Written and maintained by CASRAI Editorial Board
Last updated
Meta-regression extends subgroup analysis into a regression framework: instead of splitting studies into two or three bins by a categorical characteristic, it models how the estimated intervention effect changes as a continuous (or multi-level) study-level covariate changes. It answers a narrower, more specific question than heterogeneity assessment alone — not just “how much do results vary?” (that is I² and τ², covered in the companion guide below) but “does a specific, pre-specified study characteristic explain part of that variation?” Used well, it turns unexplained heterogeneity into a testable hypothesis. Used carelessly — with too few studies, or with covariates chosen after seeing which ones look significant — it manufactures false positives that read as findings.
What meta-regression actually tests
In a standard pairwise meta-analysis, the unit of analysis is the individual study’s effect estimate, pooled by fixed-effect or random-effects weighting. Meta-regression keeps that same unit of analysis — one data point per study — but regresses the effect estimate on one or more study-level covariates: mean participant age, publication year, dose, risk-of-bias score, region, or any other characteristic that varies across the included studies and was recorded (or can be extracted) for each one.
Because residual heterogeneity almost always remains after fitting the covariate (few, if any, single characteristics explain all of a review’s between-study variation), meta-regression is normally run as a random-effects meta-regression: the model estimates a regression slope for the covariate while also estimating a residual between-study variance term for whatever heterogeneity the covariate does not explain. A fixed-effect meta-regression, which assumes the covariate explains all of it, is rarely defensible in practice.
Meta-regression is the continuous-and-multivariable generalization of subgroup analysis, not a separate technique with a separate purpose. A subgroup analysis splitting trials into “pediatric” vs. “adult” is mathematically a meta-regression with a single binary covariate; meta-regression becomes necessary once the moderator of interest is continuous (mean age, follow-up duration, publication year) or once more than one covariate needs to be considered simultaneously.
The ecological fallacy: the risk that is specific to meta-regression
Every covariate in a meta-regression is a study-level average, not an individual-level measurement. If a trial reports a mean participant age of 62, that number describes the trial’s sample as a whole — it says nothing about whether the trial’s own effect estimate would differ for a 45-year-old versus an 80-year-old enrolled in it. A meta-regression that finds a significant association between mean age and effect size is showing an association across trials, not a within-trial, individual-level effect-modification finding.
Treating a study-level association as if it were an individual-level one is the ecological fallacy (also called aggregation bias in this context), and it is a well-documented, structural limitation of meta-regression, not a fixable data-quality problem. Two trials can have the same true within-trial age-effect relationship and still produce a spurious across-trial age association purely from how their sample compositions and effect sizes happen to co-vary — or the reverse: a real individual-level effect-modification pattern can be invisible or reversed at the trial-average level (Simpson’s-paradox-style). This is exactly why an individual patient data (IPD) meta-analysis, which regresses on each participant’s own characteristics rather than each trial’s average, is treated as a stronger design for genuinely testing effect modification — aggregate-data meta-regression is a hypothesis-generating tool for that question, not a substitute for it.
Practically, this means a meta-regression finding should be reported as “trials with a higher mean X tended to show effect Y,” never as “patients with higher X respond differently” — the second sentence is a claim the aggregate-level analysis cannot support.
The 10-studies-per-covariate rule of thumb
Meta-regression has few data points by construction — one per study, not one per participant — and a review with a dozen or two included studies is fitting a regression with a very small effective sample size. Higgins and Thompson’s simulation study on this exact problem (Statistics in Medicine, 2004) found that meta-regression with too few studies relative to the number of covariates produces an inflated Type I error rate: apparent “significant” moderators that are really noise, at a rate well above the nominal 5%.
The resulting, widely cited rule of thumb — echoed in the Cochrane Handbook’s own guidance on investigating heterogeneity — is that a simple (single-covariate) meta-regression needs at least ten studies for the one covariate being modeled, and a multivariable meta-regression needs roughly ten studies per covariate, meaning a two-covariate model wants at least twenty studies to be reasonably powered and not badly overfit. Below that threshold, both statistical power to detect a real association and the reliability of the estimated slope drop sharply.
Two caveats the rule of thumb itself does not capture: it assumes the covariate is reasonably evenly distributed across the included studies — ten studies clustered at only two distinct covariate values behave more like a two-group subgroup analysis than a ten-point regression, and give correspondingly less information. And it is a floor, not a target: clearing ten studies makes a single pre-specified covariate defensible to test, it does not license testing every covariate that happens to be recorded until one comes back significant.
