Metaregression

A meta-analysis technique that tests whether study-level characteristics — for example mean age, drug dose, publication year, baseline risk, or risk-of-bias — explain the heterogeneity in effect sizes across studies. It is essentially a regression in which each study is one data point: the study’s effect size is regressed on the covariate, usually weighted by study size.

A bubble plot: the x-axis is a study-level covariate (here, drug dose), the y-axis is each study’s effect size, and each bubble is one study sized by its weight. A regression line is fitted through the bubbles. Here the line slopes downward — studies using a higher dose tend to report a larger benefit — suggesting the covariate accounts for some of the between-study heterogeneity.

A non-flat slope means the covariate is associated with effect size and may explain part of the heterogeneity; a flat slope means it does not. But the unit on the x-axis is the study, not the patient — so this is an ecological (group-level) relationship. It cannot tell you how the covariate behaves within an individual patient.

Metaregression is the tool examiners expect when a meta-analysis asks why studies disagree rather than just reporting I². Its great trap is the ecological fallacy: a study-level association (higher-dose trials show more benefit) does not prove a dose–response within patients, because study-level confounders may drive the trend.

• Regression with each study = one data point
• Tests if a covariate explains heterogeneity in effect sizes
• Rough rule: ≥ ~10 studies per covariate to avoid over-fitting

Thrombolysis time-to-treatment in acute stroke — meta-analyses of alteplase trials used metaregression of treatment effect against onset-to-treatment time across studies, supporting the “time is brain” principle that earlier treatment yields greater benefit. The relationship is plotted with each trial as a bubble and a fitted slope. This is a study-level trend. Because it aggregates across trials it cannot, by itself, prove the individual-level time–benefit curve — that needs individual-patient-data analysis, which the IST-3/Emberson 2014 pooled analysis later provided.

• The ecological fallacy — inferring an individual-patient relationship from a study-level one.
• Running metaregression with too few studies for the number of covariates (over-fitting, false positives).
• Aggregation / confounding by study-level factors that travel with the covariate.