Rockwell et al.'s Bayesian mediation framework explains why treatment rules work

Individualized treatment rules (ITRs) map patient characteristics to treatment decisions, promising tailored care. Yet their value—the expected outcome under the rule—is typically a black-box number that says nothing about why the rule works. Rockwell et al. introduce a causal mediation framework that decomposes the value contrast between a candidate ITR and a clinically meaningful reference rule into natural direct and indirect effects. Using rule-specific nested potential outcomes, they isolate the benefit that operates directly on the outcome from the benefit that flows through a specified mediator, and provide identification conditions via a rule-level mediation g-formula. This lets clinicians answer not just "does the rule improve outcomes?" but "through which biological or behavioral pathway does it deliver that improvement?"

For estimation, the authors adapt Bayesian causal mediation forests to obtain posterior inference for the value contrast and its path-specific components. Simulations across varied direct and mediated contributions show the estimator achieves near-nominal credible interval coverage, with bias and root mean squared error shrinking as sample size grows. They further demonstrate the method on the TRIUMPH trial, decomposing the cognitive benefit of a lifestyle intervention rule through candidate neurovascular, cardiorespiratory, and behavioral mediators. This framework complements optimal ITR learning by adding mechanistic explanation, offering a natural approach for evaluating not just whether personalized rules work, but how they generate clinical value.