Subregressors. A casual-operational definition of mechanistic disease variables

This methodological preprint introduces the target-relative mechanistic subregressor: a versioned biological state or intervention-coherent macrovariable that lies outside the constitutive boundary of a declared target at the same resolution, reaches the target through a time-respecting causal path, and exerts a directionally material effect on its future trajectory within a specified context and horizon. A supported subregressor must be tied to an explicit causal estimand or valid bound, satisfy intervention-coherence and identification requirements, and carry traceable evidence, uncertainty, provenance, and transport status. Observability, practical identifiability, actionability, target engagement, safety, package sufficiency, and certificate eligibility are treated as separate operational dimensions rather than as parts of the core causal definition.

The framework is proposed as an ontology and audit layer, not as a new causal calculus. It builds on structural causal models, Pearl–Robins causal inference, causal abstraction, systems biology, measurement theory, pharmacology, and evidence calibration. The paper develops a ten-axiom foundation, a strong-Kleene three-valued gate system, direct, indirect-only, and mixed subregressor roles, target-relative abstraction and refinement rules, package-level sufficiency and minimality logic, and machine-readable ontology patterns using BFO/OBO, OWL, SHACL, and related standards.

The work also formally distinguishes subregressors from risk factors, biomarkers, mediators, therapeutic targets, disease drivers, hallmarks, endotypes, treatable traits, disease modules, control-theoretic driver nodes, and digital-twin states. Worked examples are provided across type 2 diabetes, atherosclerotic cardiovascular disease, severe asthma, cancer, HFpEF, depression, COPD, chronic kidney disease, and Alzheimer disease. The proposed term is explicitly falsifiable and should be adopted only if prospective studies show that it reduces semantic ambiguity, annotation error, retrieval failure, and invalid mechanistic inference enough to outweigh training, migration, and lexical-collision costs.