Estimand Alignment and Implementation Sensitivity in Causal Mediation Analysis of Environmental Mixtures: Simulation, Plasmode, and Applied Study

Causal mediation analysis of environmental mixtures is challenging because exposures are correlated and mediation effects depend on joint exposure intervention. We compared Bayesian kernel machine regression causal mediation analysis (BKMR-CMA) with Bayesian additive regression trees causal mediation analysis (BART-CMA) for a simultaneous shift of mixture components from the 25th to the 75th percentile in seventeen synthetic scenarios and two plasmode simulations based on National Health and Nutrition Examination Survey (NHANES) data. We evaluated DeepMed separately because its binary treatment formulation could not target the same continuous mixture intervention. With an exposure–mediator interaction, a binary outcome, or additional effect-carrying components, BKMR-CMA had smaller indirect-effect bias and better interval coverage than BART-CMA; without interaction, BART-CMA was approximately unbiased. BART-CMA’s bias did not grow with the number of components when the added components were inert but increased markedly when they carried additional effects. In the NHANES-based plasmode, the ordering reversed: BART-CMA was more accurate, BKMR-CMA’s mean indirect effect had the wrong sign, and neither method reliably detected the imposed indirect effect. DeepMed performed poorly against both the continuous-mixture truth and the truth for its own binary contrast, so its error reflected estimator performance, not estimand mismatch alone. Among 1498 adults aged 20 years or older, BART-CMA estimated near-zero indirect effects of a PFAS–metal mixture on six liver outcomes through the Dietary Inflammatory Index. No estimator was uniformly superior: performance depended on estimand alignment, interaction structure, the number of effect-carrying components, the outcome scale, cross-world support, and implementation.