Multivariate Meta-Analysis: Jointly Modeling Correlated Outcomes for More Precise Evidence Synthesis

Most systematic reviews still analyze each outcome in a separate univariate meta-analysis. This approach is simple and familiar, but it discards information: when outcomes measured in the same studies are correlated — as they almost always are — the univariate analyses ignore that correlation, lose precision, and make it impossible to answer joint questions such as whether an intervention improves both efficacy and safety simultaneously, or whether two outcomes move together across studies. Multivariate meta-analysis (MVMA) was developed to address exactly these limitations, and its use has grown rapidly with accessible software and clear methodological guidance.

The central idea is straightforward. Instead of synthesizing each outcome in isolation, MVMA stacks the outcome-specific estimates from each study into a vector and models the full vector with a multivariate random-effects distribution. The correlation between outcomes — both within studies and across studies — becomes part of the model, so that information from studies reporting only one outcome still contributes to the estimation of the other outcome through the estimated correlation. This mechanism, sometimes called borrowing of strength, is the engine behind most of the practical gains of the approach.

1. Why univariate analysis is often not enough

Univariate meta-analysis treats each outcome as if it were independent of the others. Three consequences follow. First, precision is lost: when some studies report only one of several outcomes, the univariate analysis of the other outcome simply drops those studies, whereas MVMA can use the correlation with reported outcomes to inform the missing one. Second, joint inference is impossible: reviewers cannot formally test whether effects on multiple outcomes differ, estimate the correlation between treatment effects on different endpoints, or combine outcomes into a single decision without ad hoc methods. Third, estimates can be inconsistent: separate analyses can produce apparently contradictory conclusions for related outcomes, and readers have no principled way to reconcile them.

The magnitude of the information loss depends on how much outcome data are missing and how strongly outcomes are correlated. When every study reports every outcome, univariate and multivariate analyses give essentially the same point estimates, though multivariate methods still provide valid joint inference and a coherent modeling framework. When missingness is substantial, as it often is in real reviews — especially for safety outcomes or secondary endpoints — the precision gains from borrowing strength can be considerable, and the reduction in outcome reporting bias can change conclusions.

2. Sources of correlation between outcomes

Correlation between outcomes arises from two distinct levels, and both must be modeled. Within-study correlation exists because outcomes are measured on the same participants: a patient who benefits on the primary endpoint may be more or less likely to benefit on a secondary endpoint or to experience an adverse event. This correlation is a property of the study data and must be recovered from individual patient data or approximated from published summary statistics. Between-study correlation describes whether studies with larger effects on one outcome also tend to have larger effects on another; it reflects the study-level relationship between treatment effects and is estimated directly by the multivariate model.

Both correlations matter for borrowing strength, but they play different roles. The between-study correlation determines how much information is borrowed for missing outcomes across studies, while the within-study correlation determines how the outcomes within a single study are weighted together. In practice, within-study correlations are often unavailable in published reports and must be imputed from external sources, sensitivity analyses, or approximation formulae; the multivariate model remains valid under reasonable assumptions, but reviewers should report what was assumed.

3. The multivariate random-effects model

The standard MVMA model is a direct generalization of the univariate random-effects model. For each study i and outcome j, the observed estimate is written as the sum of the true study-specific effect, a random deviation from the average treatment effect, and sampling error, with the random deviations and sampling errors correlated across outcomes. The average treatment effects, the between-study covariance matrix, and the within-study covariance matrices are estimated jointly, typically by restricted maximum likelihood (REML) or by Bayesian methods with a Wishart prior on the between-study covariance matrix.

Software implementation is mature. The mvmeta package in Stata and R, the metafor package in R, and the mvmeta R package provide REML-based fitting, while Bayesian implementations are available in WinBUGS, JAGS, and Stan (Wei and Higgins, Stat Med 2013). A key practical requirement is that the within-study covariance matrices — or at least the within-study correlations — must be supplied, which is the single most common obstacle for applied reviewers. When correlations are unknown, analysts typically assume a fixed correlation (for example 0.5), conduct sensitivity analyses over a range of plausible values, or use approximation methods to reconstruct covariances from marginal summaries.

The between-study covariance matrix carries the substantive information about how treatment effects on different outcomes relate to each other. Estimating it reliably requires a reasonable number of studies; with few studies the estimate can be unstable, and Bayesian approaches with informative priors or restricted models may be preferable. Reviewers should always report the estimated between-study correlations, as they are directly interpretable and clinically informative.

4. Borrowing strength and missing outcomes

The most celebrated property of MVMA is borrowing of strength: outcomes missing from some studies are estimated more precisely than they would be in a univariate analysis restricted to the studies that report them, because the model uses the estimated correlation with outcomes that are reported. Simulation studies (Jackson et al., Stat Med 2011) have shown that multivariate synthesis reduces mean squared error for outcomes with incomplete reporting, with the gains largest when correlations are moderate to strong and missingness is substantial.

This property also has a bias-reduction role. Kirkham and colleagues (Stat Med 2012) demonstrated that multivariate meta-analysis can reduce the impact of outcome reporting bias — the selective reporting of favorable outcomes — by exploiting the information in reported outcomes to inform the missing ones. While this is not a substitute for obtaining the missing data, it is a principled sensitivity tool that complements the publication-bias methods discussed in a previous post in this series.

Two cautions are important. Borrowing strength does not create information from nothing: if outcomes are uncorrelated, there is nothing to borrow, and the multivariate estimates converge to the univariate ones. And if the missingness mechanism is informative in ways the model cannot capture — for example, if studies selectively omit unfavorable outcomes in a way that is correlated with the true effect — borrowing strength can propagate rather than remove bias. The multivariate analysis should therefore always be accompanied by missingness diagnostics and sensitivity analyses.

