically an estimated effect, not the individual participants' raw outcomes.
A study-level association is not automatically an explanation of causation. Studies may differ in several related ways, and a characteristic that predicts their effect estimates may be a proxy for another difference. The analysis should distinguish a proposed explanation from an established mechanism.

How it works

Start with a defensible meta-analysis dataset containing comparable effect estimates and their sampling variances. Define a small number of plausible moderators, ideally before examining the results, and record how they are measured or coded.
A simple model writes a study effect estimate as an intercept plus a moderator coefficient times the study characteristic, plus sampling error. A mixed-effects version also includes residual variation in the true effects across studies. Under a basic independent-study model, weights commonly involve the inverse of sampling variance plus residual between-study variance.
Fit the specified model and examine the coefficient, uncertainty, residual heterogeneity, and influential studies. A continuous moderator coefficient describes the association per unit of that study characteristic on the chosen effect scale. Categorical moderators require a defined reference category.
The number of studies, not merely their total participant count, limits the information available for study-level regression. A small set of studies cannot reliably support many moderator terms. Sparse categories, missing covariates, and correlated moderators can make estimates unstable.
Account for dependent effect estimates where the same study or participants contribute several results. Check functional form and restrict interpretation to the observed range. A model predicting study effects is not automatically valid for extrapolation to new populations or much longer follow-up.

Three practical examples

AI research synthesis: analysts examine whether estimated workflow benefits vary with study duration, while recognizing that duration may correlate with user experience or task type.
Learning evidence: a review investigates whether the reported effect of a training intervention is associated with its scheduled intensity across studies.
Evaluation review: researchers compare effect estimates by study design and consider whether methodological differences explain part of the observed heterogeneity.

Think of it this way

Think of a map where each point represents a whole study, with its effect on one axis and a study characteristic on the other. Meta-regression fits a relationship among those study points, giving attention to their uncertainty. It does not reveal the individual-level pattern hidden inside each point.

Worked example

For an arithmetic illustration, three invented independent studies have moderator values 0, 1, and 2. Their estimated effects are 1, 2, and 4 outcome units. Suppose sampling variances are 1, 0.5, and 1, with no additional between-study variance included in this demonstration. The weights are therefore 1, 2, and 1.
The weighted moderator mean is one, and the weighted effect mean is (1 + 4 + 4)/4 = 2.25. The weighted cross-product of centered moderator and effect is three. The weighted squared moderator deviation is two, so the slope is 3/2 = 1.5. The intercept is 2.25 - 1.5 × 1 = 0.75.
The fitted line is effect = 0.75 + 1.5 × moderator. It predicts 0.75, 2.25, and 3.75 at the three observed moderator values. Residuals are 0.25, -0.25, and 0.25, and their weighted squared sum is 0.25.
This fitted association says that a one-unit study-level moderator difference corresponds to 1.5 effect units in this constructed regression. It does not prove that increasing an individual's exposure by one unit causes that change. Three points illustrate the calculation only; no confidence interval, significance claim, or adequate-study-count claim is made.

Illustrative case study

Fictional case: Hazel Reviews synthesizes studies of an AI tutoring tool. Longer studies appear to report larger gains. A meta-regression identifies an association with duration, but longer studies also tend to provide more instructor support.
The team reports the association cautiously, examines the limited overlap in study designs, and identifies a need for stronger comparisons. It avoids claiming that extending deployment alone will reproduce the estimated gain. This fictional scenario illustrates confounding among study characteristics rather than a real finding about tutoring technology.

Common misconceptions and mistakes

Do not infer individual-level effect modification from a relationship between study averages. For example, an association with study mean age need not represent how treatment effects vary by age within each study. This is an ecological-inference problem, and directions can even differ.
Do not conclude that two subgroups differ merely because one has a significant pooled effect and the other does not. The difference itself needs an appropriate comparison. Do not search many moderators and report only the most favorable association without acknowledging multiplicity and exploration.
A reduction in estimated heterogeneity does not prove that the moderator causally explains it. Nor does a nonsignificant slope establish absence of effect modification; imprecision, measurement error, and restricted moderator range can obscure relationships.

Business use

Meta-regression can help organize heterogeneity and generate focused research questions. It is most useful when the candidate moderators are meaningful, measured consistently, and supported by sufficient study-level information.
Report the effect scale, moderator coding, weights, variance model, uncertainty method, number of studies, missing data, and whether hypotheses were prespecified. For AI-assisted evidence reviews, verify extracted study characteristics and avoid treating a language model's inferred description as a measured moderator. Interpret the resulting association alongside design differences and risk of bias.

Check your understanding

what does a continuous moderator slope describe?

Answer: change in the modeled study effect per unit of the study characteristic.

does a study-mean-age association prove the same relationship within participants?

Answer: no.

can several correlated moderators be difficult to distinguish?

Answer: yes; their separate associations may be weakly identified.

Related terms

Meta-analysis
Inverse-variance weighting
Confounding
Moderator
Heterogeneity
Mixed-effects model
Ecological fallacy
Effect modification

Learn next

Next, compare study-level and individual-participant-data analyses of effect modification. Then study influence diagnostics, dependent effects, and the assumptions behind residual heterogeneity estimates.

Sources

Cochrane Handbook, Chapter 10, sections on meta-regression and subgroup interpretation: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-10

Previous: Meta-optimization Next: Metaclass (knowledge representation)