Reducing false positives: pre-specification and covariate count
The Cochrane Handbook’s guidance on subgroup analysis and meta-regression is explicit that explorations of heterogeneity devised after heterogeneity is already observed — rather than specified in the protocol beforehand — should be treated as hypothesis-generating at best, not confirmatory. This matters more for meta-regression than for a single subgroup split, because a review with data on ten or fifteen recorded study characteristics offers ten or fifteen chances to find a “significant” moderator by chance if each is tested separately without correction.
The practical mitigations, consistent across the methodological literature: specify the covariates to be tested in the protocol, before results are pooled; keep the number of covariates tested small relative to the number of included studies (the 10-studies rule above, applied per covariate); and treat any post-hoc, exploratory covariate finding as a hypothesis for a future, purpose-designed study rather than as an established effect-modification finding on its own. Higgins and Thompson’s 2004 paper also proposed a permutation-test approach specifically to give a more honest significance level for a meta-regression finding than the standard regression p-value provides, precisely because the standard p-value does not account for how many moderators were implicitly available to be tested.
Meta-regression vs. subgroup analysis vs. heterogeneity assessment
| Technique | What it answers | When it applies |
|---|---|---|
| Heterogeneity assessment (I², τ²) | How much do effect estimates vary across studies, beyond what sampling error alone would predict? | Always, on any meta-analysis with more than one study |
| Subgroup analysis | Does the pooled effect differ between two or a few predefined categories of studies? | A categorical, low-cardinality moderator with enough studies in each subgroup |
| Meta-regression | Does the effect change as a continuous (or multi-category) study-level covariate changes? | A continuous moderator, or more than one moderator to consider jointly, with roughly 10+ studies per covariate |
| Individual patient data (IPD) meta-analysis | Does the effect differ by participant-level characteristics, avoiding the ecological fallacy entirely? | Raw participant-level data is obtainable from the original trial investigators |
Reporting a meta-regression
PRISMA 2020 (see the PRISMA and systematic review methodology guide) does not carve out meta-regression as a separate checklist item beyond the general requirement to report methods for any additional analyses, including subgroup analyses and meta-regression, and to state which were pre-specified. In practice this means a transparent write-up names every covariate tested (not only the ones that came back significant), states whether each was specified in the protocol or added post-hoc, reports the number of studies contributing to each covariate, and presents the regression coefficient with its confidence interval rather than only a p-value.
Related concepts
Meta-regression sits alongside several other tools for understanding why a pooled result varies or whether it can be trusted: heterogeneity in meta-analysis for quantifying the variation meta-regression tries to explain, fixed-effect vs. random-effects meta-analysis for the pooling model meta-regression builds on, standardized mean difference and Hedges’ g for the effect measure most often regressed, publication bias detection and the funnel plot for a separate small-study-effects check that is sometimes confused with meta-regression, network meta-analysis for comparing three or more interventions at once, and multicollinearity and VIF for the same covariate-correlation problem meta-regression inherits whenever more than one moderator is fit at once. The Cochrane Handbook is the primary methodological reference for all of the guidance above.
Frequently asked questions
Is meta-regression the same as subgroup analysis?
No, but they are closely related. A subgroup analysis compares a small number of categories; meta-regression handles continuous covariates and multiple covariates simultaneously within one regression model. A two-category subgroup split is a special case of a single-covariate meta-regression.
How many studies do I need for a meta-regression?
The commonly cited floor is at least ten studies for a single covariate, and roughly ten studies per covariate if fitting more than one — so a two-covariate model wants at least twenty studies. This comes from Higgins and Thompson’s 2004 simulation work on Type I error rates in meta-regression, and is echoed in the Cochrane Handbook. Below this, both power and reliability drop off sharply, and an uneven distribution of covariate values across studies makes the effective sample size even smaller than the study count alone suggests.
What is the ecological fallacy in meta-regression, in plain terms?
It is the mistake of reading a trial-level association (e.g., “trials enrolling older patients tended to show a smaller effect”) as if it were an individual-level finding (“older patients respond less well”). The covariate is a study average, not a per-patient measurement, so the two claims are not interchangeable — only individual patient data meta-analysis can test the individual-level claim directly.
Can I test every study characteristic I recorded and report the significant ones?
That approach inflates the false-positive rate, because each additional covariate tested is another chance to find a spurious association by chance alone. Pre-specifying a small number of covariates in the protocol, and clearly labelling any additional covariate as post-hoc/exploratory, is the accepted mitigation in the methodological literature and in Cochrane Handbook guidance.