5. Multivariate meta-regression and multi-parameter syntheses

MVMA extends naturally beyond multiple outcomes to any setting with multiple correlated parameters per study. Multivariate meta-regression models study-level covariates jointly for all outcomes, allowing tests of whether covariate effects differ across outcomes. Dose-response meta-analysis uses multivariate methods to pool the multiple regression coefficients that describe a nonlinear dose-response curve (for example, a restricted cubic spline), correctly accounting for the correlation between coefficients within each study. Gasparrini and colleagues (Stat Med 2012) formalized this as multivariate meta-analysis for non-linear and multi-parameter associations, providing the methodological foundation now used across environmental and nutritional epidemiology.

Multiple-arm trials provide another natural application. A trial comparing several treatments yields multiple treatment effects sharing a common control group; the effects are correlated, and analyzing them as independent comparisons inflates the variance and double-counts the control arm. A multivariate analysis stacks the correlated comparisons and estimates them jointly, an approach that connects directly to network meta-analysis (NMA), where the same principle is used to model the correlation structure induced by multi-arm trials and to estimate all pairwise comparisons coherently.

6. Relationship to network meta-analysis and diagnostic accuracy

Two other methods in this series are special cases of the multivariate framework. Network meta-analysis synthesizes multiple treatments simultaneously; its core statistical models can be expressed as multivariate meta-analyses of the treatment effects from multi-arm trials, and Riley and colleagues (BMJ 2017) present MVMA and NMA within a single conceptual framework of "multivariate and network meta-analysis of multiple outcomes and multiple treatments." The bivariate random-effects model for diagnostic test accuracy — covered in a recent post on diagnostic test accuracy meta-analysis — is exactly a two-outcome MVMA of logit-transformed sensitivity and specificity, where the correlation between the two measures within each study is explicitly modeled. Recognizing these connections helps reviewers choose the right model and borrow software and concepts across settings.

The unifying principle is that any synthesis problem with multiple correlated estimates per study is a candidate for multivariate modeling. Whether the multiple estimates are outcomes, treatments, time points, or regression coefficients, the same machinery — a vector of study-specific estimates, a within-study covariance structure, and a between-study covariance to be estimated — applies.

7. Practical implementation steps

Implementing MVMA requires careful data preparation. The minimal data for each study are the outcome-specific effect estimates and their variances; the within-study covariances (or correlations) are needed for full efficiency. When they are missing, reviewers must either extract them from individual patient data, reconstruct them from reported summary statistics, or assume and sensitivity-test them. The between-study covariance is estimated by the model and should be inspected for plausibility, with the estimated correlation matrix reported.

Model choice depends on the estimand and the data. Frequentist REML estimation is the default in most software; Bayesian estimation is attractive when studies are few, when prior information on correlations is available, or when the model is extended to non-standard settings. The number of outcomes should be modest relative to the number of studies, because the between-study covariance has k(k+1)/2 parameters for k outcomes; with many outcomes and few studies, dimension reduction or restricted structures may be needed.

Reporting should follow PRISMA 2020 principles: state which outcomes were modeled jointly, how within-study correlations were obtained or assumed, what missingness was present, and how the multivariate results differ from univariate sensitivity analyses. Transparency about assumptions is what makes a multivariate analysis reproducible and credible.

8. Evidence summary table

9. Actionable Steps: Run a Multivariate Meta-Analysis

10. Common mistakes in multivariate meta-analysis

The most frequent error is applying MVMA without any within-study correlation data and without acknowledging the assumption, silently treating the analysis as if the missing covariances did not matter. A second error is reporting only the multivariate estimates while hiding the univariate comparison, obscuring how much of the result depends on borrowed strength. A third error is fitting a model with too many outcomes for the number of studies, producing an unstable between-study covariance matrix that is presented as precise; dimension reduction or Bayesian priors should be considered. Finally, some reviewers use MVMA to "fix" outcome reporting bias without checking the missingness mechanism — borrowing strength cannot rescue data that are missing for reasons correlated with the true effects, and the analysis should be framed as a sensitivity tool, not a cure.

11. Relation to other evidence-synthesis methods

MVMA completes the evidence-synthesis toolkit covered in this series. Univariate random-effects models remain appropriate for single-outcome questions; heterogeneity assessment and meta-regression explain between-study variation; network meta-analysis handles multiple treatments; diagnostic test accuracy meta-analysis uses the bivariate model as a special case; trial sequential analysis controls random-error risk; and publication-bias assessment handles missing results. Multivariate meta-analysis is the unifying framework for correlated parameters across all of these settings, and reviewers who understand it can move fluidly between outcomes, treatments, time points, and regression coefficients.

Conclusion

Multivariate meta-analysis turns the correlation structure that univariate syntheses discard into a source of precision, coherence, and insight. By jointly modeling correlated outcomes, it borrows strength across endpoints, accommodates partially missing outcomes, supports valid joint inference, and connects naturally to network meta-analysis and diagnostic accuracy modeling. The approach demands discipline — within-study correlations must be obtained or justified, missingness must be diagnosed, and sensitivity analyses must be reported — but the payoff is a synthesis that uses all available information and answers the joint clinical questions that single-outcome analyses cannot. For reviews of interventions with multiple important outcomes, multivariate meta-analysis is no longer an advanced extra; it is increasingly the standard against which evidence synthesis will be judged.