I have linked here a preprint in review which is the latest of a trilogy (the second of the three has just been accepted).
These articles focus on RCT in complex clinical environments. This latest article A new structural analysis of randomized trials: three estimands for trial design and safe guideline development
presents a new conceptual and methodological analysis of the transportability of randomized trial estimates under different enrollment structures. Building on prior structural work contrasting cause agnostic randomized controlled trials (CARs), characterized by broad enrollment gates, with causal-integrity randomized controlled trials (CIRs), defined by narrow, disease-specific criteria, it introduces a three-layer framework that makes the structure of the trial estimand explicit and clarifies the roles of the selection mechanism and covariate vectors.
At Layer 1 (E₁), the target is the individual causal effect for a patient in a specific disease state. At Layer 2 (E₂), a causal-integrity randomized controlled trial (CIR) estimates the disease-level average treatment effect across patients within a single disease, selected by a disease-specific enrollment gate. Within this layer, covariate vectors act as within-disease effect modifiers, indexing variation across patient states within a shared causal system.
In contrast, at Layer 3 (E₃), a cause agnostic randomized controlled trial (CAR) produces a fundamentally different estimand: a disease- and mechanism-agnostic gate selects a mixture of distinct diseases, each representing a parallel causal system with its own covariate structure. The observed effect is therefore a weighted average across distinct causal systems, with weights determined by the enrollment gate.
This framework suggests that this structure may produce unstable effect estimates and a “cause–mixture paradox,” in which observed treatment effects reverse to the null or opposite sign across trials as the disease mixture πᵢ(S = 1) shifts. This provides a structural explanation for reversals of RCT findings and a new formal structural basis for assessing the evidentiary strength of trials under guideline consideration.
I look forward to comments criticisms, ect.