Semiparametric Bayesian inference for causal mediation in cluster randomized trials

Abstract: Cluster randomized trials (CRTs) are frequently used to evaluate interventions, yet conducting causal mediation analysis in these settings remains challenging, particularly when the mediator is measured at the cluster level and the number of clusters is small. Standard inference methods often rely on asymptotic assumptions that fail in finite-sample settings, leading to biased variance estimation and invalid confidence intervals. In this paper, we propose a robust inference framework for causal mediation analysis in CRTs. We utilize parametric Bayesian models for the outcome and mediator to ensure computational efficiency and interpretability. Crucially, to quantify uncertainty, we specify a novel similarity-weighted Bayesian bootstrap (SWBB) with a distance' metric between clusters; this avoids the need for restrictive parametric assumptions and allows the model to borrow more information fromcloser' clusters. By combining observed data models with causal assumptions, our approach accurately estimates natural direct and indirect effects even with limited clusters. Simulation studies demonstrate that our method achieves nominal coverage probability across diverse scenarios. We illustrate the practical utility of our approach by assessing mediation in a CRT in Kenya.

• The paper presents a semiparametric Bayesian framework incorporating a similarity-weighted Bayesian bootstrap (SWBB) for causal mediation analysis in CRTs.
• The method improves estimation precision by reducing RMSE and credible interval length compared to standard BB and HBB techniques, especially in sparse cluster settings.
• Simulation studies and empirical analysis from the BIGPIC trial validate its unbiased, computationally tractable approach to estimating direct and indirect effects in clustered data.

Semiparametric Bayesian Causal Mediation for CRTs with Cluster-Level Mediators

Introduction and Context

Cluster randomized trials (CRTs) are the gold standard for evaluating interventions implemented at the level of groups (e.g., facilities, villages). Mediation analysis in this context aims to decompose the effect of an intervention into direct and indirect components via intermediate (mediator) variables. This task is particularly complex when mediators are measured at the cluster level and the number of clusters is small, resulting in issues with standard asymptotic inference, biased variance estimates, and unreliable confidence intervals.

The paper "Semiparametric Bayesian inference for causal mediation in cluster randomized trials" (2606.13305) introduces a semiparametric Bayesian framework for mediation analysis in CRTs. It leverages parametric modeling for outcomes and mediators to facilitate tractable inference, but crucially, introduces a similarity-weighted Bayesian bootstrap (SWBB) for flexible, nonparametric modeling of confounder distributions and to enable adaptive information borrowing across clusters. This approach addresses methodological and computational limitations of existing mediation methods in small-sample, sparse CRT settings.

Structural Causal Framework and Identification

The paper rigorously formalizes the causal estimands—natural direct effects (NDE), natural indirect effects (NIE), and total effects (ATE)—at individual, cluster, and population levels, all in the potential outcomes model. This framework explicitly characterizes the role of both cluster-level and individual-level confounders and introduces a bipartite structure reflecting the hierarchical design of CRTs. The identification strategy relies on cluster-level SUTVA, sequential ignorability (SI), and standard positivity assumptions, partly justified by randomization.

A notable distinction is made between cluster-level and individual-level mediation mechanisms. Conditional and marginal effects are defined via specific marginalizations over confounder distributions, aligning with existing mediation theory but extending it to the nuanced CRT context with cluster-level mediators.

Methodology: Similarity-Weighted Bayesian Bootstrap (SWBB)

The methodological centerpiece is the SWBB, a two-stage Bayesian nonparametric model for estimating the joint distribution of cluster- and individual-level confounders, agnostic to specific parametric forms. Unlike standard BB or HBB methods, the SWBB employs a distance-based metric (incorporating both cluster-level and individual-level confounders) to weight clusters by similarity when “borrowing” information to stabilize estimates.

Parametric models are fit for the outcome (typically with cluster-level random effects) and the mediator (via fixed or mixed effects), while the SWBB governs inference about confounder distributions—which acts as a nuisance parameter from the causal point of view. The conjugacy and modularity of this specification allow for plug-in replacement with nonparametric models (e.g., BART), though this is not the focus due to limited cluster sample size.

At each MCMC iteration, the posterior for mediation estimands is computed by integrating over the SWBB-reconstructed confounder joint PMF. The approach is computationally tractable, requiring only Gibbs updates for Dirichlet weights at each hierarchy, and achieves robust uncertainty quantification.

Simulation and Empirical Results

Extensive simulation studies evaluate the SWBB, HBB, and BB under increasing cross-level dependency between cluster- and individual-level confounders, across 1,000 replications per scenario. Key findings include:

• Bias: All methods are unbiased under correct parametric outcome/mediator models and SI.
• Precision (RMSE): SWBB demonstrates substantial efficiency gains over BB/HBB—up to a 5% reduction in RMSE for NIE and ATE in independence settings, and similar improvements in dependent settings, with performance stabilizing at higher information borrowing levels.
• Coverage: Empirical coverage for SWBB remains at nominal levels (0.95–0.97), indicating valid posterior uncertainty.

Empirical analysis using the BIGPIC trial (Western Kenya, $J=24$ clusters) demonstrates practical utility. No statistically significant mediation is detected for network-based mediators, but the SWBB achieves 18-22% reductions in credible interval length for ATE compared to BB/HBB, confirming variance reduction and sharper inference in finite samples.

Conditional analyses (by community or individual stratum) achieve improved posterior concentration, although CIs are naturally wider due to reduced sample sizes; SWBB remains superior to alternatives.

Theoretical Implications and Extensions

The SWBB generalizes the hierarchical Bayesian bootstrap for clustered data by decoupling information borrowing from strict sample size and anchoring it to empirical similarity—a property not addressed by existing methods [oganisian2022hierarchical]. This is particularly relevant in CRTs with sparse or heterogeneous clusters, preserving flexibility without over-shrinking toward the global empirical distribution.

From a theoretical standpoint, the paper sharply distinguishes bias reduction (the field of correct modeling and identification) from variance reduction (driven by the SWBB’s adaptive pooling). In the context of mediation, this distinction is crucial: the framework allows robust estimation of NIE and ATE despite limited data, and offers the possibility to scale to complex hierarchical confounder structures or incorporate flexible outcome/mediator models.

Possible future directions include:

• Nonparametric modeling for mediator/outcome to address model misspecification, although data sparseness remains challenging with small $J$.
• Application to multiple or time-varying mediators.
• Integration of "clever covariates" or targeted machine learning features for semiparametric efficiency and double robustness.
• Sensitivity analyses for SI violations, extending beyond point identification.

Conclusion

This work provides a robust and adaptable methodology for causal mediation analysis in CRTs with cluster-level mediators and small $J$, filling a methodological gap in both Bayesian and frequentist literature. The SWBB enables adaptive variance reduction through similarity-based information pooling, delivering reliable inference where standard asymptotics or sample size-based pooling fail. Practically, the approach incurs no restrictive modeling overhead, is computationally tractable, and preserves theoretical validity under minimal assumptions, with wide applicability for cluster-structured mediation questions in medical and social science RCTs. Future developments in integrating nonparametric outcome modeling and sensitivity analysis will further strengthen its utility in real-world complex CRTs.