Distributional regression models for meta-analysis

Abstract: Meta-analyses are regarded as the highest level in the hierarchy of evidence, yet standard models traditionally concentrated on estimating the mean effect size, often under restrictive assumptions about the underlying distribution, such as homogeneous variance, symmetric shapes. We introduce a distributional regression framework for meta-analysis that generalizes these conventional models by allowing all parameters of the effect size distribution, such as location, scale, and shape, to be modelled as functions of explanatory variables. This unified framework accommodates a wide range of existing models, including random-effects, multilevel, multivariate, location-scale, and outlier-robust meta-analyses, as special cases. We provide an illustrative example, using 67,393 meta-analyses from the Cochrane Database of Systematic Reviews, employing location-scale models to investigate whether smaller studies tend to report larger effect sizes (i.e., small-study effects) and exhibit greater heterogeneity. We discuss implementation strategies using existing software, considerations for model selection and pre-registration, and the need for further methodological development. By moving beyond the mean effect size, distributional regression enables researchers to explore systematic variation in distributional structure, facilitating the joint test of new hypotheses corresponding to multiple distributional parameters.

• The paper introduces a statistical framework that models location, scale, and shape parameters as functions of explanatory variables.
• It unifies traditional meta-analytic models, such as random-effects and multivariate approaches, under a single distributional regression framework.
• Empirical findings show the approach can reveal subtle small-study effects and heterogeneity, improving the robustness of inferences in meta-analyses.

Distributional Regression Models for Meta-Analysis: A Comprehensive Framework

Introduction

This paper, "Distributional regression models for meta-analysis" (2604.00424), proposes a unifying statistical framework for meta-analytic modelling that generalizes conventional approaches by allowing all parameters of the effect size distribution—location, scale, and shape—to be modelled as functions of explanatory variables. Departing from traditional meta-analytic paradigms, which typically concentrate on mean effect size estimation under restrictive assumptions (e.g., homogeneous variance, symmetry), the distributional regression framework accommodates a spectrum of advanced models and addresses fundamental assumption violations pervasive in empirical research. The framework not only subsumes random-effects, multilevel, multivariate, location-scale, and robust models as special cases but also equips researchers with the capability to jointly interrogate hypotheses about multiple distributional dimensions.

Limitations of Standard Meta-Analytic Models

Standard random-effects (RE) and mixed-effects meta-analytic models are predicated on assumptions such as normality of within-study errors, independence of effect sizes, homogeneous between-study heterogeneity ($\tau^2$), and the normality of between-study random effects. These constraints often fail in practice, particularly in preclinical meta-analysis, leading to model misspecification, inflated Type I errors, biased parameter estimates, and compromised inference validity.

• Non-normal sampling errors: Binary or count data, prevalent in preclinical studies, violate normality assumptions. GLMMs have emerged as viable solutions by allowing for the correct mean-variance relationship for such outcomes.
• Dependence of effect sizes: Hierarchical and multivariate dependencies, resulting from multiple outcomes or repeated measures within studies, require multilevel and multivariate models.
• Heteroscedasticity: Uniform $\tau^2$ is untenable when between-study variance varies systematically with study characteristics. Location-scale models address heterogeneity as a function of moderators.
• Outliers and distributional misspecification: Heavy-tailed or skewed true effect size distributions, and structural zeros in rare event data, necessitate the adoption of robust and zero-inflated mixture models.

While these extensions improve flexibility, they remain ad hoc and compartmentalized, motivating the need for a unified, extensible modelling framework.

The Distributional Regression Framework

The core innovation is the formulation of distributional regression for meta-analysis, in which all parameters of the chosen effect size distribution (e.g., mean, variance, skewness, kurtosis) are specified as (potentially hierarchical/multilevel) regression models with covariate and random effect components:

• For observed effect sizes $y_{ij}$, the density $f(y_{ij}|\theta_{ij})$ has parameter vector $$\theta_{ij} = (\theta^{(1)}_{ij}, ..., \theta^{(p)}_{ij})$$.
• Each parameter $\theta^{(m)}_{ij}$ is related to covariates through a regression with link function $g_m$, fixed and random effects: $$g_m(\theta^{(m)}_{ij}) = X_{ij}\beta^{(m)} + Z_{ij}u^{(m)}_{i}$$.

