Consistent Bayesian meta-analysis on subgroup specific effects and interactions

Commonly, clinical trials report effects not only for the full study population but also for patient subgroups. Meta-analyses of subgroup-specific effects and treatment-by-subgroup interactions may be inconsistent, especially when trials apply different subgroup weightings. We show that meta-regression can, in principle, with a contribution adjustment, recover the same interaction inference regardless of whether interaction data or subgroup data are used. Our Bayesian framework for subgroup-data interaction meta-analysis inherently (i) adjusts for varying relative subgroup contribution, quantified by the information fraction (IF) within a trial; (ii) is robust to prevalence imbalance and variation; (iii) provides a self-contained, model-based approach; and (iv) can be used to incorporate prior information into interaction meta-analyses with few studies.The method is demonstrated using an example with as few as seven trials of disease-modifying therapies in relapsing-remitting multiple sclerosis. The Bayesian Contribution-adjusted Meta-analysis by Subgroup (CAMS) indicates a stronger treatment-by-disability interaction (relapse rate reduction) in patients with lower disability (EDSS <= 3.5) compared with the unadjusted model, while results for younger patients (age < 40 years) are unchanged.By controlling subgroup contribution while retaining subgroup interpretability, this approach enables reliable interaction decision-making when published subgroup data are available.Although the proposed CAMS approach is presented in a Bayesian context, it can also be implemented in frequentist or likelihood frameworks.

💡 Research Summary

The paper addresses a pervasive problem in meta‑analysis of subgroup‑specific treatment effects: when individual trials report different subgroup prevalences, meta‑analyses of subgroup effects and of treatment‑by‑subgroup interactions can yield inconsistent conclusions. Traditional approaches include (i) Bayesian Interaction Meta‑analysis (BIM), which directly meta‑analyses within‑trial contrasts (the difference between subgroup treatment effects) but does not provide subgroup‑specific pooled estimates, and (ii) Bayesian Meta‑analysis by Subgroup (BMS), which pools each subgroup’s effect separately and then compares them. BMS, however, suffers from ecological bias because the overall effect in each trial is a weighted average of subgroup means, and the weights differ across trials when subgroup sizes vary. This mismatch leads to Simpson’s paradox‑type distortions and undermines the validity of interaction inference, especially in few‑study settings.

To resolve this, the authors propose the Contribution‑adjusted Meta‑analysis by Subgroup (CAMS), a Bayesian hierarchical model that explicitly incorporates the relative contribution of each subgroup within a trial. The contribution is quantified by the Information Fraction (IF), defined as π_j = σ_Bj⁻² / (σ_Aj⁻² + σ_Bj⁻²), where σ_Aj and σ_Bj are the standard errors of the subgroup effect estimates in study j. Under the common assumption of a constant Unit Information Standard Deviation (UISD) within a trial, π_j approximates the actual proportion of participants in subgroup B (p_j). By treating π_j as a known weight, the model defines the trial‑specific overall effect as m_j = (1‑π_j)·y_Aj + π_j·y_Bj, where y_Aj and y_Bj are the true subgroup effects. The hierarchical structure then models both the within‑study variability (through the observed standard errors) and the between‑study heterogeneity of the main effects and the interaction term.

CAMS extends the normal‑normal hierarchical model (NNHM) to a bivariate random‑effects formulation, orthogonalising the main effect and interaction components. Weakly informative normal priors are placed on the overall interaction mean (γ) and on subgroup main effects, while half‑normal priors are used for heterogeneity variances (τ_γ, τ_m). Crucially, the authors prove that when the contribution weights π_j are correctly specified, the posterior distribution of the interaction parameter γ under CAMS coincides exactly with that obtained from BIM. Thus, CAMS delivers the same interaction inference as BIM while simultaneously providing coherent subgroup‑specific pooled estimates and a properly weighted overall effect. This unifies the two previously disparate approaches and satisfies the SWADA (Same Weighting Across Different Analyses) principle within a Bayesian framework, where explicit weighting of the likelihood is not feasible.

The methodology is illustrated with a real‑world example: seven randomized controlled trials of disease‑modifying therapies in relapsing‑remitting multiple sclerosis (RRMS). Two subgroup variables are examined: age (<40 vs ≥40 years) and baseline disability measured by the Expanded Disability Status Scale (EDSS ≤3.5 vs >3.5). For each trial, subgroup‑specific rate ratios (RR) and their standard errors are extracted, the IFs are computed, and the CAMS model is fitted. Results show that, after contribution adjustment, the treatment‑by‑disability interaction becomes stronger and its credible interval narrows compared with the unadjusted model, indicating that patients with lower disability derive greater benefit. In contrast, the age‑by‑treatment interaction remains essentially unchanged, reflecting the relatively balanced age distribution across trials. The overall treatment effect estimates also become more consistent across studies when weighted by the IFs.

Beyond the Bayesian implementation, the authors discuss how CAMS can be translated into frequentist or likelihood‑based settings (e.g., using restricted maximum likelihood with pre‑specified weights). They provide reproducible R code, facilitating adoption by meta‑analysts who only have access to published aggregate subgroup data. The paper emphasizes that CAMS mitigates aggregation bias, Simpson’s paradox, and ecological fallacy without requiring individual participant data, making it especially valuable in contexts with few studies or limited data availability.

In conclusion, the Contribution‑adjusted Meta‑analysis by Subgroup offers a principled, model‑based solution to the inconsistency between subgroup‑specific and interaction meta‑analyses. By adjusting for varying subgroup contributions via the information fraction, it yields coherent estimates of subgroup effects, overall effects, and treatment‑by‑subgroup interactions, enhancing the reliability of decision‑making in clinical guideline development and regulatory evaluation.