Consistent Bayesian meta-analysis on subgroup specific effects and interactions

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Consistent Bayesian meta-analysis on subgroup specific effects and interactions".

Jane: Here's the gist of the paper. Commonly clinical trials do not only report effects on the full study population but also on subgroups of patients The Bayesian Contribution-adjusted Meta-analysis by Subgroup (CAMS) indicates a stronger treatment-by-disability interaction (relapse rate reduction) in lower disability patients (EDSS ≤ 3.5) compared with the unadjusted model By controlling subgroup contribution while retaining subgroup interpretability, this approach allows for reliable interaction decision-making when published subgroup data are available Core Problem Addressed The paper addresses the inconsistency that arises when meta-analyses of subgroup-specific effects and treatment-by-subgroup interactions might not be consistent, especially when trials yield varying subgroup weighting This inconsistency is motivated by the fact that small studies may exhibit not just a small-study effect, that is, an outlying interaction estimate, but also a poor representation of subgroup prevalence in the target population Methodological Framework The authors introduce the Contribution-adjusted Meta-analysis by Subgroup (CAMS) in the Bayesian framework, which inherently adjusts for varying relative subgroup contribution quantified by the information fraction (IF) within trial The method demonstrates that CAMS reproduces standard Bayesian Interaction Meta-analysis (BIM), and consequently also achieves the interaction weights SWADA goal of preserving interaction inference Key Analytical Concepts The paper outlines several key concepts for reconciling subgroup and interaction estimates)&subgroup contribution is quantified as the information fraction (IF) within trial, CAMS)&the framework can be implemented in a frequentist or likelihood framework, CAMS)&CAMS accounts for varying study-specific prevalences and isolates the interaction, by incorporating study-specific contributions into the analysis through the IF (pij), CAMS Findings and Implications The CAMS approach yields within-trial interaction estimates via a model-based approach which allows both frequentist and Bayesian inference types The CAMS formulation prevents borrowing of information between subgroup means and the interaction, making the within-trial contrast gj = yB j − yA j sufficient for (𛾽, τγ) The CAMS model can reproduce BIM’s unconfounded interaction estimates, typically drawn from interaction data This approach reconciles pooled (BMS) and contrast-based (BIM) analyses and provides a transparent diagnostic tool through prevalence-specific visualizations Reporting Strategies The paper discusses several ways of handling dependence on the information fraction (proxy: prevalence) pi)&report inference at a single value of pi that is considered relevant, CAMS)&average inference with respect to an assumed or estimated distribution p(pij) for the proxy prevalences (e.g., from external evidence or a previous meta-analysis), thereby propagating uncertainty in pi, CAMS)&choose pi such that the overall effect estimate is recovered (which requires a within-trial collapsibility assumption), CAMS The authors recommend subgroup effects to be presented under this IF/prevalence, since it uniquely bridges the subgroup-adjusted and overall analyses while maintaining confounding-free inference on interaction The adjusted aggregate-data analyses can deliver BIM-equivalent interaction inference and coherent subgroup estimates, offering a practical, privacy-preserving alternative when IPD are unavailable CAMS extends earlier approaches such as SWADA by ensuring coherent uncertainty quantification and accommodating prior information CAMS addresses common pitfalls of traditional subgroup analyses, including aggregation bias, difficulties in estimating interaction heterogeneity, and unequal subgroup sizes, while maintaining interpretability The IF adjustment is based on the normal approximations commonly employed for meta-analysis, so that the approach is generalizable to a range of outcome scales such as risk ratios, odds ratios, hazard ratios, standardized mean differences CAMS may also be applied within a frequentist framework, where the number of studies needs to be sufficiently large for accurate estimation of the variance components Appendix Details The implementation details show that the CAMS model can be fitted using brms for Bayesian implementations and metafor for effect measure computations The data used consists of 14 rows and 11 columns, representing results from multiple studies analysing treatment-by-subgroup effects and prevalence estimates The dataset enables detailed modelling of subgroup interactions and variance components to investigate treatment effects with Bayesian meta-analytic framework The authors used MCMC to estimate between-trial variance components, following recommended practice on prior choices in meta-analysis The dataset consists of 14 rows and 11 columns, representing results from multiple studies analysing treatment-by-subgroup effects and prevalence estimates The columns include study.name (character), which identifies the study, and numerical variables contrast.esti and contrast.se that capture the treatment-by-subgroup interaction effect measures and associated standard errors for each study Subgroup-specific effect measures and their standard errors are recorded in est and se, respectively Additionally, ifrac provides the IF for the non-reference subgroup (subgroup B), while subgroup12 serves as the subgroup identifier, coded as −0.5 and 0.5 to reflect subgroup contrasts and capture heterogeneity structures across trials ifrac2 = subgroup12 + 0.5-ifrac provides the IF contrast for the interaction heterogeneity matrix in (3.

Paper summary: Tom: Exactly. So what they claim is that by controlling this subgroup contribution—which they call the information fraction, or IF—they can maintain reliability when looking at both the subgroup effects and the interaction itself, even if the published data has varying levels of subgroup weighting across studies.

Jane: They specifically highlight that in lower disability patients, like those with an EDSS score less than or equal to three point five, there's a stronger treatment-by-disability interaction effect shown in their analysis compared to the unadjusted model. That's a concrete finding they highlight when discussing CAMS.

Lu: The methodology relies on incorporating this information fraction, which is basically how much each study contributes to the overall data pool, into the Bayesian framework. This lets them adjust for those varying study-specific prevalences, which is crucial because studies aren't always representative of the target population they are trying to model.

Meng: If you're running a production system and you have one hundred studies contributing to your meta-analysis, and some of those studies only looked at a tiny fraction of your actual user base, this adjustment seems necessary to avoid overfitting on those small, potentially unrepresentative subgroups <ref:2512.18785#pg1>.

Lalam: I see how that connects to my training data. When I process different demographic slices of text, the way I weight those slices matters; if the weighting is inconsistent across sources, my overall representation will be noisy. CAMS seems like a principled way to handle that noise in meta-analysis.

Conclusion: Tom: So, to wrap up this paper "Consistent Bayesian meta-analysis on subgroup specific effects and interactions", they've essentially created a tool that lets you get interaction inference just as reliably, even when you're looking at subgroups defined differently in each study. The authors recommend using the overall IF when reporting both subgroup and overall effect estimates.

Jane: It’s about providing a consistent way to decide on those interactions. They show that this CAMS approach can reproduce the interaction inference you get from standard Bayesian Interaction Meta-analysis, which is really good because it bridges two different ways of looking at the data: pooled analyses versus contrast-based analyses.

Lu: The implication here is that for researchers in this field, it offers an alternative way to find matching subgroup and interaction meta-analyses when detailed patient data isn't available. They also offer a generic framework that can be applied to other types of outcomes or even network meta-analysis, as noted in the introduction.

Meng: It’s interesting from an engineering standpoint because it provides a diagnostic tool through prevalence-specific visualizations, which helps you see *why* your interaction estimate might be behaving unexpectedly based on how much data each study actually contributed. That transparency is something I value when shipping models.

Lalam: I think the main thing for listeners to grasp is that this method addresses common pitfalls like aggregation bias and Simpson's paradox by ensuring the weights assigned to subgroups are specified identically for both the subgroup effect analysis and the interaction analysis, which guarantees consistency between those two different inferences.