All prevalent meta-analytic models are special cases under constraints on $f$, $p$, the regression structure, and parameterization.

Unification of Meta-Analytic Models

The framework demonstrates how conventional and advanced models arise as constrained instances:

• Random-Effects Model: The effect size distribution is univariate Gaussian with mean and variance parameters. The model reduces to $\tau^2$0, $\tau^2$1, with all parameters fixed apart from the mean.
• GLMMs for Non-Normal Data: Binary or count outcomes are modelled via binomial or Poisson distributions, with logit/log links and moderator effects, integrating random effects for residual heterogeneity.
• Multilevel Meta-Analysis: Multiple nested levels (e.g., outcome within study within lab) are captured by additive random effects, each reflecting variability at distinct hierarchical scales.
• Multivariate Meta-Analysis: Correlated effect sizes (e.g., multiple endpoints per study) are modelled through joint, typically Gaussian, distributions with structured variance-covariance matrices.
• Location-Scale Models: Between-study variance is linked to study-level moderators, enabling heteroscedasticity modelling: $\tau^2$2.
• Robust and Mixture Models: Heavy-tailed (e.g., $\tau^2$3-distribution), skewed, or mixture distributions are permitted for random effects, accommodating outliers, skewness, and latent subpopulations.

This generalization allows comprehensive hypothesis testing about any aspect of the effect size distribution, conditional on substantive predictors.

Empirical Case Study: Small-Study Effects and Heterogeneity

A large-scale illustration leverages 67,393 meta-analyses from the Cochrane Database. Here, location-scale models are used to assess whether smaller studies (proxied by larger standard errors) not only inflate effect size estimates (traditionally interpreted as publication bias or small-study effects) but also display amplified heterogeneity.

• Results: 13.0% (8,766/67,393) of meta-analyses exhibited significantly larger effects in smaller studies via Egger's regression test. The location-scale model identified 10.2% (6,897/67,393) for inflated effects and 2.3% (1,519/67,393) for increased heterogeneity among small studies.
• Significance: These results underscore the capacity of distributional regression to interrogate multi-parameter hypotheses, revealing patterns invisible to univariate models. However, limitations include low statistical power in meta-analyses with few studies and potentially unmodelled sources of bias and heterogeneity.

Software Implementation and Model Selection

Current implementations of distributional regression models are not meta-analysis-specific but can be adapted via flexible modelling environments (e.g., brms, gamlss, bamlss in R). Special cases are supported by meta-analytic tools such as metafor, blsmeta, and metaplus. Model specification and selection demand rigorous pre-registration, a priori hypothesis formulation, and diagnostic evaluation, given the increased risk of overfitting and parameter proliferation.

Limitations and Forward Directions

Despite its generality, the practical adoption of distributional regression for meta-analysis is curtailed by the prevailing focus on average effects, the lack of dedicated software infrastructure, and limited formal guidance on model selection/validation. Addressing these challenges necessitates:

• Development of user-friendly packages supporting full distributional regression in meta-analysis with proper weighting for known sampling variances,
• Methodological advances in inference and diagnostic tools for high-dimensional multi-parameter models,
• Increased theoretical and empirical research on modelling variance, skewness, and higher-order moments in applied fields,
• Community education to shift hypotheses and reporting norms beyond mean-centric inference.

Conclusion

Distributional regression models represent a substantial generalization of meta-analytic methodology, unifying and extending prevailing models to enable the study of all effect size distributional parameters as functions of explanatory variables. This approach paves the way for a deeper mechanistic and methodological understanding of effect heterogeneity, outlier behaviour, and the structure of scientific evidence. Its broad implementation could fundamentally alter practices in evidence synthesis, motivating richer hypotheses and facilitating more robust and informative policy, clinical, and scientific decision-making. Future work should prioritize practical tools, robust criteria for model selection, and the empirical validation of these models across diverse domains.