Meta Regression

Meta regression is a statistical extension of meta analysis that examines whether study level characteristics (covariates) explain variation in effect sizes across studies. It models the relationship between the effect estimate and one or more study features, such as mean participant age, intervention dose, study quality score, or study design, and tests whether these features account for observed heterogeneity.

When a meta analysis reveals substantial heterogeneity (for example, high I² values), the pooled estimate alone may not tell the full story. Meta regression helps explain why studies produced different results, moving beyond the observation that heterogeneity exists to understanding its sources. This information is valuable for guideline development (identifying which patient populations benefit most), for future research planning (identifying gaps in the evidence), and for clinical decision making (understanding how treatment effects vary with dose, duration, or population characteristics). The Cochrane Handbook describes it as one way to investigate heterogeneity, with cautions: it should generally not be considered with fewer than ten studies, its findings are observational, and the characteristics to be examined should be prespecified where possible.

Meta regression is analogous to standard regression analysis but operates at the study level rather than the individual level. The dependent variable is the effect estimate from each study (such as the log odds ratio, the mean difference, or the standardized mean difference). The independent variables are study level characteristics (moderators) hypothesized to explain variation in effects. The model estimates the association between each moderator and the effect size, controlling for the precision of each study through inverse variance weighting.

The standard approach uses a random effects meta regression model, which accounts for both within study sampling error and residual between study variability (the heterogeneity not explained by the included covariates). The proportion of heterogeneity explained by the model is often reported as R² analog or adjusted R².

Inference in meta regression is challenging because the number of data points (studies) is typically small relative to the number of potential moderators. Overfitting is a serious risk: with 20 studies and 5 candidate moderators, the analysis is akin to a regression with 20 observations and 5 predictors, which can produce spurious associations. The Cochrane Handbook cites the typical advice that at least ten studies should be available for each characteristic modeled (and notes that even this is too few when the covariates are unevenly distributed across studies), and it advises prespecifying the moderators to be tested on clinical or methodological grounds.

Permutation tests provide more accurate p values than standard methods when the number of studies is small, because the standard normal approximation may be unreliable. Knapp Hartung adjustments to confidence intervals and p values are also recommended to account for the uncertainty in estimating between study variance.

Because meta regression operates on aggregate (study level) data, it is susceptible to ecological bias (also called ecological fallacy or aggregation bias). An association between mean participant age and effect size at the study level does not necessarily mean that age modifies the treatment effect at the individual level. IPD meta analysis is needed to examine individual level effect modification without this risk.

Use meta regression when a meta analysis reveals substantial heterogeneity and you have a priori hypotheses about which study characteristics might explain it. It is most informative when there is meaningful variation in the moderator across studies, when the number of studies is sufficient (at least 10 studies, though more are desirable), and when the moderators are prespecified in the protocol. It adds little when there is little between study variation to explain (though a low I squared from few or imprecise studies does not rule out real differences), and it is not appropriate when the number of studies is very small or when the analysis is purely exploratory without prespecified hypotheses.

A common mistake is testing too many moderators relative to the number of studies, which leads to overfitting and spurious findings. Another error is interpreting meta regression results causally, when the analysis is observational and subject to confounding by other study characteristics that correlate with the moderator being tested. Some researchers fail to use random effects meta regression, using fixed effect models that do not account for residual heterogeneity. Interpreting a nonsignificant meta regression coefficient as evidence that the moderator has no influence (rather than as a failure to detect an influence with limited power) is another common mistake. Finally, committing ecological fallacy by drawing individual level conclusions from study level associations can be seriously misleading.

A meta analysis of 30 RCTs examining the effect of exercise interventions on blood pressure reduction shows I² = 71%, indicating substantial heterogeneity. The reviewers prespecify three potential moderators: mean participant age, intervention duration in weeks, and exercise intensity (moderate versus vigorous). Meta regression reveals that intervention duration is significantly associated with effect size (beta = negative 0.3 mmHg per additional week, p = 0.01), explaining 38% of the between study variance. Mean age and exercise intensity are not significantly associated with effect size, which, given the limited power, does not show that they have no influence. The reviewers conclude that trials of longer exercise programs reported greater blood pressure reductions, while noting that this is an observational, study level association: duration may be confounded with other trial characteristics, and the association may not directly translate to individual patient recommendations.